docs(F153): Phase I spec — Step Summary (Agent Loop count) - #2
Merged
Conversation
Define a first-class "step length" view for the Hub observability plane. - Step = one agent loop, boundary anchor = cat_cafe.llm_call span - Length × Width model (depth + avg tools per loop) - Reuse existing telemetry; add 1 new counter (a2a.dispatch.count) - Live vs Restored visual distinction (— marker for unrecoverable subcounts) - AC-I1..AC-I7 - KD-31 (step = agent loop), KD-32 (descriptive only, no quality score) Naming aligns with industry: LangChain AgentExecutor, Anthropic SDK agent loop, AutoGPT step, OpenAI run step, DeepEval StepEfficiencyMetric all use agent loop / step as the basic unit — not LLM call, not tool call. Refs zts212653#721 [宪宪/Opus-47🐾] Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
7 tasks
P1-1 (Finding 1): Change agent loop anchor from cat_cafe.llm_call span to assistant message done event. llm_call span only created when durationApiMs available (invoke-single-cat.ts:1408-1410); Codex/Gemini/ Kimi don't produce it — would yield 0/missing loops for those providers. done event path is provider-agnostic. P1-2 (Finding 2): Rename aggregate counter to cat_cafe.a2a.dispatch.count and remove invocationId from its attributes. metric-allowlist.ts:8-9 forbids high-cardinality attributes on metrics. Per-route a2a_dispatch_count now derives from cat_cafe.mention_dispatch span count instead. P2-1 (Finding 3): tool_call_count goes dual-track — child cat_cafe.tool_use spans (MCP/business) + invocationSpan tool.basic_call_count attribute (basic tools, per span-helpers.ts:81-96 design to avoid trace tree flooding). P2-2 (Finding 4): token_total sources from cat_cafe.route span ROUTE_TOTAL_TOKENS attribute (route-serial.ts:1900), not the token.usage metric instrument (no route dimension on metrics). Add KD-33/34/35 capturing the data-source corrections. [宪宪/Opus-47🐾] Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
P1 (Finding 1 v2): Loop boundary is NOT observable from done event. done event yields once per CLI invocation (ClaudeAgentService.ts:476, CodexAgentService.ts:709), and cat_cafe.llm_call span also goes through the done path (msg.metadata.usage is consumed only in msg.type === 'done' branch, invoke-single-cat.ts:1308/1387) — both are per-invocation, not per-loop. Spec now requires per-provider stream parser to identify LLM call boundaries and emit a unified cat_cafe.agent_loop marker; providers without an identifiable boundary signal show '—' (NOT degraded to invocation count). P2 (Finding 2): cat_cafe.a2a.dispatch.count bounded labels reduced to AGENT_ID only. CALLBACK_TOOL/CALLBACK_REASON are callback auth failure semantics, unrelated to dispatch; if dispatch.source/status labels are needed during implementation, the metric-allowlist must be extended first. Update KD-31 (anchor = stream marker, not done event), KD-33 (marker strategy), KD-34 (label set). Reflect new marker/counter in section header. [宪宪/Opus-47🐾] Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
P3 (Maine Coon review 3rd round): AC-I7 explicitly requires test coverage for the "live provider without cat_cafe.agent_loop marker shows '—', not invocation count" case — Phase I's most critical non-degradation boundary. [宪宪/Opus-47🐾] Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Merged
5 tasks
bouillipx
added a commit
to zts212653/clowder-ai
that referenced
this pull request
May 20, 2026
F153 Phase I implementation. Closes #721. Surfaces agent_loop_count as the primary "step length" per cat_cafe.route in the Hub Traces tab, with Length × Width visualization. Descriptive only (per KD-16/KD-32, no efficiency or quality scoring). - Register cat_cafe.a2a.dispatch.count counter (AGENT_ID-only attribute per KD-34, metric-allowlist compliant) wired at both dispatch sites (route-serial in-process + dispatch-span callback) with lazy creation so phantom dispatches under depth-limit / dedup / ping-pong / fallback don't pollute counts. - recordAgentLoop(invocationSpan) helper using WeakMap + setAttribute pattern (analogous to tool.basic_call_count). Provider-agnostic, does NOT depend on