Sync ctx 0f61777

GitHub commit: 0f617779a5a1f11f8714c8f61087eb80c1c492b4

.github/workflows/test.yml

@@ -816,13 +816,11 @@ jobs:
           lfs: false
 
       - name: Enforce test-coverage-per-PR policy
-        # Policy: any product or CI/package contract change
-        # must be accompanied by at least one test file change in the
-        # same PR. Exemptions:
-        #   - Pure docs / comment changes (matched by a trivial heuristic)
-        #   - Config JSON (src/config.json) — schema changes get tests
-        #     for the accessor, not the JSON itself.
-        #   - Changes ONLY inside __pycache__, .pyc files, etc.
+        # Policy: product or CI/package contract changes must include
+        # at least one src/tests/ change in the same PR. Exemptions are
+        # intentionally narrow and implemented in scripts/ci_no_test_policy.py:
+        # release metadata-only changes, or a maintainer-applied
+        # no-tests-needed label for changes where tests are genuinely not useful.
         run: |
           set -euo pipefail
           BASE="${{ github.event.pull_request.base.sha }}"

README.md

@@ -18,7 +18,7 @@ tags:
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
 [![Python 3.11+](https://img.shields.io/badge/Python-3.11+-green.svg)](https://python.org)
 [![PyPI](https://img.shields.io/pypi/v/claude-ctx.svg)](https://pypi.org/project/claude-ctx/)
-[![Tests](https://img.shields.io/badge/Tests-4584_inventory-brightgreen.svg)](https://github.com/stevesolun/ctx/actions/workflows/test.yml)
+[![Tests](https://img.shields.io/badge/Tests-4606_inventory-brightgreen.svg)](https://github.com/stevesolun/ctx/actions/workflows/test.yml)
 [![Graph](https://img.shields.io/badge/Graph-79%2C958_nodes_/_1%2C778%2C069_edges-red.svg)](https://stevesolun.github.io/ctx/knowledge-graph/)
 [![Skills](https://img.shields.io/badge/Skills-68%2C494-blue.svg)](https://stevesolun.github.io/ctx/catalog/?type=skill)
 [![Agents](https://img.shields.io/badge/Agents-467-purple.svg)](https://stevesolun.github.io/ctx/catalog/?type=agent)

docs/dashboard.md

@@ -225,6 +225,7 @@ per-process monitor token injected into the rendered page.
 | `/` | Home: seven stat cards (loaded, sidecars, wiki entities, graph nodes, runtime checks, audit events, sessions), grade distribution pills, recent sessions table, recent audit events |
 | `/loaded` | **Currently-loaded skills, agents, MCP servers, and installed harness records** from `~/.claude/skill-manifest.json` plus `~/.claude/harness-installs/*.json`; skill/agent/MCP rows expose supported live actions |
 | `/skills` | Every sidecar as a filterable **card grid**: left sidebar (search by slug, grade checkboxes, skill/agent/MCP toggle, hide-floored), card shows grade pill + raw score + links to sidecar/wiki/graph |
+| `/skillspector` | SkillSpector audit tab for ctx-run static scan results over skill bodies, with status, severity, tag, graph-family, query, and limit filters plus a link to `/api/skillspector.json` |
 | `/skill/<slug>` | Full sidecar breakdown: four-signal score (telemetry · intake · graph · routing), hard-floor reason, computed_at timestamp, per-skill audit timeline |
 | `/wiki` | **Wiki entity index** - bounded card-grid sample of up to 500 pages per dashboard-supported entity type, merging installed wiki-pack pages with local-only entity files not shadowed by a pack page or tombstone. Includes sharded MCP server pages and flat harness pages. Left sidebar: text search over the visible sample (slug, description, tag), skill/agent/MCP/harness checkboxes. |
 | `/wiki/<slug>?type=<entity>` | Dashboard-supported wiki entity page rendered from the merged pack/local source: markdown body + full frontmatter table + grade banner + deep links to sidecar and graph-neighborhood views. The optional `type` query disambiguates duplicate slugs such as `langgraph`; omitted types are inferred from frontmatter or path when possible. |
@@ -250,10 +251,13 @@ per-process monitor token injected into the rendered page.
 | `GET /api/sessions.json` | All sessions with aggregated counts |
 | `GET /api/manifest.json` | Raw `skill-manifest.json` passthrough |
 | `GET /api/status.json` | `{queue, artifacts, telemetry}` payload: durable queue counts/recent jobs, graph/wiki artifact file status and promotion metadata, plus telemetry spool/export health including malformed records and exporter errors |
+| `GET /api/grades.json` | Grade distribution used by the home page: `{grades: {A, B, C, D, F}, total}`, reusing the KPI summary when available and falling back to sidecar counts |
+| `GET /api/sidecars.json?q=<text>&type=<entity>&grade=A&hide_floor=1&limit=50` | Filtered sidecar page payload for `/skills`: matching items, totals, pagination bounds, active type/grade filters, and hide-floor state |
 | `GET /api/skill/<slug>.json` | Raw sidecar for one slug; missing or unsafe slugs return JSON `{detail: ...}` with HTTP 404 |
 | `GET /api/graph/<slug>.json?type=<entity>&hops=1&limit=40` | Dashboard-shaped skill/agent/MCP/harness `{nodes, edges, center}`; `type` is optional but recommended for duplicate slugs, `hops` is [1, 3], `limit` is [5, 150]. |
 | `GET /api/kpi.json` | `DashboardSummary` passthrough — `{total, by_subject, grade_counts, lifecycle_counts, category_breakdown, hard_floor_counts, low_quality_candidates, archived, generated_at}`. Returns `{total: 0, detail: "no sidecars yet"}` when the quality directory is empty |
 | `GET /api/runtime.json` | Runtime lifecycle summary: source path, validation count, failed/error count, open-escalation count, latest validation, recent validations, open escalations, session IDs, `tool_selection`, `token_usage`, `token_usage_history`, and `recent_tool_usage`. |
+| `GET /api/skillspector.json?q=<text>&status=<status>&severity=<severity>&tag=<tag>&family=<family>&limit=100` | SkillSpector audit payload for `/skillspector`: summary counts, available filters, filtered records, audit path, and audit availability |
 | `GET /api/config.json` | Effective/default/user config payload used by the Config tab. |
 | `GET /api/entities/search.json?q=<text>&type=<entity>&limit=80` | Wiki entity search results for Manage, Config, and entity picker flows, with wiki-relative `path` values. |
 | `GET /api/entity/<slug>.json?type=<entity>` | Frontmatter, wiki-relative `path`, and Markdown body for one wiki entity from the same merged pack/local source as `/wiki/<slug>`. |

