fix: address PR review comments (round 2)

Human reviewer (zmanian):
- H1: Accept configurable embedding dimension in LanceDbVectorStore::new()
  instead of hardcoding 1536. Dimension is sourced from
  EmbeddingProvider::dimension() at init time.
- H2: Skip double-write of embeddings to DB when external vector store
  is active (pass None to insert_chunk for embedding column).
- H3: Update PR title from "refactor" to "feat" (net-new feature).
- H4: Document non-atomic update_embedding in struct doc comment.

Bot reviewer (Copilot):
- Cache LanceDB table handle via tokio::sync::OnceCell (avoid
  open_table per operation).
- Cache Arc<Schema> in struct (avoid rebuilding per insert).
- Fix error variants: ChunkingFailed → EmbeddingFailed for LanceDB
  store/delete operations.
- Propagate store_embedding errors in reindex_document instead of
  warn-only (prevents silent data loss).
- Prefetch document metadata map in backfill_embeddings to avoid N+1
  queries.
- Add lancedb feature + protoc to CI test matrix so LanceDB tests
  actually run on Linux.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

.env.example

@@ -2,9 +2,28 @@
 DATABASE_URL=postgres://localhost/ironclaw
 DATABASE_POOL_SIZE=10
 
+# Vector store for workspace memory (optional)
+# When set to "lancedb", uses LanceDB for semantic search instead of pgvector/libsql
+# VECTOR_BACKEND=builtin   # default: use database's built-in index (pgvector or libsql_vector_idx); "pgvector" is also accepted as an alias for "builtin"
+# VECTOR_BACKEND=lancedb
+# LANCEDB_PATH=~/.ironclaw/lancedb   # path for LanceDB when VECTOR_BACKEND=lancedb
+
 # LLM Provider
 # LLM_BACKEND=nearai           # default
 # Possible values: nearai, ollama, openai_compatible, openai, anthropic, tinfoil
+# LLM_REQUEST_TIMEOUT_SECS=120  # Increase for local LLMs (Ollama, vLLM, LM Studio)
+
+# === Anthropic Direct ===
+# Two auth modes:
+#   1. API key: Set ANTHROPIC_API_KEY (from console.anthropic.com/settings/keys)
+#   2. OAuth token: Set ANTHROPIC_OAUTH_TOKEN (from `claude login`)
+#      OAuth tokens use Authorization: Bearer instead of x-api-key header.
+# ANTHROPIC_API_KEY=sk-ant-...
+# ANTHROPIC_OAUTH_TOKEN=sk-ant-oat01-...   # from `claude login` credentials
+# ANTHROPIC_MODEL=claude-sonnet-4-20250514
+
+# === OpenAI Direct ===
+# OPENAI_API_KEY=sk-...
 
 # === NEAR AI (Chat Completions API) ===
 # Two auth modes:

.github/workflows/code_style.yml

@@ -36,6 +36,11 @@ jobs:
       uses: dtolnay/rust-toolchain@stable
       with:
         components: clippy
+    - name: Install protoc (for lancedb)
+      if: matrix.name == 'all-features'
+      uses: arduino/setup-protoc@v3
+      with:
+        repo-token: ${{ secrets.GITHUB_TOKEN }}
     - uses: Swatinem/rust-cache@v2
       with:
         key: clippy-${{ matrix.name }}
@@ -62,6 +67,11 @@ jobs:
       uses: dtolnay/rust-toolchain@stable
       with:
         components: clippy
+    - name: Install protoc (for lancedb)
+      if: matrix.name == 'all-features'
+      uses: arduino/setup-protoc@v3
+      with:
+        repo-token: ${{ secrets.GITHUB_TOKEN }}
     - uses: Swatinem/rust-cache@v2
       with:
         key: clippy-windows-${{ matrix.name }}

.github/workflows/coverage.yml

@@ -1,3 +1,31 @@
+# Code Coverage Workflow
+#
+# This workflow runs test coverage analysis and uploads reports to Codecov.
+# Coverage reports help identify untested code paths and maintain code quality.
+#
+# What it does:
+# - Runs unit and integration tests with coverage instrumentation
+# - Runs E2E tests with coverage instrumentation
+# - Uploads coverage reports to Codecov (https://codecov.io/gh/nearai/ironclaw)
+#
+# Viewing coverage reports:
+# - PRs automatically get coverage comments showing changes in coverage
+# - Visit https://codecov.io/gh/nearai/ironclaw for detailed coverage reports
+# - Coverage reports are generated for three configurations:
+#   1. all-features: Full feature set
+#   2. default: Default features
+#   3. libsql-only: Minimal libSQL-only configuration
+# - E2E coverage tracks end-to-end test coverage separately
+#
+# Coverage files:
+# - Unit/integration: lcov.info (uploaded to Codecov with "unit" flag)
+# - E2E: e2e-coverage.info (uploaded to Codecov with "e2e" flag)
+#
+# Requirements:
+# - Uses cargo-llvm-cov for coverage instrumentation
+# - Requires PostgreSQL for integration tests (pgvector/pgvector:pg16)
+# - E2E tests require Python 3.12 and Playwright
+
 name: Code Coverage
 on:
   push:

.github/workflows/test.yml

@@ -14,7 +14,7 @@ jobs:
       matrix:
         include:
           - name: all-features
-            flags: "--features postgres,libsql,html-to-markdown"
+            flags: "--features postgres,libsql,lancedb,html-to-markdown"
           - name: default
             flags: ""
           - name: libsql-only
@@ -26,6 +26,11 @@ jobs:
         uses: dtolnay/rust-toolchain@stable
         with:
           targets: wasm32-wasip2
+      - name: Install protoc (for lancedb)
+        if: matrix.name == 'all-features'
+        uses: arduino/setup-protoc@v3
+        with:
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
       - uses: Swatinem/rust-cache@v2
         with:
           key: ${{ matrix.name }}
@@ -66,6 +71,11 @@ jobs:
         uses: actions/checkout@v6
       - name: Install Rust
         uses: dtolnay/rust-toolchain@stable
+      - name: Install protoc (for lancedb)
+        if: matrix.name == 'all-features'
+        uses: arduino/setup-protoc@v3
+        with:
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
       - uses: Swatinem/rust-cache@v2
         with:
           key: windows-${{ matrix.name }}
@@ -82,6 +92,10 @@ jobs:
         uses: dtolnay/rust-toolchain@stable
         with:
           targets: wasm32-wasip2
+      - name: Install protoc (for lancedb)
+        uses: arduino/setup-protoc@v3
+        with:
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
       - uses: Swatinem/rust-cache@v2
         with:
           key: wasm-extensions

.gitignore

@@ -4,8 +4,9 @@
 .env.*
 !.env.example
 
-# Claude Code worktrees
+# Claude Code worktrees and lock files
 .claude/worktrees/
+.claude/scheduled_tasks.lock
 
 # Sidecar tool data
 .sidecar/
@@ -22,3 +23,4 @@ bench-results/
 # WASM build artifacts (loaded from disk, not bundled)
 *.wasm
 
+trace_*.json

CLAUDE.md

@@ -387,12 +387,27 @@ Dead code behind the wrong `#[cfg]` gate will only show up when building with a
 
 **Zero clippy warnings policy:** Fix ALL clippy warnings before committing, including pre-existing ones in files you didn't change. Never leave warnings behind — treat `cargo clippy` output as a zero-tolerance gate.
 
+**Transaction safety:** Multi-step database operations (INSERT+INSERT, UPDATE+DELETE, read-then-write) MUST be wrapped in a transaction. Never assume sequential calls are atomic. Before committing DB code, ask: "If this crashes between step N and N+1, is the database consistent?" If not, wrap in a transaction. This applies to both postgres and libsql backends.
+
+**UTF-8 string safety:** Never use byte-index slicing (`&s[..n]`) on user-supplied or external strings — it panics on multi-byte characters. Use `is_char_boundary()` to walk backwards from the desired length, or iterate with `char_indices()`. Grep for `[..` in changed files to catch violations.
+
+**Case-insensitive comparisons:** When comparing user-supplied strings (file paths, media types, extension names), always normalize to lowercase first with `.to_ascii_lowercase()`. On case-insensitive filesystems (macOS, Windows), path comparisons must be case-insensitive. File extension checks (`.png`, `.jpg`) and media type checks (`image/jpeg`) are common offenders.
+
+**Decorator/wrapper trait delegation:** When adding a new method to `LlmProvider` (or any trait with decorator wrappers), you MUST update ALL wrapper types to delegate to their inner provider. Grep for `impl LlmProvider for` to find all implementations. Add a test that exercises the method through the full provider chain (`build_provider_chain()`), not just the base impl.
+
+**Sensitive data in logs & events:** Tool parameters and outputs MUST be redacted before logging or broadcasting via SSE/WebSocket. Use `redact_params()` before any `tracing::info!`, `JobEvent`, or SSE emission that includes tool call data. Never log raw parameters from tool calls.
+
+**Test temporary files:** Use the `tempfile` crate for test directories/files. Never hardcode `/tmp/...` paths — they collide in parallel test runs and break on non-Unix platforms.
+
+**Trust boundaries in multi-process architecture:** Data from worker containers is untrusted. The orchestrator MUST validate: tool domain (never execute `Container`-domain tools on the host), nesting depth (server-side tracking, not client-supplied), and parameter sensitivity (redact before logging/broadcasting).
+
 **Mechanical verification before committing:** Run these checks on changed files before committing:
 - `cargo clippy --all --benches --tests --examples --all-features` -- zero warnings
 - `grep -rnE '\.unwrap\(|\.expect\(' <files>` -- no panics in production
 - `grep -rn 'super::' <files>` -- use `crate::` imports
 - If you fixed a pattern bug, `grep` for other instances of that pattern across `src/`
 - Fix commits must include regression tests (enforced by `commit-msg` hook; bypass with `[skip-regression-check]`)
+- Run `scripts/pre-commit-safety.sh` to catch UTF-8, case-sensitivity, hardcoded /tmp, and logging issues
 
 ## Configuration
 

Cargo.toml

@@ -56,7 +56,7 @@ rustls = { version = "0.23", optional = true, default-features = false }
 rustls-native-certs = { version = "0.8", optional = true }
 
 # Database - libSQL/Turso (optional embedded database)
-libsql = { version = "0.6", optional = true, default-features = false, features = ["core", "replication"] }
+libsql = { version = "0.6", optional = true, default-features = false, features = ["core", "replication", "remote", "tls"] }
 
 # Error handling
 thiserror = "2"
@@ -73,6 +73,8 @@ toml = "0.8"
 # Core types
 uuid = { version = "1", features = ["v4", "v5", "serde"] }
 chrono = { version = "0.4", features = ["serde"] }
+chrono-tz = "0.10"
+iana-time-zone = "0.1"
 rust_decimal = { version = "1", features = ["serde", "serde-with-str", "maths"] }
 rust_decimal_macros = "1"
 
@@ -123,6 +125,11 @@ open = "5"
 # The postgres feature provides ToSql/FromSql for postgres-types (shared by tokio-postgres)
 pgvector = { version = "0.4", features = ["postgres"], optional = true }
 
+# LanceDB vector store (optional alternative to pgvector/libsql for workspace search)
+lancedb = { version = "0.26", optional = true }
+arrow-array = { version = "57", optional = true }
+arrow-schema = { version = "57", optional = true }
+
 # WASM sandbox for untrusted tool execution
 wasmtime = { version = "28", features = ["component-model"] }
 wasmtime-wasi = "28"  # WASI support for component model
@@ -187,6 +194,7 @@ insta = "1.46.3"
 
 [features]
 default = ["postgres", "libsql", "html-to-markdown"]
+lancedb = ["dep:lancedb", "dep:arrow-array", "dep:arrow-schema"]
 postgres = [
     "dep:deadpool-postgres",
     "dep:tokio-postgres",

FEATURE_PARITY.md

@@ -39,7 +39,7 @@ This document tracks feature parity between IronClaw (Rust implementation) and O
 | Network modes (loopback/LAN/remote) | ✅ | 🚧 | HTTP only |
 | OpenAI-compatible HTTP API | ✅ | ✅ | /v1/chat/completions, per-request `model` override |
 | Canvas hosting | ✅ | ❌ | Agent-driven UI |
-| Gateway lock (PID-based) | ✅ | ❌ | |
+| Gateway lock (PID-based) | ✅ | ✅ | `fs4` flock-based, acquired in `main.rs` before agent startup |
 | launchd/systemd integration | ✅ | ❌ | |
 | Bonjour/mDNS discovery | ✅ | ❌ | |
 | Tailscale integration | ✅ | ❌ | |
@@ -215,9 +215,13 @@ This document tracks feature parity between IronClaw (Rust implementation) and O
 | NEAR AI | ✅ | ✅ | - | Primary provider |
 | Anthropic (Claude) | ✅ | 🚧 | - | Via NEAR AI proxy; Opus 4.5, Sonnet 4, Sonnet 4.6 |
 | OpenAI | ✅ | 🚧 | - | Via NEAR AI proxy |
-| AWS Bedrock | ✅ | ❌ | P3 | |
-| Google Gemini | ✅ | ❌ | P3 | |
-| NVIDIA API | ✅ | ❌ | P3 | New provider |
+| AWS Bedrock | ✅ | ✅ | P3 | Via `openai_compatible` adapter (e.g. LiteLLM) |
+| Google Gemini | ✅ | ✅ | P3 | Via `gemini` adapter |
+| io.net | ✅ | ✅ | P3 | Via `ionet` adapter |
+| Mistral | ✅ | ✅ | P3 | Via `mistral` adapter |
+| Yandex AI Studio | ✅ | ✅ | P3 | Via `yandex` adapter |
+| Cloudflare Workers AI | ✅ | ✅ | P3 | Via `cloudflare` adapter |
+| NVIDIA API | ✅ | ✅ | P3 | Via `nvidia` adapter and `providers.json` |
 | OpenRouter | ✅ | ✅ | - | Via OpenAI-compatible provider (RigAdapter) |
 | Tinfoil | ❌ | ✅ | - | Private inference provider (IronClaw-only) |
 | OpenAI-compatible | ❌ | ✅ | - | Generic OpenAI-compatible endpoint (RigAdapter) |
@@ -336,7 +340,7 @@ This document tracks feature parity between IronClaw (Rust implementation) and O
 | Gemini embeddings | ✅ | ❌ | |
 | Local embeddings | ✅ | ❌ | |
 | SQLite-vec backend | ✅ | ❌ | IronClaw uses PostgreSQL |
-| LanceDB backend | ✅ | ❌ | Configurable auto-capture max length |
+| LanceDB backend | ✅ | ✅ | VectorStore trait + LanceDbVectorStore (configured via VECTOR_BACKEND=lancedb) |
 | QMD backend | ✅ | ❌ | |
 | Atomic reindexing | ✅ | ✅ | |
 | Embeddings batching | ✅ | ✅ | `embed_batch` on EmbeddingProvider trait |

docker-compose.yml

@@ -3,7 +3,7 @@ services:
   postgres:
     image: pgvector/pgvector:pg16
     ports:
-      - "5432:5432"
+      - "127.0.0.1:5432:5432"
     environment:
       POSTGRES_DB: ironclaw
       POSTGRES_USER: ironclaw

docs/LLM_PROVIDERS.md

@@ -11,6 +11,12 @@ configurations.
 | NEAR AI | `nearai` | OAuth (browser) | Default; multi-model |
 | Anthropic | `anthropic` | `ANTHROPIC_API_KEY` | Claude models |
 | OpenAI | `openai` | `OPENAI_API_KEY` | GPT models |
+| Google Gemini | `gemini` | `GEMINI_API_KEY` | Gemini models |
+| AWS Bedrock | `bedrock` | `BEDROCK_ACCESS_KEY` | Requires OpenAI proxy (e.g. LiteLLM) |
+| io.net | `ionet` | `IONET_API_KEY` | Intelligence API |
+| Mistral | `mistral` | `MISTRAL_API_KEY` | Mistral models |
+| Yandex AI Studio | `yandex` | `YANDEX_API_KEY` | YandexGPT models |
+| Cloudflare Workers AI | `cloudflare` | `CLOUDFLARE_API_KEY` | Access to Workers AI |
 | Ollama | `ollama` | No | Local inference |
 | OpenRouter | `openai_compatible` | `LLM_API_KEY` | 300+ models |
 | Together AI | `openai_compatible` | `LLM_API_KEY` | Fast inference |

migrations/V11__conversation_unique_indexes.sql

@@ -0,0 +1,13 @@
+-- Partial unique indexes to prevent duplicate singleton conversations.
+-- These guard against TOCTOU races in get_or_create_routine_conversation
+-- and get_or_create_heartbeat_conversation.
+
+-- One routine conversation per user per routine_id.
+CREATE UNIQUE INDEX IF NOT EXISTS uq_conv_routine
+ON conversations (user_id, (metadata->>'routine_id'))
+WHERE metadata->>'routine_id' IS NOT NULL;
+
+-- One heartbeat conversation per user.
+CREATE UNIQUE INDEX IF NOT EXISTS uq_conv_heartbeat
+ON conversations (user_id)
+WHERE metadata->>'thread_type' = 'heartbeat';

providers.json

@@ -1,7 +1,9 @@
 [
   {
     "id": "openai",
-    "aliases": ["open_ai"],
+    "aliases": [
+      "open_ai"
+    ],
     "protocol": "open_ai_completions",
     "api_key_env": "OPENAI_API_KEY",
     "api_key_required": true,
@@ -19,7 +21,9 @@
   },
   {
     "id": "anthropic",
-    "aliases": ["claude"],
+    "aliases": [
+      "claude"
+    ],
     "protocol": "anthropic",
     "api_key_env": "ANTHROPIC_API_KEY",
     "api_key_required": true,
@@ -52,7 +56,10 @@
   },
   {
     "id": "openai_compatible",
-    "aliases": ["openai-compatible", "compatible"],
+    "aliases": [
+      "openai-compatible",
+      "compatible"
+    ],
     "protocol": "open_ai_completions",
     "base_url_env": "LLM_BASE_URL",
     "base_url_required": true,
@@ -89,7 +96,9 @@
   },
   {
     "id": "openrouter",
-    "aliases": ["open_router"],
+    "aliases": [
+      "open_router"
+    ],
     "protocol": "open_ai_completions",
     "default_base_url": "https://openrouter.ai/api/v1",
     "api_key_env": "OPENROUTER_API_KEY",
@@ -126,7 +135,10 @@
   },
   {
     "id": "nvidia",
-    "aliases": ["nvidia_nim", "nim"],
+    "aliases": [
+      "nvidia_nim",
+      "nim"
+    ],
     "protocol": "open_ai_completions",
     "default_base_url": "https://integrate.api.nvidia.com/v1",
     "api_key_env": "NVIDIA_API_KEY",
@@ -144,7 +156,10 @@
   },
   {
     "id": "venice",
-    "aliases": ["venice_ai", "veniceai"],
+    "aliases": [
+      "venice_ai",
+      "veniceai"
+    ],
     "protocol": "open_ai_completions",
     "default_base_url": "https://api.venice.ai/api/v1",
     "api_key_env": "VENICE_API_KEY",
@@ -162,7 +177,10 @@
   },
   {
     "id": "together",
-    "aliases": ["together_ai", "togetherai"],
+    "aliases": [
+      "together_ai",
+      "togetherai"
+    ],
     "protocol": "open_ai_completions",
     "default_base_url": "https://api.together.xyz/v1",
     "api_key_env": "TOGETHER_API_KEY",
@@ -180,7 +198,9 @@
   },
   {
     "id": "fireworks",
-    "aliases": ["fireworks_ai"],
+    "aliases": [
+      "fireworks_ai"
+    ],
     "protocol": "open_ai_completions",
     "default_base_url": "https://api.fireworks.ai/inference/v1",
     "api_key_env": "FIREWORKS_API_KEY",
@@ -198,7 +218,9 @@
   },
   {
     "id": "deepseek",
-    "aliases": ["deep_seek"],
+    "aliases": [
+      "deep_seek"
+    ],
     "protocol": "open_ai_completions",
     "default_base_url": "https://api.deepseek.com/v1",
     "api_key_env": "DEEPSEEK_API_KEY",
@@ -234,7 +256,9 @@
   },
   {
     "id": "sambanova",
-    "aliases": ["samba_nova"],
+    "aliases": [
+      "samba_nova"
+    ],
     "protocol": "open_ai_completions",
     "default_base_url": "https://api.sambanova.ai/v1",
     "api_key_env": "SAMBANOVA_API_KEY",
@@ -249,5 +273,131 @@
       "display_name": "SambaNova",
       "can_list_models": false
     }
+  },
+  {
+    "id": "gemini",
+    "aliases": [
+      "google_gemini",
+      "google"
+    ],
+    "protocol": "open_ai_completions",
+    "default_base_url": "https://generativelanguage.googleapis.com/v1beta/openai",
+    "api_key_env": "GEMINI_API_KEY",
+    "api_key_required": true,
+    "model_env": "GEMINI_MODEL",
+    "default_model": "gemini-2.5-flash",
+    "description": "Google Gemini (via OpenAI-compatible endpoint)",
+    "setup": {
+      "kind": "api_key",
+      "secret_name": "llm_gemini_api_key",
+      "key_url": "https://aistudio.google.com/app/apikey",
+      "display_name": "Google Gemini",
+      "can_list_models": true
+    }
+  },
+  {
+    "id": "bedrock",
+    "aliases": [
+      "aws_bedrock",
+      "aws"
+    ],
+    "protocol": "open_ai_completions",
+    "api_key_env": "BEDROCK_ACCESS_KEY",
+    "api_key_required": false,
+    "base_url_env": "BEDROCK_BASE_URL",
+    "model_env": "BEDROCK_MODEL",
+    "default_model": "anthropic.claude-3-5-sonnet-20241022-v2:0",
+    "description": "AWS Bedrock (requires LiteLLM or OpenAI-compatible proxy)",
+    "setup": {
+      "kind": "open_ai_compatible",
+      "secret_name": "llm_bedrock_api_key",
+      "display_name": "AWS Bedrock",
+      "can_list_models": false
+    }
+  },
+  {
+    "id": "ionet",
+    "aliases": [
+      "io_net",
+      "io.net"
+    ],
+    "protocol": "open_ai_completions",
+    "default_base_url": "https://api.intelligence.io.solutions/api/v1",
+    "api_key_env": "IONET_API_KEY",
+    "api_key_required": true,
+    "model_env": "IONET_MODEL",
+    "default_model": "deepseek-coder-v2-instruct",
+    "description": "io.net Intelligence API",
+    "setup": {
+      "kind": "api_key",
+      "secret_name": "llm_ionet_api_key",
+      "key_url": "https://cloud.io.net/intelligence",
+      "display_name": "io.net",
+      "can_list_models": true
+    }
+  },
+  {
+    "id": "mistral",
+    "aliases": [
+      "mistral_ai",
+      "mistralai"
+    ],
+    "protocol": "open_ai_completions",
+    "default_base_url": "https://api.mistral.ai/v1",
+    "api_key_env": "MISTRAL_API_KEY",
+    "api_key_required": true,
+    "model_env": "MISTRAL_MODEL",
+    "default_model": "mistral-large-latest",
+    "description": "Mistral AI API",
+    "setup": {
+      "kind": "api_key",
+      "secret_name": "llm_mistral_api_key",
+      "key_url": "https://console.mistral.ai/api-keys",
+      "display_name": "Mistral",
+      "can_list_models": true
+    }
+  },
+  {
+    "id": "yandex",
+    "aliases": [
+      "yandex_ai_studio",
+      "yandexgpt",
+      "yandex_gpt"
+    ],
+    "protocol": "open_ai_completions",
+    "default_base_url": "https://ai.api.cloud.yandex.net/v1",
+    "api_key_env": "YANDEX_API_KEY",
+    "api_key_required": true,
+    "model_env": "YANDEX_MODEL",
+    "extra_headers_env": "YANDEX_EXTRA_HEADERS",
+    "default_model": "yandexgpt-lite",
+    "description": "Yandex AI Studio (YandexGPT)",
+    "setup": {
+      "kind": "api_key",
+      "secret_name": "llm_yandex_api_key",
+      "key_url": "https://aistudio.yandex.ru/platform/folders/",
+      "display_name": "Yandex AI Studio",
+      "can_list_models": true
+    }
+  },
+  {
+    "id": "cloudflare",
+    "aliases": [
+      "cloudflare_ai",
+      "cf_ai"
+    ],
+    "protocol": "open_ai_completions",
+    "api_key_env": "CLOUDFLARE_API_KEY",
+    "api_key_required": true,
+    "base_url_env": "CLOUDFLARE_BASE_URL",
+    "model_env": "CLOUDFLARE_MODEL",
+    "default_model": "@cf/meta/llama-3.3-70b-instruct-fp8-fast",
+    "description": "Cloudflare Workers AI",
+    "setup": {
+      "kind": "open_ai_compatible",
+      "secret_name": "llm_cloudflare_api_key",
+      "display_name": "Cloudflare Workers AI",
+      "can_list_models": false
+    }
   }
-]
+]

registry/channels/discord.json

@@ -19,7 +19,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/discord-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "030707431717bca3411a48f311c6ab5f92a45c747de26cafe4f6e3e23a8b3b2d"
     }
   },
   "auth_summary": {

registry/channels/slack.json

@@ -19,7 +19,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/slack-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "6ed36077b67ac70a041f06f760f93ba79b33269885413c3c3f2c8c87ee60807e"
     }
   },
   "auth_summary": {

registry/channels/telegram.json

@@ -19,7 +19,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/telegram-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "98c86895a9c4b0a1e19fe8a47f1ccbfe7e972e112b05e584bc897130dc32283a"
     }
   },
   "auth_summary": {

registry/channels/whatsapp.json

@@ -19,7 +19,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/whatsapp-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "bd35cad18d87292ea8d2f52db9b514ed9f814a414de910f59073d475c26c4c14"
     }
   },
   "auth_summary": {

registry/tools/github.json

@@ -20,7 +20,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/github-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "6fcd32719a4ff15641a4b50fff8984686550f0c491dce60518f4126857d0c544"
     }
   },
   "auth_summary": {

registry/tools/gmail.json

@@ -19,7 +19,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/gmail-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "023da7000b17568bf0e64b2e5013c8a042b2f323c85f1632339231c73d500e39"
     }
   },
   "auth_summary": {

registry/tools/google-calendar.json

@@ -19,7 +19,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/google-calendar-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "fc42277b65881d6e9bcc5403dc54c7f5b3ddeaaaf04617fce2c5da05d76325f0"
     }
   },
   "auth_summary": {

registry/tools/google-docs.json

@@ -19,7 +19,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/google-docs-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "385c04abd1e6b8011ccc330e1f4bd7ce58577e488959b51594aa04eb26cbe7cc"
     }
   },
   "auth_summary": {

registry/tools/google-drive.json

@@ -19,7 +19,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/google-drive-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "1b107d575a5d52cc8c76d9a681802190f4373fb485f7f54f445533f097fa37c0"
     }
   },
   "auth_summary": {

registry/tools/google-sheets.json

@@ -19,7 +19,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/google-sheets-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "c4f6b1e8c5126ac2c8a4b98e4283a3afa32223d2488fc3c3a609758c0c9beb90"
     }
   },
   "auth_summary": {

registry/tools/google-slides.json

@@ -18,7 +18,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/google-slides-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "7110b8565340c888e51f99e9c013bf4de8f8a7f7b33bace00eb8fc47831ff20b"
     }
   },
   "auth_summary": {

registry/tools/slack.json

@@ -18,7 +18,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/slack-tool-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "6ed36077b67ac70a041f06f760f93ba79b33269885413c3c3f2c8c87ee60807e"
     }
   },
   "auth_summary": {

registry/tools/telegram.json

@@ -19,7 +19,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/telegram-mtproto-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "98c86895a9c4b0a1e19fe8a47f1ccbfe7e972e112b05e584bc897130dc32283a"
     }
   },
   "auth_summary": {

registry/tools/web-search.json

@@ -19,7 +19,7 @@
   "artifacts": {
     "wasm32-wasip2": {
       "url": "https://github.com/nearai/ironclaw/releases/latest/download/web-search-wasm32-wasip2.tar.gz",
-      "sha256": null
+      "sha256": "66cb2b9b00652385e9f30f17c74902b9222c17c53e9d3bd1ef42f5cab705bcf6"
     }
   },
   "auth_summary": {

scripts/dev-setup.sh

@@ -51,9 +51,11 @@ echo "[6/6] Installing git hooks..."
 HOOKS_DIR=$(git rev-parse --git-path hooks 2>/dev/null) || true
 if [ -n "$HOOKS_DIR" ]; then
     mkdir -p "$HOOKS_DIR"
-    SCRIPT_ABS="$(cd "$(dirname "$0")" && pwd)/commit-msg-regression.sh"
-    ln -sf "$SCRIPT_ABS" "$HOOKS_DIR/commit-msg"
+    SCRIPTS_ABS="$(cd "$(dirname "$0")" && pwd)"
+    ln -sf "$SCRIPTS_ABS/commit-msg-regression.sh" "$HOOKS_DIR/commit-msg"
     echo "  commit-msg hook installed (regression test enforcement)"
+    ln -sf "$SCRIPTS_ABS/pre-commit-safety.sh" "$HOOKS_DIR/pre-commit"
+    echo "  pre-commit hook installed (UTF-8, case-sensitivity, /tmp, redaction checks)"
 else
     echo "  Skipped: not a git repository"
 fi

scripts/pre-commit-safety.sh

@@ -0,0 +1,136 @@
+#!/usr/bin/env bash
+# Pre-commit safety checks for common issues caught by AI code reviewers.
+#
+# Can be run standalone: bash scripts/pre-commit-safety.sh
+# Or installed as a git pre-commit hook via dev-setup.sh.
+#
+# Checks staged .rs files for:
+#   1. Unsafe UTF-8 byte slicing (panics on multi-byte chars)
+#   2. Case-sensitive file extension comparisons
+#   3. Hardcoded /tmp paths in tests (flaky in parallel runs)
+#   4. Tool parameters logged without redaction (secret leaks)
+#   5. Multi-step DB operations without transaction wrapping
+#
+# Suppress individual lines with an inline "// safety: <reason>" comment.
+
+set -euo pipefail
+
+# Determine a suitable base ref for standalone diffs.
+resolve_base_ref() {
+    local candidates=(
+        "@{upstream}"
+        "origin/HEAD"
+        "origin/main"
+        "origin/master"
+        "main"
+        "master"
+    )
+
+    for ref in "${candidates[@]}"; do
+        if git rev-parse --verify --quiet "$ref" >/dev/null 2>&1; then
+            echo "$ref"
+            return 0
+        fi
+    done
+
+    echo "pre-commit-safety: could not determine a base Git ref for diff (tried: ${candidates[*]})." >&2
+    echo "pre-commit-safety: ensure your repository has an upstream or a local main/master branch." >&2
+    exit 1
+}
+
+# Support both pre-commit hook (staged files) and standalone (all changed vs base)
+if git diff --cached --quiet 2>/dev/null; then
+    # No staged changes -- compare working tree against a resolved base ref
+    BASE_REF="$(resolve_base_ref)"
+    DIFF_OUTPUT=$(git diff "$BASE_REF" -- '*.rs' 2>/dev/null || true)
+else
+    DIFF_OUTPUT=$(git diff --cached -U0 -- '*.rs' 2>/dev/null || true)
+fi
+
+# Early exit if there are no relevant .rs changes
+if [ -z "$DIFF_OUTPUT" ]; then
+    exit 0
+fi
+
+WARNINGS=0
+
+warn() {
+    if [ "$WARNINGS" -eq 0 ]; then
+        echo ""
+        echo "=== Pre-commit Safety Checks ==="
+        echo ""
+    fi
+    WARNINGS=$((WARNINGS + 1))
+    echo "  [$1] $2"
+}
+
+# 1. Unsafe UTF-8 byte slicing: &s[..N] or &s[..some_var] on strings
+#    Safe patterns: is_char_boundary, char_indices, // safety:
+if echo "$DIFF_OUTPUT" | grep -nE '^\+' | grep -E '\[\.\..*\]' | grep -vE 'is_char_boundary|char_indices|// safety:|as_bytes|Vec<|&\[u8\]|\[u8\]|bytes\(\)|&bytes' | head -3 | grep -q .; then
+    warn "UTF8" "Possible unsafe byte-index string slicing. Use is_char_boundary() or char_indices()."
+    echo "$DIFF_OUTPUT" | grep -nE '^\+' | grep -E '\[\.\..*\]' | grep -vE 'is_char_boundary|char_indices|// safety:|as_bytes|Vec<|&\[u8\]|\[u8\]|bytes\(\)|&bytes' | head -3 | sed 's/^/    /'
+fi
+
+# 2. Case-sensitive file extension checks
+#    Match: .ends_with(".png") without prior to_lowercase
+if echo "$DIFF_OUTPUT" | grep -nE '^\+.*ends_with\("\.([pP][nN][gG]|[jJ][pP][eE]?[gG]|[gG][iI][fF]|[wW][eE][bB][pP]|[mM][dD])"\)' | grep -vE 'to_lowercase|to_ascii_lowercase|// safety:' | head -3 | grep -q .; then
+    warn "CASE" "Case-sensitive file extension comparison. Normalize to lowercase first."
+    echo "$DIFF_OUTPUT" | grep -nE '^\+.*ends_with\("\.([pP][nN][gG]|[jJ][pP][eE]?[gG]|[gG][iI][fF]|[wW][eE][bB][pP]|[mM][dD])"\)' | grep -vE 'to_lowercase|to_ascii_lowercase|// safety:' | head -3 | sed 's/^/    /'
+fi
+
+# 3. Hardcoded /tmp paths in test files
+if echo "$DIFF_OUTPUT" | grep -nE '^\+.*"/tmp/' | grep -vE 'tempfile|tempdir|// safety:' | head -3 | grep -q .; then
+    warn "TMPDIR" "Hardcoded /tmp path. Use tempfile::tempdir() for parallel-safe tests."
+    echo "$DIFF_OUTPUT" | grep -nE '^\+.*"/tmp/' | grep -vE 'tempfile|tempdir|// safety:' | head -3 | sed 's/^/    /'
+fi
+
+# 4. Logging tool parameters without redaction
+if echo "$DIFF_OUTPUT" | grep -nE '^\+.*tracing::(info|debug|warn|error).*param' | grep -vE 'redact|// safety:' | head -3 | grep -q .; then
+    warn "REDACT" "Logging tool parameters without redaction. Use redact_params() first."
+    echo "$DIFF_OUTPUT" | grep -nE '^\+.*tracing::(info|debug|warn|error).*param' | grep -vE 'redact|// safety:' | head -3 | sed 's/^/    /'
+fi
+
+# 5. Multi-step DB operations without transaction
+#    Uses -W (function context) to reduce false positives from existing transactions.
+#    Suppressible with "// safety:" in the hunk.
+DIFF_W_OUTPUT=$(git diff --cached -W -- '*.rs' 2>/dev/null || git diff "$(resolve_base_ref)" -W -- '*.rs' 2>/dev/null || true)
+if [ -n "$DIFF_W_OUTPUT" ]; then
+    HUNK_COUNT=$(echo "$DIFF_W_OUTPUT" | awk '
+        /^@@/ {
+            if (count >= 2 && !has_tx && !has_safety) found++
+            count=0; has_tx=0; has_safety=0
+        }
+        /^\+.*\.(execute|query)\(/ { count++ }
+        /^\+.*(transaction|\.tx\.|\.begin\()/ { has_tx=1 }
+        / .*(transaction|\.tx\.|\.begin\()/ { has_tx=1 }
+        /\/\/ safety:/ { has_safety=1 }
+        END {
+            if (count >= 2 && !has_tx && !has_safety) found++
+            print found+0
+        }
+    ')
+    if [ "$HUNK_COUNT" -gt 0 ]; then
+        warn "TX" "Multiple DB operations in same function without transaction. Wrap in a transaction for atomicity."
+        echo "$DIFF_W_OUTPUT" | awk '
+            /^@@/ {
+                if (count >= 2 && !has_tx && !has_safety) { print buf }
+                buf=""; count=0; has_tx=0; has_safety=0
+            }
+            /^\+.*\.(execute|query)\(/ { count++ }
+            /^\+.*(transaction|\.tx\.|\.begin\()/ { has_tx=1 }
+            / .*(transaction|\.tx\.|\.begin\()/ { has_tx=1 }
+            /\/\/ safety:/ { has_safety=1 }
+            { buf = buf "\n" $0 }
+            END {
+                if (count >= 2 && !has_tx && !has_safety) { print buf }
+            }
+        ' | grep -E '^\+.*\.(execute|query)\(' | head -4 | sed 's/^/    /'
+    fi
+fi
+
+if [ "$WARNINGS" -gt 0 ]; then
+    echo ""
+    echo "Found $WARNINGS potential issue(s). Fix them or add '// safety: <reason>' to suppress."
+    echo ""
+    exit 1
+fi

skills/review-checklist/SKILL.md

@@ -0,0 +1,54 @@
+---
+name: review-checklist
+version: 0.1.0
+description: Pre-merge review checklist based on recurring AI reviewer feedback patterns
+activation:
+  patterns:
+    - "review.*checklist"
+    - "ready to merge"
+    - "pre-merge check"
+    - "check.*before.*merge"
+  keywords:
+    - review
+    - checklist
+    - merge
+    - pre-merge
+  max_context_tokens: 1500
+---
+
+# Pre-Merge Review Checklist
+
+Before merging, verify these items. They represent the most common issues caught by automated code reviewers (Copilot, Gemini) on IronClaw PRs.
+
+## Database Operations
+- [ ] Multi-step DB operations are wrapped in transactions (INSERT+INSERT, UPDATE+DELETE, read-modify-write)
+- [ ] Both postgres AND libsql backends updated for any new Database trait methods
+- [ ] Migrations are atomic (SQL execution + version recording in same transaction)
+
+## Security & Data Safety
+- [ ] Tool parameters are redacted via `redact_params()` before logging or SSE/WebSocket broadcast
+- [ ] URL validation resolves DNS before checking for private/loopback IPs (anti-SSRF via DNS rebinding)
+- [ ] Destructive tools have `requires_approval()` returning `Always` or `UnlessAutoApproved`
+- [ ] Data from worker containers is treated as untrusted (tool domain checks, server-side nesting depth)
+- [ ] No secrets or credentials in error messages, logs, or SSE events
+
+## String Safety
+- [ ] No byte-index slicing (`&s[..n]`) on external/user strings -- use `is_char_boundary()` or `char_indices()`
+- [ ] File extension and media type comparisons are case-insensitive (`.to_ascii_lowercase()` before matching)
+- [ ] Path comparisons are case-insensitive where needed (macOS/Windows filesystems)
+
+## Trait Wrappers & Decorator Chain
+- [ ] New `LlmProvider` trait methods are delegated in ALL wrapper types (grep `impl LlmProvider for`)
+- [ ] New trait methods are tested through the full decorator/provider chain, not just the base impl
+- [ ] Default trait method implementations are intentional -- wrappers that silently return defaults are bugs
+
+## Tests
+- [ ] Temporary files/dirs use `tempfile` crate, no hardcoded `/tmp/` paths
+- [ ] Tests don't mutate global statics without synchronization (use per-test state or `serial_test`)
+- [ ] Tests don't make real network requests (use mocks, stubs, or RFC 5737 TEST-NET IPs like 192.0.2.1)
+- [ ] Test names and comments match actual test behavior and assertions
+
+## Comments & Documentation
+- [ ] Code comments match actual behavior (especially route paths, tool names, function semantics)
+- [ ] Spec/README files updated if module behavior changed
+- [ ] Error messages are clear and non-redundant (don't nest tool name inside tool error that already contains it)

src/agent/agent_loop.rs

@@ -96,6 +96,9 @@ pub struct Agent {
     pub(super) heartbeat_config: Option<HeartbeatConfig>,
     pub(super) hygiene_config: Option<crate::config::HygieneConfig>,
     pub(super) routine_config: Option<RoutineConfig>,
+    /// Optional slot to expose the routine engine to the gateway for manual triggering.
+    pub(super) routine_engine_slot:
+        Option<Arc<tokio::sync::RwLock<Option<Arc<crate::agent::routine_engine::RoutineEngine>>>>>,
 }
 
 impl Agent {
@@ -148,9 +151,18 @@ impl Agent {
             heartbeat_config,
             hygiene_config,
             routine_config,
+            routine_engine_slot: None,
         }
     }
 
+    /// Set the routine engine slot for exposing the engine to the gateway.
+    pub fn set_routine_engine_slot(
+        &mut self,
+        slot: Arc<tokio::sync::RwLock<Option<Arc<crate::agent::routine_engine::RoutineEngine>>>>,
+    ) {
+        self.routine_engine_slot = Some(slot);
+    }
+
     // Convenience accessors
 
     /// Get the scheduler (for external wiring, e.g. CreateJobTool).
@@ -342,8 +354,19 @@ impl Agent {
         let heartbeat_handle = if let Some(ref hb_config) = self.heartbeat_config {
             if hb_config.enabled {
                 if let Some(workspace) = self.workspace() {
-                    let config = AgentHeartbeatConfig::default()
+                    let mut config = AgentHeartbeatConfig::default()
                         .with_interval(std::time::Duration::from_secs(hb_config.interval_secs));
+                    config.quiet_hours_start = hb_config.quiet_hours_start;
+                    config.quiet_hours_end = hb_config.quiet_hours_end;
+                    config.timezone = hb_config
+                        .timezone
+                        .clone()
+                        .or_else(|| Some(self.config.default_timezone.clone()));
+                    if let (Some(user), Some(channel)) =
+                        (&hb_config.notify_user, &hb_config.notify_channel)
+                    {
+                        config = config.with_notify(user, channel);
+                    }
 
                     // Set up notification channel
                     let (notify_tx, mut notify_rx) =
@@ -394,8 +417,8 @@ impl Agent {
                         hygiene,
                         workspace.clone(),
                         self.cheap_llm().clone(),
-                        self.safety().clone(),
                         Some(notify_tx),
+                        self.store().map(Arc::clone),
                     ))
                 } else {
                     tracing::warn!("Heartbeat enabled but no workspace available");
@@ -486,6 +509,11 @@ impl Agent {
                     // SAFETY: self is consumed by run(), we can smuggle the engine in
                     // via a local to use in the message loop below.
 