durationApiMs or done event (both per-invocation, not per-loop — KD-33). - cat_cafe.agent_loop stream marker emitted by Claude provider at message_stop event (the true per-LLM-call boundary in Claude NDJSON stream). Other providers (Codex/Gemini/Kimi/Antigravity/OpenCode/DARE/ A2A) don't emit yet — Step Summary shows '—' per AC-I2/I7 non-degradation rule. - agent_loop is telemetry-only at invoke-single-cat: returned via return outputs (empty array), never pushed to outputs/transcript/ downstream forwarding. - GET /api/telemetry/step-summary route — session-authed, reads full buffer via traceStore.stats().maxSpans, returns descriptive metrics + Length × Width derivation + is_restored flag. - computeStepSummary aggregation: dual-track tool count (MCP child spans + tool.basic_call_count attribute per KD-35), span-derived a2a_dispatch_count (KD-34), ROUTE_TOTAL_TOKENS token total (KD-35), null-propagation for restored / no-marker cases. - HubTraceTree.tsx StepSummaryPanel inside expanded TraceCard: 6 metrics grid + Length × Width line + "Restored (history)" badge + '—' rendering for null/missing/restored sub-counts (never '0'). - 21 source-level + behavioral tests: * Counter wiring (instruments + dispatch-span + route-serial) * recordAgentLoop helper + Claude parser message_stop emit * AgentMessageType extension + invoke-single-cat telemetry-only branch * step-summary route registration + lazy worklist regression * Behavioral: live aggregation, restored returns null, missing marker returns null (NOT 1 — non-degradation), no normative fields, error_count by status code, 600-span long-trace non-truncation. This PR was developed and reviewed on clowder-labs/clowder-ai (internal fork) before being submitted upstream: - clowder-labs#2: spec PR (3-round Maine Coon review) - clowder-labs#3: implementation PR (3 review rounds — P1×2 + P2-1 + P2-2 all addressed; 砚砚 final approval "放行延续到 0de3046. No findings.") - clowder-labs#4: governance split (dir-exceptions extension) — owner gpt52 + ADR-010 reviewer 砚砚 both approved. - KD-31: step = one agent loop, anchored on cat_cafe.agent_loop stream marker - KD-32: Phase I descriptive only, no efficiency/quality score - KD-33: done event / llm_call span are per-invocation; marker required - KD-34: metric counter only AGENT_ID; route dim via span count - KD-35: tool dual-track (MCP span + basic counter attr); token via ROUTE_TOTAL_TOKENS - Other providers' agent_loop marker emit (Codex/Gemini/Kimi/etc) — will follow in subsequent phases; non-emitting providers display '—'. - Task-level step length — needs cross-invocation task_id primitive. - Step efficiency / quality eval — descriptive-plane boundary (KD-16). [宪宪/Opus-47🐾] Co-Authored-By: Maine Coon Codex (砚砚, GPT-5.5) <noreply@anthropic.com> Co-Authored-By: Maine Coon GPT-5.4 (gpt52) <noreply@anthropic.com> Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
zts212653
pushed a commit
to zts212653/clowder-ai
that referenced
this pull request
May 22, 2026
* feat(F153): Phase I — Step Summary (Agent Loop count) — community PR F153 Phase I implementation. Closes #721. Surfaces agent_loop_count as the primary "step length" per cat_cafe.route in the Hub Traces tab, with Length × Width visualization. Descriptive only (per KD-16/KD-32, no efficiency or quality scoring). - Register cat_cafe.a2a.dispatch.count counter (AGENT_ID-only attribute per KD-34, metric-allowlist compliant) wired at both dispatch sites (route-serial in-process + dispatch-span callback) with lazy creation so phantom dispatches under depth-limit / dedup / ping-pong / fallback don't pollute counts. - recordAgentLoop(invocationSpan) helper using WeakMap + setAttribute pattern (analogous to tool.basic_call_count). Provider-agnostic, does NOT depend on durationApiMs or done event (both per-invocation, not per-loop — KD-33). - cat_cafe.agent_loop stream marker emitted by Claude provider at message_stop event (the true per-LLM-call boundary in Claude NDJSON stream). Other providers (Codex/Gemini/Kimi/Antigravity/OpenCode/DARE/ A2A) don't emit yet — Step Summary shows '—' per AC-I2/I7 non-degradation rule. - agent_loop is