7 Commits

Include renames

Author	SHA1	Message	Date
4 people	30790439ee	perf: build system prompt once per turn, skip tools on force-text (#583) ... * perf: build system prompt once per turn, skip tools on force-text, fix nudge role (#565) Three fixes to agentic loop prompt handling: 1. Build system prompt once per turn instead of every tool iteration. `build_system_prompt_with_tools` is now pub; callers pass the result via `ReasoningContext::system_prompt` to avoid rebuilding ~1,500 tokens per iteration. 2. Skip `## Available Tools` section when `force_text = true`. The dispatcher passes a no-tools prompt variant on the final iteration, saving ~460 tokens and removing misleading instructions. 3. Change nudge message from `Role::System` to `Role::User`. A second system message mid-conversation is unsupported by most providers. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * fix: revert nudge role change to keep ChatMessage::system Copilot review correctly identified that using Role::User for the nudge breaks compact_messages_for_retry, which uses rposition for Role::User to find the last real user message. Role::Assistant would cause back-to-back assistant messages. Since no production issues were reported with the original system role, revert to ChatMessage::system. [skip-regression-check] Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * fix: address PR review — omit tool guidance when tools empty, rename shadowed var - Conditionalize "Call tools…" guidelines and "## Tool Call Style" section in the system prompt so they are only included when tools are non-empty. Previously the force-text (no-tools) prompt still contained misleading tool-calling instructions. (Copilot review comment) - Rename `system_prompt` → `cached_prompt` in dispatcher to avoid shadowing the earlier workspace identity `system_prompt` variable. (Copilot review) - Add regression tests: `test_system_prompt_with_tools_contains_tool_guidance` and extended assertions in `test_system_prompt_without_tools_omits_tools_section`. Co-Authored-By: Claude Opus 4.6 <[email protected]> --------- Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]> Co-authored-by: [email protected] <[email protected]>	2026-03-07 09:15:00 +00:00
5 people	424a0366a9	feat: enable Anthropic prompt caching via automatic cache_control injection (#660) ... * feat(llm): add Anthropic prompt caching and cache token tracking - Inject cache_control via additional_params for Claude models in rig_adapter - Add cache_read_input_tokens and cache_creation_input_tokens to CompletionResponse and ToolCompletionResponse - Extract cached_input_tokens from rig-core unified Usage - Add is_anthropic_model() detection helper with provider prefix support - Log prompt cache hits at debug level (consistent with response_cache) - Add 7 unit tests for cache injection and model detection - Update all mock providers and test fixtures with new fields * feat(cost): apply 90% cache discount to prompt-cached tokens in CostGuard - Add cache_read_input_tokens to TokenUsage so cache counts flow from CompletionResponse through the reasoning layer to the dispatcher - Update CostGuard::record_llm_call() to accept cache_read_input_tokens: cached tokens are billed at 10% of the normal input rate - Thread cache_read_input_tokens from dispatcher into CostGuard - Add test_cache_discount_reduces_cost verifying exact savings match 90% of input cost for fully-cached requests - Update all existing test callers with zero-cache parameter * refactor(cache): scope cache_control to Anthropic backend and validate model support - Replace model-name-based is_anthropic_model() with explicit enable_prompt_cache flag on RigAdapter, set only for the direct Anthropic backend via with_prompt_cache(true) - Add supports_prompt_cache() to validate model names per Anthropic docs: only Claude 3+ models support caching; claude-2 and claude-instant are excluded to prevent 400 errors - Warn when caching is enabled but model does not support it - Replace is_anthropic_model tests with flag-based and model validation tests * fix(cache): validate model at construction and propagate cache metrics through proxy - Move supports_prompt_cache() check into with_prompt_cache() so unsupported models are detected once at construction, not per request - Add cache_read_input_tokens and cache_creation_input_tokens to ProxyCompletionResponse and ProxyToolCompletionResponse with serde(default) for backward compatibility - Pass cache metrics through orchestrator proxy instead of zeroing - Use claude-opus-4-6 in cache discount test to match Anthropic semantics * feat(llm): add configurable cache retention with write surcharge - Add CacheRetention enum (none/short/long) to AnthropicDirectConfig - Parse ANTHROPIC_CACHE_RETENTION env var (default: short) - Inject TTL-aware cache_control (short=5m ephemeral, long=1h) - Extract cache_creation_input_tokens