docs/harness/attaching-to-hosts.md

@@ -38,6 +38,11 @@ Claude on the next turn, alongside runtime lifecycle tools such as
 `ctx__load_entity`, `ctx__mark_entity_used`, and `ctx__session_state`.
 Ask "What skills help with FastAPI auth?" and it will call them.
 
+For permissioned adapters that should expose only read/query tools, start the
+same server with `--allow-tools` and `--entity-types`. For example,
+`ctx-mcp-server --allow-tools ctx__recommend_bundle,ctx__wiki_search,ctx__wiki_get --entity-types skill,mcp-server`
+limits tool discovery and read results to the named tools and entity types.
+
 ### Claude Agent SDK (Python)
 
 ```python
@@ -288,8 +293,8 @@ The adapter emits a JSON contract with:
   engine;
 - `related_recommendations` after the loop passes selected and rejected
   recommendation IDs;
-- optional harness recommendations only when the loop declares a user-owned,
-  API, or local model.
+- optional harness recommendations only when the loop gives explicit
+  user-owned/API/local model consent with `--own-llm` or `own_llm=True`.
 
 For LoopFlow, keep the `.loop` file in charge and call ctx before the plan:
 
@@ -299,6 +304,10 @@ python -m ctx.adapters.loopflow \
   --permissions skills,agents,mcps
 ```
 