telemetry-only at invoke-single-cat: returned via return outputs (empty array), never pushed to outputs/transcript/ downstream forwarding. - GET /api/telemetry/step-summary route — session-authed, reads full buffer via traceStore.stats().maxSpans, returns descriptive metrics + Length × Width derivation + is_restored flag. - computeStepSummary aggregation: dual-track tool count (MCP child spans + tool.basic_call_count attribute per KD-35), span-derived a2a_dispatch_count (KD-34), ROUTE_TOTAL_TOKENS token total (KD-35), null-propagation for restored / no-marker cases. - HubTraceTree.tsx StepSummaryPanel inside expanded TraceCard: 6 metrics grid + Length × Width line + "Restored (history)" badge + '—' rendering for null/missing/restored sub-counts (never '0'). - 21 source-level + behavioral tests: * Counter wiring (instruments + dispatch-span + route-serial) * recordAgentLoop helper + Claude parser message_stop emit * AgentMessageType extension + invoke-single-cat telemetry-only branch * step-summary route registration + lazy worklist regression * Behavioral: live aggregation, restored returns null, missing marker returns null (NOT 1 — non-degradation), no normative fields, error_count by status code, 600-span long-trace non-truncation. This PR was developed and reviewed on clowder-labs/clowder-ai (internal fork) before being submitted upstream: - clowder-labs#2: spec PR (3-round Maine Coon review) - clowder-labs#3: implementation PR (3 review rounds — P1×2 + P2-1 + P2-2 all addressed; 砚砚 final approval "放行延续到 0de3046. No findings.") - clowder-labs#4: governance split (dir-exceptions extension) — owner gpt52 + ADR-010 reviewer 砚砚 both approved. - KD-31: step = one agent loop, anchored on cat_cafe.agent_loop stream marker - KD-32: Phase I descriptive only, no efficiency/quality score - KD-33: done event / llm_call span are per-invocation; marker required - KD-34: metric counter only AGENT_ID; route dim via span count - KD-35: tool dual-track (MCP span + basic counter attr); token via ROUTE_TOTAL_TOKENS - Other providers' agent_loop marker emit (Codex/Gemini/Kimi/etc) — will follow in subsequent phases; non-emitting providers display '—'. - Task-level step length — needs cross-invocation task_id primitive. - Step efficiency / quality eval — descriptive-plane boundary (KD-16). [宪宪/Opus-47🐾] Co-Authored-By: Maine Coon Codex (砚砚, GPT-5.5) <noreply@anthropic.com> Co-Authored-By: Maine Coon GPT-5.4 (gpt52) <noreply@anthropic.com> Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * fix(F153 Phase I): cover F185 deferred A2A path with mention_dispatch span + counter + trace ctx 砚砚 R1 P2: route-serial 的 F185 fairness-gate deferred branch 只调 deferA2AEnqueue 入队， 跳过了 inline path 的 mention_dispatch span 创建 / a2aDispatchCount.add / caller trace context 注入，造成两个观测断点： - Step Summary 的 a2a_dispatch_count 在 fairness gate 触发时低报 - 通过 fairness gate 进入 InvocationQueue 的 child route 失去跨 route causality 修复： - route-helpers.ts: deferA2AEnqueue entry type 增加 callerTraceContext 字段 - route-serial.ts: deferred path 镜像 inline path :1661-1675 — lazy 创建 mention_dispatch span（dispatch.source='text-scan-deferred' 区分 inline/callback），立刻 end span （child route 在另一个 process loop 跑），把 span ctx 通过 entry 传到 QueueProcessor， 由后者作为 caller trace context 重建 remote parent 新增 5 个 regression tests 覆盖： - source-string × 4：span 创建、counter+1、entry callerTraceContext 字段、entry type 契约 - behavioral × 1：完整 dispatch → queue handoff → 重建 → 重新 parenting 链路 [宪宪/Opus-47🐾] Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * style(F153): biome format mention-dispatch-trace test for CI Lint Per 砚砚 R2: biome collapses assert.ok(...) and otelTracer.startSpan(...) to single-line. Tests still 37/37 pass. [宪宪/Opus-47🐾] Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * fix(F153): scope Step Summary to per-route subtree + mixed-provider coverage Maintainer inbound review (#732 comment): 1. Route vs trace scope — counts were trace-wide while duration/tokens came from one route span. Now computeStepSummary scopes to a single cat_cafe.route subtree (collectSubtree BFS on parentSpanId). API accepts optional routeSpanId; UI auto-detects root route. 