+                    // Expose engine to gateway for manual triggering
+                    if let Some(ref slot) = self.routine_engine_slot {
+                        *slot.write().await = Some(Arc::clone(&engine));
+                    }
+
                     tracing::info!(
                         "Routines enabled: cron ticker every {}s, max {} concurrent",
                         rt_config.cron_check_interval_secs,

src/agent/commands.rs

@@ -345,7 +345,6 @@ impl Agent {
             crate::workspace::hygiene::HygieneConfig::default(),
             workspace.clone(),
             self.llm().clone(),
-            self.safety().clone(),
         );
 
         match runner.check_heartbeat().await {
@@ -406,7 +405,7 @@ impl Agent {
             .with_max_tokens(512)
             .with_temperature(0.3);
 
-        let reasoning = Reasoning::new(self.llm().clone(), self.safety().clone());
+        let reasoning = Reasoning::new(self.llm().clone());
         match reasoning.complete(request).await {
             Ok((text, _usage)) => Ok(SubmissionResult::response(format!(
                 "Thread Summary:\n\n{}",
@@ -454,7 +453,7 @@ impl Agent {
             .with_max_tokens(512)
             .with_temperature(0.5);
 
-        let reasoning = Reasoning::new(self.llm().clone(), self.safety().clone());
+        let reasoning = Reasoning::new(self.llm().clone());
         match reasoning.complete(request).await {
             Ok((text, _usage)) => Ok(SubmissionResult::response(format!(
                 "Suggested Next Steps:\n\n{}",
@@ -663,10 +662,14 @@ impl Agent {
                     }
 
                     match self.llm().set_model(requested) {
-                        Ok(()) => Ok(SubmissionResult::response(format!(
-                            "Switched model to: {}",
-                            requested
-                        ))),
+                        Ok(()) => {
+                            // Persist the model choice so it survives restarts.
+                            self.persist_selected_model(requested).await;
+                            Ok(SubmissionResult::response(format!(
+                                "Switched model to: {}",
+                                requested
+                            )))
+                        }
                         Err(e) => Ok(SubmissionResult::error(format!(
                             "Failed to switch model: {}",
                             e
@@ -822,4 +825,42 @@ impl Agent {
             _ => Ok(None),
         }
     }
+
+    /// Persist the selected model to the settings store (DB and/or TOML config).
+    ///
+    /// Best-effort: logs warnings on failure but does not propagate errors,
+    /// since the in-memory model switch already succeeded.
+    async fn persist_selected_model(&self, model: &str) {
+        // 1. Persist to DB if available.
+        if let Some(store) = self.store() {
+            let value = serde_json::Value::String(model.to_string());
+            if let Err(e) = store.set_setting("default", "selected_model", &value).await {
+                tracing::warn!("Failed to persist model to DB: {}", e);
+            }
+        }
+
+        // 2. Update TOML config file if it exists (sync I/O in spawn_blocking).
+        let model_owned = model.to_string();
+        if let Err(e) = tokio::task::spawn_blocking(move || {
+            let toml_path = crate::settings::Settings::default_toml_path();
+            match crate::settings::Settings::load_toml(&toml_path) {
+                Ok(Some(mut settings)) => {
+                    settings.selected_model = Some(model_owned);
+                    if let Err(e) = settings.save_toml(&toml_path) {
+                        tracing::warn!("Failed to persist model to config.toml: {}", e);
+                    }
+                }
+                Ok(None) => {
+                    // No config file on disk; nothing to update.
+                }
+                Err(e) => {
+                    tracing::warn!("Failed to load config.toml for model persistence: {}", e);
+                }
+            }
+        })
+        .await
+        {
+            tracing::warn!("Model TOML persistence task failed: {}", e);
+        }
+    }
 }

src/agent/compaction.rs

@@ -13,7 +13,6 @@ use crate::agent::context_monitor::{CompactionStrategy, ContextBreakdown};
 use crate::agent::session::Thread;
 use crate::error::Error;
 use crate::llm::{ChatMessage, CompletionRequest, LlmProvider, Reasoning};
-use crate::safety::SafetyLayer;
 use crate::workspace::Workspace;
 
 /// Result of a compaction operation.
@@ -34,13 +33,12 @@ pub struct CompactionResult {
 /// Compacts conversation context to stay within limits.
 pub struct ContextCompactor {
     llm: Arc<dyn LlmProvider>,
-    safety: Arc<SafetyLayer>,
 }
 
 impl ContextCompactor {
     /// Create a new context compactor.
-    pub fn new(llm: Arc<dyn LlmProvider>, safety: Arc<SafetyLayer>) -> Self {
-        Self { llm, safety }
+    pub fn new(llm: Arc<dyn LlmProvider>) -> Self {
+        Self { llm }
     }
 
     /// Compact a thread's context using the given strategy.
@@ -233,7 +231,7 @@ Be brief but capture all important details. Use bullet points."#,
             .with_max_tokens(1024)
             .with_temperature(0.3);
 
-        let reasoning = Reasoning::new(self.llm.clone(), self.safety.clone());
+        let reasoning = Reasoning::new(self.llm.clone());
         let (text, _) = reasoning.complete(request).await?;
         Ok(text)
     }
@@ -346,17 +344,11 @@ mod tests {
     // === QA Plan - Compaction strategy tests ===
 
     use crate::agent::context_monitor::CompactionStrategy;
-    use crate::config::SafetyConfig;
-    use crate::safety::SafetyLayer;
     use crate::testing::StubLlm;
 
     /// Helper: build a `ContextCompactor` with the given `StubLlm`.
     fn make_compactor(llm: Arc<StubLlm>) -> ContextCompactor {
-        let safety = Arc::new(SafetyLayer::new(&SafetyConfig {
-            max_output_length: 100_000,
-            injection_check_enabled: false,
-        }));
-        ContextCompactor::new(llm, safety)
+        ContextCompactor::new(llm)
     }
 
     /// Helper: build a thread with `n` completed turns.

src/agent/cost_guard.rs

@@ -131,10 +131,12 @@ impl CostGuard {
         // Check hourly rate
         if let Some(limit) = self.config.max_actions_per_hour {
             let mut window = self.action_window.lock().await;
-            let cutoff = Instant::now() - std::time::Duration::from_secs(3600);
-            // Drain expired entries
-            while window.front().is_some_and(|t| *t < cutoff) {
-                window.pop_front();
+            // checked_sub avoids panic when system uptime < 1 hour (Windows)
+            if let Some(cutoff) = Instant::now().checked_sub(std::time::Duration::from_secs(3600)) {
+                // Drain expired entries
+                while window.front().is_some_and(|t| *t < cutoff) {
+                    window.pop_front();
+                }
             }
             let count = window.len() as u64;
             if count >= limit {
@@ -260,9 +262,11 @@ impl CostGuard {
     /// Number of actions in the current hourly window.
     pub async fn actions_this_hour(&self) -> u64 {
         let mut window = self.action_window.lock().await;
-        let cutoff = Instant::now() - std::time::Duration::from_secs(3600);
-        while window.front().is_some_and(|t| *t < cutoff) {
-            window.pop_front();
+        // checked_sub avoids panic when system uptime < 1 hour (Windows)
+        if let Some(cutoff) = Instant::now().checked_sub(std::time::Duration::from_secs(3600)) {
+            while window.front().is_some_and(|t| *t < cutoff) {
+                window.pop_front();
+            }
         }
         window.len() as u64
     }
@@ -621,4 +625,35 @@ mod tests {
             "surcharge should be 100% of input cost for 1h cache writes"
         );
     }
+
+    /// Regression test for #657: Instant::now() - Duration panics on Windows
+    /// when system uptime is less than the subtracted duration.
+    #[tokio::test]
+    async fn test_checked_sub_no_panic_on_fresh_guard() {
+        // A fresh CostGuard with rate limits should not panic even if
+        // checked_sub returns None (simulating short uptime).
+        let guard = CostGuard::new(CostGuardConfig {
+            max_cost_per_day_cents: None,
+            max_actions_per_hour: Some(100),
+        });
+
+        // These must not panic regardless of system uptime
+        assert!(guard.check_allowed().await.is_ok());
+        assert_eq!(guard.actions_this_hour().await, 0);
+
+        // Record some actions and verify again
+        guard
+            .record_llm_call("gpt-4o", 10, 10, 0, 0, Decimal::ONE, Decimal::ONE, None)
+            .await;
+        assert!(guard.check_allowed().await.is_ok());
+        assert_eq!(guard.actions_this_hour().await, 1);
+    }
+
+    /// Verify that checked_sub itself behaves as expected for the pattern we use.
+    #[test]
+    fn test_instant_checked_sub_returns_none_for_overflow() {
+        // Duration::MAX will always exceed uptime, so checked_sub must return None
+        let result = Instant::now().checked_sub(std::time::Duration::MAX);
+        assert!(result.is_none());
+    }
 }

src/agent/dispatcher.rs

@@ -50,8 +50,18 @@ impl Agent {
 
         // Load workspace system prompt (identity files: AGENTS.md, SOUL.md, etc.)
         // In group chats, MEMORY.md is excluded to prevent leaking personal context.
+        // Resolve the user's timezone
+        let user_tz = crate::timezone::resolve_timezone(
+            message.timezone.as_deref(),
+            None, // user setting lookup can be added later
+            &self.config.default_timezone,
+        );
+
         let system_prompt = if let Some(ws) = self.workspace() {
-            match ws.system_prompt_for_context(is_group_chat).await {
+            match ws
+                .system_prompt_for_context_tz(is_group_chat, user_tz)
+                .await
+            {
                 Ok(prompt) if !prompt.is_empty() => Some(prompt),
                 Ok(_) => None,
                 Err(e) => {
@@ -103,7 +113,7 @@ impl Agent {
             None
         };
 
-        let mut reasoning = Reasoning::new(self.llm().clone(), self.safety().clone())
+        let mut reasoning = Reasoning::new(self.llm().clone())
             .with_channel(message.channel.clone())
             .with_model_name(self.llm().active_model_name())
             .with_group_chat(is_group_chat);
@@ -130,6 +140,7 @@ impl Agent {
         let mut job_ctx =
             JobContext::with_user(&message.user_id, "chat", "Interactive chat session");
         job_ctx.http_interceptor = self.deps.http_interceptor.clone();
+        job_ctx.user_timezone = user_tz.name().to_string();
 
         // Build system prompts once for this turn. Two variants: with tools
         // (normal iterations) and without (force_text final iteration).
@@ -785,6 +796,7 @@ impl Agent {
                             tool_call_id: tc.id.clone(),
                             context_messages: context_messages.clone(),
                             deferred_tool_calls: tool_calls[approval_idx + 1..].to_vec(),
+                            user_timezone: Some(user_tz.name().to_string()),
                         };
 
                         return Ok(AgenticLoopResult::NeedApproval { pending });
@@ -1146,6 +1158,7 @@ mod tests {
                 max_actions_per_hour: None,
                 max_tool_iterations: 50,
                 auto_approve_tools: false,
+                default_timezone: "UTC".to_string(),
             },
             deps,
             Arc::new(ChannelManager::new()),
@@ -1248,6 +1261,7 @@ mod tests {
                     arguments: serde_json::json!({"message": "done"}),
                 },
             ],
+            user_timezone: None,
         };
 
         let json = serde_json::to_string(&pending).expect("serialize");
@@ -1595,12 +1609,8 @@ mod tests {
         use crate::testing::StubLlm;
 
         let stub = Arc::new(StubLlm::failing_non_transient("ctx-bomb"));
-        let safety = Arc::new(SafetyLayer::new(&SafetyConfig {
-            max_output_length: 100_000,
-            injection_check_enabled: false,
-        }));
 
-        let reasoning = Reasoning::new(stub.clone(), safety);
+        let reasoning = Reasoning::new(stub.clone());
 
         // Build a fat context with lots of history.
         let messages = vec![
@@ -1710,11 +1720,7 @@ mod tests {
         use crate::llm::{Reasoning, ReasoningContext, RespondResult, ToolDefinition};
 
         let provider = Arc::new(AlwaysToolCallProvider);
-        let safety = Arc::new(SafetyLayer::new(&SafetyConfig {
-            max_output_length: 100_000,
-            injection_check_enabled: false,
-        }));
-        let reasoning = Reasoning::new(provider, safety);
+        let reasoning = Reasoning::new(provider);
 
         let tool_def = ToolDefinition {
             name: "echo".to_string(),
@@ -1900,6 +1906,7 @@ mod tests {
                 max_actions_per_hour: None,
                 max_tool_iterations,
                 auto_approve_tools: true,
+                default_timezone: "UTC".to_string(),
             },
             deps,
             Arc::new(ChannelManager::new()),
@@ -2015,6 +2022,7 @@ mod tests {
                     max_actions_per_hour: None,
                     max_tool_iterations: max_iter,
                     auto_approve_tools: true,
+                    default_timezone: "UTC".to_string(),
                 },
                 deps,
                 Arc::new(ChannelManager::new()),

src/agent/heartbeat.rs

@@ -29,8 +29,8 @@ use std::time::Duration;
 use tokio::sync::mpsc;
 
 use crate::channels::OutgoingResponse;
+use crate::db::Database;
 use crate::llm::{ChatMessage, CompletionRequest, LlmProvider, Reasoning};
-use crate::safety::SafetyLayer;
 use crate::workspace::Workspace;
 use crate::workspace::hygiene::HygieneConfig;
 
@@ -47,6 +47,12 @@ pub struct HeartbeatConfig {
     pub notify_user_id: Option<String>,
     /// Channel to notify on heartbeat findings.
     pub notify_channel: Option<String>,
+    /// Hour (0-23) when quiet hours start.
+    pub quiet_hours_start: Option<u32>,
+    /// Hour (0-23) when quiet hours end.
+    pub quiet_hours_end: Option<u32>,
+    /// Timezone for quiet hours evaluation (IANA name).
+    pub timezone: Option<String>,
 }
 
 impl Default for HeartbeatConfig {
@@ -57,6 +63,9 @@ impl Default for HeartbeatConfig {
             max_failures: 3,
             notify_user_id: None,
             notify_channel: None,
+            quiet_hours_start: None,
+            quiet_hours_end: None,
+            timezone: None,
         }
     }
 }
@@ -74,6 +83,26 @@ impl HeartbeatConfig {
         self
     }
 
+    /// Check whether the current time falls within configured quiet hours.
+    pub fn is_quiet_hours(&self) -> bool {
+        use chrono::Timelike;
+        let (Some(start), Some(end)) = (self.quiet_hours_start, self.quiet_hours_end) else {
+            return false;
+        };
+        let tz = self
+            .timezone
+            .as_deref()
+            .and_then(crate::timezone::parse_timezone)
+            .unwrap_or(chrono_tz::UTC);
+        let now_hour = crate::timezone::now_in_tz(tz).hour();
+        if start <= end {
+            now_hour >= start && now_hour < end
+        } else {
+            // Wraps midnight, e.g. 22..06
+            now_hour >= start || now_hour < end
+        }
+    }
+
     /// Set the notification target.
     pub fn with_notify(mut self, user_id: impl Into<String>, channel: impl Into<String>) -> Self {
         self.notify_user_id = Some(user_id.into());
@@ -101,8 +130,8 @@ pub struct HeartbeatRunner {
     hygiene_config: HygieneConfig,
     workspace: Arc<Workspace>,
     llm: Arc<dyn LlmProvider>,
-    safety: Arc<SafetyLayer>,
     response_tx: Option<mpsc::Sender<OutgoingResponse>>,
+    store: Option<Arc<dyn Database>>,
     consecutive_failures: u32,
 }
 
@@ -113,15 +142,14 @@ impl HeartbeatRunner {
         hygiene_config: HygieneConfig,
         workspace: Arc<Workspace>,
         llm: Arc<dyn LlmProvider>,
-        safety: Arc<SafetyLayer>,
     ) -> Self {
         Self {
             config,
             hygiene_config,
             workspace,
             llm,
-            safety,
             response_tx: None,
+            store: None,
             consecutive_failures: 0,
         }
     }
@@ -132,6 +160,12 @@ impl HeartbeatRunner {
         self
     }
 
+    /// Set the database store for persistent heartbeat conversations.
+    pub fn with_store(mut self, store: Arc<dyn Database>) -> Self {
+        self.store = Some(store);
+        self
+    }
+
     /// Run the heartbeat loop.
     ///
     /// This runs forever, checking periodically based on the configured interval.
@@ -153,6 +187,12 @@ impl HeartbeatRunner {
         loop {
             interval.tick().await;
 
+            // Skip during quiet hours
+            if self.config.is_quiet_hours() {
+                tracing::debug!("Heartbeat skipped: quiet hours");
+                continue;
+            }
+
             // Run memory hygiene in the background so it never delays the
             // heartbeat checklist. Failures are logged inside run_if_due.
             let hygiene_workspace = Arc::clone(&self.workspace);
@@ -263,7 +303,7 @@ impl HeartbeatRunner {
             .with_max_tokens(max_tokens)
             .with_temperature(0.3);
 
-        let reasoning = Reasoning::new(self.llm.clone(), self.safety.clone());
+        let reasoning = Reasoning::new(self.llm.clone());
         let (content, _usage) = match reasoning.complete(request).await {
             Ok(r) => r,
             Err(e) => return HeartbeatResult::Failed(format!("LLM call failed: {}", e)),
@@ -292,9 +332,32 @@ impl HeartbeatRunner {
             return;
         };
 
+        let user_id = self.config.notify_user_id.as_deref().unwrap_or("default");
+
+        // Persist to heartbeat conversation and get thread_id
+        let thread_id = if let Some(ref store) = self.store {
+            match store.get_or_create_heartbeat_conversation(user_id).await {
+                Ok(conv_id) => {
+                    if let Err(e) = store
+                        .add_conversation_message(conv_id, "assistant", message)
+                        .await
+                    {
+                        tracing::error!("Failed to persist heartbeat message: {}", e);
+                    }
+                    Some(conv_id.to_string())
+                }
+                Err(e) => {
+                    tracing::error!("Failed to get heartbeat conversation: {}", e);
+                    None
+                }
+            }
+        } else {
+            None
+        };
+
         let response = OutgoingResponse {
             content: format!("🔔 *Heartbeat Alert*\n\n{}", message),
-            thread_id: None,
+            thread_id,
             attachments: Vec::new(),
             metadata: serde_json::json!({
                 "source": "heartbeat",
@@ -354,13 +417,16 @@ pub fn spawn_heartbeat(
     hygiene_config: HygieneConfig,
     workspace: Arc<Workspace>,
     llm: Arc<dyn LlmProvider>,
-    safety: Arc<SafetyLayer>,
     response_tx: Option<mpsc::Sender<OutgoingResponse>>,
+    store: Option<Arc<dyn Database>>,
 ) -> tokio::task::JoinHandle<()> {
-    let mut runner = HeartbeatRunner::new(config, hygiene_config, workspace, llm, safety);
+    let mut runner = HeartbeatRunner::new(config, hygiene_config, workspace, llm);
     if let Some(tx) = response_tx {
         runner = runner.with_response_channel(tx);
     }
+    if let Some(s) = store {
+        runner = runner.with_store(s);
+    }
 
     tokio::spawn(async move {
         runner.run().await;
@@ -495,4 +561,98 @@ mod tests {
         let content = "<!-- comment -->\nActual task here";
         assert!(!is_effectively_empty(content));
     }
+
+    // ==================== quiet hours ====================
+
+    #[test]
+    fn test_quiet_hours_inside() {
+        use chrono::{Timelike, Utc};
+
+        let now_utc = Utc::now();
+        let hour = now_utc.hour();
+        let start = hour;
+        let end = (hour + 1) % 24;
+
+        let config = HeartbeatConfig {
+            quiet_hours_start: Some(start),
+            quiet_hours_end: Some(end),
+            timezone: Some("UTC".to_string()),
+            ..HeartbeatConfig::default()
+        };
+        // Current UTC hour is inside [start, end) by construction
+        assert!(config.is_quiet_hours());
+    }
+
+    #[test]
+    fn test_quiet_hours_outside() {
+        use chrono::{Timelike, Utc};
+
+        let now_utc = Utc::now();
+        let hour = now_utc.hour();
+        let start = (hour + 1) % 24;
+        let end = (hour + 2) % 24;
+
+        let config = HeartbeatConfig {
+            quiet_hours_start: Some(start),
+            quiet_hours_end: Some(end),
+            timezone: Some("UTC".to_string()),
+            ..HeartbeatConfig::default()
+        };
+        // Current UTC hour is outside [start, end) by construction
+        assert!(!config.is_quiet_hours());
+    }
+
+    #[test]
+    fn test_quiet_hours_wraparound_excludes_now() {
+        use chrono::{Timelike, Utc};
+
+        let now_utc = Utc::now();
+        let hour = now_utc.hour();
+        // Window covers all hours except the current one
+        let start = (hour + 1) % 24;
+        let end = hour;
+
+        let config = HeartbeatConfig {
+            quiet_hours_start: Some(start),
+            quiet_hours_end: Some(end),
+            timezone: Some("UTC".to_string()),
+            ..HeartbeatConfig::default()
+        };
+        assert!(!config.is_quiet_hours());
+    }
+
+    #[test]
+    fn test_quiet_hours_none_configured() {
+        let config = HeartbeatConfig::default();
+        assert!(!config.is_quiet_hours());
+    }
+
+    #[test]
+    fn test_quiet_hours_same_start_end() {
+        let config = HeartbeatConfig {
+            quiet_hours_start: Some(10),
+            quiet_hours_end: Some(10),
+            timezone: Some("UTC".to_string()),
+            ..HeartbeatConfig::default()
+        };
+        // start == end means zero-width window, should be false
+        assert!(!config.is_quiet_hours());
+    }
+
+    #[test]
+    fn test_spawn_heartbeat_accepts_store_param() {
+        // Regression: spawn_heartbeat must accept an optional Database store
+        // for persisting heartbeat notifications to a dedicated conversation.
+        // Compile-time check: the 7th parameter is `Option<Arc<dyn Database>>`.
+        #[allow(clippy::type_complexity)]
+        let _fn_ptr: fn(
+            HeartbeatConfig,
+            HygieneConfig,
+            Arc<crate::workspace::Workspace>,
+            Arc<dyn crate::llm::LlmProvider>,
+            Option<tokio::sync::mpsc::Sender<crate::channels::OutgoingResponse>>,
+            Option<Arc<dyn crate::db::Database>>,
+        ) -> tokio::task::JoinHandle<()> = spawn_heartbeat;
+        let _ = _fn_ptr;
+    }
 }

src/agent/routine.rs

@@ -57,7 +57,11 @@ pub struct Routine {
 #[serde(tag = "type", rename_all = "snake_case")]
 pub enum Trigger {
     /// Fire on a cron schedule (e.g. "0 9 * * MON-FRI" or "every 2h").
-    Cron { schedule: String },
+    Cron {
+        schedule: String,
+        #[serde(default)]
+        timezone: Option<String>,
+    },
     /// Fire when a channel message matches a pattern.
     Event {
         /// Optional channel filter (e.g. "telegram", "slack").
@@ -99,7 +103,21 @@ impl Trigger {
                         field: "schedule".into(),
                     })?
                     .to_string();
-                Ok(Trigger::Cron { schedule })
+                let timezone = config
+                    .get("timezone")
+                    .and_then(|v| v.as_str())
+                    .and_then(|tz| {
+                        if crate::timezone::parse_timezone(tz).is_some() {
+                            Some(tz.to_string())
+                        } else {
+                            tracing::warn!(
+                                "Ignoring invalid timezone '{}' from DB for cron trigger",
+                                tz
+                            );
+                            None
+                        }
+                    });
+                Ok(Trigger::Cron { schedule, timezone })
             }
             "event" => {
                 let pattern = config
@@ -137,7 +155,10 @@ impl Trigger {
     /// Serialize trigger-specific config to JSON for DB storage.
     pub fn to_config_json(&self) -> serde_json::Value {
         match self {
-            Trigger::Cron { schedule } => serde_json::json!({ "schedule": schedule }),
+            Trigger::Cron { schedule, timezone } => serde_json::json!({
+                "schedule": schedule,
+                "timezone": timezone,
+            }),
             Trigger::Event { channel, pattern } => serde_json::json!({
                 "pattern": pattern,
                 "channel": channel,
@@ -415,12 +436,25 @@ pub fn content_hash(content: &str) -> u64 {
 }
 
 /// Parse a cron expression and compute the next fire time from now.
-pub fn next_cron_fire(schedule: &str) -> Result<Option<DateTime<Utc>>, RoutineError> {
+///
+/// When `timezone` is provided and valid, the schedule is evaluated in that
+/// timezone and the result is converted back to UTC. Otherwise UTC is used.
+pub fn next_cron_fire(
+    schedule: &str,
+    timezone: Option<&str>,
+) -> Result<Option<DateTime<Utc>>, RoutineError> {
     let cron_schedule =
         cron::Schedule::from_str(schedule).map_err(|e| RoutineError::InvalidCron {
             reason: e.to_string(),
         })?;
-    Ok(cron_schedule.upcoming(Utc).next())
+    if let Some(tz) = timezone.and_then(crate::timezone::parse_timezone) {
+        Ok(cron_schedule
+            .upcoming(tz)
+            .next()
+            .map(|dt| dt.with_timezone(&Utc)))
+    } else {
+        Ok(cron_schedule.upcoming(Utc).next())
+    }
 }
 
 #[cfg(test)]
@@ -433,10 +467,11 @@ mod tests {
     fn test_trigger_roundtrip() {
         let trigger = Trigger::Cron {
             schedule: "0 9 * * MON-FRI".to_string(),
+            timezone: None,
         };
         let json = trigger.to_config_json();
         let parsed = Trigger::from_db("cron", json).expect("parse cron");
-        assert!(matches!(parsed, Trigger::Cron { schedule } if schedule == "0 9 * * MON-FRI"));
+        assert!(matches!(parsed, Trigger::Cron { schedule, .. } if schedule == "0 9 * * MON-FRI"));
     }
 
     #[test]
@@ -509,16 +544,58 @@ mod tests {
     #[test]
     fn test_next_cron_fire_valid() {
         // Every minute should always have a next fire
-        let next = next_cron_fire("* * * * * *").expect("valid cron");
+        let next = next_cron_fire("* * * * * *", None).expect("valid cron");
         assert!(next.is_some());
     }
 
     #[test]
     fn test_next_cron_fire_invalid() {
-        let result = next_cron_fire("not a cron");
+        let result = next_cron_fire("not a cron", None);
         assert!(result.is_err());
     }
 
+    #[test]
+    fn test_trigger_cron_timezone_roundtrip() {
+        let trigger = Trigger::Cron {
+            schedule: "0 9 * * MON-FRI".to_string(),
+            timezone: Some("America/New_York".to_string()),
+        };
+        let json = trigger.to_config_json();
+        let parsed = Trigger::from_db("cron", json).expect("parse cron");
+        assert!(matches!(parsed, Trigger::Cron { schedule, timezone }
+                if schedule == "0 9 * * MON-FRI"
+                && timezone.as_deref() == Some("America/New_York")));
+    }
+
+    #[test]
+    fn test_trigger_cron_no_timezone_backward_compat() {
+        let json = serde_json::json!({"schedule": "0 9 * * *"});
+        let parsed = Trigger::from_db("cron", json).expect("parse cron");
+        assert!(matches!(parsed, Trigger::Cron { timezone, .. } if timezone.is_none()));
+    }
+
+    #[test]
+    fn test_trigger_cron_invalid_timezone_coerced_to_none() {
+        let json = serde_json::json!({"schedule": "0 9 * * *", "timezone": "Fake/Zone"});
+        let parsed = Trigger::from_db("cron", json).expect("parse cron");
+        assert!(
+            matches!(parsed, Trigger::Cron { timezone, .. } if timezone.is_none()),
+            "invalid timezone should be coerced to None"
+        );
+    }
+
+    #[test]
+    fn test_next_cron_fire_with_timezone() {
+        let next_utc = next_cron_fire("0 0 9 * * * *", None)
+            .expect("valid cron")
+            .expect("has next");
+        let next_est = next_cron_fire("0 0 9 * * * *", Some("America/New_York"))
+            .expect("valid cron")
+            .expect("has next");
+        // EST is UTC-5 (or EDT UTC-4), so the UTC result should differ
+        assert_ne!(next_utc, next_est, "timezone should shift the fire time");
+    }
+
     #[test]
     fn test_guardrails_default() {
         let g = RoutineGuardrails::default();
@@ -531,7 +608,8 @@ mod tests {
     fn test_trigger_type_tag() {
         assert_eq!(
             Trigger::Cron {
-                schedule: String::new()
+                schedule: String::new(),
+                timezone: None,
             }
             .type_tag(),
             "cron"

src/agent/routine_engine.rs

@@ -170,7 +170,7 @@ impl RoutineEngine {
                 continue;
             }
 
-            let detail = if let Trigger::Cron { ref schedule } = routine.trigger {
+            let detail = if let Trigger::Cron { ref schedule, .. } = routine.trigger {
                 Some(schedule.clone())
             } else {
                 None
@@ -184,7 +184,11 @@ impl RoutineEngine {
     ///
     /// Bypasses cooldown checks (those only apply to cron/event triggers).
     /// Still enforces enabled check and concurrent run limit.
-    pub async fn fire_manual(&self, routine_id: Uuid) -> Result<Uuid, RoutineError> {
+    pub async fn fire_manual(
+        &self,
+        routine_id: Uuid,
+        user_id: Option<&str>,
+    ) -> Result<Uuid, RoutineError> {
         let routine = self
             .store
             .get_routine(routine_id)
@@ -194,6 +198,13 @@ impl RoutineEngine {
             })?
             .ok_or(RoutineError::NotFound { id: routine_id })?;
 
+        // Enforce ownership when a user_id is provided (gateway calls).
+        if let Some(uid) = user_id
+            && routine.user_id != uid
+        {
+            return Err(RoutineError::NotAuthorized { id: routine_id });
+        }
+
         if !routine.enabled {
             return Err(RoutineError::Disabled {
                 name: routine.name.clone(),
@@ -369,8 +380,12 @@ async fn execute_routine(ctx: EngineContext, routine: Routine, run: RoutineRun)
 
     // Update routine runtime state
     let now = Utc::now();
-    let next_fire = if let Trigger::Cron { ref schedule } = routine.trigger {
-        next_cron_fire(schedule).unwrap_or(None)
+    let next_fire = if let Trigger::Cron {
+        ref schedule,
+        ref timezone,
+    } = routine.trigger
+    {
+        next_cron_fire(schedule, timezone.as_deref()).unwrap_or(None)
     } else {
         None
     };
@@ -396,6 +411,39 @@ async fn execute_routine(ctx: EngineContext, routine: Routine, run: RoutineRun)
         tracing::error!(routine = %routine.name, "Failed to update runtime state: {}", e);
     }
 
+    // Persist routine result to its dedicated conversation thread
+    let thread_id = match ctx
+        .store
+        .get_or_create_routine_conversation(routine.id, &routine.name, &routine.user_id)
+        .await
+    {
+        Ok(conv_id) => {
+            tracing::debug!(
+                routine = %routine.name,
+                routine_id = %routine.id,
+                conversation_id = %conv_id,
+                "Resolved routine conversation thread"
+            );
+            // Record the run result as a conversation message
+            let msg = match (&summary, status) {
+                (Some(s), _) => format!("[{}] {}: {}", run.trigger_type, status, s),
+                (None, _) => format!("[{}] {}", run.trigger_type, status),
+            };
+            if let Err(e) = ctx
+                .store
+                .add_conversation_message(conv_id, "assistant", &msg)
+                .await
+            {
+                tracing::error!(routine = %routine.name, "Failed to persist routine message: {}", e);
+            }
+            Some(conv_id.to_string())
+        }
+        Err(e) => {
+            tracing::error!(routine = %routine.name, "Failed to get routine conversation: {}", e);
+            None
+        }
+    };
+
     // Send notifications based on config
     send_notification(
         &ctx.notify_tx,
@@ -403,6 +451,7 @@ async fn execute_routine(ctx: EngineContext, routine: Routine, run: RoutineRun)
         &routine.name,
         status,
         summary.as_deref(),
+        thread_id.as_deref(),
     )
     .await;
 }
@@ -443,18 +492,13 @@ async fn execute_full_job(
             reason: "scheduler not available".to_string(),
         })?;
 
-    // Set the message tool's default channel/target from the routine's notify config
-    // so the LLM can send results without triggering cross-channel approval.
-    // TODO: This mutates shared global state and can race with concurrent jobs.
-    // Move notify config into JobContext metadata and apply per-job instead.
+    let mut metadata = serde_json::json!({ "max_iterations": max_iterations });
+    // Carry the routine's notify config in job metadata so the message tool
+    // can resolve channel/target per-job without global state mutation.
     if let Some(channel) = &routine.notify.channel {
-        scheduler
-            .tools()
-            .set_message_tool_context(Some(channel.clone()), Some(routine.notify.user.clone()))
-            .await;
+        metadata["notify_channel"] = serde_json::json!(channel);
     }
-
-    let metadata = serde_json::json!({ "max_iterations": max_iterations });
+    metadata["notify_user"] = serde_json::json!(&routine.notify.user);
 
     // Build approval context: UnlessAutoApproved tools are auto-approved for routines;
     // Always tools require explicit listing in tool_permissions.
@@ -611,6 +655,7 @@ async fn send_notification(
     routine_name: &str,
     status: RunStatus,
     summary: Option<&str>,
+    thread_id: Option<&str>,
 ) {
     let should_notify = match status {
         RunStatus::Ok => notify.on_success,
@@ -637,7 +682,7 @@ async fn send_notification(
 
     let response = OutgoingResponse {
         content: message,
-        thread_id: None,
+        thread_id: thread_id.map(String::from),
         attachments: Vec::new(),
         metadata: serde_json::json!({
             "source": "routine",

src/agent/session.rs

@@ -164,6 +164,10 @@ pub struct PendingApproval {
     /// executed yet when approval was requested.
     #[serde(default)]
     pub deferred_tool_calls: Vec<ToolCall>,
+    /// User timezone at the time the approval was requested, so it persists
+    /// through the approval flow even if the approval message lacks timezone.
+    #[serde(default)]
+    pub user_timezone: Option<String>,
 }
 
 /// A conversation thread within a session.
@@ -976,6 +980,7 @@ mod tests {
             tool_call_id: "call_123".to_string(),
             context_messages: vec![ChatMessage::user("do it")],
             deferred_tool_calls: vec![],
+            user_timezone: None,
         };
 
         thread.await_approval(approval);
@@ -1001,6 +1006,7 @@ mod tests {
             tool_call_id: "call_456".to_string(),
             context_messages: vec![],
             deferred_tool_calls: vec![],
+            user_timezone: None,
         };
 
         thread.await_approval(approval);

src/agent/thread_ops.rs

@@ -230,7 +230,7 @@ impl Agent {
                     )
                     .await;
 
-                let compactor = ContextCompactor::new(self.llm().clone(), self.safety().clone());
+                let compactor = ContextCompactor::new(self.llm().clone());
                 if let Err(e) = compactor
                     .compact(thread, strategy, self.workspace().map(|w| w.as_ref()))
                     .await
@@ -627,7 +627,7 @@ impl Agent {
                 crate::agent::context_monitor::CompactionStrategy::Summarize { keep_recent: 5 },
             );
 
-        let compactor = ContextCompactor::new(self.llm().clone(), self.safety().clone());
+        let compactor = ContextCompactor::new(self.llm().clone());
         match compactor
             .compact(thread, strategy, self.workspace().map(|w| w.as_ref()))
             .await
@@ -746,6 +746,16 @@ impl Agent {
             let mut job_ctx =
                 JobContext::with_user(&message.user_id, "chat", "Interactive chat session");
             job_ctx.http_interceptor = self.deps.http_interceptor.clone();
+            // Prefer a valid timezone from the approval message, fall back to the
+            // resolved timezone stored when the approval was originally requested.
+            let tz_candidate = message
+                .timezone
+                .as_deref()
+                .filter(|tz| crate::timezone::parse_timezone(tz).is_some())
+                .or(pending.user_timezone.as_deref());
+            if let Some(tz) = tz_candidate {
+                job_ctx.user_timezone = tz.to_string();
+            }
 
             let _ = self
                 .channels
@@ -1111,6 +1121,8 @@ impl Agent {
                     tool_call_id: tc.id.clone(),
                     context_messages: context_messages.clone(),
                     deferred_tool_calls: deferred_tool_calls[approval_idx + 1..].to_vec(),
+                    // Carry forward the resolved timezone from the original pending approval
+                    user_timezone: pending.user_timezone.clone(),
                 };
 
                 let request_id = new_pending.request_id;

src/agent/worker.rs

@@ -15,7 +15,8 @@ use crate::db::Database;
 use crate::error::Error;
 use crate::hooks::HookRegistry;
 use crate::llm::{
-    ActionPlan, ChatMessage, LlmProvider, Reasoning, ReasoningContext, RespondResult, ToolSelection,
+    ActionPlan, ChatMessage, LlmProvider, Reasoning, ReasoningContext, RespondResult, ToolCall,
+    ToolSelection,
 };
 use crate::safety::SafetyLayer;
 use crate::tools::rate_limiter::RateLimitResult;
@@ -211,7 +212,7 @@ impl Worker {
         let job_ctx = self.context_manager().get_context(self.job_id).await?;
 
         // Create reasoning engine
-        let reasoning = Reasoning::new(self.llm().clone(), self.safety().clone());
+        let reasoning = Reasoning::new(self.llm().clone());
 
         // Build initial reasoning context (tool definitions refreshed each iteration in execution_loop)
         let mut reason_ctx = ReasoningContext::new().with_job(&job_ctx.description);
@@ -576,37 +577,54 @@ Report when the job is complete or if you encounter issues you cannot resolve."#
                         }
                     }
                 }
-            } else if selections.len() == 1 {
-                consecutive_tool_intent_nudges = 0;
-                // Single tool: execute directly
-                let selection = &selections[0];
-                tracing::debug!(
-                    "Job {} selecting tool: {} - {}",
-                    self.job_id,
-                    selection.tool_name,
-                    selection.reasoning
-                );
-
-                let result = self
-                    .execute_tool(&selection.tool_name, &selection.parameters)
-                    .await;
-
-                self.process_tool_result(reason_ctx, selection, result)
-                    .await?;
             } else {
-                // Multiple tools: execute in parallel
-                tracing::debug!(
-                    "Job {} executing {} tools in parallel",
-                    self.job_id,
-                    selections.len()
-                );
+                consecutive_tool_intent_nudges = 0;
 
-                let results = self.execute_tools_parallel(&selections).await;
+                // Record the assistant tool_calls message so that tool_result
+                // messages have a matching parent (prevents orphaned rewrites).
+                let tool_calls: Vec<ToolCall> = selections
+                    .iter()
+                    .map(|s| ToolCall {
+                        id: s.tool_call_id.clone(),
+                        name: s.tool_name.clone(),
+                        arguments: s.parameters.clone(),
+                    })
+                    .collect();
+                reason_ctx
+                    .messages
+                    .push(ChatMessage::assistant_with_tool_calls(None, tool_calls));
 
-                // Process all results
-                for (selection, result) in selections.iter().zip(results) {
-                    self.process_tool_result(reason_ctx, selection, result.result)
+                if selections.len() == 1 {
+                    // Single tool: execute directly
+                    let selection = &selections[0];
+                    tracing::debug!(
+                        "Job {} selecting tool: {} - {}",
+                        self.job_id,
+                        selection.tool_name,
+                        selection.reasoning
+                    );
+
+                    let result = self
+                        .execute_tool(&selection.tool_name, &selection.parameters)
+                        .await;
+
+                    self.process_tool_result(reason_ctx, selection, result)
                         .await?;
+                } else {
+                    // Multiple tools: execute in parallel
+                    tracing::debug!(
+                        "Job {} executing {} tools in parallel",
+                        self.job_id,
+                        selections.len()
+                    );
+
+                    let results = self.execute_tools_parallel(&selections).await;
+
+                    // Process all results
+                    for (selection, result) in selections.iter().zip(results) {
+                        self.process_tool_result(reason_ctx, selection, result.result)
+                            .await?;
+                    }
                 }
             }
 
@@ -1087,11 +1105,6 @@ Report when the job is complete or if you encounter issues you cannot resolve."#
                 action.reasoning
             );
 
-            // Execute the planned tool
-            let result = self
-                .execute_tool(&action.tool_name, &action.parameters)
-                .await;
-
             // Create a synthetic ToolSelection for process_tool_result.
             // Plan actions don't originate from an LLM tool_call response so
             // there is no real tool_call_id; generate a unique one.
@@ -1103,6 +1116,24 @@ Report when the job is complete or if you encounter issues you cannot resolve."#
                 tool_call_id: format!("plan_{}_{}", self.job_id, i),
             };
 
+            // Record the assistant tool_calls message so that the tool_result
+            // has a matching parent (prevents orphaned rewrites).
+            reason_ctx
+                .messages
+                .push(ChatMessage::assistant_with_tool_calls(
+                    None,
+                    vec![ToolCall {
+                        id: selection.tool_call_id.clone(),
+                        name: selection.tool_name.clone(),
+                        arguments: selection.parameters.clone(),
+                    }],
+                ));
+
+            // Execute the planned tool
+            let result = self
+                .execute_tool(&action.tool_name, &action.parameters)
+                .await;
+
             // Process the result
             let completed = self
                 .process_tool_result(reason_ctx, &selection, result)

src/app.rs

@@ -244,11 +244,31 @@ impl AppBuilder {
         let master_key = match self.config.secrets.master_key() {
             Some(k) => k,
             None => {
+                // No secrets DB available, but we can still load tokens from
+                // OS credential stores (e.g., Anthropic OAuth via Claude Code's
+                // macOS Keychain / Linux ~/.claude/.credentials.json).
+                crate::config::inject_os_credentials();
+
                 // Consume unused handles
                 #[cfg(feature = "libsql")]
                 {
                     self.libsql_db.take();
                 }
+
+                // Re-resolve only the LLM config with OS credentials.
+                let store: Option<&(dyn crate::db::SettingsStore + Sync)> =
+                    self.db.as_ref().map(|db| db.as_ref() as _);
+                let toml_path = self.toml_path.as_deref();
+                if let Err(e) = self
+                    .config
+                    .re_resolve_llm(store, "default", toml_path)
+                    .await
+                {
+                    tracing::warn!(
+                        "Failed to re-resolve LLM config after OS credential injection: {e}"
+                    );
+                }
+
                 return Ok(());
             }
         };
@@ -291,18 +311,16 @@ impl AppBuilder {
             // Inject LLM API keys from encrypted storage
             crate::config::inject_llm_keys_from_secrets(secrets.as_ref(), "default").await;
 
-            // Re-resolve config with newly available keys
-            if let Some(ref db) = self.db {
-                let toml_path = self.toml_path.as_deref();
-                match Config::from_db_with_toml(db.as_ref(), "default", toml_path).await {
-                    Ok(refreshed) => {
-                        self.config = refreshed;
-                        tracing::debug!("LlmConfig re-resolved after secret injection");
-                    }
-                    Err(e) => {
-                        tracing::warn!("Failed to re-resolve config after secret injection: {}", e);
-                    }
-                }
+            // Re-resolve only the LLM config with newly available keys.
+            let store: Option<&(dyn crate::db::SettingsStore + Sync)> =
+                self.db.as_ref().map(|db| db.as_ref() as _);
+            let toml_path = self.toml_path.as_deref();
+            if let Err(e) = self
+                .config
+                .re_resolve_llm(store, "default", toml_path)
+                .await
+            {
+                tracing::warn!("Failed to re-resolve LLM config after secret injection: {e}");
             }
         }
 
@@ -368,12 +386,53 @@ impl AppBuilder {
             .embeddings
             .create_provider(&self.config.llm.nearai.base_url, self.session.clone());
 