from raw Anthropic response - Add cache_write_multiplier() to LlmProvider trait (1.25x short, 2.0x long) - Pipe dynamic write multiplier through dispatcher to CostGuard - Add TokenUsage.cache_creation_input_tokens field - Add tests for Long TTL injection, 5m and 1h write surcharges - Document ANTHROPIC_CACHE_RETENTION in .env.example * docs: fix stale cache_retention field comment * fix: resolve CI failures after upstream merge - Add missing cost_per_token arg to cache test callsites - Apply cargo fmt to long lines in tests and tracing macros * fix: address Copilot review feedback - Use saturating_add for cache token sum to prevent u32 overflow - Tighten supports_prompt_cache to explicitly match claude-3+/claude-4+ and named families (claude-sonnet/claude-opus/claude-haiku) * fix: adapt prompt caching to registry architecture and add missing cache fields - Resolve merge conflicts: adapt CacheRetention and cache injection to the declarative provider registry (RegistryProviderConfig replaces AnthropicDirectConfig) - Parse ANTHROPIC_CACHE_RETENTION env var in create_anthropic_from_registry() - Use Anthropic automatic caching via top-level cache_control in additional_params (rig-core #[serde(flatten)] places it at request root) - Add cache_read/creation_input_tokens fields to all mock LlmProviders added on main after PR #291 branched (response_cache, dispatcher, provider_chaos, trace_llm) - Suppress clippy::too_many_arguments on record_llm_call and build_rig_request - Add regression tests for cache injection (short/long/none) and cache_write_multiplier values Co-Authored-By: Canvinus <[email protected]> * fix: delegate cache_write_multiplier through provider wrappers and make cache_read_discount configurable The 6 decorator providers (Retry, CircuitBreaker, Failover, SmartRouting, CachedProvider, RecordingLlm) did not delegate cache_write_multiplier() to their inner provider, causing it to always return 1.0 instead of the actual 1.25x/2.0x from RigAdapter. This fix adds delegation for both cache_write_multiplier() and the new cache_read_discount() method. Also makes the cache read discount per-provider instead of hardcoding Anthropic's 90% discount (÷10). OpenAI uses 50% (÷2), so the discount is now returned by each provider via the LlmProvider trait. Addresses review feedback on PR #660. Co-Authored-By: Claude Opus 4.6 <[email protected]> * style: cargo fmt Co-Authored-By: Claude Opus 4.6 <[email protected]> * test: add CacheRetention FromStr/Display unit tests Tests cover primary values, aliases (off/disabled/5m/ephemeral/1h), case-insensitivity, invalid input error, and Display round-trip. Addresses Copilot review feedback on PR #660. Co-Authored-By: Claude Opus 4.6 <[email protected]> --------- Co-authored-by: Andrey <[email protected]> Co-authored-by: Andrey Gruzdev <[email protected]> Co-authored-by: Claude Opus 4.6 <[email protected]>	2026-03-07 09:10:05 +00:00
3 people	cf96a3253c	fix(tests): replace hardcoded /tmp paths with tempdir + add 300 unit tests (#659) ... * test: add unit tests across 20 modules for coverage push Add 300+ unit tests covering config, context, evaluation, extensions, LLM, secrets, tools/builder, and tools/mcp modules. All tests are pure unit tests (no mocks) exercising serde roundtrips, edge cases, error paths, and business logic. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix(tests): replace hardcoded /tmp paths with tempfile::tempdir The e2e_metrics_test::test_metrics_collected_from_tool_trace test was failing because setup_test_dir() created /tmp/ironclaw_metrics_test but the fixture referenced /tmp/ironclaw_e2e_test/hello.txt (path mismatch). Added LlmTrace::replace_paths() to substitute fixture paths at runtime, then converted all 12 test files from hardcoded /tmp/ironclaw_* paths to tempfile::tempdir(). Tests are now isolated, parallel-safe, and leave no debris on disk. Regression test: test_metrics_collected_from_tool_trace now passes consistently regardless of prior /tmp state. Co-Authored-By: Claude Opus 4.6 <[email protected]> --------- Co-authored-by: Claude Opus 4.6 <[email protected]>	2026-03-07 08:24:24 +00:00
3 people	ae89a52ac2	feat(routines): approval context for autonomous job execution (#577) ... * feat(routines): add approval context for autonomous job execution Routines and background jobs were unable to use any tools that required approval (file ops, shell, message, http), making them effectively useless. This adds an ApprovalContext system that lets autonomous jobs pre-authorize tools at dispatch time. - Add ApprovalContext enum with Autonomous variant that auto-approves UnlessAutoApproved tools and optionally pre-authorizes Always tools - Add tool_permissions field to RoutineAction::FullJob for pre-authorizing Always-gated tools (e.g. destructive shell, cross-channel messaging) - Add Scheduler::dispatch_job_with_context() to thread approval context through