2. Mixed-provider coverage — new agent_loop_partial flag (true when some live invocations have agent_loop.count and others don't). UI renders count as "N+" to indicate lower bound. 3. Issue gate — added triaged + feature:F153 labels to #721. 4. F027 anchor normalization — fixed F27 → F027 in test comment. Tests: +4 behavioral (multi-route subtree, auto-detect root, mixed partial, full coverage), updated frontend source test. [宪宪/Opus-46🐾] Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(F153): root-route auto-detect checks trace ancestry, not route-only Maintainer re-review P1: child route parented under mention_dispatch was misidentified as root (parent not a route ≠ parent absent from trace). Fix: auto-detect now picks the route span whose parentSpanId is absent from the entire trace span set, with earliest startTimeMs tiebreaker. UI no longer passes routeSpanId — relies on corrected backend. Regression test updated to real A2A shape (route2.parentSpanId = dispatch, not route1) + newest-first ordering case proving auto-detect stability. [宪宪/Opus-46🐾] Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Maine Coon Codex (砚砚, GPT-5.5) <noreply@anthropic.com>
bouillipx
pushed a commit
that referenced
this pull request
May 25, 2026
…ed wiring (zts212653#695) * fix(embedding): batch splitting + reprobe + collection embed deps Bug #1: Extract embedIndexedItems from IndexBuilder into shared embed-utils.ts with 64-item batch splitting. IndexBuilder's private version sent all items at once — embed-api MAX_BATCH_SIZE=64 caused >64 items to fail with HTTP 400, silently swallowed by fail-open. Bug #2: Add reprobeIfNeeded() to EmbeddingService. load() probes /health once at startup; if embed-api starts later, isReady() stays false forever. reprobeIfNeeded() re-probes when not ready. Bug #6: Wire embedDeps into CollectionIndexBuilder and library.ts rebuild route. Collection rebuild now produces vector embeddings (was FTS-only). Also pass force param from request body (was ignored). Upstream: zts212653#693 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> [宪宪/Opus-46🐾] * fix(embedding): wire production call sites for reprobeIfNeeded and per-collection VectorStore P1 fixes from gpt55 review: - embed-utils.ts: call reprobeIfNeeded() before isReady() check so all callers (IndexBuilder + CollectionIndexBuilder) get late-start recovery - library.ts: accept embeddingService instead of static embedDeps; dynamically load sqlite-vec and create per-collection VectorStore at rebuild time (prevents cross-DB vector pollution) - index.ts: pass memoryServices.embeddingService to libraryRoutes Tests: add reprobe recovery test, update all embedding mocks with reprobeIfNeeded method. 68/68 embedding tests pass. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(embedding): wire setEmbedDeps on collection store for search-time hybrid P1 from gpt55 R2: rebuild writes vectors via CollectionIndexBuilder but the collection's SqliteEvidenceStore never gets setEmbedDeps, so search() stays lexical-only. Now the rebuild route calls store.setEmbedDeps() after constructing the per-collection VectorStore, using the system embedMode. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: tianyiliang <tianyiliang@tianyiliangdeMacBook-Pro.local> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Lysander Su <773678591@qq.com>
buproof
pushed a commit
to buproof/clowder-ai
that referenced
this pull request
Jun 29, 2026
* feat(F153): Phase I — Step Summary (Agent Loop count) — community PR F153 Phase I implementation. Closes zts212653#721. Surfaces agent_loop_count as the primary "step length" per cat_cafe.route in the Hub Traces tab, with Length × Width visualization. Descriptive only (per KD-16/KD-32, no efficiency or quality scoring). - Register cat_cafe.a2a.dispatch.count counter (AGENT_ID-only attribute per KD-34, metric-allowlist compliant) wired at both dispatch sites (route-serial in-process + dispatch-span callback) with lazy creation so phantom dispatches under depth-limit / dedup / ping-pong / fallback don't pollute counts. - recordAgentLoop(invocationSpan) helper using WeakMap + setAttribute pattern (analogous to tool.basic_call_count). Provider-agnostic, does NOT depend on durationApiMs or done event (both per-invocation, not per-loop — KD-33). - cat_cafe.agent_loop stream marker emitted by Claude provider at message_stop event (the true per-LLM-call boundary in Claude NDJSON stream). Other providers (Codex/Gemini/Kimi/Antigravity/OpenCode/DARE/ A2A) don't emit yet — Step Summary shows '—' per AC-I2/I7 non-degradation rule. - agent_loop is telemetry-only at invoke-single-cat: returned via return outputs (empty array), never pushed to outputs/transcript/ downstream forwarding. - GET /api/telemetry/step-summary route — session-authed, reads full buffer via traceStore.stats().maxSpans, returns descriptive metrics + Length × Width derivation + is_restored flag. - computeStepSummary aggregation: dual-track tool count (MCP child spans + tool.basic_call_count attribute per KD-35), span-derived a2a_dispatch_count (KD-34), ROUTE_TOTAL_TOKENS token total (KD-35), null-propagation for restored / no-marker cases. - HubTraceTree.tsx StepSummaryPanel inside expanded TraceCard: 6 metrics grid + Length × Width line + "Restored (history)" badge + '—' rendering for null/missing/restored sub-counts (never '0'). - 21 source-level + behavioral tests: * Counter wiring (instruments + dispatch-span + route-serial) * recordAgentLoop helper + Claude parser message_stop emit * AgentMessageType extension + invoke-single-cat telemetry-only branch * step-summary route registration + lazy worklist regression * Behavioral: live aggregation, restored returns null, missing marker returns null (NOT 1 — non-degradation), no normative fields, error_count by status code, 600-span long-trace non-truncation. This PR was developed and reviewed on clowder-labs/clowder-ai (internal fork) before being submitted upstream: - clowder-labs#2: spec PR (3-round Maine Coon review) - clowder-labs#3: implementation PR (3 review rounds — P1×2 + P2-1 + P2-2 all addressed; 砚砚 final approval "放行延续到 0de3046. No findings.") - clowder-labs#4: governance split (dir-exceptions extension) — owner gpt52 + ADR-010 reviewer 砚砚 both approved. - KD-31: step = one agent loop, anchored on cat_cafe.agent_loop stream marker - KD-32: Phase I descriptive only, no efficiency/quality score - KD-33: done event / llm_call span are per-invocation; marker required - KD-34: metric counter only AGENT_ID; route dim via span count - KD-35: tool dual-track (MCP span + basic counter attr); token via ROUTE_TOTAL_TOKENS - Other providers' agent_loop marker emit (Codex/Gemini/Kimi/etc) — will follow in subsequent phases; non-emitting providers display '—'. - Task-level step length — needs cross-invocation task_id primitive. - Step efficiency / quality eval — descriptive-plane boundary (KD-16). [宪宪/Opus-47🐾] Co-Authored-By: Maine Coon Codex (砚砚, GPT-5.5) <noreply@anthropic.com> Co-Authored-By: Maine Coon GPT-5.4 (gpt52) <noreply@anthropic.com> Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * fix(F153 Phase I): cover F185 deferred A2A path with mention_dispatch span + counter + trace ctx 砚砚 R1 P2: route-serial 的 F185 fairness-gate deferred branch 只调 deferA2AEnqueue 入队， 跳过了 inline path 的 mention_dispatch span 创建 / a2aDispatchCount.add / caller trace context 注入，造成两个观测断点： - Step Summary 的 a2a_dispatch_count 在 fairness gate 触发时低报 - 通过 fairness gate 进入 InvocationQueue 的 child route 失去跨 route causality 修复： - route-helpers.ts: deferA2AEnqueue entry type 增加 callerTraceContext 字段 - route-serial.ts: deferred path 镜像 inline path :1661-1675 — lazy 创建 mention_dispatch span（dispatch.source='text-scan-deferred' 区分 inline/callback），立刻 end span （child route 在另一个 process loop 跑），把 span ctx 通过 entry 传到 QueueProcessor， 由后者作为 caller trace context 重建 remote parent 新增 5 个 regression tests 覆盖： - source-string × 4：span 创建、counter+1、entry callerTraceContext 字段、entry type 契约 - behavioral × 1：完整 dispatch → queue handoff → 重建 → 重新 parenting 链路 [宪宪/Opus-47🐾] Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * style(F153): biome format mention-dispatch-trace test for CI Lint Per 砚砚 R2: biome collapses assert.ok(...) and otelTracer.startSpan(...) to single-line. Tests still 37/37 pass. [宪宪/Opus-47🐾] Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * fix(F153): scope Step Summary to per-route subtree + mixed-provider coverage Maintainer inbound review (zts212653#732 comment): 1. Route vs trace scope — counts were trace-wide while duration/tokens came from one route span. Now computeStepSummary scopes to a single cat_cafe.route subtree (collectSubtree BFS on parentSpanId). API accepts optional routeSpanId; UI auto-detects root route. 