+        // Create optional external vector store for workspace semantic search
+        let vector_store: Option<Arc<dyn crate::workspace::VectorStore>> = {
+            #[cfg(feature = "lancedb")]
+            {
+                if self.config.database.vector_backend == crate::config::VectorBackend::LanceDb {
+                    let path = self
+                        .config
+                        .database
+                        .lancedb_path
+                        .clone()
+                        .unwrap_or_else(crate::config::default_lancedb_path);
+                    let dim = embeddings.as_ref().map(|p| p.dimension());
+                    match crate::workspace::LanceDbVectorStore::new(path, dim).await {
+                        Ok(store) => {
+                            tracing::info!("LanceDB vector store connected for workspace search");
+                            Some(Arc::new(store) as Arc<dyn crate::workspace::VectorStore>)
+                        }
+                        Err(e) => {
+                            tracing::warn!("Failed to initialize LanceDB: {}", e);
+                            None
+                        }
+                    }
+                } else {
+                    None
+                }
+            }
+            #[cfg(not(feature = "lancedb"))]
+            {
+                if self.config.database.vector_backend == crate::config::VectorBackend::LanceDb {
+                    tracing::warn!(
+                        "VECTOR_BACKEND=lancedb but 'lancedb' feature not enabled; \
+                         falling back to built-in vector search"
+                    );
+                }
+                None
+            }
+        };
+
         // Register memory tools if database is available
         let workspace = if let Some(ref db) = self.db {
             let mut ws = Workspace::new_with_db("default", db.clone());
             if let Some(ref emb) = embeddings {
                 ws = ws.with_embeddings(emb.clone());
             }
+            if let Some(ref vs) = vector_store {
+                ws = ws.with_vector_store(vs.clone());
+            }
             let ws = Arc::new(ws);
             tools.register_memory_tools(Arc::clone(&ws));
             Some(ws)
@@ -386,11 +445,7 @@ impl AppBuilder {
             && (self.config.agent.allow_local_tools || !self.config.sandbox.enabled)
         {
             tools
-                .register_builder_tool(
-                    llm.clone(),
-                    safety.clone(),
-                    Some(self.config.builder.to_builder_config()),
-                )
+                .register_builder_tool(llm.clone(), Some(self.config.builder.to_builder_config()))
                 .await;
             tracing::info!("Builder mode enabled");
         }
@@ -665,6 +720,17 @@ impl AppBuilder {
         self.init_database().await?;
         self.init_secrets().await?;
 
+        // Post-init validation: if a non-nearai backend was selected but
+        // credentials were never resolved (deferred resolution found no keys),
+        // fail early with a clear error instead of a confusing runtime failure.
+        if self.config.llm.backend != "nearai" && self.config.llm.provider.is_none() {
+            let backend = &self.config.llm.backend;
+            anyhow::bail!(
+                "LLM_BACKEND={backend} is configured but no credentials were found. \
+                 Set the appropriate API key environment variable or run the setup wizard."
+            );
+        }
+
         let (llm, cheap_llm, recording_handle) = if let Some(llm) = self.llm_override.take() {
             (llm, None, None)
         } else {

src/bootstrap.rs

@@ -414,10 +414,103 @@ pub enum MigrationError {
     Io(String),
 }
 
+// ── PID Lock ──────────────────────────────────────────────────────────────
+
+/// Path to the PID lock file: `~/.ironclaw/ironclaw.pid`.
+pub fn pid_lock_path() -> PathBuf {
+    ironclaw_base_dir().join("ironclaw.pid")
+}
+
+/// A PID-based lock that prevents multiple IronClaw instances from running
+/// simultaneously.
+///
+/// Uses `fs4::try_lock_exclusive()` for atomic locking (no TOCTOU race),
+/// then writes the current PID into the locked file for diagnostics.
+/// The OS-level lock is held for the lifetime of this struct and
+/// automatically released on drop (along with the PID file cleanup).
+#[derive(Debug)]
+pub struct PidLock {
+    path: PathBuf,
+    /// Held open to maintain the OS-level exclusive lock.
+    _file: std::fs::File,
+}
+
+/// Errors from PID lock acquisition.
+#[derive(Debug, thiserror::Error)]
+pub enum PidLockError {
+    #[error("Another IronClaw instance is already running (PID {pid})")]
+    AlreadyRunning { pid: u32 },
+    #[error("Failed to acquire PID lock: {0}")]
+    Io(#[from] std::io::Error),
+}
+
+impl PidLock {
+    /// Try to acquire the PID lock.
+    ///
+    /// Uses an exclusive file lock (`flock`/`LockFileEx`) so that two
+    /// concurrent processes cannot both acquire the lock — no TOCTOU race.
+    /// If the lock file exists but the holding process is gone (stale),
+    /// the lock is reclaimed automatically by the OS.
+    pub fn acquire() -> Result<Self, PidLockError> {
+        Self::acquire_at(pid_lock_path())
+    }
+
+    /// Acquire at a specific path (for testing).
+    fn acquire_at(path: PathBuf) -> Result<Self, PidLockError> {
+        use fs4::FileExt;
+        use std::fs::OpenOptions;
+        use std::io::Write;
+
+        // Ensure parent directory exists
+        if let Some(parent) = path.parent() {
+            std::fs::create_dir_all(parent)?;
+        }
+
+        // Open (or create) the lock file
+        let mut file = OpenOptions::new()
+            .read(true)
+            .write(true)
+            .create(true)
+            .truncate(false)
+            .open(&path)?;
+
+        // Try non-blocking exclusive lock — if another process holds it,
+        // this fails immediately instead of blocking.
+        if let Err(e) = file.try_lock_exclusive() {
+            if e.kind() == std::io::ErrorKind::WouldBlock {
+                // Lock held by another process — read its PID for the error message
+                let pid = std::fs::read_to_string(&path)
+                    .ok()
+                    .and_then(|s| s.trim().parse::<u32>().ok())
+                    .unwrap_or(0);
+                return Err(PidLockError::AlreadyRunning { pid });
+            }
+            // Other errors (permissions, unsupported filesystem, etc.)
+            return Err(PidLockError::Io(e));
+        }
+
+        // We hold the exclusive lock — write our PID
+        file.set_len(0)?; // truncate
+        write!(file, "{}", std::process::id())?;
+
+        Ok(PidLock { path, _file: file })
+    }
+}
+
+impl Drop for PidLock {
+    fn drop(&mut self) {
+        // Remove the PID file; the OS-level lock is released when _file is dropped.
+        let _ = std::fs::remove_file(&self.path);
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
+    use std::process::Command;
     use std::sync::Mutex;
+    use std::thread;
+    use std::time::{Duration, Instant};
     use tempfile::tempdir;
 
     static ENV_MUTEX: Mutex<()> = Mutex::new(());
@@ -986,4 +1079,162 @@ INJECTED="pwned"#;
             unsafe { std::env::remove_var("IRONCLAW_BASE_DIR") };
         }
     }
+
+    // ── PID Lock tests ───────────────────────────────────────────────
+
+    #[test]
+    fn test_pid_lock_acquire_and_drop() {
+        let dir = tempdir().unwrap();
+        let pid_path = dir.path().join("ironclaw.pid");
+
+        // Acquire lock
+        let lock = PidLock::acquire_at(pid_path.clone()).unwrap();
+        assert!(pid_path.exists());
+
+        // PID file should contain our PID
+        let contents = std::fs::read_to_string(&pid_path).unwrap();
+        assert_eq!(contents.trim().parse::<u32>().unwrap(), std::process::id());
+
+        // Drop should remove the file
+        drop(lock);
+        assert!(!pid_path.exists());
+    }
+
+    #[test]
+    fn test_pid_lock_rejects_second_acquire() {
+        let dir = tempdir().unwrap();
+        let pid_path = dir.path().join("ironclaw.pid");
+
+        // First lock succeeds
+        let _lock1 = PidLock::acquire_at(pid_path.clone()).unwrap();
+
+        // Second acquire on same file must fail (exclusive flock held)
+        let result = PidLock::acquire_at(pid_path.clone());
+        assert!(result.is_err());
+        match result.unwrap_err() {
+            PidLockError::AlreadyRunning { pid } => {
+                assert_eq!(pid, std::process::id());
+            }
+            other => panic!("expected AlreadyRunning, got: {}", other),
+        }
+    }
+
+    #[test]
+    fn test_pid_lock_reclaims_after_drop() {
+        let dir = tempdir().unwrap();
+        let pid_path = dir.path().join("ironclaw.pid");
+
+        // Acquire and release
+        let lock = PidLock::acquire_at(pid_path.clone()).unwrap();
+        drop(lock);
+
+        // Should succeed — OS lock was released on drop
+        let lock2 = PidLock::acquire_at(pid_path).unwrap();
+        drop(lock2);
+    }
+
+    #[test]
+    fn test_pid_lock_reclaims_stale_file_without_flock() {
+        let dir = tempdir().unwrap();
+        let pid_path = dir.path().join("ironclaw.pid");
+
+        // Write a stale PID file manually (no flock held)
+        std::fs::write(&pid_path, "4294967294").unwrap();
+
+        // Should succeed because no OS lock is held on the file
+        let lock = PidLock::acquire_at(pid_path.clone()).unwrap();
+        let contents = std::fs::read_to_string(&pid_path).unwrap();
+        assert_eq!(contents.trim().parse::<u32>().unwrap(), std::process::id());
+        drop(lock);
+    }
+
+    #[test]
+    fn test_pid_lock_handles_corrupt_pid_file() {
+        let dir = tempdir().unwrap();
+        let pid_path = dir.path().join("ironclaw.pid");
+
+        // Write garbage (no flock held)
+        std::fs::write(&pid_path, "not-a-number").unwrap();
+
+        // Should succeed — no OS lock held, file is reclaimed
+        let lock = PidLock::acquire_at(pid_path).unwrap();
+        drop(lock);
+    }
+
+    #[test]
+    fn test_pid_lock_creates_parent_dirs() {
+        let dir = tempdir().unwrap();
+        let pid_path = dir.path().join("nested").join("deep").join("ironclaw.pid");
+
+        let lock = PidLock::acquire_at(pid_path.clone()).unwrap();
+        assert!(pid_path.exists());
+        drop(lock);
+    }
+
+    #[test]
+    fn test_pid_lock_child_helper_holds_lock() {
+        if std::env::var("IRONCLAW_PID_LOCK_CHILD").ok().as_deref() != Some("1") {
+            return;
+        }
+
+        let pid_path = PathBuf::from(
+            std::env::var("IRONCLAW_PID_LOCK_PATH").expect("IRONCLAW_PID_LOCK_PATH missing"),
+        );
+        let hold_ms = std::env::var("IRONCLAW_PID_LOCK_HOLD_MS")
+            .ok()
+            .and_then(|s| s.parse::<u64>().ok())
+            .unwrap_or(3000);
+
+        let _lock = PidLock::acquire_at(pid_path).expect("child failed to acquire pid lock");
+        thread::sleep(Duration::from_millis(hold_ms));
+    }
+
+    #[test]
+    fn test_pid_lock_rejects_lock_held_by_other_process() {
+        let dir = tempdir().unwrap();
+        let pid_path = dir.path().join("ironclaw.pid");
+
+        let current_exe = std::env::current_exe().unwrap();
+        let mut child = Command::new(current_exe)
+            .args([
+                "--exact",
+                "bootstrap::tests::test_pid_lock_child_helper_holds_lock",
+                "--nocapture",
+                "--test-threads=1",
+            ])
+            .env("IRONCLAW_PID_LOCK_CHILD", "1")
+            .env("IRONCLAW_PID_LOCK_PATH", pid_path.display().to_string())
+            .env("IRONCLAW_PID_LOCK_HOLD_MS", "3000")
+            .spawn()
+            .unwrap();
+
+        let started = Instant::now();
+        while started.elapsed() < Duration::from_secs(2) {
+            if pid_path.exists() {
+                break;
+            }
+            if let Some(status) = child.try_wait().unwrap() {
+                panic!("child exited before acquiring lock: {}", status);
+            }
+            thread::sleep(Duration::from_millis(20));
+        }
+        assert!(
+            pid_path.exists(),
+            "child did not create lock file in time: {}",
+            pid_path.display()
+        );
+
+        let result = PidLock::acquire_at(pid_path.clone());
+        match result.unwrap_err() {
+            PidLockError::AlreadyRunning { .. } => {}
+            other => panic!("expected AlreadyRunning, got: {}", other),
+        }
+
+        let status = child.wait().unwrap();
+        assert!(status.success(), "child process failed: {}", status);
+
+        // After the child exits, lock should be released and reacquirable.
+        let lock = PidLock::acquire_at(pid_path).unwrap();
+        drop(lock);
+    }
 }

src/channels/channel.rs

@@ -79,6 +79,8 @@ pub struct IncomingMessage {
     pub received_at: DateTime<Utc>,
     /// Channel-specific metadata.
     pub metadata: serde_json::Value,
+    /// IANA timezone string from the client (e.g. "America/New_York").
+    pub timezone: Option<String>,
     /// File or media attachments on this message.
     pub attachments: Vec<IncomingAttachment>,
 }
@@ -99,6 +101,7 @@ impl IncomingMessage {
             thread_id: None,
             received_at: Utc::now(),
             metadata: serde_json::Value::Null,
+            timezone: None,
             attachments: Vec::new(),
         }
     }
@@ -121,6 +124,12 @@ impl IncomingMessage {
         self
     }
 
+    /// Set the client timezone.
+    pub fn with_timezone(mut self, tz: impl Into<String>) -> Self {
+        self.timezone = Some(tz.into());
+        self
+    }
+
     /// Set attachments.
     pub fn with_attachments(mut self, attachments: Vec<IncomingAttachment>) -> Self {
         self.attachments = attachments;
@@ -454,4 +463,10 @@ mod tests {
             panic!("expected ToolCompleted variant");
         }
     }
+
+    #[test]
+    fn test_incoming_message_with_timezone() {
+        let msg = IncomingMessage::new("test", "user1", "hello").with_timezone("America/New_York");
+        assert_eq!(msg.timezone.as_deref(), Some("America/New_York"));
+    }
 }

src/channels/repl.rs

@@ -18,7 +18,7 @@
 //! - `Esc` - Interrupt current operation
 
 use std::borrow::Cow;
-use std::io::{self, Write};
+use std::io::{self, IsTerminal, Write};
 use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, Ordering};
 
@@ -297,10 +297,15 @@ impl Channel for ReplChannel {
         let esc_interrupt_triggered_for_thread = Arc::new(AtomicBool::new(false));
 
         std::thread::spawn(move || {
+            let sys_tz = crate::timezone::detect_system_timezone().name().to_string();
+
             // Single message mode: send it and return
             if let Some(msg) = single_message {
-                let incoming = IncomingMessage::new("repl", "default", &msg);
+                let incoming = IncomingMessage::new("repl", "default", &msg).with_timezone(&sys_tz);
                 let _ = tx.blocking_send(incoming);
+                // Ensure the agent exits after handling exactly one turn in -m mode,
+                // even when other channels (gateway/http) are enabled.
+                let _ = tx.blocking_send(IncomingMessage::new("repl", "default", "/quit"));
                 return;
             }
 
@@ -361,7 +366,8 @@ impl Channel for ReplChannel {
                             "/quit" | "/exit" => {
                                 // Forward shutdown command so the agent loop exits even
                                 // when other channels (e.g. web gateway) are still active.
-                                let msg = IncomingMessage::new("repl", "default", "/quit");
+                                let msg = IncomingMessage::new("repl", "default", "/quit")
+                                    .with_timezone(&sys_tz);
                                 let _ = tx.blocking_send(msg);
                                 break;
                             }
@@ -382,7 +388,8 @@ impl Channel for ReplChannel {
                             _ => {}
                         }
 
-                        let msg = IncomingMessage::new("repl", "default", line);
+                        let msg =
+                            IncomingMessage::new("repl", "default", line).with_timezone(&sys_tz);
                         if tx.blocking_send(msg).is_err() {
                             break;
                         }
@@ -390,21 +397,29 @@ impl Channel for ReplChannel {
                     Err(ReadlineError::Interrupted) => {
                         if esc_interrupt_triggered_for_thread.swap(false, Ordering::Relaxed) {
                             // Esc: interrupt current operation and keep REPL open.
-                            let msg = IncomingMessage::new("repl", "default", "/interrupt");
+                            let msg = IncomingMessage::new("repl", "default", "/interrupt")
+                                .with_timezone(&sys_tz);
                             if tx.blocking_send(msg).is_err() {
                                 break;
                             }
                         } else {
                             // Ctrl+C (VINTR): request graceful shutdown.
-                            let msg = IncomingMessage::new("repl", "default", "/quit");
+                            let msg = IncomingMessage::new("repl", "default", "/quit")
+                                .with_timezone(&sys_tz);
                             let _ = tx.blocking_send(msg);
                             break;
                         }
                     }
                     Err(ReadlineError::Eof) => {
-                        // Ctrl+D: send /quit so the agent loop runs graceful shutdown
-                        let msg = IncomingMessage::new("repl", "default", "/quit");
-                        let _ = tx.blocking_send(msg);
+                        // Ctrl+D in interactive mode: graceful shutdown.
+                        // In daemon mode (stdin = /dev/null, no TTY), EOF arrives
+                        // immediately — just drop the REPL thread silently so other
+                        // channels (gateway, telegram, …) keep running.
+                        if std::io::stdin().is_terminal() {
+                            let msg = IncomingMessage::new("repl", "default", "/quit")
+                                .with_timezone(&sys_tz);
+                            let _ = tx.blocking_send(msg);
+                        }
                         break;
                     }
                     Err(e) => {
@@ -614,3 +629,29 @@ impl Channel for ReplChannel {
         Ok(())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use futures::StreamExt;
+
+    use super::*;
+
+    #[tokio::test]
+    async fn single_message_mode_sends_message_then_quit() {
+        let repl = ReplChannel::with_message("hi".to_string());
+        let mut stream = repl.start().await.expect("repl start should succeed");
+
+        let first = stream.next().await.expect("first message missing");
+        assert_eq!(first.channel, "repl");
+        assert_eq!(first.content, "hi");
+
+        let second = stream.next().await.expect("quit message missing");
+        assert_eq!(second.channel, "repl");
+        assert_eq!(second.content, "/quit");
+
+        assert!(
+            stream.next().await.is_none(),
+            "stream should end after /quit"
+        );
+    }
+}

src/channels/web/handlers/chat.rs

@@ -426,7 +426,7 @@ pub async fn chat_threads_handler(
             .map_err(|e| (StatusCode::INTERNAL_SERVER_ERROR, e.to_string()))?;
 
         if let Ok(summaries) = store
-            .list_conversations_with_preview(&state.user_id, "gateway", 50)
+            .list_conversations_all_channels(&state.user_id, 50)
             .await
         {
             let mut assistant_thread = None;
@@ -441,6 +441,7 @@ pub async fn chat_threads_handler(
                     updated_at: s.last_activity.to_rfc3339(),
                     title: s.title.clone(),
                     thread_type: s.thread_type.clone(),
+                    channel: Some(s.channel.clone()),
                 };
 
                 if s.id == assistant_id {
@@ -460,6 +461,7 @@ pub async fn chat_threads_handler(
                     updated_at: chrono::Utc::now().to_rfc3339(),
                     title: None,
                     thread_type: Some("assistant".to_string()),
+                    channel: Some("gateway".to_string()),
                 });
             }
 
@@ -472,9 +474,10 @@ pub async fn chat_threads_handler(
     }
 
     // Fallback: in-memory only (no assistant thread without DB)
-    let threads: Vec<ThreadInfo> = sess
-        .threads
-        .values()
+    let mut sorted_threads: Vec<_> = sess.threads.values().collect();
+    sorted_threads.sort_by(|a, b| b.updated_at.cmp(&a.updated_at));
+    let threads: Vec<ThreadInfo> = sorted_threads
+        .into_iter()
         .map(|t| ThreadInfo {
             id: t.id,
             state: format!("{:?}", t.state),
@@ -483,6 +486,7 @@ pub async fn chat_threads_handler(
             updated_at: t.updated_at.to_rfc3339(),
             title: None,
             thread_type: None,
+            channel: Some("gateway".to_string()),
         })
         .collect();
 
@@ -502,38 +506,39 @@ pub async fn chat_new_thread_handler(
     ))?;
 
     let session = session_manager.get_or_create_session(&state.user_id).await;
-    let mut sess = session.lock().await;
-    let thread = sess.create_thread();
-    let thread_id = thread.id;
-    let info = ThreadInfo {
-        id: thread.id,
-        state: format!("{:?}", thread.state),
-        turn_count: thread.turns.len(),
-        created_at: thread.created_at.to_rfc3339(),
-        updated_at: thread.updated_at.to_rfc3339(),
-        title: None,
-        thread_type: Some("thread".to_string()),
+    let (thread_id, info) = {
+        let mut sess = session.lock().await;
+        let thread = sess.create_thread();
+        let id = thread.id;
+        let info = ThreadInfo {
+            id: thread.id,
+            state: format!("{:?}", thread.state),
+            turn_count: thread.turns.len(),
+            created_at: thread.created_at.to_rfc3339(),
+            updated_at: thread.updated_at.to_rfc3339(),
+            title: None,
+            thread_type: Some("thread".to_string()),
+            channel: Some("gateway".to_string()),
+        };
+        (id, info)
     };
 
-    // Persist the empty conversation row with thread_type metadata
+    // Persist the empty conversation row with thread_type metadata synchronously
+    // so that the subsequent loadThreads() call from the frontend sees it.
     if let Some(ref store) = state.store {
-        let store = Arc::clone(store);
-        let user_id = state.user_id.clone();
-        tokio::spawn(async move {
-            if let Err(e) = store
-                .ensure_conversation(thread_id, "gateway", &user_id, None)
-                .await
-            {
-                tracing::warn!("Failed to persist new thread: {}", e);
-            }
-            let metadata_val = serde_json::json!("thread");
-            if let Err(e) = store
-                .update_conversation_metadata_field(thread_id, "thread_type", &metadata_val)
-                .await
-            {
-                tracing::warn!("Failed to set thread_type metadata: {}", e);
-            }
-        });
+        if let Err(e) = store
+            .ensure_conversation(thread_id, "gateway", &state.user_id, None)
+            .await
+        {
+            tracing::warn!("Failed to persist new thread: {}", e);
+        }
+        let metadata_val = serde_json::json!("thread");
+        if let Err(e) = store
+            .update_conversation_metadata_field(thread_id, "thread_type", &metadata_val)
+            .await
+        {
+            tracing::warn!("Failed to set thread_type metadata: {}", e);
+        }
     }
 
     Ok(Json(info))

src/channels/web/handlers/routines.rs

@@ -10,9 +10,9 @@ use axum::{
 use serde::Deserialize;
 use uuid::Uuid;
 
-use crate::channels::IncomingMessage;
 use crate::channels::web::server::GatewayState;
 use crate::channels::web::types::*;
+use crate::error::RoutineError;
 
 pub async fn routines_list_handler(
     State(state): State<Arc<GatewayState>>,
@@ -133,56 +133,27 @@ pub async fn routines_trigger_handler(
     State(state): State<Arc<GatewayState>>,
     Path(id): Path<String>,
 ) -> Result<Json<serde_json::Value>, (StatusCode, String)> {
-    let store = state.store.as_ref().ok_or((
-        StatusCode::SERVICE_UNAVAILABLE,
-        "Database not available".to_string(),
-    ))?;
+    // Clone the Arc out of the lock to avoid holding the RwLock across .await.
+    let engine = {
+        let guard = state.routine_engine.read().await;
+        guard.as_ref().cloned().ok_or((
+            StatusCode::SERVICE_UNAVAILABLE,
+            "Routine engine not available".to_string(),
+        ))?
+    };
 
     let routine_id = Uuid::parse_str(&id)
         .map_err(|_| (StatusCode::BAD_REQUEST, "Invalid routine ID".to_string()))?;
 
-    let routine = store
-        .get_routine(routine_id)
+    let run_id = engine
+        .fire_manual(routine_id, Some(&state.user_id))
         .await
-        .map_err(|e| (StatusCode::INTERNAL_SERVER_ERROR, e.to_string()))?
-        .ok_or((StatusCode::NOT_FOUND, "Routine not found".to_string()))?;
-
-    if routine.user_id != state.user_id {
-        return Err((StatusCode::FORBIDDEN, "Access denied".to_string()));
-    }
-
-    // Send the routine prompt through the message pipeline as a manual trigger.
-    let prompt = match &routine.action {
-        crate::agent::routine::RoutineAction::Lightweight { prompt, .. } => prompt.clone(),
-        crate::agent::routine::RoutineAction::FullJob {
-            title, description, ..
-        } => format!("{}: {}", title, description),
-    };
-
-    let content = format!("[routine:{}] {}", routine.name, prompt);
-    let thread_id = format!(
-        "routine-{}-{}",
-        routine_id,
-        chrono::Utc::now().timestamp_millis()
-    );
-    let msg = IncomingMessage::new("gateway", &state.user_id, content).with_thread(thread_id);
-
-    let tx_guard = state.msg_tx.read().await;
-    let tx = tx_guard.as_ref().ok_or((
-        StatusCode::SERVICE_UNAVAILABLE,
-        "Channel not started".to_string(),
-    ))?;
-
-    tx.send(msg).await.map_err(|_| {
-        (
-            StatusCode::INTERNAL_SERVER_ERROR,
-            "Channel closed".to_string(),
-        )
-    })?;
+        .map_err(|e| (routine_error_status(&e), e.to_string()))?;
 
     Ok(Json(serde_json::json!({
         "status": "triggered",
         "routine_id": routine_id,
+        "run_id": run_id,
     })))
 }
 
@@ -293,7 +264,7 @@ pub async fn routines_runs_handler(
 /// Convert a Routine to the trimmed RoutineInfo for list display.
 fn routine_to_info(r: &crate::agent::routine::Routine) -> RoutineInfo {
     let (trigger_type, trigger_summary) = match &r.trigger {
-        crate::agent::routine::Trigger::Cron { schedule } => {
+        crate::agent::routine::Trigger::Cron { schedule, .. } => {
             ("cron".to_string(), format!("cron: {}", schedule))
         }
         crate::agent::routine::Trigger::Event {
@@ -337,3 +308,13 @@ fn routine_to_info(r: &crate::agent::routine::Routine) -> RoutineInfo {
         status: status.to_string(),
     }
 }
+
+/// Map `RoutineError` variants to appropriate HTTP status codes.
+fn routine_error_status(err: &RoutineError) -> StatusCode {
+    match err {
+        RoutineError::NotFound { .. } => StatusCode::NOT_FOUND,
+        RoutineError::NotAuthorized { .. } => StatusCode::FORBIDDEN,
+        RoutineError::Disabled { .. } | RoutineError::MaxConcurrent { .. } => StatusCode::CONFLICT,
+        _ => StatusCode::INTERNAL_SERVER_ERROR,
+    }
+}

src/channels/web/mod.rs

@@ -99,6 +99,7 @@ impl GatewayChannel {
             chat_rate_limiter: server::RateLimiter::new(30, 60),
             registry_entries: Vec::new(),
             cost_guard: None,
+            routine_engine: Arc::new(tokio::sync::RwLock::new(None)),
             startup_time: std::time::Instant::now(),
         });
 
@@ -134,6 +135,7 @@ impl GatewayChannel {
             chat_rate_limiter: server::RateLimiter::new(30, 60),
             registry_entries: self.state.registry_entries.clone(),
             cost_guard: self.state.cost_guard.clone(),
+            routine_engine: Arc::clone(&self.state.routine_engine),
             startup_time: self.state.startup_time,
         };
         mutate(&mut new_state);
@@ -281,7 +283,15 @@ impl Channel for GatewayChannel {
         msg: &IncomingMessage,
         response: OutgoingResponse,
     ) -> Result<(), ChannelError> {
-        let thread_id = msg.thread_id.clone().unwrap_or_default();
+        let thread_id = match &msg.thread_id {
+            Some(tid) => tid.clone(),
+            None => {
+                tracing::warn!(
+                    "Gateway respond with no thread_id — skipping (clients would drop it)"
+                );
+                return Ok(());
+            }
+        };
 
         self.state.sse.broadcast(SseEvent::Response {
             content: response.content,
@@ -387,9 +397,18 @@ impl Channel for GatewayChannel {
         _user_id: &str,
         response: OutgoingResponse,
     ) -> Result<(), ChannelError> {
+        let thread_id = match response.thread_id {
+            Some(tid) => tid,
+            None => {
+                tracing::warn!(
+                    "Gateway broadcast with no thread_id — skipping (clients would drop it)"
+                );
+                return Ok(());
+            }
+        };
         self.state.sse.broadcast(SseEvent::Response {
             content: response.content,
-            thread_id: String::new(),
+            thread_id,
         });
         Ok(())
     }

src/channels/web/server.rs

@@ -57,6 +57,10 @@ pub type PromptQueue = Arc<
     >,
 >;
 
+/// Slot for the routine engine, filled at runtime after the agent starts.
+pub type RoutineEngineSlot =
+    Arc<tokio::sync::RwLock<Option<Arc<crate::agent::routine_engine::RoutineEngine>>>>;
+
 /// Simple sliding-window rate limiter.
 ///
 /// Tracks the number of requests in the current window. Resets when the window expires.
@@ -165,6 +169,8 @@ pub struct GatewayState {
     pub registry_entries: Vec<crate::extensions::RegistryEntry>,
     /// Cost guard for token/cost tracking.
     pub cost_guard: Option<Arc<crate::agent::cost_guard::CostGuard>>,
+    /// Routine engine slot for manual routine triggering (filled at runtime).
+    pub routine_engine: RoutineEngineSlot,
     /// Server startup time for uptime calculation.
     pub startup_time: std::time::Instant,
 }
@@ -604,6 +610,7 @@ async fn oauth_callback_handler(
 
 async fn chat_send_handler(
     State(state): State<Arc<GatewayState>>,
+    headers: axum::http::HeaderMap,
     Json(req): Json<SendMessageRequest>,
 ) -> Result<(StatusCode, Json<SendMessageResponse>), (StatusCode, String)> {
     tracing::debug!(
@@ -620,6 +627,14 @@ async fn chat_send_handler(
     }
 
     let mut msg = IncomingMessage::new("gateway", &state.user_id, &req.content);
+    // Prefer timezone from JSON body, fall back to X-Timezone header
+    let tz = req
+        .timezone
+        .as_deref()
+        .or_else(|| headers.get("X-Timezone").and_then(|v| v.to_str().ok()));
+    if let Some(tz) = tz {
+        msg = msg.with_timezone(tz);
+    }
 
     if let Some(ref thread_id) = req.thread_id {
         msg = msg.with_thread(thread_id);
@@ -1037,7 +1052,7 @@ async fn chat_threads_handler(
             .map_err(|e| (StatusCode::INTERNAL_SERVER_ERROR, e.to_string()))?;
 
         if let Ok(summaries) = store
-            .list_conversations_with_preview(&state.user_id, "gateway", 50)
+            .list_conversations_all_channels(&state.user_id, 50)
             .await
         {
             let mut assistant_thread = None;
@@ -1052,6 +1067,7 @@ async fn chat_threads_handler(
                     updated_at: s.last_activity.to_rfc3339(),
                     title: s.title.clone(),
                     thread_type: s.thread_type.clone(),
+                    channel: Some(s.channel.clone()),
                 };
 
                 if s.id == assistant_id {
@@ -1071,6 +1087,7 @@ async fn chat_threads_handler(
                     updated_at: chrono::Utc::now().to_rfc3339(),
                     title: None,
                     thread_type: Some("assistant".to_string()),
+                    channel: Some("gateway".to_string()),
                 });
             }
 
@@ -1083,9 +1100,10 @@ async fn chat_threads_handler(
     }
 
     // Fallback: in-memory only (no assistant thread without DB)
-    let threads: Vec<ThreadInfo> = sess
-        .threads
-        .values()
+    let mut sorted_threads: Vec<_> = sess.threads.values().collect();
+    sorted_threads.sort_by(|a, b| b.updated_at.cmp(&a.updated_at));
+    let threads: Vec<ThreadInfo> = sorted_threads
+        .into_iter()
         .map(|t| ThreadInfo {
             id: t.id,
             state: format!("{:?}", t.state),
@@ -1094,6 +1112,7 @@ async fn chat_threads_handler(
             updated_at: t.updated_at.to_rfc3339(),
             title: None,
             thread_type: None,
+            channel: Some("gateway".to_string()),
         })
         .collect();
 
@@ -1113,38 +1132,39 @@ async fn chat_new_thread_handler(
     ))?;
 
     let session = session_manager.get_or_create_session(&state.user_id).await;
-    let mut sess = session.lock().await;
-    let thread = sess.create_thread();
-    let thread_id = thread.id;
-    let info = ThreadInfo {
-        id: thread.id,
-        state: format!("{:?}", thread.state),
-        turn_count: thread.turns.len(),
-        created_at: thread.created_at.to_rfc3339(),
-        updated_at: thread.updated_at.to_rfc3339(),
-        title: None,
-        thread_type: Some("thread".to_string()),
+    let (thread_id, info) = {
+        let mut sess = session.lock().await;
+        let thread = sess.create_thread();
+        let id = thread.id;
+        let info = ThreadInfo {
+            id: thread.id,
+            state: format!("{:?}", thread.state),
+            turn_count: thread.turns.len(),
+            created_at: thread.created_at.to_rfc3339(),
+            updated_at: thread.updated_at.to_rfc3339(),
+            title: None,
+            thread_type: Some("thread".to_string()),
+            channel: Some("gateway".to_string()),
+        };
+        (id, info)
     };
 
-    // Persist the empty conversation row with thread_type metadata
+    // Persist the empty conversation row with thread_type metadata synchronously
+    // so that the subsequent loadThreads() call from the frontend sees it.
     if let Some(ref store) = state.store {
-        let store = Arc::clone(store);
-        let user_id = state.user_id.clone();
-        tokio::spawn(async move {
-            if let Err(e) = store
-                .ensure_conversation(thread_id, "gateway", &user_id, None)
-                .await
-            {
-                tracing::warn!("Failed to persist new thread: {}", e);
-            }
-            let metadata_val = serde_json::json!("thread");
-            if let Err(e) = store
-                .update_conversation_metadata_field(thread_id, "thread_type", &metadata_val)
-                .await
-            {
-                tracing::warn!("Failed to set thread_type metadata: {}", e);
-            }
-        });
+        if let Err(e) = store
+            .ensure_conversation(thread_id, "gateway", &state.user_id, None)
+            .await
+        {
+            tracing::warn!("Failed to persist new thread: {}", e);
+        }
+        let metadata_val = serde_json::json!("thread");
+        if let Err(e) = store
+            .update_conversation_metadata_field(thread_id, "thread_type", &metadata_val)
+            .await
+        {
+            tracing::warn!("Failed to set thread_type metadata: {}", e);
+        }
     }
 
     Ok(Json(info))
@@ -1965,47 +1985,35 @@ async fn routines_trigger_handler(
     State(state): State<Arc<GatewayState>>,
     Path(id): Path<String>,
 ) -> Result<Json<serde_json::Value>, (StatusCode, String)> {
-    let store = state.store.as_ref().ok_or((
-        StatusCode::SERVICE_UNAVAILABLE,
-        "Database not available".to_string(),
-    ))?;
+    let engine = {
+        let guard = state.routine_engine.read().await;
+        guard.as_ref().cloned().ok_or((
+            StatusCode::SERVICE_UNAVAILABLE,
+            "Routine engine not available".to_string(),
+        ))?
+    };
 
     let routine_id = Uuid::parse_str(&id)
         .map_err(|_| (StatusCode::BAD_REQUEST, "Invalid routine ID".to_string()))?;
 
-    let routine = store
-        .get_routine(routine_id)
+    let run_id = engine
+        .fire_manual(routine_id, Some(&state.user_id))
         .await
-        .map_err(|e| (StatusCode::INTERNAL_SERVER_ERROR, e.to_string()))?
-        .ok_or((StatusCode::NOT_FOUND, "Routine not found".to_string()))?;
-
-    // Send the routine prompt through the message pipeline as a manual trigger.
-    let prompt = match &routine.action {
-        crate::agent::routine::RoutineAction::Lightweight { prompt, .. } => prompt.clone(),
-        crate::agent::routine::RoutineAction::FullJob {
-            title, description, ..
-        } => format!("{}: {}", title, description),
-    };
-
-    let content = format!("[routine:{}] {}", routine.name, prompt);
-    let msg = IncomingMessage::new("gateway", &state.user_id, content);
-
-    let tx_guard = state.msg_tx.read().await;
-    let tx = tx_guard.as_ref().ok_or((
-        StatusCode::SERVICE_UNAVAILABLE,
-        "Channel not started".to_string(),
-    ))?;
-
-    tx.send(msg).await.map_err(|_| {
-        (
-            StatusCode::INTERNAL_SERVER_ERROR,
-            "Channel closed".to_string(),
-        )
-    })?;
+        .map_err(|e| {
+            let status = match &e {
+                crate::error::RoutineError::NotFound { .. } => StatusCode::NOT_FOUND,
+                crate::error::RoutineError::NotAuthorized { .. } => StatusCode::FORBIDDEN,
+                crate::error::RoutineError::Disabled { .. }
+                | crate::error::RoutineError::MaxConcurrent { .. } => StatusCode::CONFLICT,
+                _ => StatusCode::INTERNAL_SERVER_ERROR,
+            };
+            (status, e.to_string())
+        })?;
 
     Ok(Json(serde_json::json!({
         "status": "triggered",
         "routine_id": routine_id,
+        "run_id": run_id,
     })))
 }
 
@@ -2116,7 +2124,7 @@ async fn routines_runs_handler(
 /// Convert a Routine to the trimmed RoutineInfo for list display.
 fn routine_to_info(r: &crate::agent::routine::Routine) -> RoutineInfo {
     let (trigger_type, trigger_summary) = match &r.trigger {
-        crate::agent::routine::Trigger::Cron { schedule } => {
+        crate::agent::routine::Trigger::Cron { schedule, .. } => {
             ("cron".to_string(), format!("cron: {}", schedule))
         }
         crate::agent::routine::Trigger::Event {
@@ -2463,6 +2471,7 @@ mod tests {
             chat_rate_limiter: RateLimiter::new(30, 60),
             registry_entries: vec![],
             cost_guard: None,
+            routine_engine: Arc::new(tokio::sync::RwLock::new(None)),
             startup_time: std::time::Instant::now(),
         })
     }
@@ -2620,7 +2629,9 @@ mod tests {
             secrets,
             sse_sender: None,
             gateway_token: None,
-            created_at: std::time::Instant::now() - std::time::Duration::from_secs(600),
+            created_at: std::time::Instant::now()
+                .checked_sub(std::time::Duration::from_secs(600))
+                .expect("System uptime is too low to run expired flow test"),
         };
 
         ext_mgr
@@ -2727,7 +2738,9 @@ mod tests {
             sse_sender: None,
             gateway_token: None,
             // Expired — handler will reject after lookup (no network I/O)
-            created_at: std::time::Instant::now() - std::time::Duration::from_secs(600),
+            created_at: std::time::Instant::now()
+                .checked_sub(std::time::Duration::from_secs(600))
+                .expect("System uptime is too low to run expired flow test"),
         };
 
         ext_mgr

src/channels/web/static/app.js

@@ -5,6 +5,7 @@ let eventSource = null;
 let logEventSource = null;
 let currentTab = 'chat';
 let currentThreadId = null;
+let currentThreadIsReadOnly = false;
 let assistantThreadId = null;
 let hasMore = false;
 let oldestTimestamp = null;
@@ -13,6 +14,8 @@ let sseHasConnectedBefore = false;
 let jobEvents = new Map(); // job_id -> Array of events
 let jobListRefreshTimer = null;
 let pairingPollInterval = null;
+let unreadThreads = new Map(); // thread_id -> unread count
+let _loadThreadsTimer = null;
 const JOB_EVENTS_CAP = 500;
 const MEMORY_SEARCH_QUERY_MAX_LENGTH = 100;
 
@@ -178,6 +181,7 @@ function confirmRestart() {
     body: {
       content: '/restart',
       thread_id: currentThreadId,
+      timezone: Intl.DateTimeFormat().resolvedOptions().timeZone,
     },
   })
     .then((response) => {
@@ -273,7 +277,13 @@ function connectSSE() {
 
   eventSource.addEventListener('response', (e) => {
     const data = JSON.parse(e.data);
-    if (!isCurrentThread(data.thread_id)) return;
+    if (!isCurrentThread(data.thread_id)) {
+      if (data.thread_id) {
+        unreadThreads.set(data.thread_id, (unreadThreads.get(data.thread_id) || 0) + 1);
+        debouncedLoadThreads();
+      }
+      return;
+    }
     finalizeActivityGroup();
     addMessage('assistant', data.content);
     enableChatInput();
@@ -288,7 +298,10 @@ function connectSSE() {
 
   eventSource.addEventListener('thinking', (e) => {
     const data = JSON.parse(e.data);
-    if (!isCurrentThread(data.thread_id)) return;
+    if (!isCurrentThread(data.thread_id)) {
+      if (data.thread_id) debouncedLoadThreads();
+      return;
+    }
     showActivityThinking(data.message);
   });
 
@@ -324,7 +337,10 @@ function connectSSE() {
 
   eventSource.addEventListener('status', (e) => {
     const data = JSON.parse(e.data);
-    if (!isCurrentThread(data.thread_id)) return;
+    if (!isCurrentThread(data.thread_id)) {
+      if (data.thread_id) debouncedLoadThreads();
+      return;
+    }
     // "Done" and "Awaiting approval" are terminal signals from the agent:
     // the agentic loop finished, so re-enable input as a safety net in case
     // the response SSE event is empty or lost.
@@ -414,9 +430,9 @@ function connectSSE() {
 }
 
 // Check if an SSE event belongs to the currently viewed thread.
-// Events without a thread_id (legacy) are always shown.
+// Events without a thread_id are dropped (prevents notification leaking).
 function isCurrentThread(threadId) {
-  if (!threadId) return true;
+  if (!threadId) return false;
   if (!currentThreadId) return true;
   return threadId === currentThreadId;
 }
@@ -439,14 +455,21 @@ function sendMessage() {
 
   apiFetch('/api/chat/send', {
     method: 'POST',
-    body: { content, thread_id: currentThreadId || undefined },
+    body: { content, thread_id: currentThreadId || undefined, timezone: Intl.DateTimeFormat().resolvedOptions().timeZone },
   }).catch((err) => {
     addMessage('system', 'Failed to send: ' + err.message);
   });
 }
 
 function enableChatInput() {
-  // no-op: input and send button are always enabled
+  if (currentThreadIsReadOnly) return;
+  const input = document.getElementById('chat-input');
+  const btn = document.getElementById('send-btn');
+  if (input) {
+    input.disabled = false;
+    input.placeholder = 'Message or / for commands...';
+  }
+  if (btn) btn.disabled = false;
 }
 
 // --- Slash Autocomplete ---
@@ -541,6 +564,13 @@ function sendApprovalAction(requestId, action) {
 
 function renderMarkdown(text) {
   if (typeof marked !== 'undefined') {
+    // Escape raw HTML error pages instead of rendering them as markup.
+    // Only triggers when the text *starts with* a doctype or <html> tag
+    // (after optional whitespace), so normal messages that mention HTML
+    // tags in prose or code fences are not affected.  See #263.
+    if (/^\s*<!doctype\s/i.test(text) || /^\s*<html[\s>]/i.test(text)) {
+      return escapeHtml(text);
+    }
     let html = marked.parse(text);
     // Sanitize HTML output to prevent XSS from tool output or LLM responses.
     html = sanitizeRenderedHtml(html);
@@ -1134,7 +1164,9 @@ function loadHistory(before) {
       // Fresh load: clear and render
       container.innerHTML = '';
       for (const turn of data.turns) {
-        addMessage('user', turn.user_input);
+        if (turn.user_input) {
+          addMessage('user', turn.user_input);
+        }
         if (turn.tool_calls && turn.tool_calls.length > 0) {
           addToolCallsSummary(turn.tool_calls);
         }
@@ -1156,8 +1188,10 @@ function loadHistory(before) {
       const savedHeight = container.scrollHeight;
       const fragment = document.createDocumentFragment();
       for (const turn of data.turns) {
-        const userDiv = createMessageElement('user', turn.user_input);
-        fragment.appendChild(userDiv);
+        if (turn.user_input) {
+          const userDiv = createMessageElement('user', turn.user_input);
+          fragment.appendChild(userDiv);
+        }
         if (turn.tool_calls && turn.tool_calls.length > 0) {
           fragment.appendChild(createToolCallsSummaryElement(turn.tool_calls));
         }
@@ -1256,6 +1290,37 @@ function removeScrollSpinner() {
 