to workers - Set message tool default channel/target from routine NotifyConfig so routines can send results without cross-channel approval - Fix Completed→Completed state transition error in worker (plan marks job completed, then direct loop or outer run() tries again) Co-Authored-By: Claude Opus 4.6 <[email protected]> * test(routines): add E2E trace for routine news digest workflow Add a 3-turn trace fixture and test that exercises: - Turn 1: routine_create with full_job mode and tool_permissions - Turn 2: Simulated digest workflow with echo + memory_write - Turn 3: Verification via memory_search [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix(test): wire RoutineEngine into test rig for routine_create E2E - Add `with_routines()` to TestRigBuilder that passes a RoutineConfig to Agent::new, enabling routine tool registration during agent startup - Add Turn 2 (routine_list) to the trace to verify routine persistence in the database after routine_create - Fix formatting issues flagged by CI (cargo fmt) [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * refactor(scheduler): deduplicate dispatch_job and dispatch_job_with_context Extract shared logic into private `dispatch_job_inner` to prevent divergence when dispatch behavior changes in the future. [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(routines): add routine_fire tool and real E2E routine execution test - Add `routine_fire` tool that calls `RoutineEngine::fire_manual` to trigger a routine on demand. Registered alongside the other 5 routine tools (now 6 total). - Rewrite the routine_news_digest E2E trace to exercise the full execution stack end-to-end: 1. routine_create (manual trigger, full_job, tool_permissions: [message]) 2. routine_fire → RoutineEngine → Scheduler::dispatch_job_with_context → autonomous Worker consuming TraceLlm steps 3. Worker calls echo → memory_write → message (broadcast to test channel) 4. Test verifies the message broadcast arrived, proving ApprovalContext correctly allowed the Always-approval message tool - Register message tools in TestRig so routines can send messages to the test channel via channel_manager.broadcast(). [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(routines): wire HttpInterceptor through scheduler for routine worker http calls Propagate http_interceptor from AgentDeps → Scheduler → WorkerDeps → JobContext so that routine workers (and any scheduler-dispatched workers) can use the ReplayingHttpInterceptor for mock HTTP responses during tests. Changes: - Add http_interceptor field to Scheduler and WorkerDeps - Set job_ctx.http_interceptor in Worker before tool execution - Add with_http_exchanges() builder method to TestRigBuilder - Replace echo tool with http tool in routine_news_digest trace - Test now exercises real http tool with mock response → memory_write → message Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: address review comments from Copilot on PR #577 - Extract `ApprovalContext::is_blocked_or_default()` helper to deduplicate approval check logic in worker.rs and scheduler.rs - Extract `parse_tool_permissions()` helper to deduplicate JSON array parsing in routine.rs and builtin/routine.rs - Fix test name: `test_mark_completed_twice_does_not_error` → `test_mark_completed_twice_returns_error` (matches actual behavior) - Fix ApprovalContext doc comment to clarify it only models autonomous mode - Fix flaky index-based assertion in routine_news_digest test — now uses content-based search instead of fixed position - Fix stale comment: echo → http in routine test header - Add TODO for subtask approval context propagation (latent, not in active code paths) - Add TODO for global message tool context race in routine_engine Co-Authored-By: Claude Opus 4.6 <[email protected]> * style: fix formatting in is_blocked_or_default test Co-Authored-By: Claude Opus 4.6 <[email protected]> * refactor(test_rig): destructure self in build() to avoid partial-move fragility Destructure TestRigBuilder at the top of build() instead of accessing self.* fields after moving self.http_exchanges. While the prior code compiled (remaining fields are Copy), it was fragile and would break if any non-Copy field were added. Co-Authored-By: Claude Opus 4.6 <[email protected]> * docs: clarify that routine_fire bypasses cooldown Manual fires are explicitly user-initiated and intentionally bypass cooldown checks (which only apply to automated cron/event triggers). Updated tool description and fire_manual docstring to make this clear. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix(routines): fix message tool approval in routine context Two fixes for message tool failures in autonomous routine jobs: 1. MessageTool::requires_approval() now returns UnlessAutoApproved when the explicit channel param matches the default channel (was Always, causing "requires authentication" errors for routine workers). 