2. Mixed-provider coverage — new agent_loop_partial flag (true when some live invocations have agent_loop.count and others don't). UI renders count as "N+" to indicate lower bound. 3. Issue gate — added triaged + feature:F153 labels to zts212653#721. 4. F027 anchor normalization — fixed F27 → F027 in test comment. Tests: +4 behavioral (multi-route subtree, auto-detect root, mixed partial, full coverage), updated frontend source test. [宪宪/Opus-46🐾] Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(F153): root-route auto-detect checks trace ancestry, not route-only Maintainer re-review P1: child route parented under mention_dispatch was misidentified as root (parent not a route ≠ parent absent from trace). Fix: auto-detect now picks the route span whose parentSpanId is absent from the entire trace span set, with earliest startTimeMs tiebreaker. UI no longer passes routeSpanId — relies on corrected backend. Regression test updated to real A2A shape (route2.parentSpanId = dispatch, not route1) + newest-first ordering case proving auto-detect stability. [宪宪/Opus-46🐾] Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Maine Coon Codex (砚砚, GPT-5.5) <noreply@anthropic.com>
bouillipx
pushed a commit
that referenced
this pull request
Jul 7, 2026
…ce (zts212653#713) * fix(zts212653#712): unify MCP config to capabilities.json single source of truth Problem: MCP server configuration was scattered across multiple provider config files (.mcp.json, config.toml, settings.json, mcp.json). Each provider had its own writer, leading to inconsistencies when toggling MCP servers — changes in capabilities.json were overwritten by stale provider configs on restart. Solution: capabilities.json is now the single source of truth for all MCP configuration. Provider config files are generated from it (not the reverse). Key changes: - Remove bidirectional provider writers; capabilities.json → provider configs is one-way generation - MCP discovery runs once (gated by discoveryVersion), not on every GET - External config files (.mcp.json etc.) are legacy artifacts, no longer authoritative - Extract shared MCP service for plugin activation path - Unify per-cat state to blockedCats (blacklist), remove legacy overrides field - Frontend MCP toggle derives state from raw data (cats/globalEnabled) matching Skills pattern, not from backend computed enabled field - Clean up all legacy `?? cap.enabled` fallbacks — globalEnabled is the sole runtime source of truth - Multi-project MCP sync management (F249): cascade global MCP state to registered projects, drift detection, per-project blockedCats Closes zts212653#712 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(test): update f041 tests to use globalEnabled + blockedCats Tests were using legacy `enabled` + `overrides` fields that `isMcpEnabledForCat` no longer reads after the zts212653#712 cleanup. At runtime, `readCapabilitiesConfig` fills `globalEnabled` from `enabled` via in-memory migration, so raw pre-migration data never reaches `resolveServersForCat` — but tests construct configs directly. - Add `globalEnabled` alongside `enabled` in config round-trip and hot-reload tests - Rewrite per-cat override tests to use `blockedCats` instead of deprecated `overrides` field - Update test names to reflect canonical model semantics Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(test): add globalEnabled to capability-orchestrator tests Two more tests in capability-orchestrator.test.js used `enabled: false` without `globalEnabled`, causing isMcpEnabledForCat and isSameRepoExternalSplit to treat them as enabled (globalEnabled ?? true). - ensureCatCafeMainServer: disabled cat-cafe-limb now has globalEnabled: false - resolveServersForCat: disabled MCP entry now has globalEnabled: false Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(codex): include dummy command in disabled MCP --config entries Codex CLI ≥0.142 validates transport for ALL mcp_servers entries, including disabled ones. A bare `enabled=false` override (no command, no type) fails with "invalid transport in mcp_servers.