 // --- Threads ---
 
+function threadTitle(thread) {
+  if (thread.title) return thread.title;
+  const ch = thread.channel || 'gateway';
+  if (thread.thread_type === 'heartbeat') return 'Heartbeat Alerts';
+  if (thread.thread_type === 'routine') return 'Routine';
+  if (ch !== 'gateway') return ch.charAt(0).toUpperCase() + ch.slice(1);
+  if (thread.turn_count === 0) return 'New chat';
+  return thread.id.substring(0, 8);
+}
+
+function relativeTime(isoStr) {
+  if (!isoStr) return '';
+  const diff = Date.now() - new Date(isoStr).getTime();
+  const mins = Math.floor(diff / 60000);
+  if (mins < 1) return 'now';
+  if (mins < 60) return mins + 'm ago';
+  const hrs = Math.floor(mins / 60);
+  if (hrs < 24) return hrs + 'h ago';
+  const days = Math.floor(hrs / 24);
+  return days + 'd ago';
+}
+
+function isReadOnlyChannel(channel) {
+  return channel && channel !== 'gateway' && channel !== 'routine' && channel !== 'heartbeat';
+}
+
+function debouncedLoadThreads() {
+  if (_loadThreadsTimer) clearTimeout(_loadThreadsTimer);
+  _loadThreadsTimer = setTimeout(() => { _loadThreadsTimer = null; loadThreads(); }, 500);
+}
+
 function loadThreads() {
   apiFetch('/api/chat/threads').then((data) => {
     // Pinned assistant thread
@@ -1264,9 +1329,13 @@ function loadThreads() {
       const el = document.getElementById('assistant-thread');
       const isActive = currentThreadId === assistantThreadId;
       el.className = 'assistant-item' + (isActive ? ' active' : '');
+      const labelEl = document.getElementById('assistant-label');
+      if (labelEl) {
+        const at = data.assistant_thread;
+        labelEl.textContent = 'Assistant';
+      }
       const meta = document.getElementById('assistant-meta');
-      const count = data.assistant_thread.turn_count || 0;
-      meta.textContent = count > 0 ? count + ' turns' : '';
+      meta.textContent = relativeTime(data.assistant_thread.updated_at);
     }
 
     // Regular threads
@@ -1275,16 +1344,38 @@ function loadThreads() {
     const threads = data.threads || [];
     for (const thread of threads) {
       const item = document.createElement('div');
-      item.className = 'thread-item' + (thread.id === currentThreadId ? ' active' : '');
+      const isActive = thread.id === currentThreadId;
+      item.className = 'thread-item' + (isActive ? ' active' : '');
+
+      // Channel badge for non-gateway threads
+      const ch = thread.channel || 'gateway';
+      if (ch !== 'gateway') {
+        const badge = document.createElement('span');
+        badge.className = 'thread-badge thread-badge-' + ch;
+        badge.textContent = ch;
+        item.appendChild(badge);
+      }
+
       const label = document.createElement('span');
       label.className = 'thread-label';
-      label.textContent = thread.title || thread.id.substring(0, 8);
-      label.title = thread.title ? thread.title + ' (' + thread.id + ')' : thread.id;
+      label.textContent = threadTitle(thread);
+      label.title = (thread.title || '') + ' (' + thread.id + ')';
       item.appendChild(label);
+
       const meta = document.createElement('span');
       meta.className = 'thread-meta';
-      meta.textContent = (thread.turn_count || 0) + ' turns';
+      meta.textContent = relativeTime(thread.updated_at);
       item.appendChild(meta);
+
+      // Unread dot
+      const unread = unreadThreads.get(thread.id) || 0;
+      if (unread > 0 && !isActive) {
+        const dot = document.createElement('span');
+        dot.className = 'thread-unread';
+        dot.textContent = unread > 9 ? '9+' : String(unread);
+        item.appendChild(dot);
+      }
+
       item.addEventListener('click', () => switchThread(thread.id));
       list.appendChild(item);
     }
@@ -1294,17 +1385,36 @@ function loadThreads() {
       switchToAssistant();
     }
 
-    // Enable chat input once a thread is available
+    // Enable/disable chat input based on channel type
     if (currentThreadId) {
-      enableChatInput();
+      const currentThread = threads.find(t => t.id === currentThreadId);
+      const ch = currentThread ? currentThread.channel : 'gateway';
+      currentThreadIsReadOnly = isReadOnlyChannel(ch);
+      if (currentThreadIsReadOnly) {
+        disableChatInputReadOnly();
+      } else {
+        enableChatInput();
+      }
     }
   }).catch(() => {});
 }
 
+function disableChatInputReadOnly() {
+  const input = document.getElementById('chat-input');
+  const btn = document.getElementById('send-btn');
+  if (input) {
+    input.disabled = true;
+    input.placeholder = 'Read-only thread (external channel)';
+  }
+  if (btn) btn.disabled = true;
+}
+
 function switchToAssistant() {
   if (!assistantThreadId) return;
   finalizeActivityGroup();
   currentThreadId = assistantThreadId;
+  currentThreadIsReadOnly = false;
+  unreadThreads.delete(assistantThreadId);
   hasMore = false;
   oldestTimestamp = null;
   loadHistory();
@@ -1314,6 +1424,7 @@ function switchToAssistant() {
 function switchThread(threadId) {
   finalizeActivityGroup();
   currentThreadId = threadId;
+  unreadThreads.delete(threadId);
   hasMore = false;
   oldestTimestamp = null;
   loadHistory();
@@ -1370,7 +1481,7 @@ chatInput.addEventListener('keydown', (e) => {
     }
   }
 
-  if (e.key === 'Enter' && !e.shiftKey) {
+  if (e.key === 'Enter' && !e.shiftKey && !e.isComposing) {
     e.preventDefault();
     hideSlashAutocomplete();
     sendMessage();

src/channels/web/static/index.html

@@ -113,12 +113,12 @@
     <div class="tab-panel active" id="tab-chat">
       <div class="thread-sidebar" id="thread-sidebar">
         <div class="thread-sidebar-header">
-          <span>Threads</span>
           <button class="thread-new-btn" onclick="createNewThread()" title="New thread (Ctrl/Cmd+N)">+</button>
+          <div class="spacer"></div>
           <button class="thread-toggle-btn" id="thread-toggle-btn" onclick="toggleThreadSidebar()" title="Toggle sidebar">&laquo;</button>
         </div>
         <div class="assistant-item" id="assistant-thread" onclick="switchToAssistant()">
-          <span class="assistant-label">Assistant</span>
+          <span class="assistant-label" id="assistant-label">Assistant</span>
           <span class="assistant-meta" id="assistant-meta"></span>
         </div>
         <div class="threads-section-header">

src/channels/web/static/style.css

@@ -3074,7 +3074,7 @@ mark {
 }
 
 .thread-sidebar {
-  width: 200px;
+  width: 240px;
   background: var(--bg-secondary);
   border-right: 1px solid var(--border);
   display: flex;
@@ -3082,6 +3082,8 @@ mark {
   flex-shrink: 0;
   transition: width 0.2s ease;
   overflow: hidden;
+  padding: 6px;
+  gap: 2px;
 }
 
 .thread-sidebar.collapsed {
@@ -3099,8 +3101,7 @@ mark {
 .thread-sidebar-header {
   display: flex;
   align-items: center;
-  padding: 10px 12px;
-  border-bottom: 1px solid var(--border);
+  padding: 10px 10px;
   font-size: 13px;
   font-weight: 600;
   gap: 8px;
@@ -3134,21 +3135,22 @@ mark {
   display: flex;
   align-items: center;
   justify-content: space-between;
-  padding: 10px 12px;
+  padding: 12px 14px;
   cursor: pointer;
   font-size: 13px;
   font-weight: 600;
   color: var(--text);
-  border-bottom: 1px solid var(--border);
-  background: var(--bg-secondary);
+  background: var(--bg-tertiary);
+  border-radius: var(--radius);
+  margin-bottom: 2px;
 }
 
 .assistant-item:hover {
-  background: var(--bg-tertiary);
+  background: rgba(255, 255, 255, 0.06);
 }
 
 .assistant-item.active {
-  background: rgba(52, 211, 153, 0.08);
+  background: rgba(52, 211, 153, 0.1);
   color: var(--accent);
   border-left: 2px solid var(--accent);
 }
@@ -3166,7 +3168,7 @@ mark {
 }
 
 .threads-section-header {
-  padding: 8px 12px 4px;
+  padding: 10px 10px 4px;
   font-size: 11px;
   font-weight: 500;
   text-transform: uppercase;
@@ -3196,11 +3198,11 @@ mark {
   display: flex;
   align-items: center;
   justify-content: space-between;
-  padding: 8px 12px;
+  padding: 10px 14px;
   cursor: pointer;
   font-size: 13px;
   color: var(--text-secondary);
-  border-bottom: 1px solid rgba(255, 255, 255, 0.03);
+  border-radius: var(--radius);
 }
 
 .thread-item:hover {
@@ -3222,6 +3224,43 @@ mark {
 .thread-meta {
   font-size: 11px;
   color: var(--text-secondary);
+  flex-shrink: 0;
+}
+
+.thread-badge {
+  display: inline-block;
+  font-size: 9px;
+  font-weight: 600;
+  text-transform: uppercase;
+  letter-spacing: 0.5px;
+  padding: 1px 5px;
+  border-radius: 3px;
+  background: rgba(255, 255, 255, 0.08);
+  color: var(--text-secondary);
+  margin-right: 6px;
+  flex-shrink: 0;
+}
+
+.thread-badge-routine { background: rgba(52, 211, 153, 0.15); color: var(--accent); }
+.thread-badge-heartbeat { background: rgba(245, 166, 35, 0.15); color: var(--warning); }
+.thread-badge-telegram { background: rgba(0, 136, 204, 0.15); color: #0088cc; }
+.thread-badge-signal { background: rgba(59, 118, 240, 0.15); color: #3b76f0; }
+.thread-badge-slack { background: rgba(74, 21, 75, 0.15); color: #e01e5a; }
+
+.thread-unread {
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  min-width: 16px;
+  height: 16px;
+  font-size: 10px;
+  font-weight: 700;
+  background: var(--accent);
+  color: var(--bg);
+  border-radius: 8px;
+  padding: 0 4px;
+  margin-left: auto;
+  flex-shrink: 0;
 }
 
 /* --- Memory editing --- */
@@ -3620,7 +3659,7 @@ mark {
     left: 0;
     top: 0;
     bottom: 0;
-    width: 200px;
+    width: 240px;
     z-index: 50;
   }
 

src/channels/web/test_helpers.rs

@@ -84,6 +84,7 @@ impl TestGatewayBuilder {
             chat_rate_limiter: RateLimiter::new(30, 60),
             registry_entries: Vec::new(),
             cost_guard: None,
+            routine_engine: Arc::new(tokio::sync::RwLock::new(None)),
             startup_time: std::time::Instant::now(),
         })
     }

src/channels/web/types.rs

@@ -9,6 +9,7 @@ use uuid::Uuid;
 pub struct SendMessageRequest {
     pub content: String,
     pub thread_id: Option<String>,
+    pub timezone: Option<String>,
 }
 
 #[derive(Debug, Serialize)]
@@ -28,6 +29,8 @@ pub struct ThreadInfo {
     pub title: Option<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub thread_type: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub channel: Option<String>,
 }
 
 #[derive(Debug, Serialize)]
@@ -611,6 +614,7 @@ pub enum WsClientMessage {
     Message {
         content: String,
         thread_id: Option<String>,
+        timezone: Option<String>,
     },
     /// Approve or deny a pending tool execution.
     #[serde(rename = "approval")]
@@ -796,7 +800,9 @@ mod tests {
         let json = r#"{"type":"message","content":"hello","thread_id":"t1"}"#;
         let msg: WsClientMessage = serde_json::from_str(json).unwrap();
         match msg {
-            WsClientMessage::Message { content, thread_id } => {
+            WsClientMessage::Message {
+                content, thread_id, ..
+            } => {
                 assert_eq!(content, "hello");
                 assert_eq!(thread_id.as_deref(), Some("t1"));
             }
@@ -809,7 +815,9 @@ mod tests {
         let json = r#"{"type":"message","content":"hi"}"#;
         let msg: WsClientMessage = serde_json::from_str(json).unwrap();
         match msg {
-            WsClientMessage::Message { content, thread_id } => {
+            WsClientMessage::Message {
+                content, thread_id, ..
+            } => {
                 assert_eq!(content, "hi");
                 assert!(thread_id.is_none());
             }
@@ -1063,4 +1071,40 @@ mod tests {
         let req: AuthCancelRequest = serde_json::from_str(json).unwrap();
         assert_eq!(req.extension_name, "telegram");
     }
+
+    // ---- ThreadInfo channel field tests ----
+
+    #[test]
+    fn test_thread_info_channel_serialized() {
+        let info = ThreadInfo {
+            id: Uuid::nil(),
+            state: "Idle".to_string(),
+            turn_count: 0,
+            created_at: "2026-01-01T00:00:00Z".to_string(),
+            updated_at: "2026-01-01T00:00:00Z".to_string(),
+            title: None,
+            thread_type: None,
+            channel: Some("telegram".to_string()),
+        };
+        let json = serde_json::to_string(&info).unwrap();
+        let parsed: serde_json::Value = serde_json::from_str(&json).unwrap();
+        assert_eq!(parsed["channel"], "telegram");
+    }
+
+    #[test]
+    fn test_thread_info_channel_omitted_when_none() {
+        let info = ThreadInfo {
+            id: Uuid::nil(),
+            state: "Idle".to_string(),
+            turn_count: 0,
+            created_at: "2026-01-01T00:00:00Z".to_string(),
+            updated_at: "2026-01-01T00:00:00Z".to_string(),
+            title: None,
+            thread_type: None,
+            channel: None,
+        };
+        let json = serde_json::to_string(&info).unwrap();
+        let parsed: serde_json::Value = serde_json::from_str(&json).unwrap();
+        assert!(parsed.get("channel").is_none());
+    }
 }

src/channels/web/util.rs

@@ -83,6 +83,19 @@ pub fn build_turns_from_db_messages(
 
             turns.push(turn);
             turn_number += 1;
+        } else if msg.role == "assistant" {
+            // Standalone assistant message (e.g. routine output, heartbeat)
+            // with no preceding user message — render as a turn with empty input.
+            turns.push(TurnInfo {
+                turn_number,
+                user_input: String::new(),
+                response: Some(msg.content.clone()),
+                state: "Completed".to_string(),
+                started_at: msg.created_at.to_rfc3339(),
+                completed_at: Some(msg.created_at.to_rfc3339()),
+                tool_calls: Vec::new(),
+            });
+            turn_number += 1;
         }
     }
 
@@ -220,6 +233,29 @@ mod tests {
         assert_eq!(turns[0].response.as_deref(), Some("Done"));
     }
 
+    #[test]
+    fn test_build_turns_standalone_assistant_messages() {
+        // Routine conversations only have assistant messages (no user messages).
+        let messages = vec![
+            make_msg("assistant", "Routine executed: all checks passed", 0),
+            make_msg("assistant", "Routine executed: found 2 issues", 5000),
+        ];
+        let turns = build_turns_from_db_messages(&messages);
+        assert_eq!(turns.len(), 2);
+        // Standalone assistant messages should have empty user_input
+        assert_eq!(turns[0].user_input, "");
+        assert_eq!(
+            turns[0].response.as_deref(),
+            Some("Routine executed: all checks passed")
+        );
+        assert_eq!(turns[0].state, "Completed");
+        assert_eq!(turns[1].user_input, "");
+        assert_eq!(
+            turns[1].response.as_deref(),
+            Some("Routine executed: found 2 issues")
+        );
+    }
+
     #[test]
     fn test_build_turns_backward_compatible() {
         let messages = vec![

src/channels/web/ws.rs

@@ -156,8 +156,15 @@ async fn handle_client_message(
     direct_tx: &mpsc::Sender<WsServerMessage>,
 ) {
     match msg {
-        WsClientMessage::Message { content, thread_id } => {
+        WsClientMessage::Message {
+            content,
+            thread_id,
+            timezone,
+        } => {
             let mut incoming = IncomingMessage::new("gateway", user_id, &content);
+            if let Some(ref tz) = timezone {
+                incoming = incoming.with_timezone(tz);
+            }
             if let Some(ref tid) = thread_id {
                 incoming = incoming.with_thread(tid);
             }
@@ -349,6 +356,7 @@ mod tests {
             WsClientMessage::Message {
                 content: "hello agent".to_string(),
                 thread_id: Some("t1".to_string()),
+                timezone: None,
             },
             &state,
             "user1",
@@ -373,6 +381,7 @@ mod tests {
             WsClientMessage::Message {
                 content: "hello".to_string(),
                 thread_id: None,
+                timezone: None,
             },
             &state,
             "user1",
@@ -493,6 +502,7 @@ mod tests {
             chat_rate_limiter: crate::channels::web::server::RateLimiter::new(30, 60),
             registry_entries: Vec::new(),
             cost_guard: None,
+            routine_engine: Arc::new(tokio::sync::RwLock::new(None)),
             startup_time: std::time::Instant::now(),
         }
     }

src/cli/status.rs

@@ -8,9 +8,35 @@ use std::path::PathBuf;
 use crate::bootstrap::ironclaw_base_dir;
 use crate::settings::Settings;
 
+/// Load settings from JSON and TOML config files, matching the runtime
+/// priority: TOML overlay > settings.json > defaults.
+///
+/// This mirrors the loading chain in `Config::from_env_with_toml()` but
+/// without resolving the full `Config` (which requires async + secrets).
+fn load_settings() -> Settings {
+    load_settings_from(&Settings::default_path(), &Settings::default_toml_path())
+}
+
+/// Inner implementation with injectable paths (testable).
+fn load_settings_from(json_path: &std::path::Path, toml_path: &std::path::Path) -> Settings {
+    let mut settings = Settings::load_from(json_path);
+
+    match Settings::load_toml(toml_path) {
+        Ok(Some(toml_settings)) => {
+            settings.merge_from(&toml_settings);
+        }
+        Ok(None) => {} // File not found — fine for default path
+        Err(e) => {
+            eprintln!("Warning: failed to parse {}: {}", toml_path.display(), e);
+        }
+    }
+
+    settings
+}
+
 /// Run the status command, printing system health info.
 pub async fn run_status_command() -> anyhow::Result<()> {
-    let settings = Settings::default();
+    let settings = load_settings();
 
     println!("IronClaw Status");
     println!("===============\n");
@@ -209,3 +235,99 @@ fn default_tools_dir() -> PathBuf {
 fn default_channels_dir() -> PathBuf {
     ironclaw_base_dir().join("channels")
 }
+
+#[cfg(test)]
+mod tests {
+    use super::load_settings_from;
+
+    /// Regression test for #354: load_settings_from must read config.toml.
+    #[test]
+    fn reads_toml_heartbeat_enabled() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let json_path = dir.path().join("settings.json");
+        let toml_path = dir.path().join("config.toml");
+
+        // No JSON file — only TOML
+        std::fs::write(
+            &toml_path,
+            "[heartbeat]\nenabled = true\ninterval_secs = 600",
+        )
+        .expect("write toml");
+
+        let settings = load_settings_from(&json_path, &toml_path);
+        assert!(settings.heartbeat.enabled);
+        assert_eq!(settings.heartbeat.interval_secs, 600);
+    }
+
+    /// Without any config files, defaults are returned.
+    #[test]
+    fn defaults_without_config_files() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let settings = load_settings_from(
+            &dir.path().join("nonexistent.json"),
+            &dir.path().join("nonexistent.toml"),
+        );
+        assert!(!settings.heartbeat.enabled);
+    }
+
+    /// settings.json is respected.
+    #[test]
+    fn reads_json_heartbeat_enabled() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let json_path = dir.path().join("settings.json");
+        let toml_path = dir.path().join("nonexistent.toml");
+
+        std::fs::write(
+            &json_path,
+            r#"{"heartbeat":{"enabled":true,"interval_secs":900}}"#,
+        )
+        .expect("write json");
+
+        let settings = load_settings_from(&json_path, &toml_path);
+        assert!(settings.heartbeat.enabled);
+        assert_eq!(settings.heartbeat.interval_secs, 900);
+    }
+
+    /// TOML overlay wins over JSON settings.
+    #[test]
+    fn toml_overlay_wins_over_json() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let json_path = dir.path().join("settings.json");
+        let toml_path = dir.path().join("config.toml");
+
+        std::fs::write(
+            &json_path,
+            r#"{"heartbeat":{"enabled":false,"interval_secs":100}}"#,
+        )
+        .expect("write json");
+        std::fs::write(
+            &toml_path,
+            "[heartbeat]\nenabled = true\ninterval_secs = 200",
+        )
+        .expect("write toml");
+
+        let settings = load_settings_from(&json_path, &toml_path);
+        assert!(settings.heartbeat.enabled);
+        assert_eq!(settings.heartbeat.interval_secs, 200);
+    }
+
+    /// Invalid TOML is warned but doesn't crash; falls back to JSON/defaults.
+    #[test]
+    fn invalid_toml_falls_back_gracefully() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let json_path = dir.path().join("settings.json");
+        let toml_path = dir.path().join("config.toml");
+
+        std::fs::write(
+            &json_path,
+            r#"{"heartbeat":{"enabled":true,"interval_secs":500}}"#,
+        )
+        .expect("write json");
+        std::fs::write(&toml_path, "this is not valid toml [[[").expect("write bad toml");
+
+        let settings = load_settings_from(&json_path, &toml_path);
+        // Should fall back to JSON values, not crash
+        assert!(settings.heartbeat.enabled);
+        assert_eq!(settings.heartbeat.interval_secs, 500);
+    }
+}

src/config/agent.rs

@@ -27,6 +27,8 @@ pub struct AgentConfig {
     pub max_tool_iterations: usize,
     /// When true, skip tool approval checks entirely. For benchmarks/CI.
     pub auto_approve_tools: bool,
+    /// Default timezone for new sessions (IANA name, e.g. "America/New_York").
+    pub default_timezone: String,
 }
 
 impl AgentConfig {
@@ -47,6 +49,7 @@ impl AgentConfig {
             max_actions_per_hour: None,
             max_tool_iterations: 10,
             auto_approve_tools: true,
+            default_timezone: "UTC".to_string(),
         }
     }
 
@@ -89,6 +92,40 @@ impl AgentConfig {
                 "AGENT_AUTO_APPROVE_TOOLS",
                 settings.agent.auto_approve_tools,
             )?,
+            default_timezone: {
+                let tz: String = parse_optional_env(
+                    "DEFAULT_TIMEZONE",
+                    settings.agent.default_timezone.clone(),
+                )?;
+                if crate::timezone::parse_timezone(&tz).is_none() {
+                    return Err(ConfigError::InvalidValue {
+                        key: "DEFAULT_TIMEZONE".into(),
+                        message: format!("invalid IANA timezone: '{tz}'"),
+                    });
+                }
+                tz
+            },
         })
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_default_timezone_rejects_invalid() {
+        let mut settings = Settings::default();
+        settings.agent.default_timezone = "Fake/Zone".to_string();
+
+        let result = AgentConfig::resolve(&settings);
+        assert!(result.is_err(), "invalid IANA timezone should be rejected");
+    }
+
+    #[test]
+    fn test_default_timezone_accepts_valid() {
+        let settings = Settings::default(); // default is "UTC"
+        let config = AgentConfig::resolve(&settings).expect("resolve");
+        assert_eq!(config.default_timezone, "UTC");
+    }
+}

src/config/database.rs

@@ -82,6 +82,31 @@ impl std::str::FromStr for SslMode {
     }
 }
 
+/// Which vector store backend to use for workspace semantic search.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum VectorBackend {
+    /// Use the database's built-in vector support (pgvector or libsql_vector_idx).
+    #[default]
+    Builtin,
+    /// Use LanceDB as an external vector store (requires `lancedb` feature).
+    LanceDb,
+}
+
+impl std::str::FromStr for VectorBackend {
+    type Err = String;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s.to_lowercase().as_str() {
+            "" | "builtin" | "pgvector" | "libsql" => Ok(Self::Builtin),
+            "lancedb" | "lance" => Ok(Self::LanceDb),
+            _ => Err(format!(
+                "invalid vector backend '{}', expected 'builtin' or 'lancedb'",
+                s
+            )),
+        }
+    }
+}
+
 /// Database configuration.
 #[derive(Debug, Clone)]
 pub struct DatabaseConfig {
@@ -101,6 +126,12 @@ pub struct DatabaseConfig {
     pub libsql_url: Option<String>,
     /// Turso auth token (required when libsql_url is set).
     pub libsql_auth_token: Option<SecretString>,
+
+    // -- Vector store fields --
+    /// Which vector store to use for workspace semantic search (default: Builtin).
+    pub vector_backend: VectorBackend,
+    /// Path to LanceDB directory (default: ~/.ironclaw/lancedb when vector_backend is LanceDb).
+    pub lancedb_path: Option<PathBuf>,
 }
 
 impl DatabaseConfig {
@@ -159,6 +190,25 @@ impl DatabaseConfig {
             });
         }
 
+        let vector_backend: VectorBackend = if let Some(s) = optional_env("VECTOR_BACKEND")? {
+            s.parse().map_err(|e| ConfigError::InvalidValue {
+                key: "VECTOR_BACKEND".to_string(),
+                message: e,
+            })?
+        } else {
+            VectorBackend::default()
+        };
+
+        let lancedb_path = optional_env("LANCEDB_PATH")?
+            .map(PathBuf::from)
+            .or_else(|| {
+                if vector_backend == VectorBackend::LanceDb {
+                    Some(default_lancedb_path())
+                } else {
+                    None
+                }
+            });
+
         Ok(Self {
             backend,
             url: SecretString::from(url),
@@ -167,6 +217,8 @@ impl DatabaseConfig {
             libsql_path,
             libsql_url,
             libsql_auth_token,
+            vector_backend,
+            lancedb_path,
         })
     }
 
@@ -195,6 +247,11 @@ pub fn default_libsql_path() -> PathBuf {
     ironclaw_base_dir().join("ironclaw.db")
 }
 
+/// Default LanceDB directory (~/.ironclaw/lancedb).
+pub fn default_lancedb_path() -> PathBuf {
+    ironclaw_base_dir().join("lancedb")
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -224,4 +281,39 @@ mod tests {
     fn ssl_mode_parse_invalid() {
         assert!("invalid".parse::<SslMode>().is_err());
     }
+
+    #[test]
+    fn vector_backend_parse() {
+        assert_eq!(
+            "builtin".parse::<VectorBackend>().unwrap(),
+            VectorBackend::Builtin
+        );
+        assert_eq!(
+            "pgvector".parse::<VectorBackend>().unwrap(),
+            VectorBackend::Builtin
+        );
+        assert_eq!(
+            "libsql".parse::<VectorBackend>().unwrap(),
+            VectorBackend::Builtin
+        );
+        assert_eq!("".parse::<VectorBackend>().unwrap(), VectorBackend::Builtin);
+
+        assert_eq!(
+            "lancedb".parse::<VectorBackend>().unwrap(),
+            VectorBackend::LanceDb
+        );
+        assert_eq!(
+            "lance".parse::<VectorBackend>().unwrap(),
+            VectorBackend::LanceDb
+        );
+
+        assert!("invalid".parse::<VectorBackend>().is_err());
+    }
+
+    #[test]
+    fn default_lancedb_path_under_ironclaw() {
+        let path = super::default_lancedb_path();
+        assert!(path.to_string_lossy().contains("ironclaw"));
+        assert!(path.to_string_lossy().ends_with("lancedb"));
+    }
 }

src/config/heartbeat.rs

@@ -1,4 +1,4 @@
-use crate::config::helpers::{optional_env, parse_bool_env, parse_optional_env};
+use crate::config::helpers::{optional_env, parse_bool_env, parse_option_env, parse_optional_env};
 use crate::error::ConfigError;
 use crate::settings::Settings;
 
@@ -13,6 +13,12 @@ pub struct HeartbeatConfig {
     pub notify_channel: Option<String>,
     /// User ID to notify on heartbeat findings.
     pub notify_user: Option<String>,
+    /// Hour (0-23) when quiet hours start.
+    pub quiet_hours_start: Option<u32>,
+    /// Hour (0-23) when quiet hours end.
+    pub quiet_hours_end: Option<u32>,
+    /// Timezone for quiet hours evaluation (IANA name).
+    pub timezone: Option<String>,
 }
 
 impl Default for HeartbeatConfig {
@@ -22,6 +28,9 @@ impl Default for HeartbeatConfig {
             interval_secs: 1800, // 30 minutes
             notify_channel: None,
             notify_user: None,
+            quiet_hours_start: None,
+            quiet_hours_end: None,
+            timezone: None,
         }
     }
 }
@@ -38,6 +47,98 @@ impl HeartbeatConfig {
                 .or_else(|| settings.heartbeat.notify_channel.clone()),
             notify_user: optional_env("HEARTBEAT_NOTIFY_USER")?
                 .or_else(|| settings.heartbeat.notify_user.clone()),
+            quiet_hours_start: parse_option_env::<u32>("HEARTBEAT_QUIET_START")?
+                .or(settings.heartbeat.quiet_hours_start)
+                .map(|h| {
+                    if h > 23 {
+                        return Err(ConfigError::InvalidValue {
+                            key: "HEARTBEAT_QUIET_START".into(),
+                            message: "must be 0-23".into(),
+                        });
+                    }
+                    Ok(h)
+                })
+                .transpose()?,
+            quiet_hours_end: parse_option_env::<u32>("HEARTBEAT_QUIET_END")?
+                .or(settings.heartbeat.quiet_hours_end)
+                .map(|h| {
+                    if h > 23 {
+                        return Err(ConfigError::InvalidValue {
+                            key: "HEARTBEAT_QUIET_END".into(),
+                            message: "must be 0-23".into(),
+                        });
+                    }
+                    Ok(h)
+                })
+                .transpose()?,
+            timezone: {
+                let tz = optional_env("HEARTBEAT_TIMEZONE")?
+                    .or_else(|| settings.heartbeat.timezone.clone());
+                if let Some(ref tz_str) = tz
+                    && crate::timezone::parse_timezone(tz_str).is_none()
+                {
+                    return Err(ConfigError::InvalidValue {
+                        key: "HEARTBEAT_TIMEZONE".into(),
+                        message: format!("invalid IANA timezone: '{tz_str}'"),
+                    });
+                }
+                tz
+            },
         })
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_quiet_hours_settings_fallback() {
+        // When env vars are not set, settings values should be used
+        let mut settings = Settings::default();
+        settings.heartbeat.quiet_hours_start = Some(22);
+        settings.heartbeat.quiet_hours_end = Some(6);
+
+        let config = HeartbeatConfig::resolve(&settings).expect("resolve");
+        assert_eq!(config.quiet_hours_start, Some(22));
+        assert_eq!(config.quiet_hours_end, Some(6));
+    }
+
+    #[test]
+    fn test_quiet_hours_rejects_invalid_hour() {
+        let mut settings = Settings::default();
+        settings.heartbeat.quiet_hours_start = Some(24);
+
+        let result = HeartbeatConfig::resolve(&settings);
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_quiet_hours_accepts_boundary_values() {
+        let mut settings = Settings::default();
+        settings.heartbeat.quiet_hours_start = Some(0);
+        settings.heartbeat.quiet_hours_end = Some(23);
+
+        let config = HeartbeatConfig::resolve(&settings).expect("resolve");
+        assert_eq!(config.quiet_hours_start, Some(0));
+        assert_eq!(config.quiet_hours_end, Some(23));
+    }
+
+    #[test]
+    fn test_heartbeat_timezone_rejects_invalid() {
+        let mut settings = Settings::default();
+        settings.heartbeat.timezone = Some("Fake/Zone".to_string());
+
+        let result = HeartbeatConfig::resolve(&settings);
+        assert!(result.is_err(), "invalid IANA timezone should be rejected");
+    }
+
+    #[test]
+    fn test_heartbeat_timezone_accepts_valid() {
+        let mut settings = Settings::default();
+        settings.heartbeat.timezone = Some("America/New_York".to_string());
+
+        let config = HeartbeatConfig::resolve(&settings).expect("resolve");
+        assert_eq!(config.timezone.as_deref(), Some("America/New_York"));
+    }
+}

src/config/helpers.rs

@@ -25,8 +25,13 @@ pub(crate) fn optional_env(key: &str) -> Result<Option<String>, ConfigError> {
     }
 
     // Fall back to thread-safe overlay (secrets injected from DB)
-    if let Some(val) = INJECTED_VARS.get().and_then(|map| map.get(key)) {
-        return Ok(Some(val.clone()));
+    if let Some(val) = INJECTED_VARS
+        .lock()
+        .unwrap_or_else(|p| p.into_inner())
+        .get(key)
+        .cloned()
+    {
+        return Ok(Some(val));
     }
 
     Ok(None)

src/config/llm.rs

@@ -9,6 +9,13 @@ use crate::llm::registry::{ProviderProtocol, ProviderRegistry};
 use crate::llm::session::SessionConfig;
 use crate::settings::Settings;
 
+/// Sentinel value used as `api_key` when only an OAuth token is present.
+///
+/// When we only have an OAuth token the provider factory in `llm/mod.rs`
+/// checks for this value and routes to `AnthropicOAuthProvider`, so this
+/// placeholder is never sent over the wire.
+pub const OAUTH_PLACEHOLDER: &str = "oauth-placeholder";
+
 /// Prompt cache retention policy for Anthropic.
 ///
 /// Controls Anthropic's automatic prompt caching via a top-level
@@ -66,6 +73,7 @@ pub struct RegistryProviderConfig {
     /// Provider identifier (e.g., "groq", "openai", "tinfoil").
     pub provider_id: String,
     /// API key (optional for some providers like Ollama).
+    /// For Anthropic OAuth, this is set to `OAUTH_PLACEHOLDER`.
     pub api_key: Option<SecretString>,
     /// Base URL for the API endpoint.
     pub base_url: String,
@@ -73,6 +81,9 @@ pub struct RegistryProviderConfig {
     pub model: String,
     /// Extra HTTP headers injected into every request.
     pub extra_headers: Vec<(String, String)>,
+    /// OAuth token for providers that support Bearer auth (e.g. Anthropic via `claude login`).
+    /// When set, the provider factory routes to the OAuth-specific provider implementation.
+    pub oauth_token: Option<SecretString>,
 }
 
 /// LLM provider configuration.
@@ -92,6 +103,10 @@ pub struct LlmConfig {
     /// Resolved provider config for registry-based providers.
     /// `None` when backend is "nearai".
     pub provider: Option<RegistryProviderConfig>,
+    /// HTTP request timeout in seconds for LLM API calls.
+    /// Default: 120. Increase for local LLMs (Ollama, vLLM, LM Studio) that
+    /// need more time for prompt evaluation on consumer hardware.
+    pub request_timeout_secs: u64,
 }
 
 /// NEAR AI configuration.
@@ -154,6 +169,7 @@ impl LlmConfig {
                 smart_routing_cascade: false,
             },
             provider: None,
+            request_timeout_secs: 120,
         }
     }
 
@@ -243,6 +259,8 @@ impl LlmConfig {
             )?)
         };
 
+        let request_timeout_secs = parse_optional_env("LLM_REQUEST_TIMEOUT_SECS", 120)?;
+
         Ok(Self {
             backend: if is_nearai {
                 "nearai".to_string()
@@ -254,6 +272,7 @@ impl LlmConfig {
             session,
             nearai,
             provider,
+            request_timeout_secs,
         })
     }
 
@@ -366,6 +385,22 @@ impl LlmConfig {
             Vec::new()
         };
 
+        // Resolve OAuth token (Anthropic-specific: `claude login` flow).
+        // Only check for OAuth token when the provider is actually Anthropic.
+        let oauth_token = if canonical_id == "anthropic" {
+            optional_env("ANTHROPIC_OAUTH_TOKEN")?.map(SecretString::from)
+        } else {
+            None
+        };
+        let api_key = if api_key.is_none() && oauth_token.is_some() {
+            // OAuth token present but no API key: use a placeholder so the
+            // config block is populated. The provider factory will route to
+            // the OAuth provider instead of rig-core's x-api-key client.
+            Some(SecretString::from(OAUTH_PLACEHOLDER.to_string()))
+        } else {
+            api_key
+        };
+
         Ok(RegistryProviderConfig {
             protocol,
             provider_id: canonical_id.to_string(),
@@ -373,6 +408,7 @@ impl LlmConfig {
             base_url,
             model,
             extra_headers,
+            oauth_token,
         })
     }
 }
@@ -677,8 +713,6 @@ mod tests {
 
     #[test]
     fn backend_alias_normalized_to_canonical_id() {
-        // When the user sets LLM_BACKEND to an alias (e.g., "open_ai"),
-        // LlmConfig.backend should resolve to the canonical ID ("openai").
         let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
         clear_openai_compatible_env();
         // SAFETY: Under ENV_MUTEX.
@@ -705,8 +739,6 @@ mod tests {
 
     #[test]
     fn unknown_backend_falls_back_to_openai_compatible() {
-        // An unrecognized LLM_BACKEND should fall back to the openai_compatible
-        // provider definition instead of erroring.
         let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
         clear_openai_compatible_env();
         // SAFETY: Under ENV_MUTEX.
@@ -717,7 +749,6 @@ mod tests {
 
         let settings = Settings::default();
         let cfg = LlmConfig::resolve(&settings).expect("resolve should succeed");
-        // Falls back to openai_compatible since "some_custom_provider" is unknown
         assert_eq!(cfg.backend, "openai_compatible");
         let provider = cfg.provider.expect("should have provider config");
         assert_eq!(provider.provider_id, "openai_compatible");
@@ -759,7 +790,6 @@ mod tests {
 
     #[test]
     fn base_url_resolution_priority() {
-        // Env var > settings > registry default
         let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
         clear_openai_compatible_env();
 
@@ -800,6 +830,119 @@ mod tests {
         }
     }
 
+    // ── OAuth resolution tests ──────────────────────────────────────
+
+    /// Clear all Anthropic-related env vars.
+    fn clear_anthropic_env() {
+        // SAFETY: Only called under ENV_MUTEX in tests.
+        unsafe {
+            std::env::remove_var("LLM_BACKEND");
+            std::env::remove_var("ANTHROPIC_API_KEY");
+            std::env::remove_var("ANTHROPIC_OAUTH_TOKEN");
+            std::env::remove_var("ANTHROPIC_MODEL");
+            std::env::remove_var("ANTHROPIC_BASE_URL");
+        }
+    }
+
+    #[test]
+    fn anthropic_oauth_token_sets_placeholder_api_key() {
+        use secrecy::ExposeSecret;
+
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        clear_anthropic_env();
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::set_var("ANTHROPIC_OAUTH_TOKEN", "sk-ant-oat01-test-token");
+        }
+
+        let settings = Settings {
+            llm_backend: Some("anthropic".to_string()),
+            ..Default::default()
+        };
+        let cfg = LlmConfig::resolve(&settings).expect("resolve should succeed");
+        let provider = cfg.provider.expect("provider config should be present");
+
+        assert_eq!(
+            provider
+                .api_key
+                .as_ref()
+                .map(|k| k.expose_secret().to_string()),
+            Some(OAUTH_PLACEHOLDER.to_string()),
+            "api_key should be the OAuth placeholder when only OAuth token is set"
+        );
+        assert!(
+            provider.oauth_token.is_some(),
+            "oauth_token should be populated"
+        );
+        assert_eq!(
+            provider.oauth_token.as_ref().unwrap().expose_secret(),
+            "sk-ant-oat01-test-token"
+        );
+
+        clear_anthropic_env();
+    }
+
+    #[test]
+    fn anthropic_api_key_takes_priority_over_oauth() {
+        use secrecy::ExposeSecret;
+
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        clear_anthropic_env();
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::set_var("ANTHROPIC_API_KEY", "sk-ant-real-key");
+            std::env::set_var("ANTHROPIC_OAUTH_TOKEN", "sk-ant-oat01-test-token");
+        }
+
+        let settings = Settings {
+            llm_backend: Some("anthropic".to_string()),
+            ..Default::default()
+        };
+        let cfg = LlmConfig::resolve(&settings).expect("resolve should succeed");
+        let provider = cfg.provider.expect("provider config should be present");
+
+        assert_eq!(
+            provider
+                .api_key
+                .as_ref()
+                .map(|k| k.expose_secret().to_string()),
+            Some("sk-ant-real-key".to_string()),
+            "real API key should take priority over OAuth placeholder"
+        );
+        assert!(
+            provider.oauth_token.is_some(),
+            "oauth_token should still be populated"
+        );
+
+        clear_anthropic_env();
+    }
+
+    #[test]
+    fn non_anthropic_provider_has_no_oauth_token() {
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        clear_anthropic_env();
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::set_var("ANTHROPIC_OAUTH_TOKEN", "sk-ant-oat01-test-token");
+        }
+
+        let settings = Settings {
+            llm_backend: Some("openai".to_string()),
+            ..Default::default()
+        };
+        let cfg = LlmConfig::resolve(&settings).expect("resolve should succeed");
+        let provider = cfg.provider.expect("provider config should be present");
+
+        assert!(
+            provider.oauth_token.is_none(),
+            "non-Anthropic providers should not pick up ANTHROPIC_OAUTH_TOKEN"
+        );
+
+        clear_anthropic_env();
+    }
+
+    // ── Cache retention tests ───────────────────────────────────────
+
     #[test]
     fn cache_retention_from_str_primary_values() {
         assert_eq!(
@@ -881,4 +1024,30 @@ mod tests {
             assert_eq!(parsed, variant, "round-trip failed for {s}");
         }
     }
+
+    #[test]
+    fn test_request_timeout_defaults_to_120() {
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::remove_var("LLM_REQUEST_TIMEOUT_SECS");
+        }
+        let config = LlmConfig::resolve(&Settings::default()).expect("resolve");
+        assert_eq!(config.request_timeout_secs, 120);
+    }
+
+    #[test]
+    fn test_request_timeout_configurable() {
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::set_var("LLM_REQUEST_TIMEOUT_SECS", "300");
+        }
+        let config = LlmConfig::resolve(&Settings::default()).expect("resolve");
+        assert_eq!(config.request_timeout_secs, 300);
+        // SAFETY: Cleanup
+        unsafe {
+            std::env::remove_var("LLM_REQUEST_TIMEOUT_SECS");
+        }
+    }
 }

src/config/mod.rs

@@ -13,7 +13,7 @@ mod embeddings;
 mod heartbeat;
 pub(crate) mod helpers;
 mod hygiene;
-mod llm;
+pub(crate) mod llm;
 mod routines;
 mod safety;
 mod sandbox;
@@ -24,7 +24,7 @@ mod tunnel;
 mod wasm;
 
 use std::collections::HashMap;
-use std::sync::OnceLock;
+use std::sync::{LazyLock, Mutex};
 
 use crate::error::ConfigError;
 use crate::settings::Settings;
@@ -33,7 +33,10 @@ use crate::settings::Settings;
 pub use self::agent::AgentConfig;
 pub use self::builder::BuilderModeConfig;
 pub use self::channels::{ChannelsConfig, CliConfig, GatewayConfig, HttpConfig, SignalConfig};
-pub use self::database::{DatabaseBackend, DatabaseConfig, SslMode, default_libsql_path};
+pub use self::database::{
+    DatabaseBackend, DatabaseConfig, SslMode, VectorBackend, default_lancedb_path,
+    default_libsql_path,
+};
 pub use self::embeddings::EmbeddingsConfig;
 pub use self::heartbeat::HeartbeatConfig;
 pub use self::hygiene::HygieneConfig;
@@ -53,7 +56,12 @@ pub use crate::llm::session::SessionConfig;
 /// Used by `inject_llm_keys_from_secrets()` to make API keys available to
 /// `optional_env()` without unsafe `set_var` calls. `optional_env()` checks
 /// real env vars first, then falls back to this overlay.
-static INJECTED_VARS: OnceLock<HashMap<String, String>> = OnceLock::new();
+///
+/// Uses `Mutex<HashMap>` instead of `OnceLock` so that both
+/// `inject_os_credentials()` and `inject_llm_keys_from_secrets()` can merge
+/// their data. Whichever runs first initialises the map; the second merges in.
+static INJECTED_VARS: LazyLock<Mutex<HashMap<String, String>>> =
+    LazyLock::new(|| Mutex::new(HashMap::new()));
 