2. routine_create tool now accepts notify_channel and notify_user params, wired into NotifyConfig. Without these, routines had channel: None, so set_message_tool_context was never called, causing "No channel specified" errors. Co-Authored-By: Claude Opus 4.6 <[email protected]> * refactor(message): remove approval requirement from message tool The message tool only sends to user-owned channels via ChannelManager::broadcast (TUI, Telegram, Slack, web gateway, etc.). It cannot reach arbitrary external services, so approval adds friction with no security benefit. This also eliminates the routine context errors entirely since approval is never checked. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: address review comments — routine_fire approval + test rename - routine_fire now returns UnlessAutoApproved since firing a routine can dispatch a full_job with pre-authorized Always-gated tools - Rename test_approval_context_never_always_passes to test_approval_context_never_is_not_blocked for clarity Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: address review nits — update stale docs and comments - Remove 'message' from tool_permissions example (no longer Always) - Reword message tool approval comment for accuracy - Clarify with_routines() docstring re: tool registration vs engine wiring Co-Authored-By: Claude Opus 4.6 <[email protected]> --------- Co-authored-by: Claude Opus 4.6 <[email protected]>	2026-03-07 05:21:58 +00:00
3 people	37bba72397	test: add 29 E2E trace tests for issues #571-575 (#593) ... * test: add 29 E2E trace tests for worker, threading, tools, workspace, and routines (#571-575) Add comprehensive E2E test coverage across five test files: - e2e_worker_coverage (7 tests): parallel tool calls, error feedback, unknown tools, invalid params, rate limiting, iteration limits, planning mode - e2e_thread_scheduling (3 tests + 2 deferred): multi-turn state, undo/redo, concurrent dispatch - e2e_builtin_tool_coverage (8 tests): time parse/diff/invalid, routine CRUD/history, job create/status/list/cancel, HTTP replay - e2e_workspace_coverage (6 tests): chunked search, multi-doc search, hybrid search, directory tree, document lifecycle, identity in system prompt - e2e_routine_heartbeat (5 tests): cron triggers, event matching, cooldown enforcement, heartbeat findings, empty checklist skip Infrastructure: extend TestRig with database/workspace/trace_llm accessors, register job and routine tools by default, add with_extra_tools() for custom stub tools. Includes 24 JSON trace fixtures across worker/, threading/, tools/, and workspace/. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: use 6-field cron format in routine_create_list fixture The cron 0.13 crate accepts both 6 and 7 fields, but the routine_create tool documents 6-field format. Align the fixture to match. [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: eliminate vacuous passes and silently-skipped assertions in E2E tests - job_create_status: replace job_status (needs dynamic UUID) with list_jobs, assert both succeed via completed() not just started() - job_list_cancel: keep cancel_job but explicitly assert it fails with invalid canned job_id "latest", verify create_job + list_jobs succeed - unknown_tool_name: add !is_empty() guard before .all() to prevent vacuous pass on empty iterator - workspace tests: change `if let Some(ws)` to `.expect()` so assertions are never silently skipped when workspace/trace_llm is available [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat: add template substitution to TraceLlm for dynamic tool result forwarding Add {{call_id.json_path}} template syntax to trace fixtures, enabling tool results from one step to flow into subsequent steps' arguments. TraceLlm extracts variables from Role::Tool messages (stripping the safety layer's <tool_output> XML wrapper and unescaping entities) and substitutes them in canned tool_call arguments before returning. This fixes job_create_status and job_list_cancel tests to properly test job_status and cancel_job with real dynamic UUIDs from create_job, instead of using invalid canned IDs that silently failed. Also adds tool result content assertions to job_create_status to verify the actual tool output contains expected data (job_id, title). [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: address PR review feedback on E2E tests - undo_redo_cycle: assert exactly 3 turns instead of >= 2 - tool_error_feedback: use tempfile::tempdir() instead of hardcoded /tmp path, patch fixture path at runtime for CI portability - worker_timeout → iteration_limit: rename to accurately describe what's tested - post_plan_work_remaining → simple_echo_flow: rename, test doesn't exercise planning - identity_in_system_prompt: seed IDENTITY.md before test, assert system prompt contains the seeded content instead of just checking Role::System exists [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: strengthen workspace E2E test assertions per PR review - write_chunk_search: assert memory_search was called and returned payment/architecture-related results - multi_document_search: assert memory_search was called for cross-document search - hybrid_search_with_embeddings: assert both memory_write and memory_search were called to confirm write-then-search pipeline - directory_tree: assert tree output contains expected alpha/beta project paths [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> --------- Co-authored-by: Claude Opus 4.6 <[email protected]>	2026-03-06 08:12:56 +00:00
