* feat: direct agentic loop for SWE-bench benchmarks
Replace the full Agent-based runner with a purpose-built agentic loop
that directly calls the LLM with tools. The old path routed through
SafetyLayer (which blocked SWE-bench prompts), dispatcher (capped at
10 iterations), approval flow (wasted iterations), and 20+ irrelevant
builtin tools (diluted the model's focus).
New architecture:
- AgenticLoop: LLM call -> tool execution -> repeat (up to 30 iters)
- Per-task tool scoping via BenchSuite::task_tools() with working dirs
- Suite-provided system prompts via BenchSuite::system_prompt()
- No safety layer, no approval flow, no sessions/threads overhead
- Configurable max_iterations in BenchConfig and TOML
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* fix: apply --model CLI override to LLM provider
The --model flag was updating matrix entry labels but not the actual
LLM provider, so requests were still sent using the model from .env.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* feat: configurable tool iterations and auto-approve for benchmarks
Add max_tool_iterations and auto_approve_tools settings to AgentConfig,
replacing the hardcoded MAX_TOOL_ITERATIONS constant. Fix shell_injection
policy rule to not block markdown backtick code snippets.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* fix: address benchmarks crate audit findings
High:
- Fix truncate_output UTF-8 panic on multi-byte char boundaries
- Fix parallel results durability (write JSONL per-task, not after all)
Medium:
- Fix --sample to use random shuffle instead of first-N
- Delegate all LlmProvider methods in InstrumentedLlm
- Fix LLM-as-judge to return fail instead of misleading 0.5
- Remove unnecessary shallow clone (always gets unshallowed)
- Replace .unwrap() with .expect() in LazyLock regex init
Low:
- Remove dead code: unused error variants, trait methods, struct fields
- Remove BenchSuite::name() (redundant with id())
- Remove TaskSubmission::conversation, ConversationTurn, TurnRole
- Remove unused methods from BenchChannel, results, config
- Clean up ChannelCapture conversation tracking
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* feat: add SWE-bench dataset and Docker scoring infrastructure
Add the SWE-bench Lite dataset (300 tasks) and Docker files for
isolated test execution and scoring of SWE-bench patches.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* chore: remove benchmarks (extracted to separate repo)
Benchmarks crate has been extracted to its own repository.
Remove the workspace member and all benchmarks/ files.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* fix: add missing AgentConfig fields in test initializer
Co-Authored-By: Claude Opus 4.6 <[email protected]>
---------
Co-authored-by: Claude Opus 4.6 <[email protected]>
* Bump MSRV to 1.92 and add GCP deployment files
rig-core 0.30 uses let_chains (stabilized post-1.87), which breaks
builds on Rust 1.85. Bump rust-version in Cargo.toml and both
Dockerfiles to 1.92 (verified working).
Add cloud deployment scaffolding:
- Dockerfile: multi-stage build for the main agent container
- deploy/cloud-sql-proxy.service: systemd unit for Cloud SQL Auth Proxy
- deploy/ironclaw.service: systemd unit for the IronClaw container
- deploy/setup.sh: VM bootstrap script (Docker, proxy, services)
- deploy/env.example: reference environment configuration
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* Address review feedback: harden deploy scaffolding
- Add comment explaining GATEWAY_HOST=0.0.0.0 and when to use 127.0.0.1
- Document /opt/ironclaw ownership model (root-owned, Docker reads as root)
- Switch cloud-sql-proxy service from User=root to DynamicUser=yes
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* fix: Resolve clippy lints (Rust 1.93) and fix CI test workflow
- Fix 97 collapsible_if warnings using let-chains syntax (auto-fixed)
- Fix ptr_arg: change &PathBuf to &Path in pairing store functions
- Fix suspicious_open_options: add .truncate(false) to OpenOptions
- Fix too_many_arguments: add clippy allow on execute_status
- Fix unnecessary_unwrap: use if-let in repository.rs hybrid_search
- Gate unused EchoTool with #[cfg(test)]
- Add PairingStore argument to ChannelStoreData::new() test call sites
- Add skip guard for bundled channel test when WASM artifacts unavailable
- Split CI test workflow to exclude PostgreSQL-dependent integration tests
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* fix: Address review feedback from ilblackdragon
- Add root check to setup.sh (exits with error if not root)
- Add warning comment to env.example about placeholder passwords
- Dockerfile.worker already uses rust:1.92 (no change needed)
- PR #41 overlap noted; will rebase after #41 merges
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* fix: resolve 47 collapsible_if clippy warnings
Collapse nested if statements across the codebase to satisfy
clippy::collapsible_if on Rust 1.93.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
---------
Co-authored-by: Claude Opus 4.6 <[email protected]>
* fix: comprehensive security hardening across all layers
Critical:
- Replace --dangerously-skip-permissions with explicit tool allowlist
via settings.json (Claude Code bridge)
- Constant-time token comparison (subtle crate) in web auth and
orchestrator auth to prevent timing attacks
High:
- Revoke tokens and clean up handles on container creation failure
- Drop SETUID/SETGID capabilities from containers (keep only CHOWN)