 /// Main configuration for the agent.
 #[derive(Debug, Clone)]
@@ -102,6 +110,8 @@ impl Config {
                 libsql_path: Some(libsql_path),
                 libsql_url: None,
                 libsql_auth_token: None,
+                vector_backend: VectorBackend::default(),
+                lancedb_path: None,
             },
             llm: LlmConfig::for_testing(),
             embeddings: EmbeddingsConfig::default(),
@@ -252,6 +262,32 @@ impl Config {
         Ok(())
     }
 
+    /// Re-resolve only the LLM config after credential injection.
+    ///
+    /// Called by `AppBuilder::init_secrets()` after injecting API keys into
+    /// the env overlay. Only rebuilds `self.llm` — all other config fields
+    /// are unaffected, preserving values from the initial config load (or
+    /// from `Config::for_testing()` in test mode).
+    pub async fn re_resolve_llm(
+        &mut self,
+        store: Option<&(dyn crate::db::SettingsStore + Sync)>,
+        user_id: &str,
+        toml_path: Option<&std::path::Path>,
+    ) -> Result<(), ConfigError> {
+        let settings = if let Some(store) = store {
+            let mut s = match store.get_all_settings(user_id).await {
+                Ok(map) => Settings::from_db_map(&map),
+                Err(_) => Settings::default(),
+            };
+            Self::apply_toml_overlay(&mut s, toml_path)?;
+            s
+        } else {
+            Settings::default()
+        };
+        self.llm = LlmConfig::resolve(&settings)?;
+        Ok(())
+    }
+
     /// Build config from settings (shared by from_env and from_db).
     async fn build(settings: &Settings) -> Result<Self, ConfigError> {
         Ok(Self {
@@ -285,6 +321,9 @@ impl Config {
 /// env-var-first resolution in `LlmConfig::resolve()`. Keys in the overlay
 /// are read by `optional_env()` before falling back to `std::env::var()`,
 /// so explicit env vars always win.
+///
+/// Also loads tokens from OS credential stores (macOS Keychain, Linux
+/// credentials files) which don't require the secrets DB.
 pub async fn inject_llm_keys_from_secrets(
     secrets: &dyn crate::secrets::SecretsStore,
     user_id: &str,
@@ -292,7 +331,10 @@ pub async fn inject_llm_keys_from_secrets(
     // Static mappings for well-known providers.
     // The registry's setup hints define secret_name -> env_var mappings,
     // so new providers added to providers.json get injection automatically.
-    let mut mappings: Vec<(&str, &str)> = vec![("llm_nearai_api_key", "NEARAI_API_KEY")];
+    let mut mappings: Vec<(&str, &str)> = vec![
+        ("llm_nearai_api_key", "NEARAI_API_KEY"),
+        ("llm_anthropic_oauth_token", "ANTHROPIC_OAUTH_TOKEN"),
+    ];
 
     // Dynamically discover secret->env mappings from the provider registry.
     // Uses selectable() which deduplicates user overrides correctly.
@@ -331,5 +373,62 @@ pub async fn inject_llm_keys_from_secrets(
         }
     }
 
-    let _ = INJECTED_VARS.set(injected);
+    inject_os_credential_store_tokens(&mut injected);
+
+    merge_injected_vars(injected);
+}
+
+/// Load tokens from OS credential stores (no DB required).
+///
+/// Called unconditionally during startup — even when the encrypted secrets DB
+/// is unavailable (no master key, no DB connection). This ensures OAuth tokens
+/// from `claude login` (macOS Keychain / Linux credentials.json)
+/// are available for config resolution.
+pub fn inject_os_credentials() {
+    let mut injected = HashMap::new();
+    inject_os_credential_store_tokens(&mut injected);
+    merge_injected_vars(injected);
+}
+
+/// Merge new entries into the global injected-vars overlay.
+///
+/// New keys are inserted; existing keys are overwritten (later callers win,
+/// e.g. fresh OS credential store tokens override stale DB copies).
+fn merge_injected_vars(new_entries: HashMap<String, String>) {
+    if new_entries.is_empty() {
+        return;
+    }
+    match INJECTED_VARS.lock() {
+        Ok(mut map) => map.extend(new_entries),
+        Err(poisoned) => poisoned.into_inner().extend(new_entries),
+    }
+}
+
+/// Inject a single key-value pair into the overlay.
+///
+/// Used by the setup wizard to make credentials available to `optional_env()`
+/// without calling `unsafe { std::env::set_var }`.
+pub fn inject_single_var(key: &str, value: &str) {
+    match INJECTED_VARS.lock() {
+        Ok(mut map) => {
+            map.insert(key.to_string(), value.to_string());
+        }
+        Err(poisoned) => {
+            poisoned
+                .into_inner()
+                .insert(key.to_string(), value.to_string());
+        }
+    }
+}
+
+/// Shared helper: extract tokens from OS credential stores into the overlay map.
+fn inject_os_credential_store_tokens(injected: &mut HashMap<String, String>) {
+    // Try the OS credential store for a fresh Anthropic OAuth token.
+    // Tokens from `claude login` expire in 8-12h, so the DB copy may be stale.
+    // A fresh extraction from macOS Keychain / Linux credentials.json wins
+    // over the (possibly expired) copy stored in the encrypted secrets DB.
+    if let Some(fresh) = crate::config::ClaudeCodeConfig::extract_oauth_token() {
+        injected.insert("ANTHROPIC_OAUTH_TOKEN".to_string(), fresh);
+        tracing::debug!("Refreshed ANTHROPIC_OAUTH_TOKEN from OS credential store");
+    }
 }

src/config/sandbox.rs

@@ -233,9 +233,14 @@ impl ClaudeCodeConfig {
 /// Expected shape: `{"claudeAiOauth": {"accessToken": "sk-ant-oat01-..."}}`
 fn parse_oauth_access_token(json: &str) -> Option<String> {
     let creds: serde_json::Value = serde_json::from_str(json).ok()?;
-    creds["claudeAiOauth"]["accessToken"]
-        .as_str()
-        .map(String::from)
+    let token = creds["claudeAiOauth"]["accessToken"].as_str()?;
+    // Validate that the token looks like a real OAuth token before using it.
+    // Claude CLI tokens start with "sk-ant-oat".
+    if !token.starts_with("sk-ant-oat") {
+        tracing::debug!("Ignoring credential store token with unexpected prefix");
+        return None;
+    }
+    Some(token.to_string())
 }
 
 #[cfg(test)]
@@ -401,14 +406,14 @@ mod tests {
     fn parse_oauth_token_nested_extra_fields() {
         let json = r#"{
             "claudeAiOauth": {
-                "accessToken": "sk-ant-real-token",
+                "accessToken": "sk-ant-oat01-real-token",
                 "refreshToken": "rt-abc",
                 "expiresAt": 1700000000
             }
         }"#;
         assert_eq!(
             parse_oauth_access_token(json),
-            Some("sk-ant-real-token".to_string())
+            Some("sk-ant-oat01-real-token".to_string())
         );
     }
 
@@ -418,6 +423,12 @@ mod tests {
         assert_eq!(parse_oauth_access_token(json), None);
     }
 
+    #[test]
+    fn parse_oauth_token_rejects_invalid_prefix() {
+        let json = r#"{"claudeAiOauth": {"accessToken": "not-an-oauth-token"}}"#;
+        assert_eq!(parse_oauth_access_token(json), None);
+    }
+
     // ── default_claude_code_allowed_tools ───────────────────────────
 
     #[test]

src/context/state.rs

@@ -164,6 +164,8 @@ pub struct JobContext {
     /// previous results by ID via `$tool_call_id` parameter syntax.
     #[serde(skip)]
     pub tool_output_stash: Arc<tokio::sync::RwLock<HashMap<String, String>>>,
+    /// User's preferred timezone (IANA name, e.g. "America/New_York"). Defaults to "UTC".
+    pub user_timezone: String,
 }
 
 impl JobContext {
@@ -203,9 +205,16 @@ impl JobContext {
             http_interceptor: None,
             metadata: serde_json::Value::Null,
             tool_output_stash: Arc::new(tokio::sync::RwLock::new(HashMap::new())),
+            user_timezone: "UTC".to_string(),
         }
     }
 
+    /// Set the user timezone on this context.
+    pub fn with_timezone(mut self, tz: impl Into<String>) -> Self {
+        self.user_timezone = tz.into();
+        self
+    }
+
     /// Transition to a new state.
     pub fn transition_to(
         &mut self,

src/db/libsql/conversations.rs

@@ -20,9 +20,10 @@ impl ConversationStore for LibSqlBackend {
     ) -> Result<Uuid, DatabaseError> {
         let conn = self.connect().await?;
         let id = Uuid::new_v4();
+        let now = fmt_ts(&Utc::now());
         conn.execute(
-            "INSERT INTO conversations (id, channel, user_id, thread_id) VALUES (?1, ?2, ?3, ?4)",
-            params![id.to_string(), channel, user_id, opt_text(thread_id)],
+            "INSERT INTO conversations (id, channel, user_id, thread_id, started_at, last_activity) VALUES (?1, ?2, ?3, ?4, ?5, ?5)",
+            params![id.to_string(), channel, user_id, opt_text(thread_id), now],
         )
         .await
         .map_err(|e| DatabaseError::Query(e.to_string()))?;
@@ -71,8 +72,8 @@ impl ConversationStore for LibSqlBackend {
         let now = fmt_ts(&Utc::now());
         conn.execute(
             r#"
-                INSERT INTO conversations (id, channel, user_id, thread_id)
-                VALUES (?1, ?2, ?3, ?4)
+                INSERT INTO conversations (id, channel, user_id, thread_id, started_at, last_activity)
+                VALUES (?1, ?2, ?3, ?4, ?5, ?5)
                 ON CONFLICT (id) DO UPDATE SET last_activity = ?5
                 "#,
             params![id.to_string(), channel, user_id, opt_text(thread_id), now],
@@ -97,6 +98,7 @@ impl ConversationStore for LibSqlBackend {
                     c.started_at,
                     c.last_activity,
                     c.metadata,
+                    c.channel,
                     (SELECT COUNT(*) FROM conversation_messages m WHERE m.conversation_id = c.id AND m.role = 'user') AS message_count,
                     (SELECT substr(m2.content, 1, 100)
                      FROM conversation_messages m2
@@ -106,7 +108,7 @@ impl ConversationStore for LibSqlBackend {
                     ) AS title
                 FROM conversations c
                 WHERE c.user_id = ?1 AND c.channel = ?2
-                ORDER BY c.last_activity DESC
+                ORDER BY datetime(c.last_activity) DESC
                 LIMIT ?3
                 "#,
                 params![user_id, channel, limit],
@@ -125,6 +127,13 @@ impl ConversationStore for LibSqlBackend {
                 .get("thread_type")
                 .and_then(|v| v.as_str())
                 .map(String::from);
+            let sql_title = get_opt_text(&row, 6);
+            let title = sql_title.or_else(|| {
+                metadata
+                    .get("routine_name")
+                    .and_then(|v| v.as_str())
+                    .map(String::from)
+            });
             results.push(ConversationSummary {
                 id: row
                     .get::<String>(0)
@@ -133,14 +142,213 @@ impl ConversationStore for LibSqlBackend {
                     .unwrap_or_default(),
                 started_at: get_ts(&row, 1),
                 last_activity: get_ts(&row, 2),
-                message_count: get_i64(&row, 4),
-                title: get_opt_text(&row, 5),
+                message_count: get_i64(&row, 5),
+                title,
                 thread_type,
+                channel: get_text(&row, 4),
             });
         }
         Ok(results)
     }
 
+    async fn list_conversations_all_channels(
+        &self,
+        user_id: &str,
+        limit: i64,
+    ) -> Result<Vec<ConversationSummary>, DatabaseError> {
+        let conn = self.connect().await?;
+        let mut rows = conn
+            .query(
+                r#"
+                SELECT
+                    c.id,
+                    c.started_at,
+                    c.last_activity,
+                    c.metadata,
+                    c.channel,
+                    (SELECT COUNT(*) FROM conversation_messages m WHERE m.conversation_id = c.id AND m.role = 'user') AS message_count,
+                    (SELECT substr(m2.content, 1, 100)
+                     FROM conversation_messages m2
+                     WHERE m2.conversation_id = c.id AND m2.role = 'user'
+                     ORDER BY m2.created_at ASC, m2.rowid ASC
+                     LIMIT 1
+                    ) AS title
+                FROM conversations c
+                WHERE c.user_id = ?1
+                ORDER BY datetime(c.last_activity) DESC
+                LIMIT ?2
+                "#,
+                params![user_id, limit],
+            )
+            .await
+            .map_err(|e| DatabaseError::Query(e.to_string()))?;
+
+        let mut results = Vec::new();
+        while let Some(row) = rows
+            .next()
+            .await
+            .map_err(|e| DatabaseError::Query(e.to_string()))?
+        {
+            let metadata = get_json(&row, 3);
+            let thread_type = metadata
+                .get("thread_type")
+                .and_then(|v| v.as_str())
+                .map(String::from);
+            let sql_title = get_opt_text(&row, 6);
+            let title = sql_title.or_else(|| {
+                metadata
+                    .get("routine_name")
+                    .and_then(|v| v.as_str())
+                    .map(String::from)
+            });
+            results.push(ConversationSummary {
+                id: row
+                    .get::<String>(0)
+                    .unwrap_or_default()
+                    .parse()
+                    .unwrap_or_default(),
+                started_at: get_ts(&row, 1),
+                last_activity: get_ts(&row, 2),
+                message_count: get_i64(&row, 5),
+                title,
+                thread_type,
+                channel: get_text(&row, 4),
+            });
+        }
+        Ok(results)
+    }
+
+    /// Uses BEGIN IMMEDIATE to serialize concurrent writers and prevent
+    /// duplicate routine conversations (TOCTOU race).
+    async fn get_or_create_routine_conversation(
+        &self,
+        routine_id: Uuid,
+        routine_name: &str,
+        user_id: &str,
+    ) -> Result<Uuid, DatabaseError> {
+        let conn = self.connect().await?;
+        let rid = routine_id.to_string();
+
+        conn.execute("BEGIN IMMEDIATE", params![])
+            .await
+            .map_err(|e| DatabaseError::Query(e.to_string()))?;
+
+        let result: Result<Uuid, DatabaseError> = async {
+            let mut rows = conn
+                .query(
+                    r#"
+                    SELECT id FROM conversations
+                    WHERE user_id = ?1 AND json_extract(metadata, '$.routine_id') = ?2
+                    LIMIT 1
+                    "#,
+                    params![user_id, rid],
+                )
+                .await
+                .map_err(|e| DatabaseError::Query(e.to_string()))?;
+
+            if let Some(row) = rows
+                .next()
+                .await
+                .map_err(|e| DatabaseError::Query(e.to_string()))?
+            {
+                let id_str: String = row.get(0).unwrap_or_default();
+                return id_str
+                    .parse()
+                    .map_err(|_| DatabaseError::Serialization("Invalid UUID".to_string()));
+            }
+
+            let id = Uuid::new_v4();
+            let now = fmt_ts(&Utc::now());
+            let metadata = serde_json::json!({
+                "thread_type": "routine",
+                "routine_id": routine_id.to_string(),
+                "routine_name": routine_name,
+            });
+            conn.execute(
+                "INSERT INTO conversations (id, channel, user_id, metadata, started_at, last_activity) VALUES (?1, ?2, ?3, ?4, ?5, ?5)",
+                params![id.to_string(), "routine", user_id, metadata.to_string(), now],
+            )
+            .await
+            .map_err(|e| DatabaseError::Query(e.to_string()))?;
+            Ok(id)
+        }
+        .await;
+
+        match &result {
+            Ok(_) => {
+                conn.execute("COMMIT", params![])
+                    .await
+                    .map_err(|e| DatabaseError::Query(e.to_string()))?;
+            }
+            Err(_) => {
+                let _ = conn.execute("ROLLBACK", params![]).await;
+            }
+        }
+        result
+    }
+
+    /// Uses BEGIN IMMEDIATE to serialize concurrent writers and prevent
+    /// duplicate heartbeat conversations (TOCTOU race).
+    async fn get_or_create_heartbeat_conversation(
+        &self,
+        user_id: &str,
+    ) -> Result<Uuid, DatabaseError> {
+        let conn = self.connect().await?;
+
+        conn.execute("BEGIN IMMEDIATE", params![])
+            .await
+            .map_err(|e| DatabaseError::Query(e.to_string()))?;
+
+        let result: Result<Uuid, DatabaseError> = async {
+            let mut rows = conn
+                .query(
+                    r#"
+                    SELECT id FROM conversations
+                    WHERE user_id = ?1 AND json_extract(metadata, '$.thread_type') = 'heartbeat'
+                    LIMIT 1
+                    "#,
+                    params![user_id],
+                )
+                .await
+                .map_err(|e| DatabaseError::Query(e.to_string()))?;
+
+            if let Some(row) = rows
+                .next()
+                .await
+                .map_err(|e| DatabaseError::Query(e.to_string()))?
+            {
+                let id_str: String = row.get(0).unwrap_or_default();
+                return id_str
+                    .parse()
+                    .map_err(|_| DatabaseError::Serialization("Invalid UUID".to_string()));
+            }
+
+            let id = Uuid::new_v4();
+            let now = fmt_ts(&Utc::now());
+            let metadata = serde_json::json!({ "thread_type": "heartbeat" });
+            conn.execute(
+                "INSERT INTO conversations (id, channel, user_id, metadata, started_at, last_activity) VALUES (?1, ?2, ?3, ?4, ?5, ?5)",
+                params![id.to_string(), "heartbeat", user_id, metadata.to_string(), now],
+            )
+            .await
+            .map_err(|e| DatabaseError::Query(e.to_string()))?;
+            Ok(id)
+        }
+        .await;
+
+        match &result {
+            Ok(_) => {
+                conn.execute("COMMIT", params![])
+                    .await
+                    .map_err(|e| DatabaseError::Query(e.to_string()))?;
+            }
+            Err(_) => {
+                let _ = conn.execute("ROLLBACK", params![]).await;
+            }
+        }
+        result
+    }
+
     async fn get_or_create_assistant_conversation(
         &self,
         user_id: &str,
@@ -174,10 +382,11 @@ impl ConversationStore for LibSqlBackend {
 
         // Create new
         let id = Uuid::new_v4();
+        let now = fmt_ts(&Utc::now());
         let metadata = serde_json::json!({"thread_type": "assistant", "title": "Assistant"});
         conn.execute(
-            "INSERT INTO conversations (id, channel, user_id, metadata) VALUES (?1, ?2, ?3, ?4)",
-            params![id.to_string(), channel, user_id, metadata.to_string()],
+            "INSERT INTO conversations (id, channel, user_id, metadata, started_at, last_activity) VALUES (?1, ?2, ?3, ?4, ?5, ?5)",
+            params![id.to_string(), channel, user_id, metadata.to_string(), now],
         )
         .await
         .map_err(|e| DatabaseError::Query(e.to_string()))?;
@@ -192,9 +401,10 @@ impl ConversationStore for LibSqlBackend {
     ) -> Result<Uuid, DatabaseError> {
         let conn = self.connect().await?;
         let id = Uuid::new_v4();
+        let now = fmt_ts(&Utc::now());
         conn.execute(
-            "INSERT INTO conversations (id, channel, user_id, metadata) VALUES (?1, ?2, ?3, ?4)",
-            params![id.to_string(), channel, user_id, metadata.to_string()],
+            "INSERT INTO conversations (id, channel, user_id, metadata, started_at, last_activity) VALUES (?1, ?2, ?3, ?4, ?5, ?5)",
+            params![id.to_string(), channel, user_id, metadata.to_string(), now],
         )
         .await
         .map_err(|e| DatabaseError::Query(e.to_string()))?;
@@ -353,3 +563,128 @@ impl ConversationStore for LibSqlBackend {
         Ok(found.is_some())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::db::Database;
+
+    #[tokio::test]
+    async fn test_get_or_create_routine_conversation_is_idempotent() {
+        let dir = tempfile::tempdir().unwrap();
+        let db_path = dir.path().join("test_routine_conv.db");
+        let backend = LibSqlBackend::new_local(&db_path).await.unwrap();
+        backend.run_migrations().await.unwrap();
+
+        let routine_id = Uuid::new_v4();
+        let user_id = "test_user";
+
+        // First call — creates the conversation
+        let id1 = backend
+            .get_or_create_routine_conversation(routine_id, "my-routine", user_id)
+            .await
+            .unwrap();
+
+        // Second call — should return the SAME conversation
+        let id2 = backend
+            .get_or_create_routine_conversation(routine_id, "my-routine", user_id)
+            .await
+            .unwrap();
+
+        assert_eq!(id1, id2, "Expected same conversation ID on repeated calls");
+
+        // Third call — still the same
+        let id3 = backend
+            .get_or_create_routine_conversation(routine_id, "my-routine", user_id)
+            .await
+            .unwrap();
+
+        assert_eq!(id1, id3);
+
+        // Different routine_id should get a different conversation
+        let other_routine_id = Uuid::new_v4();
+        let id4 = backend
+            .get_or_create_routine_conversation(other_routine_id, "other-routine", user_id)
+            .await
+            .unwrap();
+
+        assert_ne!(
+            id1, id4,
+            "Different routines should get different conversations"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_routine_conversation_persists_across_messages() {
+        let dir = tempfile::tempdir().unwrap();
+        let db_path = dir.path().join("test_routine_persist.db");
+        let backend = LibSqlBackend::new_local(&db_path).await.unwrap();
+        backend.run_migrations().await.unwrap();
+
+        let routine_id = Uuid::new_v4();
+        let user_id = "test_user";
+
+        // First invocation: create conversation and add a message
+        let id1 = backend
+            .get_or_create_routine_conversation(routine_id, "my-routine", user_id)
+            .await
+            .unwrap();
+
+        backend
+            .add_conversation_message(id1, "assistant", "[cron] Completed: all good")
+            .await
+            .unwrap();
+
+        // Second invocation: should find existing conversation
+        let id2 = backend
+            .get_or_create_routine_conversation(routine_id, "my-routine", user_id)
+            .await
+            .unwrap();
+
+        assert_eq!(id1, id2, "Second invocation should reuse same conversation");
+
+        backend
+            .add_conversation_message(id2, "assistant", "[cron] Completed: still good")
+            .await
+            .unwrap();
+
+        // Verify only one routine conversation exists (not two)
+        let convs = backend
+            .list_conversations_all_channels(user_id, 50)
+            .await
+            .unwrap();
+
+        let routine_convs: Vec<_> = convs.iter().filter(|c| c.channel == "routine").collect();
+        assert_eq!(
+            routine_convs.len(),
+            1,
+            "Should have exactly 1 routine conversation, found {}",
+            routine_convs.len()
+        );
+    }
+
+    #[tokio::test]
+    async fn test_get_or_create_heartbeat_conversation_is_idempotent() {
+        let dir = tempfile::tempdir().unwrap();
+        let db_path = dir.path().join("test_heartbeat_conv.db");
+        let backend = LibSqlBackend::new_local(&db_path).await.unwrap();
+        backend.run_migrations().await.unwrap();
+
+        let user_id = "test_user";
+
+        let id1 = backend
+            .get_or_create_heartbeat_conversation(user_id)
+            .await
+            .unwrap();
+
+        let id2 = backend
+            .get_or_create_heartbeat_conversation(user_id)
+            .await
+            .unwrap();
+
+        assert_eq!(
+            id1, id2,
+            "Expected same heartbeat conversation on repeated calls"
+        );
+    }
+}

src/db/libsql/jobs.rs

@@ -121,6 +121,9 @@ impl JobStore for LibSqlBackend {
                     tool_output_stash: std::sync::Arc::new(tokio::sync::RwLock::new(
                         std::collections::HashMap::new(),
                     )),
+                    // TODO(#661): persist user_timezone in agent_jobs table so
+                    // background/routine jobs retain the session's timezone context.
+                    user_timezone: "UTC".to_string(),
                 }))
             }
             None => Ok(None),

src/db/libsql/mod.rs

@@ -118,15 +118,37 @@ impl LibSqlBackend {
     /// Sets `PRAGMA busy_timeout = 5000` on every connection so concurrent
     /// writers wait up to 5 seconds instead of failing instantly with
     /// "database is locked".
+    ///
+    /// Retries up to 3 times with exponential backoff to handle transient
+    /// "unable to open database file" errors from concurrent connection
+    /// creation (e.g. cron ticker vs main thread).
     pub async fn connect(&self) -> Result<Connection, DatabaseError> {
-        let conn = self
-            .db
-            .connect()
-            .map_err(|e| DatabaseError::Pool(format!("Failed to create connection: {}", e)))?;
-        conn.query("PRAGMA busy_timeout = 5000", ())
-            .await
-            .map_err(|e| DatabaseError::Pool(format!("Failed to set busy_timeout: {}", e)))?;
-        Ok(conn)
+        let mut last_err = None;
+        for attempt in 0..3u32 {
+            match self.db.connect() {
+                Ok(conn) => {
+                    conn.query("PRAGMA busy_timeout = 5000", ())
+                        .await
+                        .map_err(|e| {
+                            DatabaseError::Pool(format!("Failed to set busy_timeout: {}", e))
+                        })?;
+                    return Ok(conn);
+                }
+                Err(e) => {
+                    last_err = Some(e);
+                    if attempt < 2 {
+                        tokio::time::sleep(std::time::Duration::from_millis(
+                            50 * 2u64.pow(attempt),
+                        ))
+                        .await;
+                    }
+                }
+            }
+        }
+        Err(DatabaseError::Pool(format!(
+            "Failed to create connection after 3 attempts: {}",
+            last_err.map(|e| e.to_string()).unwrap_or_default()
+        )))
     }
 }
 
@@ -459,4 +481,33 @@ mod tests {
         let count: i64 = row.get(0).unwrap();
         assert_eq!(count, 20);
     }
+
+    #[tokio::test]
+    async fn test_connect_retry_succeeds_on_valid_db() {
+        // Verify connect() works with retry logic on a file-backed DB
+        // (exercises the retry path even though transient failures are hard
+        // to reproduce deterministically).
+        let dir = tempfile::tempdir().unwrap();
+        let db_path = dir.path().join("test_retry.db");
+        let backend = LibSqlBackend::new_local(&db_path).await.unwrap();
+        backend.run_migrations().await.unwrap();
+
+        // Multiple concurrent connect() calls should all succeed
+        let mut handles = Vec::new();
+        for _ in 0..10 {
+            let b = LibSqlBackend {
+                db: backend.shared_db(),
+            };
+            handles.push(tokio::spawn(async move { b.connect().await }));
+        }
+
+        for handle in handles {
+            let result = handle.await.unwrap();
+            assert!(
+                result.is_ok(),
+                "concurrent connect failed: {:?}",
+                result.err()
+            );
+        }
+    }
 }

src/db/libsql_migrations.rs

@@ -45,6 +45,15 @@ CREATE INDEX IF NOT EXISTS idx_conversations_channel ON conversations(channel);
 CREATE INDEX IF NOT EXISTS idx_conversations_user ON conversations(user_id);
 CREATE INDEX IF NOT EXISTS idx_conversations_last_activity ON conversations(last_activity);
 
+-- Partial unique indexes to prevent duplicate singleton conversations.
+CREATE UNIQUE INDEX IF NOT EXISTS uq_conv_routine
+ON conversations (user_id, json_extract(metadata, '$.routine_id'))
+WHERE json_extract(metadata, '$.routine_id') IS NOT NULL;
+
+CREATE UNIQUE INDEX IF NOT EXISTS uq_conv_heartbeat
+ON conversations (user_id)
+WHERE json_extract(metadata, '$.thread_type') = 'heartbeat';
+
 CREATE TABLE IF NOT EXISTS conversation_messages (
     id TEXT PRIMARY KEY,
     conversation_id TEXT NOT NULL REFERENCES conversations(id) ON DELETE CASCADE,

src/db/mod.rs

@@ -125,6 +125,21 @@ pub trait ConversationStore: Send + Sync {
         channel: &str,
         limit: i64,
     ) -> Result<Vec<ConversationSummary>, DatabaseError>;
+    async fn list_conversations_all_channels(
+        &self,
+        user_id: &str,
+        limit: i64,
+    ) -> Result<Vec<ConversationSummary>, DatabaseError>;
+    async fn get_or_create_routine_conversation(
+        &self,
+        routine_id: Uuid,
+        routine_name: &str,
+        user_id: &str,
+    ) -> Result<Uuid, DatabaseError>;
+    async fn get_or_create_heartbeat_conversation(
+        &self,
+        user_id: &str,
+    ) -> Result<Uuid, DatabaseError>;
     async fn get_or_create_assistant_conversation(
         &self,
         user_id: &str,

src/db/postgres.rs

@@ -116,6 +116,36 @@ impl ConversationStore for PgBackend {
             .await
     }
 
+    async fn list_conversations_all_channels(
+        &self,
+        user_id: &str,
+        limit: i64,
+    ) -> Result<Vec<ConversationSummary>, DatabaseError> {
+        self.store
+            .list_conversations_all_channels(user_id, limit)
+            .await
+    }
+
+    async fn get_or_create_routine_conversation(
+        &self,
+        routine_id: Uuid,
+        routine_name: &str,
+        user_id: &str,
+    ) -> Result<Uuid, DatabaseError> {
+        self.store
+            .get_or_create_routine_conversation(routine_id, routine_name, user_id)
+            .await
+    }
+
+    async fn get_or_create_heartbeat_conversation(
+        &self,
+        user_id: &str,
+    ) -> Result<Uuid, DatabaseError> {
+        self.store
+            .get_or_create_heartbeat_conversation(user_id)
+            .await
+    }
+
     async fn get_or_create_assistant_conversation(
         &self,
         user_id: &str,

src/error.rs

@@ -401,6 +401,9 @@ pub enum RoutineError {
     #[error("Routine not found: {id}")]
     NotFound { id: Uuid },
 
+    #[error("Not authorized to trigger routine {id}")]
+    NotAuthorized { id: Uuid },
+
     #[error("Routine {name} at max concurrent runs")]
     MaxConcurrent { name: String },
 

src/evaluation/success.rs

@@ -1,13 +1,10 @@
 //! Success evaluation for jobs.
 
-use std::sync::Arc;
-
 use async_trait::async_trait;
 use serde::{Deserialize, Serialize};
 
 use crate::context::{ActionRecord, JobContext};
 use crate::error::EvaluationError;
-use crate::llm::LlmProvider;
 
 /// Result of evaluating job success.
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -64,233 +61,132 @@ pub trait SuccessEvaluator: Send + Sync {
     ) -> Result<EvaluationResult, EvaluationError>;
 }
 
-/// Rule-based success evaluator.
-pub struct RuleBasedEvaluator {
-    /// Minimum success rate for actions.
-    min_action_success_rate: f64,
-    /// Maximum allowed failures.
-    max_failures: u32,
-}
-
-impl RuleBasedEvaluator {
-    /// Create a new rule-based evaluator.
-    pub fn new() -> Self {
-        Self {
-            min_action_success_rate: 0.8,
-            max_failures: 3,
-        }
-    }
-
-    /// Set minimum action success rate.
-    #[allow(dead_code)] // Public API for configuring evaluation threshold
-    pub fn with_min_success_rate(mut self, rate: f64) -> Self {
-        self.min_action_success_rate = rate;
-        self
-    }
-
-    /// Set maximum failures.
-    #[allow(dead_code)] // Public API for configuring failure tolerance
-    pub fn with_max_failures(mut self, max: u32) -> Self {
-        self.max_failures = max;
-        self
-    }
-}
-
-impl Default for RuleBasedEvaluator {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
-#[async_trait]
-impl SuccessEvaluator for RuleBasedEvaluator {
-    async fn evaluate(
-        &self,
-        job: &JobContext,
-        actions: &[ActionRecord],
-        _output: Option<&str>,
-    ) -> Result<EvaluationResult, EvaluationError> {
-        let mut issues = Vec::new();
-
-        // Check if there were any actions
-        if actions.is_empty() {
-            return Ok(EvaluationResult::failure(
-                "No actions were taken",
-                vec!["No actions recorded".to_string()],
-            ));
-        }
-
-        // Calculate action success rate
-        let successful = actions.iter().filter(|a| a.success).count();
-        let total = actions.len();
-        let success_rate = successful as f64 / total as f64;
-
-        if success_rate < self.min_action_success_rate {
-            issues.push(format!(
-                "Action success rate {:.1}% below threshold {:.1}%",
-                success_rate * 100.0,
-                self.min_action_success_rate * 100.0
-            ));
-        }
-
-        // Count failures
-        let failures = actions.iter().filter(|a| !a.success).count() as u32;
-        if failures > self.max_failures {
-            issues.push(format!(
-                "Too many failures: {} (max {})",
-                failures, self.max_failures
-            ));
-        }
-
-        // Check for critical errors
-        for action in actions.iter().filter(|a| !a.success) {
-            if let Some(ref error) = action.error
-                && (error.to_lowercase().contains("critical")
-                    || error.to_lowercase().contains("fatal"))
-            {
-                issues.push(format!("Critical error in {}: {}", action.tool_name, error));
-            }
-        }
-
-        // Check job state
-        if job.state != crate::context::JobState::Completed
-            && job.state != crate::context::JobState::Submitted
-        {
-            issues.push(format!("Job not in completed state: {:?}", job.state));
-        }
-
-        // Calculate quality score
-        let quality_score = if issues.is_empty() {
-            let base_score = (success_rate * 80.0) as u32;
-            let completion_bonus = if job.state == crate::context::JobState::Completed {
-                20
-            } else {
-                0
-            };
-            (base_score + completion_bonus).min(100)
-        } else {
-            ((success_rate * 50.0) as u32).min(50)
-        };
-
-        if issues.is_empty() {
-            Ok(EvaluationResult::success(
-                format!(
-                    "Job completed successfully with {}/{} actions succeeding ({:.1}%)",
-                    successful,
-                    total,
-                    success_rate * 100.0
-                ),
-                quality_score,
-            ))
-        } else {
-            Ok(EvaluationResult {
-                success: false,
-                confidence: 0.85,
-                reasoning: format!("Job had {} issues", issues.len()),
-                issues,
-                suggestions: vec![
-                    "Review failed actions for common patterns".to_string(),
-                    "Consider adjusting retry logic".to_string(),
-                ],
-                quality_score,
-            })
-        }
-    }
-}
-
-/// LLM-based success evaluator for more nuanced evaluation.
-pub struct LlmEvaluator {
-    llm: Arc<dyn LlmProvider>,
-}
-
-impl LlmEvaluator {
-    /// Create a new LLM-based evaluator.
-    #[allow(dead_code)] // Public API for LLM-based evaluation
-    pub fn new(llm: Arc<dyn LlmProvider>) -> Self {
-        Self { llm }
-    }
-}
-
-#[async_trait]
-impl SuccessEvaluator for LlmEvaluator {
-    async fn evaluate(
-        &self,
-        job: &JobContext,
-        actions: &[ActionRecord],
-        output: Option<&str>,
-    ) -> Result<EvaluationResult, EvaluationError> {
-        // Build evaluation prompt
-        let actions_summary: Vec<String> = actions
-            .iter()
-            .map(|a| {
-                format!(
-                    "- {}: {} ({})",
-                    a.tool_name,
-                    if a.success { "success" } else { "failed" },
-                    a.error.as_deref().unwrap_or("ok")
-                )
-            })
-            .collect();
-
-        let prompt = format!(
-            r#"Evaluate if this job was completed successfully.
-
-Job: {}
-Description: {}
-State: {:?}
-
-Actions taken:
-{}
-
-{}
-
-Respond in JSON format:
-{{
-    "success": true/false,
-    "confidence": 0.0-1.0,
-    "reasoning": "...",
-    "issues": ["..."],
-    "suggestions": ["..."],
-    "quality_score": 0-100
-}}"#,
-            job.title,
-            job.description,
-            job.state,
-            actions_summary.join("\n"),
-            output
-                .map(|o| format!("Output:\n{}", o))
-                .unwrap_or_default()
-        );
-
-        let request =
-            crate::llm::CompletionRequest::new(vec![crate::llm::ChatMessage::user(prompt)])
-                .with_max_tokens(1024)
-                .with_temperature(0.1);
-
-        let response = self
-            .llm
-            .complete(request)
-            .await
-            .map_err(|e| EvaluationError::Failed {
-                job_id: job.job_id,
-                reason: e.to_string(),
-            })?;
-
-        // Parse the response
-        let result: EvaluationResult =
-            serde_json::from_str(&response.content).map_err(|e| EvaluationError::Failed {
-                job_id: job.job_id,
-                reason: format!("Failed to parse LLM evaluation: {}", e),
-            })?;
-
-        Ok(result)
-    }
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::context::JobContext;
+    use crate::context::{ActionRecord, JobContext};
+    use crate::error::EvaluationError;
+
+    /// Rule-based success evaluator (test-only; no production callers).
+    struct RuleBasedEvaluator {
+        min_action_success_rate: f64,
+        max_failures: u32,
+    }
+
+    impl RuleBasedEvaluator {
+        fn new() -> Self {
+            Self {
+                min_action_success_rate: 0.8,
+                max_failures: 3,
+            }
+        }
+
+        fn with_min_success_rate(mut self, rate: f64) -> Self {
+            self.min_action_success_rate = rate;
+            self
+        }
+
+        fn with_max_failures(mut self, max: u32) -> Self {
+            self.max_failures = max;
+            self
+        }
+    }
+
+    impl Default for RuleBasedEvaluator {
+        fn default() -> Self {
+            Self::new()
+        }
+    }
+
+    #[async_trait::async_trait]
+    impl SuccessEvaluator for RuleBasedEvaluator {
+        async fn evaluate(
+            &self,
+            job: &JobContext,
+            actions: &[ActionRecord],
+            _output: Option<&str>,
+        ) -> Result<EvaluationResult, EvaluationError> {
+            let mut issues = Vec::new();
+
+            if actions.is_empty() {
+                return Ok(EvaluationResult::failure(
+                    "No actions were taken",
+                    vec!["No actions recorded".to_string()],
+                ));
+            }
+
+            let successful = actions.iter().filter(|a| a.success).count();
+            let total = actions.len();
+            let success_rate = successful as f64 / total as f64;
+
+            if success_rate < self.min_action_success_rate {
+                issues.push(format!(
+                    "Action success rate {:.1}% below threshold {:.1}%",
+                    success_rate * 100.0,
+                    self.min_action_success_rate * 100.0
+                ));
+            }
+
+            let failures = actions.iter().filter(|a| !a.success).count() as u32;
+            if failures > self.max_failures {
+                issues.push(format!(
+                    "Too many failures: {} (max {})",
+                    failures, self.max_failures
+                ));
+            }
+
+            for action in actions.iter().filter(|a| !a.success) {
+                if let Some(ref error) = action.error
+                    && (error.to_lowercase().contains("critical")
+                        || error.to_lowercase().contains("fatal"))
+                {
+                    issues.push(format!("Critical error in {}: {}", action.tool_name, error));
+                }
+            }
+
+            if job.state != crate::context::JobState::Completed
+                && job.state != crate::context::JobState::Submitted
+            {
+                issues.push(format!("Job not in completed state: {:?}", job.state));
+            }
+
+            let quality_score = if issues.is_empty() {
+                let base_score = (success_rate * 80.0) as u32;
+                let completion_bonus = if job.state == crate::context::JobState::Completed {
+                    20
+                } else {
+                    0
+                };
+                (base_score + completion_bonus).min(100)
+            } else {
+                ((success_rate * 50.0) as u32).min(50)
+            };
+
+            if issues.is_empty() {
+                Ok(EvaluationResult::success(
+                    format!(
+                        "Job completed successfully with {}/{} actions succeeding ({:.1}%)",
+                        successful,
+                        total,
+                        success_rate * 100.0
+                    ),
+                    quality_score,
+                ))
+            } else {
+                Ok(EvaluationResult {
+                    success: false,
+                    confidence: 0.85,
+                    reasoning: format!("Job had {} issues", issues.len()),
+                    issues,
+                    suggestions: vec![
+                        "Review failed actions for common patterns".to_string(),
+                        "Consider adjusting retry logic".to_string(),
+                    ],
+                    quality_score,
+                })
+            }
+        }
+    }
 
     #[tokio::test]
     async fn test_rule_based_evaluator_success() {

src/extensions/manager.rs

@@ -1405,38 +1405,6 @@ impl ExtensionManager {
         Ok(())
     }
 
-    #[allow(dead_code)] // Used by upcoming hot-activation flow
-    async fn install_bundled_channel_from_artifacts(
-        &self,
-        name: &str,
-    ) -> Result<InstallResult, ExtensionError> {
-        // Check if already installed
-        let channel_wasm = self.wasm_channels_dir.join(format!("{}.wasm", name));
-        if channel_wasm.exists() {
-            return Err(ExtensionError::AlreadyInstalled(name.to_string()));
-        }
-
-        crate::channels::wasm::install_bundled_channel(name, &self.wasm_channels_dir, false)
-            .await
-            .map_err(ExtensionError::InstallFailed)?;
-
-        tracing::info!(
-            "Installed bundled channel '{}' to {}",
-            name,
-            self.wasm_channels_dir.display()
-        );
-
-        Ok(InstallResult {
-            name: name.to_string(),
-            kind: ExtensionKind::WasmChannel,
-            message: format!(
-                "Channel '{}' installed. \
-                 Run tool_auth('{}') to configure authentication, then activate.",
-                name, name,
-            ),
-        })
-    }
-
     /// Install a WASM extension from local build artifacts (WasmBuildable source).
     ///
     /// Resolves the build directory (relative to `CARGO_MANIFEST_DIR` or absolute),

src/history/store.rs

@@ -241,6 +241,9 @@ impl Store {
                     tool_output_stash: std::sync::Arc::new(tokio::sync::RwLock::new(
                         std::collections::HashMap::new(),
                     )),
+                    // TODO(#661): persist user_timezone in agent_jobs table so
+                    // background/routine jobs retain the session's timezone context.
+                    user_timezone: "UTC".to_string(),
                 }))
             }
             None => Ok(None),
@@ -1377,6 +1380,8 @@ pub struct ConversationSummary {
     pub last_activity: DateTime<Utc>,
     /// Thread type extracted from metadata (e.g. "assistant", "thread").
     pub thread_type: Option<String>,
+    /// Channel that owns this conversation (e.g. "gateway", "telegram", "routine").
+    pub channel: String,
 }
 