3 people	470de5bd2d	feat: merge http/web_fetch tools, add tool output stash for large responses (#578) ... * feat: merge http/web_fetch tools, add tool output stash for large responses Merge `web_fetch` into `http` tool with smart approval: plain GETs (no headers, no body) run without approval and follow redirects with SSRF re-validation per hop; all other requests require approval as before. Add `tool_output_stash` on JobContext so full tool outputs are preserved before safety-layer truncation. The `json` tool gains a `source_tool_call_id` parameter to reference stashed outputs, enabling reliable parsing of large API responses that exceed the 100KB context limit. Other improvements: - Descriptive User-Agent header using CARGO_PKG_VERSION - Truncation now keeps partial data + hint about source_tool_call_id - System prompt reinforces tool_calls over narration - json tool query/stringify handle pre-parsed (non-string) data [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * chore: delete dead web_fetch.rs (merged into http tool) Co-Authored-By: Claude Opus 4.6 <[email protected]> * style: fix rustfmt formatting Co-Authored-By: Claude Opus 4.6 <[email protected]> * style: rename shadowed data binding for clarity in json tool Address PR review: rename owned `data` to `data_value` before re-binding as `let data = &data_value` to make ownership explicit. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix(ci): mark network-dependent trace tests as #[ignore] The weather_sf and baseball_stats tests hit live external APIs (wttr.in, ESPN) which are unreliable in CI. Mark them #[ignore] so they don't block the pipeline. Run locally with `--ignored` to include them. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: replay recorded HTTP exchanges in trace tests instead of hitting live APIs Wire ReplayingHttpInterceptor into TestRig when the trace fixture contains http_exchanges. This replays recorded responses instead of making live network calls, making tests deterministic and CI-stable. Add captured HTTP responses to weather_sf.json (wttr.in) and baseball_stats.json (ESPN API) fixtures. Revert #[ignore] on both tests — they now run offline. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: recover inline bracket-format tool calls from LLM text responses When flatten_tool_messages converts tool calls to text like `[Called tool `http` with arguments: {...}]` for NEAR AI compatibility, the LLM sometimes echoes this format back in its text responses instead of using proper tool_calls. Add recovery for this bracket format in recover_tool_calls_from_content and strip it in clean_response so users don't see raw tool call syntax. Co-Authored-By: Claude Opus 4.6 <[email protected]> --------- Co-authored-by: Claude Opus 4.6 <[email protected]>	2026-03-06 00:49:10 +00:00
4 people	b4b19738a8	Trajectory benchmarks and e2e trace test rig (#553) ... * refactor: extract shared assertion helpers to support/assertions.rs Move 5 assertion helpers from e2e_spot_checks.rs to a shared module. Add assert_all_tools_succeeded and assert_tool_succeeded for eliminating false positives in E2E tests. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat: add tool output capture via tool_results() accessor Extract (name, preview) from ToolResult status events in TestChannel and TestRig, enabling content assertions on tool outputs. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: correct tool parameters in 3 broken trace fixtures - tool_time.json: add missing "operation": "now" for time tool - robust_correct_tool.json: same fix - memory_full_cycle.json: change "path" to "target" for memory_write Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: add tool success and output assertions to eliminate false positives Every E2E test that exercises tools now calls assert_all_tools_succeeded. Added tool output content assertions where tool results are predictable (time year, read_file content, memory_read content). Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat: capture per-tool timing from ToolStarted/ToolCompleted events Record Instant on ToolStarted and compute elapsed duration on ToolCompleted, wiring real timing data into collect_metrics() instead of hardcoded zeros. Co-Authored-By: Claude Opus 4.6 <[email protected]> * refactor: add RAII CleanupGuard for temp file/dir cleanup in tests Replace manual cleanup_test_dir() calls and inline remove_file() with Drop-based CleanupGuard that ensures cleanup even if a test panics. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: add Drop impl and graceful shutdown for TestRig Wrap agent_handle in Option so Drop can abort leaked tasks. Signal the channel shutdown before aborting for future cooperative shutdown. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: replace agent startup sleep with oneshot ready signal Use a oneshot channel fired in Channel::start() instead of a fixed 100ms sleep, eliminating the race condition on slow systems. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: replace fragile string-matching iteration limit with count-based detection Use tool completion count vs max_tool_iterations instead of scanning status messages for "iteration"/"limit" substrings. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: use assert_all_tools_succeeded for memory_full_cycle test Remove incorrect comment about memory_tree failing with empty path (it actually succeeds). Omit empty path from fixture and use the standard assert_all_tools_succeeded instead of per-tool assertions. Co-Authored-By: Claude Opus 4.6 <[email protected]> * refactor: promote benchmark metrics types to library code Move TraceMetrics, ScenarioResult, RunResult, MetricDelta, and compare_runs() from tests/support/metrics.rs to src/benchmark/metrics.rs. Existing tests use re-export for backward compatibility. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat: add Scenario and Criterion types for agent benchmarking Scenario defines a task with input, success criteria, and resource limits. Criterion is an enum of programmatic checks (tool_used, response_contains, etc.) evaluated without LLM judgment. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat: add initial benchmark scenario suite (12 scenarios across 5 categories) Scenarios cover tool_selection, tool_chaining, error_recovery, efficiency, and memory_operations. All loaded from JSON with deserialization validation test. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat: add benchmark runner with BenchChannel and InstrumentedLlm BenchChannel is a minimal Channel implementation for benchmarks. InstrumentedLlm wraps any LlmProvider to capture per-call metrics. Runner creates a fresh agent per scenario, evaluates success criteria, and produces RunResult with timing, token, and cost metrics. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat: add baseline management, reports, and benchmark entry point - baseline.rs: load/save/promote benchmark results - report.rs: format comparison reports with regression detection - benchmark_runner.rs: integration test with real LLM (feature-gated) - Add benchmark feature flag to Cargo.toml Co-Authored-By: Claude Opus 4.6 <[email protected]> * style: apply cargo fmt to benchmark module Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): add multi-turn scenario types with setup, judge, ResponseNotContains Add BenchScenario, Turn, TurnAssertions, JudgeConfig, ScenarioSetup, WorkspaceSetup, SeedDocument types for multi-turn benchmark scenarios. Add ResponseNotContains criterion variant. Add TurnAssertions::to_criteria() converter for backward compat with existing evaluation engine. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): add JSON scenario loader with recursive discovery and tag filter Add load_bench_scenarios() for the new BenchScenario format with recursive directory traversal and tag-based filtering. Create 4 initial trajectory scenarios across tool-selection, multi-turn, and efficiency categories. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): multi-turn runner with workspace seeding and per-turn metrics Add run_bench_scenario() that loops over BenchScenario turns, seeds workspace documents, collects per-turn metrics (tokens, tool calls, wall time), and evaluates per-turn assertions. Add TurnMetrics to metrics.rs and clear_for_next_turn() to BenchChannel. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): add LLM-as-judge scoring with prompt formatting and score parsing Create judge.rs with format_judge_prompt, parse_judge_score, and judge_turn. Wire into run_bench_scenario for turns with judge config -- scores below min_score fail the turn. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): add CLI subcommand (ironclaw benchmark) Add BenchmarkCommand with --tags, --scenario, --no-judge, --timeout, --update-baseline flags. Wire into Command enum and main.rs dispatch. Feature-gated behind benchmark flag. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): per-scenario JSON output with full trajectory Add save_scenario_results() that writes per-scenario JSON files alongside the run summary. Each scenario gets its own file with turn_metrics trajectory. Update CLI to use new output format. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): add ToolRegistry::retain_only and wire tool filtering in scenarios Add a retain_only() method to ToolRegistry that filters tools down to a given allowlist. Wire this into run_bench_scenario() so that when a scenario specifies a tools list in its setup, only those tools are available during the benchmark run. Includes two tests for the new method: one verifying filtering works and one verifying empty input is a no-op. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): wire identity overrides into workspace before agent start Add seed_identity() helper that writes identity files (IDENTITY.md, USER.md, etc.) into the workspace before the agent starts, so that workspace.system_prompt() picks them up. Wire it into run_bench_scenario() after workspace seeding. Include a test that verifies identity files are written and readable. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): add --parallel and --max-cost CLI flags Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix(benchmark): use feature-conditional snapshot names for CLI help tests Prevents snapshot conflicts between default (no benchmark) and all-features (with benchmark) builds by using separate snapshot names per feature set. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): parallel execution with JoinSet and budget cap enforcement Replace sequential loop in run_all_bench() with parallel execution using JoinSet + semaphore when config.parallel > 1. Add budget cap enforcement that skips remaining scenarios when max_total_cost_usd is exceeded. Track skipped count in RunResult.skipped_scenarios and display it in format_report(). Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): add tool restriction and identity override test scenarios Co-Authored-By: Claude Opus 4.6 <[email protected]> * chore: fix formatting for Phase 3 Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): add SkillRegistry::retain_only and wire skill filtering in scenarios Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(benchmark): add --json flag for machine-readable output Co-Authored-By: Claude Opus 4.6 <[email protected]> * ci: add GitHub Actions benchmark workflow (manual trigger) Co-Authored-By: Claude Opus 4.6 <[email protected]> * refactor(benchmark): remove in-tree benchmark harness, keep retain_only utilities Move benchmark-specific code out of ironclaw in preparation for the nearai/benchmarks trajectory adapter. This removes: - src/benchmark/ (runner, scenarios, metrics, judge, report, etc.) - src/cli/benchmark.rs and the Benchmark CLI subcommand - benchmarks/ data directory (scenarios + trajectories) - .github/workflows/benchmark.yml - The "benchmark" Cargo feature flag What remains: - ToolRegistry::retain_only() and SkillRegistry::retain_only() - Test support types (TraceMetrics, InstrumentedLlm) inlined into tests/support/ instead of re-exporting from the deleted module Co-Authored-By: Claude Opus 4.6 <[email protected]> * docs: add README for LLM trace fixture format Documents the trajectory JSON format, response types, request hints, directory structure, and how to write new traces. Co-Authored-By: Claude Opus 4.6 <[email protected]> * feat(test): unify trace format around turns, add multi-turn support Introduce TraceTurn type that groups user_input with LLM response steps, making traces self-contained conversation trajectories. Add run_trace() to TestRig for automatic multi-turn replay. Backward-compatible: flat "steps" JSON is deserialized as a single turn transparently. Includes all trace fixtures (spot, coverage, advanced), plan docs, and new e2e tests for steering, error recovery, long chains, memory, and prompt injection resilience. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix(test): fix CI failures after merging main - Fix tool_json fixture: use "data" parameter (not "input") to match JsonTool schema - Fix status_events test: remove assertion for "time" tool that isn't in the fixture (only "echo" calls are used) - Allow dead_code in test support metrics/instrumented_llm modules (utilities for future benchmark tests) [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * Working on recording traces and testing them * feat(test): add declarative expects to trace fixtures, split infra tests Add TraceExpects struct with 9 optional assertion fields (response_contains, tools_used, all_tools_succeeded, etc.) that can be declared in fixture JSON instead of hand-written Rust. Add verify_expects() and run_recorded_trace() so recorded trace tests become one-liners. Split trace infra tests (deserialization, backward compat) into tests/trace_format.rs which doesn't require the libsql feature gate. Co-Authored-By: Claude Opus 4.6 <[email protected]> * refactor(test): add expects to all trace fixtures, simplify e2e tests Add declarative expects blocks to all 19 trace fixture JSONs across spot/, coverage/, advanced/, and root directories. Update all 8 e2e test files to use verify_trace_expects() / run_and_verify_trace(), replacing ~270 lines of hand-written assertions with fixture-driven verification. Tests that check things beyond expects (file content on disk, metrics, event ordering) keep those extra assertions alongside the declarative ones. Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix(test): adapt tests to AppBuilder refactor, fix formatting Update test files to work with refactored TestRigBuilder that uses AppBuilder::build_all() (removing with_tools/with_workspace methods). Update telegram_check fixture to use tool_list instead of echo. Fix cargo fmt issues in src/llm/mod.rs and src/llm/recording.rs. Co-Authored-By: Claude Opus 4.6 <[email protected]> * refactor(test): deduplicate support unit tests into single binary Support modules (assertions, cleanup, test_channel, test_rig, trace_llm) had #[cfg(test)] mod tests blocks that were compiled and run 12 times — once per e2e test binary that declares `mod support;`. Extracted all 29 support unit tests into a dedicated `tests/support_unit_tests.rs` so they run exactly once. [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * style: fix trailing newlines in support files Co-Authored-By: Claude Opus 4.6 <[email protected]> * refactor(test): unify trace types and fix recorded multi-turn replay Import shared types (TraceStep, TraceResponse, TraceToolCall, RequestHint, ExpectedToolResult, MemorySnapshotEntry, HttpExchange*) from ironclaw::llm::recording instead of redefining them in trace_llm.rs. Fix the flat-steps deserializer to split at UserInput boundaries into multiple turns, instead of filtering them out and wrapping everything into a single turn. This enables recorded multi-turn traces to be replayed as proper multi-turn conversations via run_trace(). [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix(test): fix CI failures - unused imports and missing struct fields - Add #[allow(unused_imports)] on pub use re-exports in trace_llm.rs (types are re-exported for downstream test files, not used locally) - Add `..` to ToolCompleted pattern in test_channel.rs to match new `error` and `parameters` fields Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix(test): fix CI failures after merging main - Add missing `error` and `parameters` fields to ToolCompleted constructors in support_unit_tests.rs - Add `..` to ToolCompleted pattern match in support_unit_tests.rs - Add #[allow(dead_code)] to CleanupGuard, LlmTrace impl, and TraceLlm impl (only used behind #[cfg(feature = "libsql")]) Co-Authored-By: Claude Opus 4.6 <[email protected]> * Adding coverage running script * fix(test): address review feedback on E2E test infrastructure - Increase wait_for_responses polling to exponential backoff (50ms-500ms) and raise default timeout from 15s to 30s to reduce CI flakiness (#1) - Strengthen prompt_injection_resilience test with positive safety layer assertion via has_safety_warnings(), enable injection_check (#2) - Add assert_tool_order() helper and tools_order field in TraceExpects for verifying tool execution ordering in multi-step traces (#3) - Document TraceLlm sequential-call assumption for concurrency (#6) - Clean up CleanupGuard with PathKind enum instead of shotgun remove_file + remove_dir_all on every path (#8) - Fix coverage.sh: default to --lib only, fix multi-filter syntax, add COV_ALL_TARGETS option - Add coverage/ to .gitignore - Remove planning docs from PR [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: address PR review - use HashSet in retain_only, improve skill test - Use HashSet for O(N+M) lookup in SkillRegistry::retain_only and ToolRegistry::retain_only instead of linear scan - Strengthen test_retain_only_empty_is_noop in SkillRegistry to pre-populate with a skill before asserting the no-op behavior [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix(test): revert incorrect safety layer assertion in injection test The safety layer sanitizes tool output, not user input. The injection test sends a malicious user message with no tools called, so the safety layer never fires. Reverted to the original test which correctly validates the LLM refuses via trace expects. Also fixed case-sensitive request hint ("ignore" -> "Ignore") to suppress noisy warning. [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix: clean stale profdata before coverage run Adds `cargo llvm-cov clean` before each run to prevent "mismatched data" warnings from stale instrumentation profiles. [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> * style: fix formatting in retain_only test [skip-regression-check] Co-Authored-By: Claude Opus 4.6 <[email protected]> --------- Co-authored-by: Claude Opus 4.6 <[email protected]> Co-authored-by: Illia Polosukhin <[email protected]>	2026-03-05 09:13:09 +00:00