<name>". Add `command="echo"` + `args=["disabled"]` to disabled entries, matching the existing legacy `cat-cafe` shim pattern that already works. This gives Codex enough info to infer stdio transport without actually starting the server. Root cause: buildCatCafeMcpArgs injected `--config mcp_servers.<name>.enabled=false` for globally-disabled capabilities (e.g. cat-cafe-memory with globalEnabled:false), but omitted command/args. Codex couldn't determine transport → validation error. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(codex): skip disabled MCP entries instead of injecting enabled=false Disabled capabilities don't need CLI --config overrides — L5 writeCodexMcpConfig already deletes disabled managed entries from .codex/config.toml. Injecting a bare `enabled=false` (no command, no type) caused Codex CLI ≥0.142 "invalid transport" validation errors. The previous commit added dummy command/args as a workaround; this commit removes the override entirely, which is simpler and correct: no stale config.toml entry to override → no override needed. The legacy `cat-cafe` shim remains as the only disabled override — user-level ~/.codex/config.toml may have old entries that L5 cannot reach. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(zts212653#712): restore legacy enabled fallback in isMcpEnabledForCat Invoke-time code paths (Codex, Claude, Kimi, OpenCode, ACP) read capabilities.json with raw readFileSync, bypassing the in-memory migration in readCapabilitiesConfig that copies enabled→globalEnabled. When a capability has enabled:false but no globalEnabled field, isMcpEnabledForCat(cap.globalEnabled ?? true) returned true — wrong. Add ?? cap.enabled fallback so legacy configs without globalEnabled are still respected. globalEnabled takes precedence when present. Includes regression test covering: - legacy enabled:false (no globalEnabled) → disabled - legacy enabled:true (no globalEnabled) → enabled - globalEnabled:false overrides legacy enabled:true Review finding: P1 #4 from maintainer review on PR zts212653#713. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * chore: add tips_exempt to F249 feature doc Design-phase spec — capability tips will be added when the feature reaches implementation. Fixes pnpm check gate failure. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(security): harden probe endpoint with local+owner gates (zts212653#712 review) P1: POST /api/mcp/:id/tools accepted ad-hoc command/args/url from request body with only resolveUserId() — a trusted browser origin without session got default-user, enabling arbitrary command execution (stdio probe) and SSRF with env-var exfiltration (HTTP probe headers). Fix: always require requireLocalCapabilityWriteRequest (blocks remote access since probe spawns child processes); ad-hoc probes (body.command or body.url) additionally require resolveOwnerGate. P2: invalid projectPath silently fell back to STARTUP_REPO_ROOT instead of returning 400 (inconsistent with drift-check route at line 123-126). Fix: return 400 on validateProjectPath failure. Failure-mode audit: both findings are "sibling route access control inconsistency" — drift-resolve/sync-all use requireMcpWriteAccess (session+local+owner) but probe only used resolveUserId. Now aligned. Review: 缅因猫 GPT-5.4 @gpt52, PR zts212653#713. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(zts212653#712): address maintainer deep review P1 findings P1 #9: drift-resolve resolutions array now validated — each entry must have string mcpId and decision fields. Prevents malicious injection. P1 #10: probeHttpMcp now validates URL scheme — only http: and https: allowed. Prevents SSRF via file:// / gopher:// etc. P2 #18: Fixed 7 remaining F240→F249 references in MCP test comments (capabilities-route.test.js + capabilities-mcp-write-route.test.js). Connector test F240 references left unchanged (correct feature ID). Review: maintainer deep review on PR zts212653#713. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(zts212653#712): review R2 quick fixes — atomic write, safe access, globalEnabled, enum validation P1 #2: writeCapabilitiesConfig now uses temp file + rename for atomic writes P1 #7: installMcpCapability replaces non-null assertion with safe index access P1 #10: drift-resolve resolutions validation adds enum check + array length cap P2 #13: removeMcpCapability sets globalEnabled=false alongside legacy enabled Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * test(zts212653#712): add behavior tests for drift/sync subsystem (review R2 P1#12 blocker) 23 tests covering mcp-drift-detector, mcp-drift-resolver, mcp-sync-engine: - Drift detector: global-new/project-orphan/config-mismatch detection - Drift resolver: add/remove/update/skip with use-global and keep-project - Sync engine: canonicalJson, hash determinism, extractMcpEntries - Sync project: add new, update changed, skip overrides, remove orphans Drift detector tests use inline config params (no file I/O). Resolver and sync engine tests use real temp dirs with capabilities.json. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(zts212653#712): update drift-resolver test for atomic write compatibility The existing test made capabilities.json read-only to block writes, but atomic write (temp file + rename) bypasses file-level chmod — rename() needs only directory write permission on POSIX. Now makes the .cat-cafe/ directory non-writable instead, correctly blocking temp file creation. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(zts212653#712): centralize drift resolution validation around resolver contract P1: Route validation enum was accept/reject/skip but resolver contract is use-global/keep-project — completely mismatched. Now both routes (drift.ts and mcp-drift.ts) validate against VALID_MCP_DRIFT_DECISIONS exported from the resolver module, ensuring route-level and resolver-level consistency. P2: Rename stale F240 anchor to F249 in acp-mcp-resolver test. Added 3 contract tests verifying VALID_MCP_DRIFT_DECISIONS enum coverage. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(zts212653#712): close validation gap in /api/drift/resolve + route-level tests P1: /api/drift/resolve silently accepted non-array resolutions (object, string, number) by only checking Array.isArray — non-array values skipped validation and fell through to default use-global behavior. Fix: - Moved validation before drift check (fail-early on malformed input) - Added non-array rejection, MAX_RESOLUTIONS cap, enum validation - Both routes (drift.ts + mcp-drift.ts) now have identical validation Route-level tests (8 cases via Fastify inject): - Valid use-global/keep-project accepted - Non-array resolutions → 400 - String resolutions → 400 - Invalid decision values → 400 with enum hint - Missing mcpId → 400 - Over-limit array → 400 - Absent resolutions passes through Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
agent_loop_count(=cat_cafe.llm_callspan count per route) as the primary step metric.a2a.dispatch.countcounter onmention_dispatchspan creation.—for restored spans where sub-step hierarchy was flattened byhydrate-traces.ts.Multi-cat alignment thread 2026-05-19: 铲屎官 + 砚砚(缅因猫) + 宪宪(布偶猫). Refs zts212653#721.
Why this naming
Industry convention picks the agent-behavior unit, not the implementation unit:
AgentExecutor, Anthropic SDK agent loop, AutoGPT step, OpenAI Assistants run step, DeepEvalStepEfficiencyMetric— all use agent loop / step.Boundary: one agent loop = one
cat_cafe.llm_call+ 0..N tool_use + tool result observations. A pure reply (width=0) still counts as one step — the decision point exists.Test plan
mention_dispatchspan path / Phase G integration assumption (AC-I3a2a.dispatch.countplacement)mcp_tool_call_countvsbasic_tool_call_countsplit (currently out-of-scope, deferred to Phase H真实执行边界)[宪宪/Opus-47🐾]