 /// A single message in a conversation.
@@ -1429,6 +1434,7 @@ impl Store {
                     c.started_at,
                     c.last_activity,
                     c.metadata,
+                    c.channel,
                     (SELECT COUNT(*) FROM conversation_messages m WHERE m.conversation_id = c.id AND m.role = 'user') AS message_count,
                     (SELECT LEFT(m2.content, 100)
                      FROM conversation_messages m2
@@ -1453,18 +1459,181 @@ impl Store {
                     .get("thread_type")
                     .and_then(|v| v.as_str())
                     .map(String::from);
+                let sql_title: Option<String> = r.get("title");
+                let title = sql_title.or_else(|| {
+                    metadata
+                        .get("routine_name")
+                        .and_then(|v| v.as_str())
+                        .map(String::from)
+                });
                 ConversationSummary {
                     id: r.get("id"),
-                    title: r.get("title"),
+                    title,
                     message_count: r.get("message_count"),
                     started_at: r.get("started_at"),
                     last_activity: r.get("last_activity"),
                     thread_type,
+                    channel: r.get("channel"),
                 }
             })
             .collect())
     }
 
+    /// List conversations across all channels with a title derived from the first user message.
+    pub async fn list_conversations_all_channels(
+        &self,
+        user_id: &str,
+        limit: i64,
+    ) -> Result<Vec<ConversationSummary>, DatabaseError> {
+        let conn = self.conn().await?;
+        let rows = conn
+            .query(
+                r#"
+                SELECT
+                    c.id,
+                    c.started_at,
+                    c.last_activity,
+                    c.metadata,
+                    c.channel,
+                    (SELECT COUNT(*) FROM conversation_messages m WHERE m.conversation_id = c.id AND m.role = 'user') AS message_count,
+                    (SELECT LEFT(m2.content, 100)
+                     FROM conversation_messages m2
+                     WHERE m2.conversation_id = c.id AND m2.role = 'user'
+                     ORDER BY m2.created_at ASC
+                     LIMIT 1
+                    ) AS title
+                FROM conversations c
+                WHERE c.user_id = $1
+                ORDER BY c.last_activity DESC
+                LIMIT $2
+                "#,
+                &[&user_id, &limit],
+            )
+            .await?;
+
+        Ok(rows
+            .iter()
+            .map(|r| {
+                let metadata: serde_json::Value = r.get("metadata");
+                let thread_type = metadata
+                    .get("thread_type")
+                    .and_then(|v| v.as_str())
+                    .map(String::from);
+                // For routine/heartbeat threads, derive title from metadata
+                // since they may have no user messages.
+                let sql_title: Option<String> = r.get("title");
+                let title = sql_title.or_else(|| {
+                    metadata
+                        .get("routine_name")
+                        .and_then(|v| v.as_str())
+                        .map(String::from)
+                });
+                ConversationSummary {
+                    id: r.get("id"),
+                    title,
+                    message_count: r.get("message_count"),
+                    started_at: r.get("started_at"),
+                    last_activity: r.get("last_activity"),
+                    thread_type,
+                    channel: r.get("channel"),
+                }
+            })
+            .collect())
+    }
+
+    /// Get or create a persistent conversation for a routine.
+    ///
+    /// Looks for a conversation where `metadata->>'routine_id' = routine_id`.
+    /// Creates one if it doesn't exist. Uses INSERT ON CONFLICT to avoid
+    /// TOCTOU races under concurrent routine executions.
+    pub async fn get_or_create_routine_conversation(
+        &self,
+        routine_id: Uuid,
+        routine_name: &str,
+        user_id: &str,
+    ) -> Result<Uuid, DatabaseError> {
+        let conn = self.conn().await?;
+        let rid = routine_id.to_string();
+
+        // Attempt insert first; the partial unique index
+        // uq_conv_routine(user_id, (metadata->>'routine_id')) prevents duplicates.
+        let new_id = Uuid::new_v4();
+        let metadata = serde_json::json!({
+            "thread_type": "routine",
+            "routine_id": routine_id.to_string(),
+            "routine_name": routine_name,
+        });
+        conn.execute(
+            r#"
+            INSERT INTO conversations (id, channel, user_id, metadata)
+            VALUES ($1, 'routine', $2, $3)
+            ON CONFLICT (user_id, (metadata->>'routine_id'))
+                WHERE metadata->>'routine_id' IS NOT NULL
+                DO NOTHING
+            "#,
+            &[&new_id, &user_id, &metadata],
+        )
+        .await?;
+
+        // Select back — always returns the winner.
+        let row = conn
+            .query_one(
+                r#"
+                SELECT id FROM conversations
+                WHERE user_id = $1 AND metadata->>'routine_id' = $2
+                LIMIT 1
+                "#,
+                &[&user_id, &rid],
+            )
+            .await?;
+
+        Ok(row.get("id"))
+    }
+
+    /// Get or create the singleton heartbeat conversation for a user.
+    ///
+    /// Looks for a conversation where `metadata->>'thread_type' = 'heartbeat'`.
+    /// Creates one if it doesn't exist. Uses INSERT ON CONFLICT to avoid
+    /// TOCTOU races under concurrent heartbeat sends.
+    pub async fn get_or_create_heartbeat_conversation(
+        &self,
+        user_id: &str,
+    ) -> Result<Uuid, DatabaseError> {
+        let conn = self.conn().await?;
+
+        // Attempt insert; the partial unique index
+        // uq_conv_heartbeat(user_id) prevents duplicates.
+        let new_id = Uuid::new_v4();
+        let metadata = serde_json::json!({
+            "thread_type": "heartbeat",
+        });
+        conn.execute(
+            r#"
+            INSERT INTO conversations (id, channel, user_id, metadata)
+            VALUES ($1, 'heartbeat', $2, $3)
+            ON CONFLICT (user_id)
+                WHERE metadata->>'thread_type' = 'heartbeat'
+                DO NOTHING
+            "#,
+            &[&new_id, &user_id, &metadata],
+        )
+        .await?;
+
+        // Select back — always returns the winner.
+        let row = conn
+            .query_one(
+                r#"
+                SELECT id FROM conversations
+                WHERE user_id = $1 AND metadata->>'thread_type' = 'heartbeat'
+                LIMIT 1
+                "#,
+                &[&user_id],
+            )
+            .await?;
+
+        Ok(row.get("id"))
+    }
+
     /// Get or create the singleton "assistant" conversation for a user+channel.
     ///
     /// Looks for a conversation where `metadata->>'thread_type' = 'assistant'`.
@@ -1928,3 +2097,40 @@ impl Store {
         Ok(count > 0)
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_conversation_summary_has_channel_field() {
+        // Regression: ConversationSummary must include a `channel` field
+        // so the gateway can distinguish thread origins.
+        let summary = ConversationSummary {
+            id: Uuid::nil(),
+            title: Some("Hello".to_string()),
+            message_count: 1,
+            started_at: Utc::now(),
+            last_activity: Utc::now(),
+            thread_type: Some("thread".to_string()),
+            channel: "telegram".to_string(),
+        };
+        assert_eq!(summary.channel, "telegram");
+    }
+
+    #[test]
+    fn test_conversation_summary_channel_various_values() {
+        for ch in ["gateway", "routine", "heartbeat", "telegram", "signal"] {
+            let summary = ConversationSummary {
+                id: Uuid::nil(),
+                title: None,
+                message_count: 0,
+                started_at: Utc::now(),
+                last_activity: Utc::now(),
+                thread_type: None,
+                channel: ch.to_string(),
+            };
+            assert_eq!(summary.channel, ch);
+        }
+    }
+}

src/lib.rs

@@ -66,6 +66,7 @@ pub mod service;
 pub mod settings;
 pub mod setup;
 pub mod skills;
+pub mod timezone;
 pub mod tools;
 pub mod tracing_fmt;
 pub mod transcription;

src/llm/anthropic_oauth.rs

@@ -0,0 +1,641 @@
+//! Anthropic OAuth provider (direct HTTP, `Authorization: Bearer`).
+//!
+//! This provider exists because the `rig-core` Anthropic client hardcodes the
+//! `x-api-key` header, which is rejected by Anthropic's OAuth tokens from
+//! `claude login`. OAuth tokens require `Authorization: Bearer <token>` instead.
+//!
+//! Pattern follows `nearai_chat.rs`: direct HTTP calls via `reqwest::Client`.
+
+use async_trait::async_trait;
+use reqwest::Client;
+use rust_decimal::Decimal;
+use secrecy::{ExposeSecret, SecretString};
+use serde::{Deserialize, Serialize};
+
+use crate::config::RegistryProviderConfig;
+use crate::error::LlmError;
+use crate::llm::costs;
+use crate::llm::provider::{
+    ChatMessage, CompletionRequest, CompletionResponse, FinishReason, LlmProvider, Role, ToolCall,
+    ToolCompletionRequest, ToolCompletionResponse,
+};
+
+const ANTHROPIC_API_URL: &str = "https://api.anthropic.com/v1/messages";
+/// OAuth beta requires 2023-06-01; the 2024-10-22 version is not valid with the beta flag.
+const ANTHROPIC_API_VERSION: &str = "2023-06-01";
+/// Required beta flag to enable OAuth Bearer auth on api.anthropic.com.
+/// Without this header, the API returns 401 "OAuth authentication is currently not supported."
+const ANTHROPIC_OAUTH_BETA: &str = "oauth-2025-04-20";
+const DEFAULT_MAX_TOKENS: u32 = 8192;
+
+/// Anthropic provider using OAuth Bearer authentication.
+pub struct AnthropicOAuthProvider {
+    client: Client,
+    token: SecretString,
+    model: String,
+    base_url: Option<String>,
+    active_model: std::sync::RwLock<String>,
+}
+
+impl AnthropicOAuthProvider {
+    pub fn new(config: &RegistryProviderConfig) -> Result<Self, LlmError> {
+        let token = config
+            .oauth_token
+            .clone()
+            .ok_or_else(|| LlmError::AuthFailed {
+                provider: "anthropic_oauth".to_string(),
+            })?;
+
+        let client = Client::builder()
+            .timeout(std::time::Duration::from_secs(120))
+            .build()
+            .map_err(|e| LlmError::RequestFailed {
+                provider: "anthropic_oauth".to_string(),
+                reason: format!("Failed to build HTTP client: {}", e),
+            })?;
+
+        let active_model = std::sync::RwLock::new(config.model.clone());
+        let base_url = if config.base_url.is_empty() {
+            None
+        } else {
+            Some(config.base_url.clone())
+        };
+
+        Ok(Self {
+            client,
+            token,
+            model: config.model.clone(),
+            base_url,
+            active_model,
+        })
+    }
+
+    fn api_url(&self) -> String {
+        if let Some(ref base) = self.base_url {
+            let base = base.trim_end_matches('/');
+            format!("{}/v1/messages", base)
+        } else {
+            ANTHROPIC_API_URL.to_string()
+        }
+    }
+
+    async fn send_request<R: for<'de> Deserialize<'de>>(
+        &self,
+        body: &AnthropicRequest,
+    ) -> Result<R, LlmError> {
+        let url = self.api_url();
+
+        tracing::debug!("Sending request to Anthropic OAuth: {}", url);
+
+        let response = self
+            .client
+            .post(&url)
+            .bearer_auth(self.token.expose_secret())
+            .header("anthropic-version", ANTHROPIC_API_VERSION)
+            .header("anthropic-beta", ANTHROPIC_OAUTH_BETA)
+            .header("Content-Type", "application/json")
+            .json(body)
+            .send()
+            .await
+            .map_err(|e| LlmError::RequestFailed {
+                provider: "anthropic_oauth".to_string(),
+                reason: e.to_string(),
+            })?;
+
+        let status = response.status();
+
+        if !status.is_success() {
+            // Parse Retry-After header before consuming the body.
+            let retry_after = response
+                .headers()
+                .get("retry-after")
+                .and_then(|v| v.to_str().ok())
+                .and_then(|v| v.parse::<u64>().ok())
+                .map(std::time::Duration::from_secs);
+
+            let response_text = response
+                .text()
+                .await
+                .unwrap_or_else(|e| format!("(failed to read error body: {e})"));
+
+            if status.as_u16() == 401 {
+                // OAuth tokens from `claude login` expire in ~8-12h. Attempt
+                // to re-extract a fresh token from the OS credential store
+                // (macOS Keychain / Linux credentials file) before giving up.
+                if let Some(fresh) = crate::config::ClaudeCodeConfig::extract_oauth_token() {
+                    let fresh_token = SecretString::from(fresh);
+                    // Retry once with the refreshed token
+                    let retry = self
+                        .client
+                        .post(&url)
+                        .bearer_auth(fresh_token.expose_secret())
+                        .header("anthropic-version", ANTHROPIC_API_VERSION)
+                        .header("anthropic-beta", ANTHROPIC_OAUTH_BETA)
+                        .header("Content-Type", "application/json")
+                        .json(body)
+                        .send()
+                        .await
+                        .map_err(|e| LlmError::RequestFailed {
+                            provider: "anthropic_oauth".to_string(),
+                            reason: e.to_string(),
+                        })?;
+                    if retry.status().is_success() {
+                        let text = retry.text().await.map_err(|e| LlmError::RequestFailed {
+                            provider: "anthropic_oauth".to_string(),
+                            reason: format!("Failed to read response body: {}", e),
+                        })?;
+                        return serde_json::from_str(&text).map_err(|e| {
+                            let truncated = crate::agent::truncate_for_preview(&text, 512);
+                            LlmError::InvalidResponse {
+                                provider: "anthropic_oauth".to_string(),
+                                reason: format!("JSON parse error: {}. Raw: {}", e, truncated),
+                            }
+                        });
+                    }
+                    tracing::warn!(
+                        "Anthropic OAuth 401 retry with refreshed token also failed ({})",
+                        retry.status()
+                    );
+                }
+                return Err(LlmError::AuthFailed {
+                    provider: "anthropic_oauth".to_string(),
+                });
+            }
+            if status.as_u16() == 429 {
+                return Err(LlmError::RateLimited {
+                    provider: "anthropic_oauth".to_string(),
+                    retry_after,
+                });
+            }
+            let truncated = crate::agent::truncate_for_preview(&response_text, 512);
+            return Err(LlmError::RequestFailed {
+                provider: "anthropic_oauth".to_string(),
+                reason: format!("HTTP {}: {}", status, truncated),
+            });
+        }
+
+        let response_text = response.text().await.map_err(|e| LlmError::RequestFailed {
+            provider: "anthropic_oauth".to_string(),
+            reason: format!("Failed to read response body: {}", e),
+        })?;
+
+        tracing::debug!(
+            "Anthropic OAuth response: status={}, bytes={}",
+            status,
+            response_text.len()
+        );
+
+        serde_json::from_str(&response_text).map_err(|e| {
+            let truncated = crate::agent::truncate_for_preview(&response_text, 512);
+            LlmError::InvalidResponse {
+                provider: "anthropic_oauth".to_string(),
+                reason: format!("JSON parse error: {}. Raw: {}", e, truncated),
+            }
+        })
+    }
+}
+
+#[async_trait]
+impl LlmProvider for AnthropicOAuthProvider {
+    async fn complete(&self, req: CompletionRequest) -> Result<CompletionResponse, LlmError> {
+        let model = req.model.unwrap_or_else(|| self.active_model_name());
+        let (system, messages) = convert_messages(req.messages);
+
+        let request = AnthropicRequest {
+            model,
+            messages,
+            system,
+            max_tokens: req.max_tokens.unwrap_or(DEFAULT_MAX_TOKENS),
+            temperature: req.temperature,
+            tools: None,
+            tool_choice: None,
+        };
+
+        let response: AnthropicResponse = self.send_request(&request).await?;
+        let (content, _tool_calls) = extract_response_content(&response);
+
+        let finish_reason = match response.stop_reason.as_deref() {
+            Some("end_turn") | Some("stop") => FinishReason::Stop,
+            Some("max_tokens") => FinishReason::Length,
+            Some("tool_use") => FinishReason::ToolUse,
+            _ => FinishReason::Unknown,
+        };
+
+        Ok(CompletionResponse {
+            content: content.unwrap_or_default(),
+            finish_reason,
+            input_tokens: response.usage.input_tokens,
+            output_tokens: response.usage.output_tokens,
+            cache_creation_input_tokens: response.usage.cache_creation_input_tokens,
+            cache_read_input_tokens: response.usage.cache_read_input_tokens,
+        })
+    }
+
+    async fn complete_with_tools(
+        &self,
+        req: ToolCompletionRequest,
+    ) -> Result<ToolCompletionResponse, LlmError> {
+        let model = req.model.unwrap_or_else(|| self.active_model_name());
+        let (system, messages) = convert_messages(req.messages);
+
+        let tools: Vec<AnthropicTool> = req
+            .tools
+            .into_iter()
+            .map(|t| AnthropicTool {
+                name: t.name,
+                description: t.description,
+                input_schema: t.parameters,
+            })
+            .collect();
+
+        // Map tool_choice from OpenAI format to Anthropic format
+        let tool_choice = req.tool_choice.map(|tc| match tc.as_str() {
+            "auto" => AnthropicToolChoice {
+                choice_type: "auto".to_string(),
+                name: None,
+            },
+            "required" => AnthropicToolChoice {
+                choice_type: "any".to_string(),
+                name: None,
+            },
+            "none" => AnthropicToolChoice {
+                choice_type: "none".to_string(),
+                name: None,
+            },
+            specific => AnthropicToolChoice {
+                choice_type: "tool".to_string(),
+                name: Some(specific.to_string()),
+            },
+        });
+
+        let request = AnthropicRequest {
+            model,
+            messages,
+            system,
+            max_tokens: req.max_tokens.unwrap_or(DEFAULT_MAX_TOKENS),
+            temperature: req.temperature,
+            tools: if tools.is_empty() { None } else { Some(tools) },
+            tool_choice,
+        };
+
+        let response: AnthropicResponse = self.send_request(&request).await?;
+        let (content, tool_calls) = extract_response_content(&response);
+
+        let finish_reason = match response.stop_reason.as_deref() {
+            Some("end_turn") | Some("stop") => FinishReason::Stop,
+            Some("max_tokens") => FinishReason::Length,
+            Some("tool_use") => FinishReason::ToolUse,
+            _ => {
+                if !tool_calls.is_empty() {
+                    FinishReason::ToolUse
+                } else {
+                    FinishReason::Unknown
+                }
+            }
+        };
+
+        Ok(ToolCompletionResponse {
+            content,
+            tool_calls,
+            finish_reason,
+            input_tokens: response.usage.input_tokens,
+            output_tokens: response.usage.output_tokens,
+            cache_creation_input_tokens: response.usage.cache_creation_input_tokens,
+            cache_read_input_tokens: response.usage.cache_read_input_tokens,
+        })
+    }
+
+    fn model_name(&self) -> &str {
+        &self.model
+    }
+
+    fn cost_per_token(&self) -> (Decimal, Decimal) {
+        let model = self.active_model_name();
+        costs::model_cost(&model).unwrap_or_else(costs::default_cost)
+    }
+
+    fn active_model_name(&self) -> String {
+        match self.active_model.read() {
+            Ok(guard) => guard.clone(),
+            Err(poisoned) => poisoned.into_inner().clone(),
+        }
+    }
+
+    fn set_model(&self, model: &str) -> Result<(), LlmError> {
+        match self.active_model.write() {
+            Ok(mut guard) => {
+                *guard = model.to_string();
+            }
+            Err(poisoned) => {
+                *poisoned.into_inner() = model.to_string();
+            }
+        }
+        Ok(())
+    }
+}
+
+// --- Anthropic Messages API types ---
+
+#[derive(Debug, Serialize)]
+struct AnthropicRequest {
+    model: String,
+    messages: Vec<AnthropicMessage>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    system: Option<String>,
+    max_tokens: u32,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    temperature: Option<f32>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    tools: Option<Vec<AnthropicTool>>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    tool_choice: Option<AnthropicToolChoice>,
+}
+
+#[derive(Debug, Serialize)]
+struct AnthropicMessage {
+    role: String,
+    content: AnthropicContent,
+}
+
+/// Anthropic content can be a simple string or a list of content blocks.
+#[derive(Debug, Serialize)]
+#[serde(untagged)]
+enum AnthropicContent {
+    Text(String),
+    Blocks(Vec<AnthropicContentBlock>),
+}
+
+#[derive(Debug, Serialize)]
+#[serde(tag = "type")]
+enum AnthropicContentBlock {
+    #[serde(rename = "text")]
+    Text { text: String },
+    #[serde(rename = "tool_use")]
+    ToolUse {
+        id: String,
+        name: String,
+        input: serde_json::Value,
+    },
+    #[serde(rename = "tool_result")]
+    ToolResult {
+        tool_use_id: String,
+        content: String,
+    },
+}
+
+#[derive(Debug, Serialize)]
+struct AnthropicTool {
+    name: String,
+    description: String,
+    input_schema: serde_json::Value,
+}
+
+#[derive(Debug, Serialize)]
+struct AnthropicToolChoice {
+    #[serde(rename = "type")]
+    choice_type: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    name: Option<String>,
+}
+
+#[derive(Debug, Deserialize)]
+struct AnthropicResponse {
+    content: Vec<AnthropicResponseBlock>,
+    #[serde(default)]
+    stop_reason: Option<String>,
+    usage: AnthropicUsage,
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(tag = "type")]
+enum AnthropicResponseBlock {
+    #[serde(rename = "text")]
+    Text { text: String },
+    #[serde(rename = "tool_use")]
+    ToolUse {
+        id: String,
+        name: String,
+        input: serde_json::Value,
+    },
+}
+
+#[derive(Debug, Deserialize)]
+struct AnthropicUsage {
+    #[serde(default)]
+    input_tokens: u32,
+    #[serde(default)]
+    output_tokens: u32,
+    #[serde(default)]
+    cache_creation_input_tokens: u32,
+    #[serde(default)]
+    cache_read_input_tokens: u32,
+}
+
+/// Convert ChatMessage list to Anthropic format.
+///
+/// Extracts system messages to the top-level `system` parameter (Anthropic
+/// doesn't allow system messages in the `messages` array). Tool-call/tool-result
+/// pairs are converted to content blocks.
+fn convert_messages(messages: Vec<ChatMessage>) -> (Option<String>, Vec<AnthropicMessage>) {
+    let mut system_parts: Vec<String> = Vec::new();
+    let mut anthropic_msgs: Vec<AnthropicMessage> = Vec::new();
+
+    for msg in messages {
+        match msg.role {
+            Role::System => {
+                if !msg.content.is_empty() {
+                    system_parts.push(msg.content);
+                }
+            }
+            Role::User => {
+                anthropic_msgs.push(AnthropicMessage {
+                    role: "user".to_string(),
+                    content: AnthropicContent::Text(msg.content),
+                });
+            }
+            Role::Assistant => {
+                if let Some(tool_calls) = msg.tool_calls {
+                    // Assistant message with tool calls → content blocks
+                    let mut blocks: Vec<AnthropicContentBlock> = Vec::new();
+                    if !msg.content.is_empty() {
+                        blocks.push(AnthropicContentBlock::Text { text: msg.content });
+                    }
+                    for tc in tool_calls {
+                        blocks.push(AnthropicContentBlock::ToolUse {
+                            id: tc.id,
+                            name: tc.name,
+                            input: tc.arguments,
+                        });
+                    }
+                    anthropic_msgs.push(AnthropicMessage {
+                        role: "assistant".to_string(),
+                        content: AnthropicContent::Blocks(blocks),
+                    });
+                } else {
+                    anthropic_msgs.push(AnthropicMessage {
+                        role: "assistant".to_string(),
+                        content: AnthropicContent::Text(msg.content),
+                    });
+                }
+            }
+            Role::Tool => {
+                let Some(tool_call_id) = msg.tool_call_id else {
+                    tracing::warn!("Skipping Tool message without tool_call_id");
+                    continue;
+                };
+                // Tool results go into a user message with tool_result blocks
+                let block = AnthropicContentBlock::ToolResult {
+                    tool_use_id: tool_call_id,
+                    content: msg.content,
+                };
+                // If the last message is already a user message with blocks,
+                // append to it (Anthropic requires consecutive tool results
+                // in one user message).
+                if let Some(last) = anthropic_msgs.last_mut()
+                    && last.role == "user"
+                    && let AnthropicContent::Blocks(ref mut blocks) = last.content
+                {
+                    blocks.push(block);
+                    continue;
+                }
+                anthropic_msgs.push(AnthropicMessage {
+                    role: "user".to_string(),
+                    content: AnthropicContent::Blocks(vec![block]),
+                });
+            }
+        }
+    }
+
+    let system = if system_parts.is_empty() {
+        None
+    } else {
+        Some(system_parts.join("\n\n"))
+    };
+
+    (system, anthropic_msgs)
+}
+
+/// Extract text content and tool calls from an Anthropic response.
+fn extract_response_content(response: &AnthropicResponse) -> (Option<String>, Vec<ToolCall>) {
+    let mut text_parts: Vec<String> = Vec::new();
+    let mut tool_calls: Vec<ToolCall> = Vec::new();
+
+    for block in &response.content {
+        match block {
+            AnthropicResponseBlock::Text { text } => {
+                text_parts.push(text.clone());
+            }
+            AnthropicResponseBlock::ToolUse { id, name, input } => {
+                tool_calls.push(ToolCall {
+                    id: id.clone(),
+                    name: name.clone(),
+                    arguments: input.clone(),
+                });
+            }
+        }
+    }
+
+    let content = if text_parts.is_empty() {
+        None
+    } else {
+        Some(text_parts.join(""))
+    };
+
+    (content, tool_calls)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_convert_messages_extracts_system() {
+        let messages = vec![
+            ChatMessage::system("You are helpful."),
+            ChatMessage::user("Hello"),
+        ];
+        let (system, msgs) = convert_messages(messages);
+        assert_eq!(system, Some("You are helpful.".to_string()));
+        assert_eq!(msgs.len(), 1);
+        assert_eq!(msgs[0].role, "user");
+    }
+
+    #[test]
+    fn test_convert_messages_multiple_systems() {
+        let messages = vec![
+            ChatMessage::system("System 1"),
+            ChatMessage::system("System 2"),
+            ChatMessage::user("Hello"),
+        ];
+        let (system, msgs) = convert_messages(messages);
+        assert_eq!(system, Some("System 1\n\nSystem 2".to_string()));
+        assert_eq!(msgs.len(), 1);
+    }
+
+    #[test]
+    fn test_convert_messages_tool_calls() {
+        let tool_calls = vec![ToolCall {
+            id: "call_1".to_string(),
+            name: "search".to_string(),
+            arguments: serde_json::json!({"q": "test"}),
+        }];
+        let messages = vec![
+            ChatMessage::user("Search for test"),
+            ChatMessage::assistant_with_tool_calls(Some("Let me search.".to_string()), tool_calls),
+            ChatMessage::tool_result("call_1", "search", "found it"),
+        ];
+        let (system, msgs) = convert_messages(messages);
+        assert!(system.is_none());
+        assert_eq!(msgs.len(), 3);
+        assert_eq!(msgs[0].role, "user");
+        assert_eq!(msgs[1].role, "assistant");
+        // Tool result should be a user message
+        assert_eq!(msgs[2].role, "user");
+    }
+
+    #[test]
+    fn test_extract_response_text_only() {
+        let response = AnthropicResponse {
+            content: vec![AnthropicResponseBlock::Text {
+                text: "Hello!".to_string(),
+            }],
+            stop_reason: Some("end_turn".to_string()),
+            usage: AnthropicUsage {
+                input_tokens: 10,
+                output_tokens: 5,
+                cache_creation_input_tokens: 0,
+                cache_read_input_tokens: 0,
+            },
+        };
+        let (content, tool_calls) = extract_response_content(&response);
+        assert_eq!(content, Some("Hello!".to_string()));
+        assert!(tool_calls.is_empty());
+    }
+
+    #[test]
+    fn test_extract_response_with_tool_use() {
+        let response = AnthropicResponse {
+            content: vec![
+                AnthropicResponseBlock::Text {
+                    text: "Let me search.".to_string(),
+                },
+                AnthropicResponseBlock::ToolUse {
+                    id: "call_1".to_string(),
+                    name: "search".to_string(),
+                    input: serde_json::json!({"q": "test"}),
+                },
+            ],
+            stop_reason: Some("tool_use".to_string()),
+            usage: AnthropicUsage {
+                input_tokens: 20,
+                output_tokens: 15,
+                cache_creation_input_tokens: 0,
+                cache_read_input_tokens: 0,
+            },
+        };
+        let (content, tool_calls) = extract_response_content(&response);
+        assert_eq!(content, Some("Let me search.".to_string()));
+        assert_eq!(tool_calls.len(), 1);
+        assert_eq!(tool_calls[0].name, "search");
+    }
+}

src/llm/mod.rs

@@ -7,6 +7,7 @@
 //! - **Ollama**: Local model inference
 //! - **OpenAI-compatible**: Any endpoint that speaks the OpenAI API
 
+mod anthropic_oauth;
 pub mod circuit_breaker;
 pub mod costs;
 pub mod failover;
@@ -57,8 +58,10 @@ pub fn create_llm_provider(
     config: &LlmConfig,
     session: Arc<SessionManager>,
 ) -> Result<Arc<dyn LlmProvider>, LlmError> {
+    let timeout = config.request_timeout_secs;
+
     if config.backend == "nearai" || config.backend == "near_ai" || config.backend == "near" {
-        return create_llm_provider_with_config(&config.nearai, session);
+        return create_llm_provider_with_config(&config.nearai, session, timeout);
     }
 
     let reg_config = config
@@ -78,6 +81,7 @@ pub fn create_llm_provider(
 pub fn create_llm_provider_with_config(
     config: &NearAiConfig,
     session: Arc<SessionManager>,
+    request_timeout_secs: u64,
 ) -> Result<Arc<dyn LlmProvider>, LlmError> {
     let auth_mode = if config.api_key.is_some() {
         "API key"
@@ -88,9 +92,14 @@ pub fn create_llm_provider_with_config(
         model = %config.model,
         base_url = %config.base_url,
         auth = auth_mode,
+        timeout_secs = request_timeout_secs,
         "Using NEAR AI (Chat Completions API)"
     );
-    Ok(Arc::new(NearAiChatProvider::new(config.clone(), session)?))
+    Ok(Arc::new(NearAiChatProvider::new_with_timeout(
+        config.clone(),
+        session,
+        request_timeout_secs,
+    )?))
 }
 
 /// Create a provider from a registry-resolved config.
@@ -178,6 +187,24 @@ fn create_openai_compat_from_registry(
 fn create_anthropic_from_registry(
     config: &RegistryProviderConfig,
 ) -> Result<Arc<dyn LlmProvider>, LlmError> {
+    // Route to OAuth provider when an OAuth token is present and no real API
+    // key was provided. When both are set, the API key takes priority (standard
+    // x-api-key auth via rig-core).
+    let api_key_is_placeholder = config
+        .api_key
+        .as_ref()
+        .is_some_and(|k| k.expose_secret() == crate::config::llm::OAUTH_PLACEHOLDER);
+    if config.oauth_token.is_some() && (config.api_key.is_none() || api_key_is_placeholder) {
+        tracing::info!(
+            provider = %config.provider_id,
+            model = %config.model,
+            base_url = if config.base_url.is_empty() { "default" } else { &config.base_url },
+            "Using Anthropic OAuth API"
+        );
+        let provider = anthropic_oauth::AnthropicOAuthProvider::new(config)?;
+        return Ok(Arc::new(provider));
+    }
+
     use crate::config::CacheRetention;
     use crate::config::helpers::optional_env;
     use rig::providers::anthropic;
@@ -346,7 +373,11 @@ pub fn build_provider_chain(
     let llm: Arc<dyn LlmProvider> = if let Some(ref cheap_model) = config.nearai.cheap_model {
         let mut cheap_config = config.nearai.clone();
         cheap_config.model = cheap_model.clone();
-        let cheap = create_llm_provider_with_config(&cheap_config, session.clone())?;
+        let cheap = create_llm_provider_with_config(
+            &cheap_config,
+            session.clone(),
+            config.request_timeout_secs,
+        )?;
         let cheap: Arc<dyn LlmProvider> = if retry_config.max_retries > 0 {
             Arc::new(RetryProvider::new(cheap, retry_config.clone()))
         } else {
@@ -378,7 +409,11 @@ pub fn build_provider_chain(
         }
         let mut fallback_config = config.nearai.clone();
         fallback_config.model = fallback_model.clone();
-        let fallback = create_llm_provider_with_config(&fallback_config, session.clone())?;
+        let fallback = create_llm_provider_with_config(
+            &fallback_config,
+            session.clone(),
+            config.request_timeout_secs,
+        )?;
         tracing::info!(
             primary = %llm.model_name(),
             fallback = %fallback.model_name(),
@@ -484,6 +519,7 @@ mod tests {
             session: SessionConfig::default(),
             nearai: test_nearai_config(),
             provider: None,
+            request_timeout_secs: 120,
         }
     }
 

src/llm/nearai_chat.rs

@@ -58,17 +58,28 @@ impl NearAiChatProvider {
     /// By default this enables tool-message flattening for compatibility with
     /// providers that reject `role: "tool"` messages.
     pub fn new(config: NearAiConfig, session: Arc<SessionManager>) -> Result<Self, LlmError> {
-        Self::new_with_flatten(config, session, true)
+        Self::new_with_options(config, session, true, 120)
     }
 
-    /// Create a chat completions provider with configurable tool-message flattening.
-    pub fn new_with_flatten(
+    /// Create a new provider with a custom request timeout.
+    pub fn new_with_timeout(
+        config: NearAiConfig,
+        session: Arc<SessionManager>,
+        request_timeout_secs: u64,
+    ) -> Result<Self, LlmError> {
+        Self::new_with_options(config, session, true, request_timeout_secs)
+    }
+
+    /// Create a chat completions provider with configurable tool-message flattening
+    /// and request timeout.
+    pub fn new_with_options(
         config: NearAiConfig,
         session: Arc<SessionManager>,
         flatten_tool_messages: bool,
+        request_timeout_secs: u64,
     ) -> Result<Self, LlmError> {
         let client = Client::builder()
-            .timeout(std::time::Duration::from_secs(120))
+            .timeout(std::time::Duration::from_secs(request_timeout_secs))
             .build()
             .map_err(|e| LlmError::RequestFailed {
                 provider: "nearai_chat".to_string(),

src/llm/provider.rs

@@ -522,4 +522,66 @@ mod tests {
         assert_eq!(messages[3].role, Role::User); // call_2 orphaned
         assert_eq!(messages[4].role, Role::User); // call_3 orphaned
     }
+
+    /// Regression: worker's select_tools/execute_plan now emit
+    /// assistant_with_tool_calls before tool_result messages.
+    /// Verify sanitize_tool_messages preserves all tool_results when
+    /// each has a matching assistant tool_call.
+    #[test]
+    fn test_sanitize_preserves_tool_results_with_matching_assistant() {
+        let tc1 = ToolCall {
+            id: "call_sel_1".to_string(),
+            name: "search".to_string(),
+            arguments: serde_json::json!({"q": "test"}),
+        };
+        let tc2 = ToolCall {
+            id: "call_sel_2".to_string(),
+            name: "http".to_string(),
+            arguments: serde_json::json!({"url": "https://example.com"}),
+        };
+        let mut messages = vec![
+            ChatMessage::system("You are a helpful assistant."),
+            ChatMessage::assistant_with_tool_calls(None, vec![tc1, tc2]),
+            ChatMessage::tool_result("call_sel_1", "search", "found 3 results"),
+            ChatMessage::tool_result("call_sel_2", "http", "200 OK"),
+        ];
+        sanitize_tool_messages(&mut messages);
+
+        // All tool_results must keep Role::Tool -- none should be rewritten.
+        assert_eq!(messages[2].role, Role::Tool);
+        assert_eq!(messages[2].tool_call_id, Some("call_sel_1".to_string()));
+        assert_eq!(messages[2].content, "found 3 results");
+
+        assert_eq!(messages[3].role, Role::Tool);
+        assert_eq!(messages[3].tool_call_id, Some("call_sel_2".to_string()));
+        assert_eq!(messages[3].content, "200 OK");
+    }
+
+    /// Regression: the OLD buggy worker code pushed tool_result messages
+    /// without a preceding assistant_with_tool_calls, causing
+    /// sanitize_tool_messages to rewrite them as orphaned user messages.
+    /// This test reproduces that buggy sequence and confirms the rewrite.
+    #[test]
+    fn test_sanitize_rewrites_orphaned_tool_results() {
+        let mut messages = vec![
+            ChatMessage::system("You are a helpful assistant."),
+            // No assistant_with_tool_calls -- mimics the old bug.
+            ChatMessage::tool_result("call_bug_1", "search", "found 3 results"),
+            ChatMessage::tool_result("call_bug_2", "http", "200 OK"),
+        ];
+        sanitize_tool_messages(&mut messages);
+
+        // Both tool_results must be rewritten to Role::User.
+        assert_eq!(messages[1].role, Role::User);
+        assert!(messages[1].content.contains("[Tool `search` returned:"));
+        assert!(messages[1].content.contains("found 3 results"));
+        assert!(messages[1].tool_call_id.is_none());
+        assert!(messages[1].name.is_none());
+
+        assert_eq!(messages[2].role, Role::User);
+        assert!(messages[2].content.contains("[Tool `http` returned:"));
+        assert!(messages[2].content.contains("200 OK"));
+        assert!(messages[2].tool_call_id.is_none());
+        assert!(messages[2].name.is_none());
+    }
 }

src/llm/reasoning.rs

@@ -8,9 +8,9 @@ use serde::{Deserialize, Serialize};
 use crate::error::LlmError;
 
 use crate::llm::{
-    ChatMessage, CompletionRequest, LlmProvider, ToolCall, ToolCompletionRequest, ToolDefinition,
+    ChatMessage, CompletionRequest, LlmProvider, Role, ToolCall, ToolCompletionRequest,
+    ToolDefinition,
 };
-use crate::safety::SafetyLayer;
 
 /// Token the agent returns when it has nothing to say (e.g. in group chats).
 /// The dispatcher should check for this and suppress the message.
@@ -342,8 +342,6 @@ pub struct RespondOutput {
 /// Reasoning engine for the agent.
 pub struct Reasoning {
     llm: Arc<dyn LlmProvider>,
-    #[allow(dead_code)] // Will be used for sanitizing tool outputs
-    safety: Arc<SafetyLayer>,
     /// Optional workspace for loading identity/system prompts.
     workspace_system_prompt: Option<String>,
     /// Optional skill context block to inject into system prompt.
@@ -361,10 +359,9 @@ pub struct Reasoning {
 
 impl Reasoning {
     /// Create a new reasoning engine.
-    pub fn new(llm: Arc<dyn LlmProvider>, safety: Arc<SafetyLayer>) -> Self {
+    pub fn new(llm: Arc<dyn LlmProvider>) -> Self {
         Self {
             llm,
-            safety,
             workspace_system_prompt: None,
             skill_context: None,
             channel: None,
@@ -460,8 +457,15 @@ impl Reasoning {
     pub async fn plan(&self, context: &ReasoningContext) -> Result<ActionPlan, LlmError> {
         let system_prompt = self.build_planning_prompt(context);
 
+        let system_prompt = merge_system_messages(system_prompt, &context.messages);
         let mut messages = vec![ChatMessage::system(system_prompt)];
-        messages.extend(context.messages.clone());
+        messages.extend(
+            context
+                .messages
+                .iter()
+                .filter(|m| m.role != Role::System)
+                .cloned(),
+        );
 
         if let Some(ref job) = context.job_description {
             messages.push(ChatMessage::user(format!(
@@ -612,8 +616,15 @@ Respond in JSON format:
             None => self.build_system_prompt_with_tools(&context.available_tools),
         };
 
+        let system_prompt = merge_system_messages(system_prompt, &context.messages);
         let mut messages = vec![ChatMessage::system(system_prompt)];
-        messages.extend(context.messages.clone());
+        messages.extend(
+            context
+                .messages
+                .iter()
+                .filter(|m| m.role != Role::System)
+                .cloned(),
+        );
 
         let effective_tools = if context.force_text {
             Vec::new()
@@ -1026,6 +1037,22 @@ pub struct SuccessEvaluation {
     pub suggestions: Vec<String>,
 }
 
+/// Merge the reasoning method's system prompt with any system messages already
+/// present in the conversation context.  Strict LLM providers (e.g. Qwen)
+/// reject conversations with system messages that are not at the very
+/// beginning, so we concatenate all system content into a single prompt.
+fn merge_system_messages(primary: String, context_messages: &[ChatMessage]) -> String {
+    let extra: Vec<&str> = context_messages
+        .iter()
+        .filter(|m| m.role == Role::System)
+        .map(|m| m.content.as_str())
+        .collect();
+    if extra.is_empty() {
+        return primary;
+    }
+    format!("{}\n\n---\n\n{}", primary, extra.join("\n\n"))
+}
+
 /// Extract JSON from text that might contain other content.
 fn extract_json(text: &str) -> Option<&str> {
     // Find the first { and last } to extract JSON
@@ -2086,15 +2113,9 @@ That's my plan."#;
     // ---- System prompt building tests (issue #565) ----
 
     fn make_test_reasoning() -> Reasoning {
-        use crate::config::SafetyConfig;
-        use crate::safety::SafetyLayer;
         use crate::testing::StubLlm;
         let llm = Arc::new(StubLlm::new("test"));
-        let safety = Arc::new(SafetyLayer::new(&SafetyConfig {
-            max_output_length: 100_000,
-            injection_check_enabled: false,
-        }));
-        Reasoning::new(llm, safety)
+        Reasoning::new(llm)
     }
 
     #[test]
@@ -2198,6 +2219,58 @@ That's my plan."#;
         assert!(cleaned.contains("Here are the results."));
     }
 