- Disable redirect following in HTTP tool and WASM wrapper (SSRF)
- Reject URL userinfo (@) in WASM allowlist parser (host confusion)
- Fix binary body bypassing leak detection (from_utf8 -> from_utf8_lossy)
- Protect identity files from LLM overwrites (prompt injection defense)
- Prevent tool shadowing: built-in tools cannot be replaced dynamically
- User-scoped job APIs: list/detail/cancel/restart/prompt/events/files
- CORS restricted to localhost origins, WebSocket origin validation
- Sandbox shell fail-closed: no silent fallback to unsandboxed execution
- Scrub secrets from log broadcaster before SSE broadcast
- XSS sanitization on rendered markdown in web UI
- WASM epoch ticker thread so timeout deadlines actually fire
Medium:
- Cap state transition history at 200 entries
- SSE/WebSocket connection limit (100 max)
- Request body size limit (1MB)
- Response body size limit enforcement in WASM HTTP
- UTF-8 safe string truncation (routine engine, shell tool)
- Fix PolicyAction::Sanitize to actually run the sanitizer
- TOCTOU fix in scheduler and context manager (hold write lock)
- Project file serving moved behind auth
- Path traversal guard on project_id
- Session file permissions set to 0600 on unix
- AtomicUsize for routine running_count (panic-safe)
- Completion detection hardened against false positives and tool injection
- Tool output no longer drives job completion (only LLM response)
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* fix: address security review findings across all layers
- Fix path traversal sandbox bypass via lexical normalization (file.rs)
- Fix SSRF via DNS rebinding with pre-request hostname resolution (http.rs)
- Add token budget enforcement on LLM calls (reasoning.rs, state.rs)
- Fix cross-user chat history leak with ownership verification (store.rs, server.rs)
- Add sliding-window rate limiter on gateway chat endpoint (server.rs)
- Harden extension install: HTTPS-only, 50MB cap, WASM magic validation (manager.rs)
- Add destructive command blocklist that overrides shell auto-approval (shell.rs)
- Add 5MB response body size cap to HTTP tool (http.rs)
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* refactor: deduplicate shared helpers and remove dead code
Extract floor_char_boundary and llm_signals_completion into src/util.rs,
unifying diverging phrase lists from agent/worker.rs and worker/runtime.rs.
Remove dead RespondResult::usage(), duplicate PROTECTED_IDENTITY_FILES
constant, double LeakDetector scanning in WebLogLayer, and invalid
0.0.0.0 origin from WebSocket allow list.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* fix: address PR review findings and CI test failures
- Fix record_failed_approve: .truncate(true) wiped the attempts file
before reading, so failed pairing attempts never accumulated and
rate limiting never triggered.
- Guard wizard WASM test: skip gracefully when channel build artifacts
are absent (CI doesn't compile wasm32-wasip2 targets).
- Fix DNS rebinding check: use port 0 instead of hardcoded 443, since
the port is irrelevant for hostname resolution.
- Remove hardcoded CORS port 3001: the dynamic addr.port() entries
already cover the actual server port.
- Require WebSocket Origin header: reject connections that omit it
entirely, since browsers always send Origin for WS upgrades and a
missing header indicates a non-browser client bypassing the check.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* fix: address second round of PR review findings
- store.rs: reintroduce file locking around read-modify-write in
record_failed_approve (concurrent callers could clobber each other).
- sse.rs: replace load+check+fetch_add with atomic fetch_update in both
subscribe_raw() and subscribe() to prevent overshooting max_connections.
- ws.rs: decrement WS tracker before early return when subscribe_raw()
returns None (connection limit reached), fixing a counter leak.
- server.rs: parse WS Origin host exactly instead of prefix matching,
preventing bypass via crafted origins like http://localhost.evil.com.
- workspace_integration.rs: skip tests gracefully when Postgres is
unreachable instead of panicking (fixes 10 CI failures).
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* fix: add Origin header to WS integration tests
The Origin header requirement added in a3b0190 broke the WS gateway
integration tests. Test clients now send Origin: http://127.0.0.1:{port}
to match the server's localhost validation.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
---------
Co-authored-by: Claude Opus 4.6 <[email protected]>
Extends the WASM sandbox with HTTP API capabilities, secrets management,
tool aliasing, and leak detection. Key security principle: WASM never
sees credentials, injection happens at host boundary.
New modules:
- secrets: AES-256-GCM encrypted storage with HKDF key derivation
- leak_detector: Aho-Corasick + regex pattern matching for secret exfiltration
- capabilities: Extended capability system (HTTP, ToolInvoke, Secrets)
- allowlist: HTTP endpoint validation with glob patterns
- credential_injector: Host-boundary credential injection
- rate_limiter: Sliding window per-tool rate limiting
- storage: WASM binary storage with BLAKE3 integrity verification
Leak detection happens at two points:
1. Before HTTP request (prevents exfiltration via URL/headers/body)
2. After response (prevents exposure in outputs returned to WASM)
Co-Authored-By: Claude Opus 4.5 <[email protected]>