+    // ---- merge_system_messages: duplicate system message regression (Bug #597) ----
+
+    #[test]
+    fn test_merge_system_messages_no_system_in_context() {
+        let messages = vec![
+            ChatMessage::user("Hello"),
+            ChatMessage::assistant("Hi there"),
+        ];
+        let result = merge_system_messages("primary prompt".into(), &messages);
+        assert_eq!(result, "primary prompt");
+    }
+
+    #[test]
+    fn test_merge_system_messages_merges_worker_system() {
+        let messages = vec![
+            ChatMessage::system("You are an autonomous agent working on a job.\n\nJob: Test Job"),
+            ChatMessage::user("Do the thing"),
+        ];
+        let result = merge_system_messages("planning prompt".into(), &messages);
+        assert!(
+            result.contains("planning prompt"),
+            "must contain the primary prompt"
+        );
+        assert!(
+            result.contains("autonomous agent"),
+            "must contain worker system text"
+        );
+        assert!(
+            result.contains("Test Job"),
+            "must contain job description from worker system message"
+        );
+    }
+
+    #[test]
+    fn test_merge_system_messages_multiple_system() {
+        let messages = vec![
+            ChatMessage::system("First system instruction"),
+            ChatMessage::system("Second system instruction"),
+            ChatMessage::user("Hello"),
+        ];
+        let result = merge_system_messages("primary".into(), &messages);
+        assert!(result.contains("primary"), "must contain primary prompt");
+        assert!(
+            result.contains("First system instruction"),
+            "must contain first system message"
+        );
+        assert!(
+            result.contains("Second system instruction"),
+            "must contain second system message"
+        );
+    }
+
     #[test]
     fn test_system_prompt_without_tools_omits_tools_section() {
         let reasoning = make_test_reasoning();

src/llm/registry.rs

@@ -450,6 +450,8 @@ mod tests {
             if def.protocol == ProviderProtocol::OpenAiCompletions
                 && def.id != "openai"
                 && def.id != "openai_compatible"
+                && def.id != "bedrock"
+                && def.id != "cloudflare"
             {
                 assert!(
                     def.default_base_url.is_some(),

src/llm/session.rs

@@ -200,9 +200,10 @@ impl SessionManager {
 
         let status = response.status();
         let body = response.text().await.unwrap_or_default();
+        let preview = crate::agent::truncate_for_preview(&body, 200);
         Err(LlmError::SessionRenewalFailed {
             provider: "nearai".to_string(),
-            reason: format!("Validation failed: HTTP {}: {}", status, body),
+            reason: format!("Validation failed: HTTP {status}: {preview}"),
         })
     }
 

src/main.rs

@@ -145,6 +145,24 @@ async fn async_main() -> anyhow::Result<()> {
         }
     }
 
+    // ── PID lock (prevent multiple instances) ────────────────────────
+    let _pid_lock = match ironclaw::bootstrap::PidLock::acquire() {
+        Ok(lock) => Some(lock),
+        Err(ironclaw::bootstrap::PidLockError::AlreadyRunning { pid }) => {
+            anyhow::bail!(
+                "Another IronClaw instance is already running (PID {}). \
+                 If this is incorrect, remove the stale PID file: {}",
+                pid,
+                ironclaw::bootstrap::pid_lock_path().display()
+            );
+        }
+        Err(e) => {
+            eprintln!("Warning: Could not acquire PID lock: {}", e);
+            eprintln!("Continuing without PID lock protection.");
+            None
+        }
+    };
+
     // ── Agent startup ──────────────────────────────────────────────────
 
     // Enhanced first-run detection
@@ -158,18 +176,20 @@ async fn async_main() -> anyhow::Result<()> {
         wizard.run().await?;
     }
 
-    // Load initial config from env + disk + optional TOML (before DB is available)
+    // Load initial config from env + disk + optional TOML (before DB is available).
+    // Credentials may be missing at this point — that's fine. LlmConfig::resolve()
+    // defers gracefully, and AppBuilder::build_all() re-resolves after loading
+    // secrets from the encrypted DB.
     let toml_path = cli.config.as_deref();
     let config = match Config::from_env_with_toml(toml_path).await {
         Ok(c) => c,
         Err(ironclaw::error::ConfigError::MissingRequired { key, hint }) => {
-            eprintln!("Configuration error: Missing required setting '{}'", key);
-            eprintln!("  {}", hint);
-            eprintln!();
-            eprintln!(
-                "Run 'ironclaw onboard' to configure, or set the required environment variables."
+            anyhow::bail!(
+                "Configuration error: Missing required setting '{}'. {}. \
+                 Run 'ironclaw onboard' to configure, or set the required environment variables.",
+                key,
+                hint
             );
-            std::process::exit(1);
         }
         Err(e) => return Err(e.into()),
     };
@@ -475,6 +495,7 @@ async fn async_main() -> anyhow::Result<()> {
     let mut sse_sender: Option<
         tokio::sync::broadcast::Sender<ironclaw::channels::web::types::SseEvent>,
     > = None;
+    let mut routine_engine_slot: Option<ironclaw::channels::web::server::RoutineEngineSlot> = None;
     if let Some(ref gw_config) = config.channels.gateway {
         let mut gw =
             GatewayChannel::new(gw_config.clone()).with_llm_provider(Arc::clone(&components.llm));
@@ -528,10 +549,11 @@ async fn async_main() -> anyhow::Result<()> {
 
         tracing::info!("Web UI: http://{}:{}/", gw_config.host, gw_config.port);
 
-        // Capture SSE sender before moving gw into channels.
+        // Capture SSE sender and routine engine slot before moving gw into channels.
         // IMPORTANT: This must come after all `with_*` calls since `rebuild_state`
         // creates a new SseManager, which would orphan this sender.
         sse_sender = Some(gw.state().sse.sender());
+        routine_engine_slot = Some(Arc::clone(&gw.state().routine_engine));
 
         channel_names.push("gateway".to_string());
         channels.add(Box::new(gw)).await;
@@ -678,7 +700,7 @@ async fn async_main() -> anyhow::Result<()> {
         )),
     };
 
-    let agent = Agent::new(
+    let mut agent = Agent::new(
         config.agent.clone(),
         deps,
         channels,
@@ -692,6 +714,11 @@ async fn async_main() -> anyhow::Result<()> {
     // Fill the scheduler slot now that Agent (and its Scheduler) exist.
     *scheduler_slot.write().await = Some(agent.scheduler());
 
+    // Give the agent the routine engine slot so it can expose the engine to the gateway.
+    if let Some(slot) = routine_engine_slot {
+        agent.set_routine_engine_slot(slot);
+    }
+
     agent.run().await?;
 
     // ── Shutdown ────────────────────────────────────────────────────────

src/settings.rs

@@ -42,6 +42,10 @@ pub struct Settings {
     #[serde(default)]
     pub secrets_master_key_source: KeySource,
 
+    /// Generated master key hex (env var mode only, written to .env by wizard).
+    #[serde(default, skip_serializing)]
+    pub secrets_master_key_hex: Option<String>,
+
     // === Step 3: Inference Provider ===
     /// LLM backend: "nearai", "anthropic", "openai", "ollama", "openai_compatible".
     #[serde(default)]
@@ -291,6 +295,18 @@ pub struct HeartbeatSettings {
     /// User ID to notify on heartbeat findings.
     #[serde(default)]
     pub notify_user: Option<String>,
+
+    /// Hour (0-23) when quiet hours start (heartbeat skipped).
+    #[serde(default)]
+    pub quiet_hours_start: Option<u32>,
+
+    /// Hour (0-23) when quiet hours end (heartbeat resumes).
+    #[serde(default)]
+    pub quiet_hours_end: Option<u32>,
+
+    /// Timezone for quiet hours evaluation (IANA name, e.g. "America/New_York").
+    #[serde(default)]
+    pub timezone: Option<String>,
 }
 
 fn default_heartbeat_interval() -> u64 {
@@ -304,6 +320,9 @@ impl Default for HeartbeatSettings {
             interval_secs: default_heartbeat_interval(),
             notify_channel: None,
             notify_user: None,
+            quiet_hours_start: None,
+            quiet_hours_end: None,
+            timezone: None,
         }
     }
 }
@@ -351,6 +370,10 @@ pub struct AgentSettings {
     /// When true, skip tool approval checks entirely. For benchmarks/CI.
     #[serde(default)]
     pub auto_approve_tools: bool,
+
+    /// Default timezone for new sessions (IANA name, e.g. "America/New_York").
+    #[serde(default = "default_timezone")]
+    pub default_timezone: String,
 }
 
 fn default_agent_name() -> String {
@@ -385,6 +408,10 @@ fn default_max_tool_iterations() -> usize {
     50
 }
 
+fn default_timezone() -> String {
+    "UTC".to_string()
+}
+
 fn default_true() -> bool {
     true
 }
@@ -402,6 +429,7 @@ impl Default for AgentSettings {
             session_idle_timeout_secs: default_session_idle_timeout(),
             max_tool_iterations: default_max_tool_iterations(),
             auto_approve_tools: false,
+            default_timezone: default_timezone(),
         }
     }
 }
@@ -498,6 +526,10 @@ pub struct SandboxSettings {
     /// Additional domains to allow through the network proxy.
     #[serde(default)]
     pub extra_allowed_domains: Vec<String>,
+
+    /// Whether Claude Code sandbox mode is enabled.
+    #[serde(default)]
+    pub claude_code_enabled: bool,
 }
 
 fn default_sandbox_policy() -> String {
@@ -531,6 +563,7 @@ impl Default for SandboxSettings {
             image: default_sandbox_image(),
             auto_pull_image: true,
             extra_allowed_domains: Vec::new(),
+            claude_code_enabled: false,
         }
     }
 }
@@ -1169,6 +1202,31 @@ mod tests {
         assert_eq!(loaded.heartbeat.interval_secs, 900);
     }
 
+    /// Regression test: /model command must persist selected_model to TOML config.
+    /// Prior to the fix, `set_model()` only changed the in-memory provider and the
+    /// choice was lost on restart.
+    #[test]
+    fn toml_selected_model_update_persists() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("config.toml");
+
+        // Start with a config that has a different model.
+        let settings = Settings {
+            selected_model: Some("old-model".to_string()),
+            ..Default::default()
+        };
+        settings.save_toml(&path).unwrap();
+
+        // Simulate what persist_selected_model does: load, update, save.
+        let mut loaded = Settings::load_toml(&path).unwrap().unwrap();
+        loaded.selected_model = Some("new-model".to_string());
+        loaded.save_toml(&path).unwrap();
+
+        // Verify the change survived a reload.
+        let reloaded = Settings::load_toml(&path).unwrap().unwrap();
+        assert_eq!(reloaded.selected_model, Some("new-model".to_string()));
+    }
+
     #[test]
     fn toml_missing_file_returns_none() {
         let result = Settings::load_toml(std::path::Path::new("/tmp/nonexistent_config.toml"));

src/setup/README.md

@@ -152,12 +152,8 @@ This is OS-level behavior we cannot prevent. To minimize pain:
   rather than triggering system dialogs.
 
 **Invariant:** After Step 2, `self.secrets_crypto` is `Some` if the user
-chose Keychain or env-var mode (both generate a key and initialize crypto
-immediately). It is `None` only if the user skipped secrets.
-
-When env-var mode is chosen, the generated key is also stored in
-`self.secrets_master_key_hex` so that `write_bootstrap_env()` can persist
-it to `~/.ironclaw/.env` automatically.
+chose Keychain or generated a new key. It may be `None` if the user chose
+env-var mode or skipped secrets.
 
 ---
 

src/setup/wizard.rs

@@ -22,6 +22,7 @@ use crate::bootstrap::ironclaw_base_dir;
 use crate::channels::wasm::{
     ChannelCapabilitiesFile, available_channel_names, install_bundled_channel,
 };
+use crate::config::llm::OAUTH_PLACEHOLDER;
 use crate::llm::{SessionConfig, SessionManager};
 use crate::secrets::{SecretsCrypto, SecretsStore};
 use crate::settings::{KeySource, Settings};
@@ -90,8 +91,6 @@ pub struct SetupWizard {
     db_backend: Option<crate::db::libsql::LibSqlBackend>,
     /// Secrets crypto (created during setup).
     secrets_crypto: Option<Arc<SecretsCrypto>>,
-    /// Generated master key hex (stored for writing to .env in env-var mode).
-    secrets_master_key_hex: Option<String>,
     /// Cached API key from provider setup (used by model fetcher without env mutation).
     llm_api_key: Option<SecretString>,
 }
@@ -108,7 +107,6 @@ impl SetupWizard {
             #[cfg(feature = "libsql")]
             db_backend: None,
             secrets_crypto: None,
-            secrets_master_key_hex: None,
             llm_api_key: None,
         }
     }
@@ -124,7 +122,6 @@ impl SetupWizard {
             #[cfg(feature = "libsql")]
             db_backend: None,
             secrets_crypto: None,
-            secrets_master_key_hex: None,
             llm_api_key: None,
         }
     }
@@ -772,25 +769,31 @@ impl SetupWizard {
                 print_success("Master key generated and stored in OS keychain");
             }
             1 => {
-                // Env var mode: generate key, initialize crypto, and persist to .env
-                print_info("Generating master key...");
+                // Env var mode — generate key, init crypto, and persist to .env
                 let key_hex = crate::secrets::keychain::generate_master_key_hex();
 
-                // Initialize crypto so subsequent steps (API key storage) work
+                // Initialize crypto so subsequent wizard steps (channel setup,
+                // API key storage) can encrypt secrets immediately.
                 self.secrets_crypto = Some(Arc::new(
                     SecretsCrypto::new(SecretString::from(key_hex.clone()))
                         .map_err(|e| SetupError::Config(e.to_string()))?,
                 ));
 
-                // Store for write_bootstrap_env to persist to ~/.ironclaw/.env
-                self.secrets_master_key_hex = Some(key_hex.clone());
+                // Make visible to optional_env() for any subsequent config resolution.
+                crate::config::inject_single_var("SECRETS_MASTER_KEY", &key_hex);
+
+                // Store hex for write_bootstrap_env to persist to ~/.ironclaw/.env.
+                self.settings.secrets_master_key_hex = Some(key_hex.clone());
 
                 println!();
-                print_info(&format!("Generated master key: {}", mask_api_key(&key_hex)));
-                print_info("This key will be saved to ~/.ironclaw/.env automatically.");
+                print_info("Master key generated and will be saved to ~/.ironclaw/.env");
+                println!();
+                println!("  SECRETS_MASTER_KEY={}", key_hex);
+                println!();
+                print_info("You can also copy this to another .env file or CI secrets.");
 
                 self.settings.secrets_master_key_source = KeySource::Env;
-                print_success("Master key generated and configured for environment variable");
+                print_success("Configured for environment variable");
             }
             _ => {
                 self.settings.secrets_master_key_source = KeySource::None;
@@ -899,6 +902,11 @@ impl SetupWizard {
             return Ok(());
         };
 
+        // Anthropic has a custom flow: API key or OAuth token from `claude login`.
+        if provider_id == "anthropic" {
+            return self.setup_anthropic().await;
+        }
+
         match setup {
             crate::llm::registry::SetupHint::ApiKey {
                 secret_name,
@@ -1004,6 +1012,113 @@ impl SetupWizard {
         Ok(())
     }
 
+    /// Anthropic provider setup: API key or OAuth token from `claude login`.
+    async fn setup_anthropic(&mut self) -> Result<(), SetupError> {
+        let options = &["Direct API Key", "OAuth Token (from `claude login`)"];
+        let choice = select_one("How do you want to authenticate with Anthropic?", options)
+            .map_err(SetupError::Io)?;
+
+        if choice == 0 {
+            // Standard API key flow
+            self.setup_api_key_provider(
+                "anthropic",
+                "ANTHROPIC_API_KEY",
+                "llm_anthropic_api_key",
+                "Anthropic API key",
+                "https://console.anthropic.com/settings/keys",
+                None,
+            )
+            .await
+        } else {
+            // OAuth token flow
+            self.setup_anthropic_oauth().await
+        }
+    }
+
+    /// Anthropic OAuth setup: extract token from `claude login` credentials.
+    async fn setup_anthropic_oauth(&mut self) -> Result<(), SetupError> {
+        // Clear model only when switching providers (old model may be invalid)
+        if self.settings.llm_backend.as_deref() != Some("anthropic") {
+            self.settings.selected_model = None;
+        }
+        self.settings.llm_backend = Some("anthropic".to_string());
+
+        // Try to extract existing OAuth token from Claude Code credentials
+        if let Some(token) = crate::config::ClaudeCodeConfig::extract_oauth_token() {
+            print_info(&format!("Found OAuth token: {}", mask_api_key(&token)));
+            if confirm("Use this token?", true).map_err(SetupError::Io)? {
+                return self.save_anthropic_oauth_token(&token).await;
+            }
+        } else {
+            print_info("No OAuth token found from `claude login`.");
+            print_info("Run `claude login` in a terminal to authenticate, then retry.");
+            println!();
+
+            if confirm("Retry after running `claude login`?", true).map_err(SetupError::Io)? {
+                // Block until the user has run `claude login` in another terminal
+                input("Press Enter after running `claude login` in another terminal...")
+                    .map_err(SetupError::Io)?;
+                if let Some(token) = crate::config::ClaudeCodeConfig::extract_oauth_token() {
+                    print_info(&format!("Found OAuth token: {}", mask_api_key(&token)));
+                    return self.save_anthropic_oauth_token(&token).await;
+                }
+                print_error("Still no OAuth token found.");
+            }
+        }
+
+        // Fallback: let user paste the token manually, or switch to API key
+        print_info("You can paste your OAuth token directly (starts with sk-ant-oat01-).");
+        print_info("Or press Enter with no input to switch to the API key flow.");
+        let token = secret_input("Anthropic OAuth token").map_err(SetupError::Io)?;
+        let token_str = token.expose_secret();
+        if token_str.is_empty() {
+            print_info("Switching to API key flow...");
+            return self
+                .setup_api_key_provider(
+                    "anthropic",
+                    "ANTHROPIC_API_KEY",
+                    "llm_anthropic_api_key",
+                    "Anthropic API key",
+                    "https://console.anthropic.com/settings/keys",
+                    None,
+                )
+                .await;
+        }
+        self.save_anthropic_oauth_token(token_str).await
+    }
+
+    /// Save an Anthropic OAuth token to secrets and set env for immediate use.
+    async fn save_anthropic_oauth_token(&mut self, token: &str) -> Result<(), SetupError> {
+        // Validate token format to catch accidentally pasted API keys
+        if !token.starts_with("sk-ant-oat") {
+            print_error("Token doesn't look like an OAuth token (expected prefix: sk-ant-oat).");
+            print_info("If you have an API key instead, use the 'Direct API Key' option.");
+            return Err(SetupError::Config("Invalid OAuth token format".to_string()));
+        }
+
+        // Store in secrets if available
+        if let Ok(ctx) = self.init_secrets_context().await {
+            let key = SecretString::from(token.to_string());
+            ctx.save_secret("llm_anthropic_oauth_token", &key)
+                .await
+                .map_err(|e| SetupError::Config(format!("Failed to save OAuth token: {e}")))?;
+            print_success("OAuth token encrypted and saved");
+        } else {
+            print_info("Secrets not available. Set ANTHROPIC_OAUTH_TOKEN in your environment.");
+        }
+
+        // Make the token visible to `optional_env()` for subsequent config
+        // resolution (model selection step). Uses the thread-safe overlay
+        // instead of `std::env::set_var` to avoid UB on multi-threaded runtimes.
+        crate::config::inject_single_var("ANTHROPIC_OAUTH_TOKEN", token);
+
+        // Cache for model fetching
+        self.llm_api_key = Some(SecretString::from(token.to_string()));
+
+        print_success("Anthropic OAuth configured");
+        Ok(())
+    }
+
     /// Shared setup flow for API-key-based providers.
     async fn setup_api_key_provider(
         &mut self,
@@ -1020,10 +1135,11 @@ impl SetupWizard {
             other => other,
         });
 
-        self.settings.llm_backend = Some(backend.to_string());
-        if self.settings.selected_model.is_some() {
+        // Clear model only when switching providers (old model may be invalid)
+        if self.settings.llm_backend.as_deref() != Some(backend) {
             self.settings.selected_model = None;
         }
+        self.settings.llm_backend = Some(backend.to_string());
 
         // Check env var first
         if let Ok(existing) = std::env::var(env_var) {
@@ -1065,6 +1181,11 @@ impl SetupWizard {
             ));
         }
 
+        // Make key visible to `optional_env()` for subsequent config resolution.
+        // Uses the thread-safe overlay instead of `std::env::set_var` to avoid
+        // UB on multi-threaded runtimes.
+        crate::config::inject_single_var(env_var, key_str);
+
         // Cache key in memory for model fetching later in the wizard
         self.llm_api_key = Some(SecretString::from(key_str.to_string()));
 
@@ -1077,10 +1198,11 @@ impl SetupWizard {
         &mut self,
         def: &crate::llm::ProviderDefinition,
     ) -> Result<(), SetupError> {
-        self.settings.llm_backend = Some(def.id.clone());
-        if self.settings.selected_model.is_some() {
+        // Clear model only when switching providers (old model may be invalid)
+        if self.settings.llm_backend.as_deref() != Some(&def.id) {
             self.settings.selected_model = None;
         }
+        self.settings.llm_backend = Some(def.id.clone());
 
         let default_url = self
             .settings
@@ -1115,10 +1237,11 @@ impl SetupWizard {
         secret_name: &str,
         display_name: &str,
     ) -> Result<(), SetupError> {
-        self.settings.llm_backend = Some(backend_id.to_string());
-        if self.settings.selected_model.is_some() {
+        // Clear model only when switching providers (old model may be invalid)
+        if self.settings.llm_backend.as_deref() != Some(backend_id) {
             self.settings.selected_model = None;
         }
+        self.settings.llm_backend = Some(backend_id.to_string());
 
         let existing_url = self
             .settings
@@ -1372,6 +1495,7 @@ impl SetupWizard {
                 smart_routing_cascade: true,
             },
             provider: None,
+            request_timeout_secs: 120,
         };
 
         match create_llm_provider(&config, session) {
@@ -2001,6 +2125,67 @@ impl SetupWizard {
             }
         }
 
+        // Claude Code sandbox sub-step (only if Docker sandbox is enabled)
+        if self.settings.sandbox.enabled {
+            self.step_claude_code_sandbox().await?;
+        }
+
+        Ok(())
+    }
+
+    /// Claude Code sandbox sub-step: enable Claude CLI inside Docker containers.
+    async fn step_claude_code_sandbox(&mut self) -> Result<(), SetupError> {
+        println!();
+        print_info("Claude Code mode lets the agent delegate complex tasks to Claude CLI");
+        print_info("running inside sandboxed Docker containers.");
+        println!();
+
+        if !confirm("Enable Claude Code sandbox mode?", false).map_err(SetupError::Io)? {
+            self.settings.sandbox.claude_code_enabled = false;
+            return Ok(());
+        }
+
+        // Check for Anthropic credentials (API key or OAuth token).
+        // Uses `optional_env()` which reads both real env vars and the
+        // injected overlay (secrets DB, wizard-set values).
+        let has_credentials = || {
+            let has_api_key = crate::config::helpers::optional_env("ANTHROPIC_API_KEY")
+                .ok()
+                .flatten()
+                .is_some_and(|v| !v.is_empty() && v != OAUTH_PLACEHOLDER);
+            let has_oauth = crate::config::ClaudeCodeConfig::extract_oauth_token().is_some()
+                || crate::config::helpers::optional_env("ANTHROPIC_OAUTH_TOKEN")
+                    .ok()
+                    .flatten()
+                    .is_some_and(|v| !v.is_empty());
+            has_api_key || has_oauth
+        };
+
+        if has_credentials() {
+            self.settings.sandbox.claude_code_enabled = true;
+            print_success("Claude Code sandbox enabled");
+        } else {
+            print_error("No Anthropic credentials found.");
+            print_info(
+                "Claude Code needs ANTHROPIC_API_KEY or an OAuth token from `claude login`.",
+            );
+            println!();
+
+            if confirm("Retry after setting up credentials?", false).map_err(SetupError::Io)? {
+                if has_credentials() {
+                    self.settings.sandbox.claude_code_enabled = true;
+                    print_success("Claude Code sandbox enabled");
+                } else {
+                    self.settings.sandbox.claude_code_enabled = false;
+                    print_info("No credentials found. Claude Code disabled for now.");
+                    print_info("Set ANTHROPIC_API_KEY or run `claude login` and enable later.");
+                }
+            } else {
+                self.settings.sandbox.claude_code_enabled = false;
+                print_info("Claude Code disabled. Enable with CLAUDE_CODE_ENABLED=true later.");
+            }
+        }
+
         Ok(())
     }
 
@@ -2094,6 +2279,12 @@ impl SetupWizard {
     ///
     /// These are the chicken-and-egg settings needed before the database is
     /// connected (DATABASE_BACKEND, DATABASE_URL, LLM_BACKEND, etc.).
+    ///
+    /// **Credentials are NOT written here.** API keys and OAuth tokens live
+    /// only in the encrypted secrets DB. `LlmConfig::resolve()` defers
+    /// gracefully when credentials are missing during early startup, and the
+    /// re-resolution in `AppBuilder::build_all()` fills them in after
+    /// `inject_llm_keys_from_secrets()` loads from encrypted storage.
     fn write_bootstrap_env(&self) -> Result<(), SetupError> {
         let registry = crate::llm::ProviderRegistry::load();
         let mut env_vars: Vec<(String, String)> = Vec::new();
@@ -2146,13 +2337,6 @@ impl SetupWizard {
             env_vars.push((base_url_env.clone(), base_url.clone()));
         }
 
-        // Persist SECRETS_MASTER_KEY when env-var mode was chosen in step 2
-        if self.settings.secrets_master_key_source == KeySource::Env
-            && let Some(ref key_hex) = self.secrets_master_key_hex
-        {
-            env_vars.push(("SECRETS_MASTER_KEY".to_string(), key_hex.clone()));
-        }
-
         // Preserve NEARAI_API_KEY if present (set by API key auth flow)
         if let Ok(api_key) = std::env::var("NEARAI_API_KEY")
             && !api_key.is_empty()
@@ -2160,12 +2344,23 @@ impl SetupWizard {
             env_vars.push(("NEARAI_API_KEY".to_string(), api_key));
         }
 
+        // Secrets master key (env var mode): write to .env so it's available
+        // on next startup before the DB is connected.
+        if let Some(ref key_hex) = self.settings.secrets_master_key_hex {
+            env_vars.push(("SECRETS_MASTER_KEY".to_string(), key_hex.clone()));
+        }
+
         // Always write ONBOARD_COMPLETED so that check_onboard_needed()
         // (which runs before the DB is connected) knows to skip re-onboarding.
         if self.settings.onboard_completed {
             env_vars.push(("ONBOARD_COMPLETED".to_string(), "true".to_string()));
         }
 
+        // Claude Code sandbox mode
+        if self.settings.sandbox.claude_code_enabled {
+            env_vars.push(("CLAUDE_CODE_ENABLED".to_string(), "true".to_string()));
+        }
+
         // Signal channel env vars (chicken-and-egg: config resolves before DB).
         if let Some(ref url) = self.settings.channels.signal_http_url {
             env_vars.push(("SIGNAL_HTTP_URL".to_string(), url.clone()));
@@ -2533,22 +2728,39 @@ async fn fetch_anthropic_models(cached_key: Option<&str>) -> Vec<(String, String
     let api_key = cached_key
         .map(String::from)
         .or_else(|| std::env::var("ANTHROPIC_API_KEY").ok())
-        .filter(|k| !k.is_empty());
+        .filter(|k| !k.is_empty() && k != crate::config::llm::OAUTH_PLACEHOLDER);
 
-    let api_key = match api_key {
-        Some(k) => k,
-        None => return static_defaults,
+    // Fall back to OAuth token if no API key
+    let oauth_token = if api_key.is_none() {
+        crate::config::helpers::optional_env("ANTHROPIC_OAUTH_TOKEN")
+            .ok()
+            .flatten()
+            .filter(|t| !t.is_empty())
+    } else {
+        None
+    };
+
+    let (key_or_token, is_oauth) = match (api_key, oauth_token) {
+        (Some(k), _) => (k, false),
+        (None, Some(t)) => (t, true),
+        (None, None) => return static_defaults,
     };
 
     let client = reqwest::Client::new();
-    let resp = match client
+    let mut request = client
         .get("https://api.anthropic.com/v1/models")
-        .header("x-api-key", &api_key)
         .header("anthropic-version", "2023-06-01")
-        .timeout(std::time::Duration::from_secs(5))
-        .send()
-        .await
-    {
+        .timeout(std::time::Duration::from_secs(5));
+
+    if is_oauth {
+        request = request
+            .bearer_auth(&key_or_token)
+            .header("anthropic-beta", "oauth-2025-04-20");
+    } else {
+        request = request.header("x-api-key", &key_or_token);
+    }
+
+    let resp = match request.send().await {
         Ok(r) if r.status().is_success() => r,
         _ => return static_defaults,
     };
@@ -3313,36 +3525,45 @@ mod tests {
         }
     }
 
-    /// Regression test for #666: env var mode in step_security must initialize
-    /// secrets_crypto (for immediate API key storage) and secrets_master_key_hex
-    /// (for persisting to ~/.ironclaw/.env via write_bootstrap_env).
+    /// Regression test for #600: re-running provider setup for the same backend
+    /// must NOT clear selected_model. Only switching to a different backend should.
     #[test]
-    fn test_env_var_mode_initializes_crypto_and_stores_key() {
+    fn test_same_provider_preserves_selected_model() {
         let mut wizard = SetupWizard::new();
-        assert!(wizard.secrets_crypto.is_none());
-        assert!(wizard.secrets_master_key_hex.is_none());
+        wizard.settings.llm_backend = Some("ollama".to_string());
+        wizard.settings.selected_model = Some("llama3".to_string());
 
-        // Simulate the env-var branch of step_security
-        let key_hex = crate::secrets::keychain::generate_master_key_hex();
+        // Simulate re-entering the same provider -- model should survive
+        // (This is the check that each setup_* function now performs)
+        if wizard.settings.llm_backend.as_deref() != Some("ollama") {
+            wizard.settings.selected_model = None;
+        }
+        wizard.settings.llm_backend = Some("ollama".to_string());
 
-        // Verify it's a valid 64-char hex string (32 bytes = AES-256)
-        assert_eq!(key_hex.len(), 64);
-        assert!(key_hex.chars().all(|c| c.is_ascii_hexdigit()));
-
-        let crypto = SecretsCrypto::new(SecretString::from(key_hex.clone()))
-            .expect("SecretsCrypto::new should succeed with generated hex key");
-
-        wizard.secrets_crypto = Some(Arc::new(crypto));
-        wizard.secrets_master_key_hex = Some(key_hex.clone());
-        wizard.settings.secrets_master_key_source = KeySource::Env;
-
-        // Verify crypto is usable for immediate secret encryption
-        assert!(wizard.secrets_crypto.is_some());
-
-        // Verify the hex key is stored for write_bootstrap_env to persist
         assert_eq!(
-            wizard.secrets_master_key_hex.as_deref(),
-            Some(key_hex.as_str())
+            wizard.settings.selected_model.as_deref(),
+            Some("llama3"),
+            "model should be preserved when re-selecting the same provider"
+        );
+    }
+
+    /// Regression test for #600: switching to a different provider must clear
+    /// selected_model since the old model may not be valid for the new backend.
+    #[test]
+    fn test_different_provider_clears_selected_model() {
+        let mut wizard = SetupWizard::new();
+        wizard.settings.llm_backend = Some("ollama".to_string());
+        wizard.settings.selected_model = Some("llama3".to_string());
+
+        // Simulate switching to a different provider -- model should be cleared
+        if wizard.settings.llm_backend.as_deref() != Some("openai") {
+            wizard.settings.selected_model = None;
+        }
+        wizard.settings.llm_backend = Some("openai".to_string());
+
+        assert!(
+            wizard.settings.selected_model.is_none(),
+            "model should be cleared when switching providers"
         );
     }
 
@@ -3383,4 +3604,30 @@ mod tests {
             "backend should be set even without setup hint"
         );
     }
+
+    /// Regression test for #666: env-var security option must initialize
+    /// secrets_crypto so subsequent steps can encrypt API keys.
+    #[test]
+    fn test_env_var_security_initializes_crypto() {
+        use crate::secrets::SecretsCrypto;
+        use secrecy::SecretString;
+
+        // Simulate what option 1 in step_security() does after the fix:
+        let key_hex = crate::secrets::keychain::generate_master_key_hex();
+
+        // The fix: create SecretsCrypto from the generated key.
+        // Before the fix, this was skipped, leaving secrets_crypto = None.
+        let crypto = SecretsCrypto::new(SecretString::from(key_hex.clone()));
+        assert!(
+            crypto.is_ok(),
+            "generated key hex must produce valid SecretsCrypto"
+        );
+
+        // Verify the key is stored for bootstrap env persistence.
+        let settings = Settings {
+            secrets_master_key_hex: Some(key_hex),
+            ..Settings::default()
+        };
+        assert!(settings.secrets_master_key_hex.is_some());
+    }
 }

src/testing.rs

@@ -1009,6 +1009,7 @@ mod tests {
             enabled: true,
             trigger: Trigger::Cron {
                 schedule: "0 * * * *".to_string(),
+                timezone: None,
             },
             action: RoutineAction::Lightweight {
                 prompt: "Check status".to_string(),

src/timezone.rs

@@ -0,0 +1,110 @@
+//! Timezone resolution and utilities.
+
+use chrono::{DateTime, NaiveDate, Utc};
+use chrono_tz::Tz;
+
+/// Resolve the effective timezone from a priority chain.
+///
+/// Priority: client_tz > user_setting > config_default > UTC
+pub fn resolve_timezone(
+    client_tz: Option<&str>,
+    user_setting: Option<&str>,
+    config_default: &str,
+) -> Tz {
+    // Try each in priority order, skipping invalid values
+    for candidate in [client_tz, user_setting, Some(config_default)] {
+        if let Some(tz) = candidate.and_then(parse_timezone) {
+            return tz;
+        }
+    }
+    Tz::UTC
+}
+
+/// Parse a timezone string (IANA name) into a `Tz`.
+pub fn parse_timezone(s: &str) -> Option<Tz> {
+    s.parse::<Tz>().ok()
+}
+
+/// Get today's date in the given timezone.
+pub fn today_in_tz(tz: Tz) -> NaiveDate {
+    Utc::now().with_timezone(&tz).date_naive()
+}
+
+/// Get the current time in the given timezone.
+pub fn now_in_tz(tz: Tz) -> DateTime<Tz> {
+    Utc::now().with_timezone(&tz)
+}
+
+/// Detect the system's timezone, falling back to UTC.
+pub fn detect_system_timezone() -> Tz {
+    iana_time_zone::get_timezone()
+        .ok()
+        .and_then(|s| parse_timezone(&s))
+        .unwrap_or(Tz::UTC)
+}
+
+#[cfg(test)]
+mod tests {
+    use chrono::Datelike;
+
+    use super::*;
+
+    #[test]
+    fn test_resolve_client_wins() {
+        let tz = resolve_timezone(Some("America/New_York"), Some("Europe/London"), "UTC");
+        assert_eq!(tz, chrono_tz::America::New_York);
+    }
+
+    #[test]
+    fn test_resolve_user_setting_fallback() {
+        let tz = resolve_timezone(None, Some("Europe/London"), "UTC");
+        assert_eq!(tz, chrono_tz::Europe::London);
+    }
+
+    #[test]
+    fn test_resolve_config_fallback() {
+        let tz = resolve_timezone(None, None, "Asia/Tokyo");
+        assert_eq!(tz, chrono_tz::Asia::Tokyo);
+    }
+
+    #[test]
+    fn test_resolve_all_none_utc() {
+        let tz = resolve_timezone(None, None, "UTC");
+        assert_eq!(tz, Tz::UTC);
+    }
+
+    #[test]
+    fn test_resolve_invalid_client_skipped() {
+        let tz = resolve_timezone(Some("Fake/Zone"), Some("Europe/London"), "UTC");
+        assert_eq!(tz, chrono_tz::Europe::London);
+    }
+
+    #[test]
+    fn test_parse_valid() {
+        assert_eq!(
+            parse_timezone("America/Chicago"),
+            Some(chrono_tz::America::Chicago)
+        );
+    }
+
+    #[test]
+    fn test_parse_invalid() {
+        assert_eq!(parse_timezone("Fake/Zone"), None);
+    }
+
+    #[test]
+    fn test_detect_system_tz() {
+        // Should always return a valid Tz (at minimum UTC)
+        let tz = detect_system_timezone();
+        let _ = now_in_tz(tz); // Should not panic
+    }
+
+    #[test]
+    fn test_today_in_tz_returns_valid_date() {
+        let date = today_in_tz(Tz::UTC);
+        // Verify it returns a valid date (year, month, day are all positive)
+        assert!(date.year() > 0);
+        assert!((1..=12).contains(&date.month()));
+        assert!((1..=31).contains(&date.day()));
+    }
+}

src/tools/builder/core.rs

@@ -43,7 +43,6 @@ use crate::error::ToolError as AgentToolError;
 use crate::llm::{
     ChatMessage, LlmProvider, Reasoning, ReasoningContext, RespondResult, ToolDefinition,
 };
-use crate::safety::SafetyLayer;
 use crate::tools::ToolRegistry;
 use crate::tools::tool::{ApprovalRequirement, Tool, ToolError, ToolOutput};
 
@@ -251,29 +250,18 @@ pub trait SoftwareBuilder: Send + Sync {
 pub struct LlmSoftwareBuilder {
     config: BuilderConfig,
     llm: Arc<dyn LlmProvider>,
-    safety: Arc<SafetyLayer>,
     tools: Arc<ToolRegistry>,
 }
 
 impl LlmSoftwareBuilder {
     /// Create a new LLM-based software builder.
-    pub fn new(
-        config: BuilderConfig,
-        llm: Arc<dyn LlmProvider>,
-        safety: Arc<SafetyLayer>,
-        tools: Arc<ToolRegistry>,
-    ) -> Self {
+    pub fn new(config: BuilderConfig, llm: Arc<dyn LlmProvider>, tools: Arc<ToolRegistry>) -> Self {
         // Ensure build directory exists
         if let Err(e) = std::fs::create_dir_all(&config.build_dir) {
             tracing::warn!("Failed to create build directory: {}", e);
         }
 
-        Self {
-            config,
-            llm,
-            safety,
-            tools,
-        }
+        Self { config, llm, tools }
     }
 
     /// Get the build tools available for the build loop.
@@ -521,7 +509,7 @@ Create alongside the .wasm file to grant capabilities:
         let mut iteration = 0;
 
         // Create reasoning engine
-        let reasoning = Reasoning::new(self.llm.clone(), self.safety.clone());
+        let reasoning = Reasoning::new(self.llm.clone());
 
         // Build initial context
         let tool_defs = self.get_build_tools().await;
@@ -822,7 +810,7 @@ Create alongside the .wasm file to grant capabilities:
 impl SoftwareBuilder for LlmSoftwareBuilder {
     async fn analyze(&self, description: &str) -> Result<BuildRequirement, AgentToolError> {
         // Use LLM to parse the description
-        let reasoning = Reasoning::new(self.llm.clone(), self.safety.clone());
+        let reasoning = Reasoning::new(self.llm.clone());
 
         let prompt = format!(
             r#"Analyze this software requirement and extract structured information.

src/tools/builtin/memory.rs

@@ -172,7 +172,7 @@ impl Tool for MemoryWriteTool {
     async fn execute(
         &self,
         params: serde_json::Value,
-        _ctx: &JobContext,
+        ctx: &JobContext,
     ) -> Result<ToolOutput, ToolError> {
         let start = std::time::Instant::now();
 
@@ -239,11 +239,12 @@ impl Tool for MemoryWriteTool {
                 paths::MEMORY.to_string()
             }
             "daily_log" => {
+                let tz = crate::timezone::parse_timezone(&ctx.user_timezone)
+                    .unwrap_or(chrono_tz::Tz::UTC);
                 self.workspace
-                    .append_daily_log(content)
+                    .append_daily_log_tz(content, tz)
                     .await
-                    .map_err(|e| ToolError::ExecutionFailed(format!("Write failed: {}", e)))?;
-                format!("daily/{}.md", chrono::Utc::now().format("%Y-%m-%d"))
+                    .map_err(|e| ToolError::ExecutionFailed(format!("Write failed: {}", e)))?
             }
             "heartbeat" => {
                 if append {

src/tools/builtin/message.rs

@@ -105,42 +105,47 @@ impl Tool for MessageTool {
     async fn execute(
         &self,
         params: serde_json::Value,
-        _ctx: &JobContext,
+        ctx: &JobContext,
     ) -> Result<ToolOutput, ToolError> {
         let start = std::time::Instant::now();
 
         let content = require_str(&params, "content")?;
 
-        // Get channel: use param or fall back to default
-        let channel = if let Some(c) = params.get("channel").and_then(|v| v.as_str()) {
-            c.to_string()
-        } else {
-            self.default_channel
+        // Get channel: use param → conversation default → job metadata → None (broadcast all)
+        let channel: Option<String> =
+            if let Some(c) = params.get("channel").and_then(|v| v.as_str()) {
+                Some(c.to_string())
+            } else if let Some(c) = self
+                .default_channel
                 .read()
                 .unwrap_or_else(|e| e.into_inner())
                 .clone()
-                .ok_or_else(|| {
-                    ToolError::ExecutionFailed(
-                        "No channel specified and no active conversation. Provide channel parameter."
-                            .to_string(),
-                    )
-                })?
-        };
+            {
+                Some(c)
+            } else {
+                ctx.metadata
+                    .get("notify_channel")
+                    .and_then(|v| v.as_str())
+                    .map(|c| c.to_string())
+            };
 
-        // Get target: use param or fall back to default
+        // Get target: use param → conversation default → job metadata
         let target = if let Some(t) = params.get("target").and_then(|v| v.as_str()) {
             t.to_string()
+        } else if let Some(t) = self
+            .default_target
+            .read()
+            .unwrap_or_else(|e| e.into_inner())
+            .clone()
+        {
+            t
+        } else if let Some(t) = ctx.metadata.get("notify_user").and_then(|v| v.as_str()) {
+            t.to_string()
         } else {
-            self.default_target
-                .read()
-                .unwrap_or_else(|e| e.into_inner())
-                .clone()
-                .ok_or_else(|| {
-                    ToolError::ExecutionFailed(
-                        "No target specified and no active conversation. Provide target parameter."
-                            .to_string(),
-                    )
-                })?
+            return Err(ToolError::ExecutionFailed(
+                "No target specified and no active conversation. Provide target parameter."
+                    .to_string(),
+            ));
         };
 
         let attachments: Vec<String> = match params.get("attachments") {
@@ -181,36 +186,79 @@ impl Tool for MessageTool {
             response = response.with_attachments(attachments);
         }
 
-        match self
-            .channel_manager
-            .broadcast(&channel, &target, response)
-            .await
-        {
-            Ok(()) => {
-                tracing::info!(
-                    message_sent = true,
-                    channel = %channel,
-                    target = %target,
-                    attachments = attachment_count,
-                    "Message sent via message tool"
-                );
-                let msg = format!("Sent message to {}:{}", channel, target);
-                Ok(ToolOutput::text(msg, start.elapsed()))
+        if let Some(ref channel) = channel {
+            // Send to a specific channel
+            match self
+                .channel_manager
+                .broadcast(channel, &target, response)
+                .await
+            {
+                Ok(()) => {
+                    tracing::info!(
+                        message_sent = true,
+                        channel = %channel,
+                        target = %target,
+                        attachments = attachment_count,
+                        "Message sent via message tool"
+                    );
+                    let msg = format!("Sent message to {}:{}", channel, target);
+                    Ok(ToolOutput::text(msg, start.elapsed()))
+                }
+                Err(e) => {
+                    let available = self.channel_manager.channel_names().await.join(", ");
+                    let err_msg = if available.is_empty() {
+                        format!(
+                            "Failed to send to {}:{}: {}. No channels connected.",
+                            channel, target, e
+                        )
+                    } else {
+                        format!(
+                            "Failed to send to {}:{}. Available channels: {}. Error: {}",
+                            channel, target, available, e
+                        )
+                    };
+                    Err(ToolError::ExecutionFailed(err_msg))
+                }
             }
-            Err(e) => {
-                let available = self.channel_manager.channel_names().await.join(", ");
-                let err_msg = if available.is_empty() {
-                    format!(
-                        "Failed to send to {}:{}: {}. No channels connected.",
-                        channel, target, e
-                    )
+        } else {
+            // No channel specified — broadcast to all channels (routine with notify.channel = None)
+            let results = self.channel_manager.broadcast_all(&target, response).await;
+            let mut succeeded = Vec::new();
+            let mut failed: Vec<&str> = Vec::new();
+            for (ch, result) in &results {
+                match result {
+                    Ok(()) => succeeded.push(ch.as_str()),
+                    Err(e) => {
+                        tracing::warn!(
+                            channel = %ch,
+                            target = %target,
+                            "broadcast_all: channel failed: {}", e
+                        );
+                        failed.push(ch.as_str());
+                    }
+                }
+            }
+            if succeeded.is_empty() {
+                let err_msg = if failed.is_empty() {
+                    "No channels connected.".to_string()
                 } else {
-                    format!(
-                        "Failed to send to {}:{}. Available channels: {}. Error: {}",
-                        channel, target, available, e
-                    )
+                    format!("All channels failed: {}", failed.join(", "))
                 };
                 Err(ToolError::ExecutionFailed(err_msg))
+            } else {
+                tracing::info!(
+                    message_sent = true,
+                    channels = ?succeeded,
+                    target = %target,
+                    attachments = attachment_count,
+                    "Message broadcast via message tool"
+                );
+                let msg = format!(
+                    "Broadcast message to {} (target: {})",
+                    succeeded.join(", "),
+                    target
+                );
+                Ok(ToolOutput::text(msg, start.elapsed()))
             }
         }
     }
@@ -576,4 +624,90 @@ mod tests {
             ApprovalRequirement::Never,
         );
     }
+
+    #[tokio::test]
+    async fn message_tool_falls_back_to_job_metadata() {
+        // Regression: when no conversation context is set (e.g. routine full-job),
+        // the message tool should fall back to notify_channel/notify_user from
+        // JobContext metadata instead of returning "No target specified".
+        let tool = MessageTool::new(Arc::new(ChannelManager::new()));
+
+        let mut ctx = crate::context::JobContext::new("routine-job", "price alert");
+        ctx.metadata = serde_json::json!({
+            "notify_channel": "telegram",
+            "notify_user": "123456789",
+        });
+
+        // No set_context called — simulates a routine full-job worker
+        let result = tool
+            .execute(serde_json::json!({"content": "NEAR price is $5"}), &ctx)
+            .await;
+
+        // Should fail at channel broadcast (no real channel), NOT at
+        // "No target specified and no active conversation"
+        assert!(result.is_err());
+        let err = result.unwrap_err().to_string();
+        assert!(
+            !err.contains("No target specified"),
+            "Should not get 'No target specified' when metadata has notify_user, got: {}",
+            err
+        );
+        assert!(
+            !err.contains("No channel specified"),
+            "Should not get 'No channel specified' when metadata has notify_channel, got: {}",
+            err
+        );
+    }
+
+    #[tokio::test]
+    async fn message_tool_no_metadata_still_errors() {
+        // When neither conversation context nor metadata is set, should still
+        // return a clear error (target resolution fails).
+        let tool = MessageTool::new(Arc::new(ChannelManager::new()));
+        let ctx = crate::context::JobContext::new("orphan-job", "no notify config");
+
+        let result = tool
+            .execute(serde_json::json!({"content": "hello"}), &ctx)
+            .await;
+
+        assert!(result.is_err());
+        let err = result.unwrap_err().to_string();
+        assert!(
+            err.contains("No target specified"),
+            "Expected 'No target specified' error, got: {}",
+            err
+        );
+    }
+
+    #[tokio::test]
+    async fn message_tool_broadcasts_all_when_no_channel() {
+        // Regression: when notify.channel is None but notify_user is set,
+        // the message tool should attempt broadcast_all instead of erroring
+        // with "No channel specified".
+        let tool = MessageTool::new(Arc::new(ChannelManager::new()));
+
+        let mut ctx = crate::context::JobContext::new("routine-job", "price alert");
+        ctx.metadata = serde_json::json!({
+            "notify_user": "123456789",
+        });
+
+        let result = tool
+            .execute(serde_json::json!({"content": "NEAR price is $5"}), &ctx)
+            .await;
+
+        // Should fail because no channels are registered (empty ChannelManager),
+        // NOT because "No channel specified".
+        assert!(result.is_err());
+        let err = result.unwrap_err().to_string();
+        assert!(
+            !err.contains("No channel specified"),
+            "Should not get 'No channel specified' when broadcasting, got: {}",
+            err
+        );
+        assert!(
+            err.contains("No channels connected") || err.contains("All channels failed"),
+            "Expected channel delivery error, got: {}",
+            err
+        );
+    }
 }

src/tools/builtin/routine.rs

@@ -107,6 +107,10 @@ impl Tool for RoutineCreateTool {
                 "notify_user": {
                     "type": "string",
                     "description": "User/target to notify (e.g. username, chat ID). Defaults to 'default'."
+                },
+                "timezone": {
+                    "type": "string",
+                    "description": "IANA timezone for cron schedule evaluation (e.g. 'America/New_York'). Defaults to UTC."
                 }
             },
             "required": ["name", "trigger_type", "prompt"]
@@ -143,12 +147,26 @@ impl Tool for RoutineCreateTool {
                                 "cron trigger requires 'schedule'".to_string(),
                             )
                         })?;
+                let timezone = params
+                    .get("timezone")
+                    .and_then(|v| v.as_str())
+                    .map(|tz| {
+                        crate::timezone::parse_timezone(tz)
+                            .map(|_| tz.to_string())
+                            .ok_or_else(|| {
+                                ToolError::InvalidParameters(format!(
+                                    "invalid IANA timezone: '{tz}'"
+                                ))
+                            })
+                    })
+                    .transpose()?;
                 // Validate cron expression
-                next_cron_fire(schedule).map_err(|e| {
+                next_cron_fire(schedule, timezone.as_deref()).map_err(|e| {
                     ToolError::InvalidParameters(format!("invalid cron schedule: {e}"))
                 })?;
                 Trigger::Cron {
                     schedule: schedule.to_string(),
+                    timezone,
                 }
             }
             "event" => {
@@ -228,8 +246,12 @@ impl Tool for RoutineCreateTool {
             .unwrap_or(300);
 
         // Compute next fire time for cron
-        let next_fire = if let Trigger::Cron { ref schedule } = trigger {
-            next_cron_fire(schedule).unwrap_or(None)
+        let next_fire = if let Trigger::Cron {
+            ref schedule,
+            ref timezone,
+        } = trigger
+        {
+            next_cron_fire(schedule, timezone.as_deref()).unwrap_or(None)
         } else {
             None
         };
@@ -412,6 +434,10 @@ impl Tool for RoutineUpdateTool {
                     "type": "string",
                     "description": "New cron schedule (for cron triggers)"
                 },
+                "timezone": {
+                    "type": "string",
+                    "description": "IANA timezone for cron schedule (e.g. 'America/New_York'). Only valid for cron triggers."
+                },
                 "description": {
                     "type": "string",
                     "description": "New description"
@@ -453,15 +479,47 @@ impl Tool for RoutineUpdateTool {
             }
         }
 
-        if let Some(schedule) = params.get("schedule").and_then(|v| v.as_str()) {
-            // Validate
-            next_cron_fire(schedule)
-                .map_err(|e| ToolError::InvalidParameters(format!("invalid cron schedule: {e}")))?;
+        // Validate timezone param if provided
+        let new_timezone = params
+            .get("timezone")
+            .and_then(|v| v.as_str())
+            .map(|tz| {
+                crate::timezone::parse_timezone(tz)
+                    .map(|_| tz.to_string())
+                    .ok_or_else(|| {
+                        ToolError::InvalidParameters(format!("invalid IANA timezone: '{tz}'"))
+                    })
+            })
+            .transpose()?;
 
-            routine.trigger = Trigger::Cron {
-                schedule: schedule.to_string(),
+        let new_schedule = params.get("schedule").and_then(|v| v.as_str());
+
+        if new_schedule.is_some() || new_timezone.is_some() {
+            // Extract existing cron fields (cloned to avoid borrow conflict)
+            let existing_cron = match &routine.trigger {
+                Trigger::Cron { schedule, timezone } => Some((schedule.clone(), timezone.clone())),
+                _ => None,
             };
-            routine.next_fire_at = next_cron_fire(schedule).unwrap_or(None);
+
+            if let Some((old_schedule, old_tz)) = existing_cron {
+                let effective_schedule = new_schedule.unwrap_or(&old_schedule);
+                let effective_tz = new_timezone.or(old_tz);
+                // Validate
+                next_cron_fire(effective_schedule, effective_tz.as_deref()).map_err(|e| {
+                    ToolError::InvalidParameters(format!("invalid cron schedule: {e}"))
+                })?;
+
+                routine.trigger = Trigger::Cron {
+                    schedule: effective_schedule.to_string(),
+                    timezone: effective_tz.clone(),
+                };
+                routine.next_fire_at =
+                    next_cron_fire(effective_schedule, effective_tz.as_deref()).unwrap_or(None);
+            } else {
+                return Err(ToolError::InvalidParameters(
+                    "Cannot update schedule or timezone on a non-cron routine.".to_string(),
+                ));
+            }
         }
 
         self.store
@@ -620,9 +678,13 @@ impl Tool for RoutineFireTool {
             .map_err(|e| ToolError::ExecutionFailed(format!("DB error: {e}")))?
             .ok_or_else(|| ToolError::ExecutionFailed(format!("routine '{}' not found", name)))?;
 
-        let run_id = self.engine.fire_manual(routine.id).await.map_err(|e| {
-            ToolError::ExecutionFailed(format!("failed to fire routine '{}': {e}", name))
-        })?;
+        let run_id = self
+            .engine
+            .fire_manual(routine.id, None)
+            .await
+            .map_err(|e| {
+                ToolError::ExecutionFailed(format!("failed to fire routine '{}': {e}", name))
+            })?;
 
         let result = serde_json::json!({
             "name": name,

src/tools/builtin/time.rs

@@ -48,7 +48,7 @@ impl Tool for TimeTool {
     async fn execute(
         &self,
         params: serde_json::Value,
-        _ctx: &JobContext,
+        ctx: &JobContext,
     ) -> Result<ToolOutput, ToolError> {
         let start = std::time::Instant::now();
 
@@ -57,10 +57,15 @@ impl Tool for TimeTool {
         let result = match operation {
             "now" => {
                 let now = Utc::now();
+                let tz =
+                    crate::timezone::parse_timezone(&ctx.user_timezone).unwrap_or(chrono_tz::UTC);
+                let local = now.with_timezone(&tz);
                 serde_json::json!({
                     "iso": now.to_rfc3339(),
                     "unix": now.timestamp(),
-                    "unix_millis": now.timestamp_millis()
+                    "unix_millis": now.timestamp_millis(),
+                    "local_iso": local.to_rfc3339(),
+                    "timezone": tz.name()
                 })
             }
             "parse" => {
@@ -112,3 +117,42 @@ impl Tool for TimeTool {
         false // Internal tool, no external data
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test]
+    async fn test_now_includes_local_time_when_timezone_set() {
+        let tool = TimeTool;
+        let mut ctx = JobContext::with_user("test", "chat", "test");
+        ctx.user_timezone = "America/New_York".to_string();
+
+        let output = tool
+            .execute(serde_json::json!({"operation": "now"}), &ctx)
+            .await
+            .expect("execute");
+        assert!(
+            output.result.get("local_iso").is_some(),
+            "should have local_iso"
+        );
+        assert_eq!(
+            output.result["timezone"].as_str(),
+            Some("America/New_York"),
+            "should report timezone"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_now_includes_utc_timezone_by_default() {
+        let tool = TimeTool;
+        let ctx = JobContext::with_user("test", "chat", "test");
+        // Default user_timezone is "UTC" which is a valid IANA timezone
+        let output = tool
+            .execute(serde_json::json!({"operation": "now"}), &ctx)
+            .await
+            .expect("execute");
+        assert!(output.result.get("iso").is_some(), "should have iso");
+        assert_eq!(output.result["timezone"].as_str(), Some("UTC"));
+    }
+}

src/tools/mcp/client.rs

@@ -261,9 +261,9 @@ impl McpClient {
         if !response.status().is_success() {
             let status = response.status();
             let body = response.text().await.unwrap_or_default();
+            let preview = sanitize_error_body(&body);
             return Err(ToolError::ExternalService(format!(
-                "MCP server returned status: {} - {}",
-                status, body
+                "MCP server returned status: {status} - {preview}",
             )));
         }
 
@@ -548,6 +548,58 @@ impl Tool for McpToolWrapper {
     }
 }
 
+/// Sanitize an HTTP error response body for safe display.
+///
+/// Detects full HTML error pages (containing `<html` or `<!DOCTYPE`) and
+/// strips all tags, collapsing whitespace.  Non-HTML bodies are left
+/// intact.  In both cases the result is truncated to 200 *characters*
+/// (char-boundary safe) so that large payloads don't bloat error messages.
+///
+/// See #263 — raw HTML error pages were propagating through the error
+/// chain into the web UI, causing a white screen.
+fn sanitize_error_body(body: &str) -> String {
+    const MAX_CHARS: usize = 200;
+
+    // Only strip tags when the body looks like a full HTML document.
+    // Plain text that happens to contain `<` / `>` (e.g. log lines,
+    // comparison expressions) is left untouched.
+    let lower = body.to_ascii_lowercase();
+    let is_html_document = lower.contains("<html") || lower.contains("<!doctype");
+
+    let text = if is_html_document {
+        let stripped = body
+            .chars()
+            .fold((String::new(), false), |(mut out, in_tag), c| {
+                if c == '<' {
+                    (out, true)
+                } else if c == '>' {
+                    (out, false)
+                } else if !in_tag {
+                    out.push(c);
+                    (out, false)
+                } else {
+                    (out, true)
+                }
+            })
+            .0;
+        stripped.split_whitespace().collect::<Vec<_>>().join(" ")
+    } else {
+        body.to_string()
+    };
+
+    // Truncate at a char boundary (safe for multi-byte UTF-8).
+    if text.chars().count() > MAX_CHARS {
+        let byte_offset = text
+            .char_indices()
+            .nth(MAX_CHARS)
+            .map(|(i, _)| i)
+            .unwrap_or(text.len());
+        format!("{}... ({} bytes total)", &text[..byte_offset], body.len())
+    } else {
+        text
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -740,4 +792,73 @@ mod tests {
         };
         assert!(!tool.requires_approval());
     }
+
+    // Regression tests for #263: HTML error bodies must not propagate raw
+    // markup through the error chain into the web UI.
+
+    #[test]
+    fn test_sanitize_error_body_strips_html_tags() {
+        let html =
+            r#"<!DOCTYPE html><html><body><h1>422 Error</h1><p>Invalid token</p></body></html>"#;
+        let result = sanitize_error_body(html);
+        assert!(!result.contains('<'), "HTML tags must be stripped");
+        assert!(!result.contains('>'), "HTML tags must be stripped");
+        assert!(result.contains("422 Error"));
+        assert!(result.contains("Invalid token"));
+    }
+
+    #[test]
+    fn test_sanitize_error_body_truncates_large_html_page() {
+        let html = format!(
+            "<html><body><p>{}</p></body></html>",
+            "error detail ".repeat(50)
+        );
+        let result = sanitize_error_body(&html);
+        assert!(result.contains("..."));
+        assert!(result.contains("bytes total)"));
+        assert!(!result.contains('<'));
+    }
+
+    #[test]
+    fn test_sanitize_error_body_passes_short_plain_text() {
+        assert_eq!(sanitize_error_body("Not Found"), "Not Found");
+    }
+
+    #[test]
+    fn test_sanitize_error_body_truncates_long_plain_text() {
+        let long = "x".repeat(300);
+        let result = sanitize_error_body(&long);
+        assert!(result.contains("..."));
+        assert!(result.contains("300 bytes total)"));
+    }
+
+    #[test]
+    fn test_sanitize_error_body_multibyte_no_panic() {
+        // 300 CJK characters = 900 bytes; truncation must land on a
+        // char boundary, not in the middle of a multi-byte sequence.
+        let cjk = "错误".repeat(150);
+        let result = sanitize_error_body(&cjk);
+        assert!(result.contains("..."));
+        // Must be valid UTF-8 (would have panicked otherwise).
+        assert!(result.is_char_boundary(result.len()));
+    }
+
+    #[test]
+    fn test_sanitize_error_body_strips_uppercase_html() {
+        let html = "<HTML><BODY><H1>500 Internal Server Error</H1></BODY></HTML>";
+        let result = sanitize_error_body(html);
+        assert!(
+            !result.contains('<'),
+            "uppercase HTML tags must be stripped"
+        );
+        assert!(result.contains("500 Internal Server Error"));
+    }
+
+    #[test]
+    fn test_sanitize_error_body_preserves_angle_brackets_in_non_html() {
+        // Text with < and > that is NOT an HTML document should be
+        // left untouched (e.g. log lines, comparison expressions).
+        let text = "value < 10 and value > 0";
+        assert_eq!(sanitize_error_body(text), text);
+    }
 }

src/tools/mcp/session.rs

@@ -204,7 +204,9 @@ mod tests {
         assert!(!session.is_stale(1800));
 
         // Manually set last_activity to the past to simulate staleness
-        session.last_activity = std::time::Instant::now() - std::time::Duration::from_secs(10);
+        session.last_activity = std::time::Instant::now()
+            .checked_sub(std::time::Duration::from_secs(10))
+            .expect("System uptime is too low to run staleness test");
         assert!(session.is_stale(5));
         assert!(!session.is_stale(15));
     }

src/tools/registry.rs

@@ -10,7 +10,6 @@ use crate::db::Database;
 use crate::extensions::ExtensionManager;
 use crate::llm::{LlmProvider, ToolDefinition};
 use crate::orchestrator::job_manager::ContainerJobManager;
-use crate::safety::SafetyLayer;
 use crate::secrets::SecretsStore;
 use crate::skills::catalog::SkillCatalog;
 use crate::skills::registry::SkillRegistry;
@@ -485,17 +484,15 @@ impl ToolRegistry {
     pub async fn register_builder_tool(
         self: &Arc<Self>,
         llm: Arc<dyn LlmProvider>,
-        safety: Arc<SafetyLayer>,
         config: Option<BuilderConfig>,
     ) {
         // First register dev tools needed by the builder
         self.register_dev_tools();
 
-        // Create the builder (arg order: config, llm, safety, tools)
+        // Create the builder (arg order: config, llm, tools)
         let builder = Arc::new(LlmSoftwareBuilder::new(
             config.unwrap_or_default(),
             llm,
-            safety,
             Arc::clone(self),
         ));
 

src/worker/runtime.rs

@@ -133,7 +133,7 @@ impl WorkerRuntime {
             .await?;
 
         // Create reasoning engine
-        let reasoning = Reasoning::new(self.llm.clone(), self.safety.clone());
+        let reasoning = Reasoning::new(self.llm.clone());
 
         // Build initial context
         let mut reason_ctx = ReasoningContext::new().with_job(&job.description);

src/workspace/chunker.rs

@@ -113,79 +113,6 @@ pub fn chunk_document(content: &str, config: ChunkConfig) -> Vec<String> {
     chunks
 }
 
-/// Split content by paragraphs first, then chunk.
-///
-/// This is better for preserving semantic boundaries.
-#[allow(dead_code)] // Alternative chunking strategy for paragraph-aware indexing
-pub fn chunk_by_paragraphs(content: &str, config: ChunkConfig) -> Vec<String> {
-    if content.is_empty() {
-        return Vec::new();
-    }
-
-    // Split by double newlines (paragraphs)
-    let paragraphs: Vec<&str> = content
-        .split("\n\n")
-        .map(|p| p.trim())
-        .filter(|p| !p.is_empty())
-        .collect();
-
-    if paragraphs.is_empty() {
-        return chunk_document(content, config);
-    }
-
-    let mut chunks = Vec::new();
-    let mut current_chunk = String::new();
-    let mut current_word_count = 0;
-
-    for paragraph in paragraphs {
-        let para_words = paragraph.split_whitespace().count();
-
-        // If this paragraph alone exceeds chunk size, chunk it separately
-        if para_words > config.chunk_size {
-            // Flush current chunk first
-            if !current_chunk.is_empty() {
-                chunks.push(current_chunk.trim().to_string());
-                current_chunk = String::new();
-                current_word_count = 0;
-            }
-            // Chunk the large paragraph
-            let para_chunks = chunk_document(paragraph, config.clone());
-            chunks.extend(para_chunks);
-            continue;
-        }
-
-        // Check if adding this paragraph would exceed chunk size
-        if current_word_count + para_words > config.chunk_size {
-            // Flush current chunk
-            if !current_chunk.is_empty() {
-                chunks.push(current_chunk.trim().to_string());
-            }
-            current_chunk = paragraph.to_string();
-            current_word_count = para_words;
-        } else {
-            // Add paragraph to current chunk
-            if !current_chunk.is_empty() {
-                current_chunk.push_str("\n\n");
-            }
-            current_chunk.push_str(paragraph);
-            current_word_count += para_words;
-        }
-    }
-
-    // Flush remaining content
-    if !current_chunk.is_empty() {
-        // If too small, merge with previous chunk if possible
-        if current_word_count < config.min_chunk_size && !chunks.is_empty() {
-            let last = chunks.pop().unwrap();
-            chunks.push(format!("{}\n\n{}", last, current_chunk.trim()));
-        } else {
-            chunks.push(current_chunk.trim().to_string());
-        }
-    }
-
-    chunks
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -253,49 +180,6 @@ mod tests {
         assert_eq!(config.step_size(), 85);
     }
 
-    #[test]
-    fn test_paragraph_chunking() {
-        let config = ChunkConfig::default().with_chunk_size(20);
-
-        let content = "First paragraph with some words.\n\nSecond paragraph with different content.\n\nThird paragraph here.";
-        let chunks = chunk_by_paragraphs(content, config);
-
-        // Should preserve paragraph boundaries
-        assert!(!chunks.is_empty());
-        for chunk in &chunks {
-            // No chunk should start or end with \n\n
-            assert!(!chunk.starts_with("\n"));
-            assert!(!chunk.ends_with("\n"));
-        }
-    }
-
-    #[test]
-    fn test_large_paragraph_handling() {
-        let config = ChunkConfig {
-            chunk_size: 10,
-            overlap_percent: 0.15,
-            min_chunk_size: 3, // Low threshold for test
-        };
-
-        // Create a paragraph with 30 words
-        let large_para = (1..=30)
-            .map(|i| format!("word{}", i))
-            .collect::<Vec<_>>()
-            .join(" ");
-        let content = format!("Short intro.\n\n{}\n\nShort outro.", large_para);
-
-        let chunks = chunk_by_paragraphs(&content, config);
-
-        // Should have multiple chunks due to large paragraph
-        // 30 words + 2 intro + 2 outro = 34 words, chunk_size=10
-        // Expect at least 3 chunks
-        assert!(
-            chunks.len() >= 3,
-            "Expected at least 3 chunks for 34 words with chunk_size=10, got {}",
-            chunks.len()
-        );
-    }
-
     #[test]
     fn test_min_chunk_size_merging() {
         let config = ChunkConfig {

src/workspace/lancedb_store.rs

@@ -0,0 +1,644 @@
+//! LanceDB-backed vector store for workspace memory chunks.
+//!
+//! Provides an alternative to pgvector/libsql for semantic search when the
+//! `lancedb` feature is enabled. Documents and metadata stay in the main
+//! database; this store holds chunk embeddings for vector similarity search.
+//!
+//! Configuration:
+//!   LANCEDB_PATH=~/.ironclaw/lancedb   # Default
+//!   VECTOR_BACKEND=lancedb             # Use LanceDB for vector search
+
+/// Default embedding dimension (text-embedding-3-small).
+/// Override by passing the actual provider dimension to `LanceDbVectorStore::new()`.
+pub const DEFAULT_EMBEDDING_DIM: i32 = 1536;
+
+#[cfg(feature = "lancedb")]
+mod impl_lancedb {
+    use std::sync::Arc;
+
+    use arrow_array::types::Float32Type;
+    use arrow_array::{Array, FixedSizeListArray, RecordBatch, RecordBatchIterator, StringArray};
+    use arrow_schema::{DataType, Field, Schema};
+    use async_trait::async_trait;
+    use futures::StreamExt;
+    use lancedb::query::{ExecutableQuery, QueryBase};
+    use uuid::Uuid;
+
+    use super::DEFAULT_EMBEDDING_DIM;
+    use crate::error::WorkspaceError;
+    use crate::workspace::search::RankedResult;
+    use crate::workspace::vector_store::VectorStore;
+
+    const TABLE_NAME: &str = "memory_chunks";
+
+    /// Escapes a string for safe use in LanceDB predicate expressions.
+    /// Uses SQL-style escaping: single quotes are doubled to prevent injection.
+    fn escape_predicate_value(s: &str) -> String {
+        s.replace('\'', "''")
+    }
+
+    /// LanceDB-backed vector store.
+    ///
+    /// The `update_embedding` method uses delete-then-insert (not atomic).
+    /// LanceDB does not support transactions, so a crash between the two
+    /// operations can lose the embedding for that chunk. This is acceptable
+    /// for personal workspace sizes where data can be reindexed.
+    pub struct LanceDbVectorStore {
+        db: Arc<lancedb::Connection>,
+        table_name: String,
+        embedding_dim: i32,
+        schema: Arc<Schema>,
+        table: tokio::sync::OnceCell<lancedb::Table>,
+    }
+
+    impl LanceDbVectorStore {
+        /// Create a new LanceDB store at the given path.
+        ///
+        /// `embedding_dim` should match `EmbeddingProvider::dimension()`.
+        /// Pass `None` to use the default (1536, text-embedding-3-small).
+        pub async fn new(
+            path: impl AsRef<std::path::Path>,
+            embedding_dim: Option<usize>,
+        ) -> Result<Self, WorkspaceError> {
+            let path_str = path
+                .as_ref()
+                .to_str()
+                .ok_or_else(|| WorkspaceError::SearchFailed {
+                    reason: "Invalid LanceDB path".to_string(),
+                })?;
+
+            let db = lancedb::connect(path_str).execute().await.map_err(|e| {
+                WorkspaceError::SearchFailed {
+                    reason: format!("Failed to connect to LanceDB: {}", e),
+                }
+            })?;
+
+            let dim = embedding_dim.unwrap_or(DEFAULT_EMBEDDING_DIM as usize) as i32;
+            let schema = Arc::new(Self::build_schema(dim));
+
+            let store = Self {
+                db: Arc::new(db),
+                table_name: TABLE_NAME.to_string(),
+                embedding_dim: dim,
+                schema,
+                table: tokio::sync::OnceCell::new(),
+            };
+
+            store.ensure_table().await?;
+            Ok(store)
+        }
+
+        async fn ensure_table(&self) -> Result<(), WorkspaceError> {
+            let tables = self.db.table_names().execute().await.map_err(|e| {
+                WorkspaceError::SearchFailed {
+                    reason: format!("Failed to list tables: {}", e),
+                }
+            })?;
+
+            if tables.iter().any(|t| t == &self.table_name) {
+                return Ok(());
+            }
+
+            self.db
+                .create_empty_table(&self.table_name, self.schema.clone())
+                .execute()
+                .await
+                .map_err(|e| WorkspaceError::SearchFailed {
+                    reason: format!("Failed to create table: {}", e),
+                })?;
+
+            // Index creation is deferred — brute-force search via
+            // bypass_vector_index() works without a pre-built index and is
+            // sufficient for personal workspace sizes.
+            Ok(())
+        }
+
+        /// Get or open the cached table handle.
+        async fn table(&self) -> Result<&lancedb::Table, WorkspaceError> {
+            self.table
+                .get_or_try_init(|| async {
+                    self.db
+                        .open_table(&self.table_name)
+                        .execute()
+                        .await
+                        .map_err(|e| WorkspaceError::SearchFailed {
+                            reason: format!("Failed to open table: {}", e),
+                        })
+                })
+                .await
+        }
+
+        fn build_schema(embedding_dim: i32) -> Schema {
+            Schema::new(vec![
+                Field::new("chunk_id", DataType::Utf8, false),
+                Field::new("document_id", DataType::Utf8, false),
+                Field::new("document_path", DataType::Utf8, false),
+                Field::new("user_id", DataType::Utf8, false),
+                Field::new("agent_id", DataType::Utf8, true),
+                Field::new("content", DataType::Utf8, false),
+                Field::new(
+                    "vector",
+                    DataType::FixedSizeList(
+                        Arc::new(Field::new("item", DataType::Float32, true)),
+                        embedding_dim,
+                    ),
+                    false,
+                ),
+            ])
+        }
+    }
+
+    #[async_trait]
+    impl VectorStore for LanceDbVectorStore {
+        async fn store_embedding(
+            &self,
+            chunk_id: Uuid,
+            document_id: Uuid,
+            document_path: &str,
+            user_id: &str,
+            agent_id: Option<Uuid>,
+            content: &str,
+            embedding: &[f32],
+        ) -> Result<(), WorkspaceError> {
+            if embedding.len() != self.embedding_dim as usize {
+                return Err(WorkspaceError::EmbeddingFailed {
+                    reason: format!(
+                        "Embedding dimension {} does not match expected {}",
+                        embedding.len(),
+                        self.embedding_dim
+                    ),
+                });
+            }
+
+            let table = self.table().await?;
+
+            let chunk_ids = StringArray::from(vec![chunk_id.to_string()]);
+            let document_ids = StringArray::from(vec![document_id.to_string()]);
+            let document_paths = StringArray::from(vec![document_path]);
+            let user_ids = StringArray::from(vec![user_id]);
+            let agent_ids = StringArray::from(vec![agent_id.map(|a| a.to_string())]);
+            let contents = StringArray::from(vec![content]);
+            let vec_values: Vec<Option<f32>> = embedding.iter().map(|&x| Some(x)).collect();
+            let vectors = FixedSizeListArray::from_iter_primitive::<Float32Type, _, _>(
+                vec![Some(vec_values)],
+                self.embedding_dim,
+            );
+
+            let batch = RecordBatch::try_new(
+                self.schema.clone(),
+                vec![
+                    Arc::new(chunk_ids),
+                    Arc::new(document_ids),
+                    Arc::new(document_paths),
+                    Arc::new(user_ids),
+                    Arc::new(agent_ids),
+                    Arc::new(contents),
+                    Arc::new(vectors),
+                ],
+            )
+            .map_err(|e| WorkspaceError::EmbeddingFailed {
+                reason: format!("Failed to create record batch: {}", e),
+            })?;
+
+            let batches =
+                RecordBatchIterator::new(vec![Ok(batch)].into_iter(), self.schema.clone());
+
+            table
+                .add(Box::new(batches) as Box<dyn arrow_array::RecordBatchReader + Send>)
+                .execute()
+                .await
+                .map_err(|e| WorkspaceError::EmbeddingFailed {
+                    reason: format!("Failed to store embedding: {}", e),
+                })?;
+
+            Ok(())
+        }
+
+        async fn update_embedding(
+            &self,
+            chunk_id: Uuid,
+            document_id: Uuid,
+            document_path: &str,
+            user_id: &str,
+            agent_id: Option<Uuid>,
+            content: &str,
+            embedding: &[f32],
+        ) -> Result<(), WorkspaceError> {
+            let table = self.table().await?;
+
+            table
+                .delete(&format!(
+                    "chunk_id = '{}'",
+                    escape_predicate_value(&chunk_id.to_string())
+                ))
+                .await
+                .map_err(|e| WorkspaceError::EmbeddingFailed {
+                    reason: format!("Failed to delete chunk for update: {}", e),
+                })?;
+
+            self.store_embedding(
+                chunk_id,
+                document_id,
+                document_path,
+                user_id,
+                agent_id,
+                content,
+                embedding,
+            )
+            .await
+        }
+
+        async fn delete_embeddings(&self, document_id: Uuid) -> Result<(), WorkspaceError> {
+            let table = self.table().await?;
+
+            table
+                .delete(&format!(
+                    "document_id = '{}'",
+                    escape_predicate_value(&document_id.to_string())
+                ))
+                .await
+                .map_err(|e| WorkspaceError::EmbeddingFailed {
+                    reason: format!("Failed to delete embeddings: {}", e),
+                })?;
+
+            Ok(())
+        }
+
+        async fn vector_search(
+            &self,
+            user_id: &str,
+            agent_id: Option<Uuid>,
+            embedding: &[f32],
+            limit: usize,
+        ) -> Result<Vec<RankedResult>, WorkspaceError> {
+            let table = self.table().await?;
+
+            let filter = if let Some(aid) = agent_id {
+                format!(
+                    "user_id = '{}' AND agent_id = '{}'",
+                    escape_predicate_value(user_id),
+                    escape_predicate_value(&aid.to_string())
+                )
+            } else {
+                format!(
+                    "user_id = '{}' AND agent_id IS NULL",
+                    escape_predicate_value(user_id)
+                )
+            };
+
+            let query = table
+                .query()
+                .nearest_to(embedding)
+                .map_err(|e| WorkspaceError::SearchFailed {
+                    reason: format!("Invalid query vector: {}", e),
+                })?
+                .only_if(&filter)
+                .bypass_vector_index()
+                .limit(limit);
+            let mut stream = ExecutableQuery::execute(&query).await.map_err(|e| {
+                WorkspaceError::SearchFailed {
+                    reason: format!("Vector search failed: {}", e),
+                }
+            })?;
+
+            let mut results = Vec::new();
+            let mut rank: u32 = 1;
+            while let Some(batch_result) = stream.next().await {
+                let batch = batch_result.map_err(|e| WorkspaceError::SearchFailed {
+                    reason: format!("Stream error: {}", e),
+                })?;
+
+                let chunk_id_col = batch.column_by_name("chunk_id").ok_or_else(|| {
+                    WorkspaceError::SearchFailed {
+                        reason: "chunk_id column missing".to_string(),
+                    }
+                })?;
+                let document_id_col = batch.column_by_name("document_id").ok_or_else(|| {
+                    WorkspaceError::SearchFailed {
+                        reason: "document_id column missing".to_string(),
+                    }
+                })?;
+                let document_path_col = batch.column_by_name("document_path").ok_or_else(|| {
+                    WorkspaceError::SearchFailed {
+                        reason: "document_path column missing".to_string(),
+                    }
+                })?;
+                let content_col = batch.column_by_name("content").ok_or_else(|| {
+                    WorkspaceError::SearchFailed {
+                        reason: "content column missing".to_string(),
+                    }
+                })?;
+
+                let chunk_ids = chunk_id_col
+                    .as_any()
+                    .downcast_ref::<StringArray>()
+                    .ok_or_else(|| WorkspaceError::SearchFailed {
+                        reason: "chunk_id wrong type".to_string(),
+                    })?;
+                let document_ids = document_id_col
+                    .as_any()
+                    .downcast_ref::<StringArray>()
+                    .ok_or_else(|| WorkspaceError::SearchFailed {
+                        reason: "document_id wrong type".to_string(),
+                    })?;
+                let document_paths = document_path_col
+                    .as_any()
+                    .downcast_ref::<StringArray>()
+                    .ok_or_else(|| WorkspaceError::SearchFailed {
+                        reason: "document_path wrong type".to_string(),
+                    })?;
+                let contents = content_col
+                    .as_any()
+                    .downcast_ref::<StringArray>()
+                    .ok_or_else(|| WorkspaceError::SearchFailed {
+                        reason: "content wrong type".to_string(),
+                    })?;
+
+                for i in 0..batch.num_rows() {
+                    let raw_chunk_id = chunk_ids.value(i);
+                    let chunk_id =
+                        raw_chunk_id
+                            .parse::<Uuid>()
+                            .map_err(|e| WorkspaceError::SearchFailed {
+                                reason: format!("Invalid chunk_id UUID '{}': {}", raw_chunk_id, e),
+                            })?;
+
+                    let raw_document_id = document_ids.value(i);
+                    let document_id = raw_document_id.parse::<Uuid>().map_err(|e| {
+                        WorkspaceError::SearchFailed {
+                            reason: format!(
+                                "Invalid document_id UUID '{}': {}",
+                                raw_document_id, e
+                            ),
+                        }
+                    })?;
+                    let document_path = document_paths.value(i).to_string();
+                    let content = contents.value(i).to_string();
+
+                    results.push(RankedResult {
+                        chunk_id,
+                        document_id,
+                        document_path,
+                        content,
+                        rank,
+                    });
+                    rank += 1;
+                }
+            }
+
+            Ok(results)
+        }
+    }
+}
+
+#[cfg(feature = "lancedb")]
+pub use impl_lancedb::LanceDbVectorStore;
+
+#[cfg(all(test, feature = "lancedb"))]
+mod tests {
+    use tempfile::TempDir;
+    use uuid::Uuid;
+
+    use super::{DEFAULT_EMBEDDING_DIM, LanceDbVectorStore};
+    use crate::workspace::vector_store::VectorStore;
+
+    fn make_embedding(seed: f32) -> Vec<f32> {
+        (0..DEFAULT_EMBEDDING_DIM as usize)
+            .map(|i| (seed * (i as f32 + 1.0)).sin())
+            .collect()
+    }
+
+    #[tokio::test]
+    async fn test_insert_and_vector_search() {
+        let dir = TempDir::new().unwrap();
+        let store = LanceDbVectorStore::new(dir.path(), None).await.unwrap();
+
+        let chunk_id = Uuid::new_v4();
+        let document_id = Uuid::new_v4();
+        let user_id = "user1";
+        let content = "Rust is a systems programming language";
+        let embedding = make_embedding(1.0);
+
+        store
+            .store_embedding(
+                chunk_id,
+                document_id,
+                "test.md",
+                user_id,
+                None,
+                content,
+                &embedding,
+            )
+            .await
+            .unwrap();
+
+        let results = store
+            .vector_search(user_id, None, &embedding, 5)
+            .await
+            .unwrap();
+
+        assert_eq!(results.len(), 1);
+        assert_eq!(results[0].chunk_id, chunk_id);
+        assert_eq!(results[0].document_id, document_id);
+        assert_eq!(results[0].content, content);
+        assert_eq!(results[0].rank, 1);
+    }
+
+    #[tokio::test]
+    async fn test_insert_multiple_and_search_returns_ordered() {
+        let dir = TempDir::new().unwrap();
+        let store = LanceDbVectorStore::new(dir.path(), None).await.unwrap();
+
+        let doc_id = Uuid::new_v4();
+        let user_id = "user1";
+
+        for (i, seed) in [1.0, 2.0, 3.0].iter().enumerate() {
+            store
+                .store_embedding(
+                    Uuid::new_v4(),
+                    doc_id,
+                    "test.md",
+                    user_id,
+                    None,
+                    &format!("content {}", i),
+                    &make_embedding(*seed),
+                )
+                .await
+                .unwrap();
+        }
+
+        let query_emb = make_embedding(2.0);
+        let results = store
+            .vector_search(user_id, None, &query_emb, 5)
+            .await
+            .unwrap();
+
+        assert_eq!(results.len(), 3);
+        let contents: Vec<_> = results.iter().map(|r| r.content.as_str()).collect();
+        assert!(contents.contains(&"content 0"));
+        assert!(contents.contains(&"content 1"));
+        assert!(contents.contains(&"content 2"));
+    }
+
+    #[tokio::test]
+    async fn test_delete_chunks() {
+        let dir = TempDir::new().unwrap();
+        let store = LanceDbVectorStore::new(dir.path(), None).await.unwrap();
+
+        let doc_id = Uuid::new_v4();
+        let user_id = "user1";
+
+        store
+            .store_embedding(
+                Uuid::new_v4(),
+                doc_id,
+                "test.md",
+                user_id,
+                None,
+                "content",
+                &make_embedding(1.0),
+            )
+            .await
+            .unwrap();
+
+        let results = store
+            .vector_search(user_id, None, &make_embedding(1.0), 5)
+            .await
+            .unwrap();
+        assert_eq!(results.len(), 1);
+
+        store.delete_embeddings(doc_id).await.unwrap();
+
+        let results_after = store
+            .vector_search(user_id, None, &make_embedding(1.0), 5)
+            .await
+            .unwrap();
+        assert!(results_after.is_empty());
+    }
+
+    #[tokio::test]
+    async fn test_update_chunk_embedding() {
+        let dir = TempDir::new().unwrap();
+        let store = LanceDbVectorStore::new(dir.path(), None).await.unwrap();
+
+        let chunk_id = Uuid::new_v4();
+        let doc_id = Uuid::new_v4();
+        let user_id = "user1";
+        let content = "original content";
+
+        store
+            .store_embedding(
+                chunk_id,
+                doc_id,
+                "test.md",
+                user_id,
+                None,
+                content,
+                &make_embedding(1.0),
+            )
+            .await
+            .unwrap();
+
+        let new_embedding = make_embedding(5.0);
+        store
+            .update_embedding(
+                chunk_id,
+                doc_id,
+                "test.md",
+                user_id,
+                None,
+                content,
+                &new_embedding,
+            )
+            .await
+            .unwrap();
+
+        let results = store
+            .vector_search(user_id, None, &new_embedding, 5)
+            .await
+            .unwrap();
+        assert_eq!(results.len(), 1);
+        assert_eq!(results[0].chunk_id, chunk_id);
+    }
+
+    #[tokio::test]
+    async fn test_vector_search_filters_by_user_and_agent() {
+        let dir = TempDir::new().unwrap();
+        let store = LanceDbVectorStore::new(dir.path(), None).await.unwrap();
+
+        let doc_id = Uuid::new_v4();
+        let embedding = make_embedding(1.0);
+
+        store
+            .store_embedding(
+                Uuid::new_v4(),
+                doc_id,
+                "test.md",
+                "user1",
+                None,
+                "user1 content",
+                &embedding,
+            )
+            .await
+            .unwrap();
+
+        store
+            .store_embedding(
+                Uuid::new_v4(),
+                doc_id,
+                "test.md",
+                "user2",
+                None,
+                "user2 content",
+                &embedding,
+            )
+            .await
+            .unwrap();
+
+        let results_user1 = store
+            .vector_search("user1", None, &embedding, 5)
+            .await
+            .unwrap();
+        assert_eq!(results_user1.len(), 1);
+        assert_eq!(results_user1[0].content, "user1 content");
+
+        let results_user2 = store
+            .vector_search("user2", None, &embedding, 5)
+            .await
+            .unwrap();
+        assert_eq!(results_user2.len(), 1);
+        assert_eq!(results_user2[0].content, "user2 content");
+
+        let results_wrong_user = store
+            .vector_search("user3", None, &embedding, 5)
+            .await
+            .unwrap();
+        assert!(results_wrong_user.is_empty());
+    }
+
+    #[tokio::test]
+    async fn test_insert_rejects_wrong_embedding_dim() {
+        let dir = TempDir::new().unwrap();
+        let store = LanceDbVectorStore::new(dir.path(), None).await.unwrap();
+
+        let wrong_dim: Vec<f32> = vec![1.0; 100];
+
+        let err = store
+            .store_embedding(
+                Uuid::new_v4(),
+                Uuid::new_v4(),
+                "test.md",
+                "user1",
+                None,
+                "content",
+                &wrong_dim,
+            )
+            .await
+            .unwrap_err();
+
+        assert!(matches!(
+            err,
+            crate::error::WorkspaceError::EmbeddingFailed { .. }
+        ));
+    }
+}