+`.loop` files can also grant ctx capabilities with
+`ctx grants: skills, mcps, harnesses`. CLI `--permissions` override those file
+grants; when neither is present, the adapter returns no capabilities.
+
 Add `--last-failure-file .loopflow/last-failure.txt` only after the loop has
 written that file; omit it on the first run. The adapter uses that failure text
 for recommendation ranking and returns only `context.last_failure_present`, not
@@ -367,11 +376,14 @@ plan_context = recommend_for_loop(
 )
 ```
 
-Load only the groups that are explicitly granted in `permissions`. If
-`harnesses` is granted without `--own-llm`, `--model-provider`, or `--model`,
-the adapter returns a warning and no harness recommendations. The ctx MCP
-command and tool list are also permission-filtered; ctx tools that can operate
-across all capability groups only appear when all of those groups are granted.
+Load only the groups that are explicitly granted in `permissions`.
+`--model-provider` and `--model` are ranking metadata, not ownership consent:
+if `harnesses` is granted without `--own-llm`, the adapter returns a warning
+and no harness recommendations. The ctx MCP command and tool list are also
+permission-filtered. A partial `mcps` grant exposes only read-only ctx tools and
+adds `mcp_server.args` with `--allow-tools` / `--entity-types`; lifecycle write
+tools appear only when `skills`, `agents`, `mcps`, and `harnesses` are all
+granted.
 
 ---
 

docs/harness/loopflow-adapter-demo.md

@@ -27,6 +27,7 @@ loop "select ctx capabilities":
   goal: mcp agent loop local ollama filesystem
   look at: the repo plan, AGENTS.md, and the last failure
   done when "pytest src/tests/test_loopflow_adapter.py -q" passes
+  ctx grants: skills, mcps, harnesses
   each cycle: plan, then act, then observe
   when it fails: reflect, then plan again
 ```
@@ -87,6 +88,7 @@ stable.
   "mcp_server": {
     "name": "ctx",
     "command": "ctx-mcp-server",
+    "args": [],
     "tools": [
       "ctx__recommend_bundle",
       "ctx__graph_query",
@@ -175,6 +177,7 @@ if ctx_payload["mcp_server"]["command"]:
     runner.register_mcp_server(
         name=ctx_payload["mcp_server"]["name"],
         command=ctx_payload["mcp_server"]["command"],
+        args=ctx_payload["mcp_server"]["args"],
     )
 
 if ctx_payload["loopflow"]["use_skills"]:
@@ -193,14 +196,28 @@ logic, and ctx refreshes recommendations based on what just failed.
 The adapter fails closed:
 
 - `--permissions skills` only returns skill recommendations.
-- `--permissions mcps` only returns MCP server recommendations unless all
-  capability groups are granted, which lets ctx MCP tools operate safely across
-  skills, agents, MCPs, and harnesses.
-- `--permissions harnesses` returns no harnesses unless `--own-llm`,
-  `--model-provider`, or `--model` is present.
+- `--permissions mcps` returns MCP server recommendations and exposes only
+  read-only ctx MCP tools: recommendation, graph query, and wiki lookup.
+- Lifecycle MCP tools such as load, mark-used, validation, escalation, unload,
+  and session tools are exposed only when all capability groups are granted:
+  `skills`, `agents`, `mcps`, and `harnesses`.
+- `--permissions harnesses` returns no harnesses unless `--own-llm` is present.
+  `--model-provider` and `--model` improve ranking and dry-run command metadata;
+  they are not user-owned model consent.
 - `agent_loop.harness_install` is always a `--dry-run` command. It shows what
   would be installed; it does not mutate the host.
 
+`.loop` files can request ctx grants with one small line:
+
+```loop
+ctx grants: skills, mcps, harnesses
+```
+
+CLI/API grants take precedence. If the command passes `--permissions`, those
+grants replace the `.loop` line. If the API caller passes `permissions=...`,
+that explicit set is authoritative. If neither is present, the `.loop` grants
+are used; if none are present anywhere, the adapter returns no capabilities.
+
 This lets a LoopFlow user decide whether a loop may use skills, agents, MCPs,
 or harnesses without giving ctx authority to bypass the loop's own gates.
 

docs/index.md

@@ -17,9 +17,10 @@ burn fewer tokens and local models waste less CPU/GPU work.
 Watches what you develop, walks a knowledge graph of **68,494 skill pages, 467 agents, 10,790 MCP servers, and 207 cataloged harnesses**, and recommends the
 right execution bundle on the fly. The live execution bundle is skills,
 agents, and MCP servers only; custom/API/local model users and external loop
-adapters get separate harness recommendations based on model choice and task
-goal. You decide what to load, install, or adopt. Powered by a Karpathy LLM wiki
-with persistent memory that gets smarter every session.
+adapters get separate harness recommendations after explicit user-owned model
+consent, ranked by model choice and task goal. You decide what to load,
+install, or adopt. Powered by a Karpathy LLM wiki with persistent memory that
+gets smarter every session.
 
 !!! tip "Install"
 
@@ -145,9 +146,9 @@ graph-based discovery:
   agents, and MCP servers in real time — **nothing loads or
   installs without your approval**.
 - During custom/API/local model onboarding and LoopFlow/agent-loop adapter
-  calls, `ctx-init`, `ctx-harness-install`, and
-  `python -m ctx.adapters.loopflow` use the same graph to recommend harnesses
-  above the configured harness match floor.
+  calls with explicit user-owned model consent, `ctx-init`,
+  `ctx-harness-install`, and `python -m ctx.adapters.loopflow` use the same
+  graph to recommend harnesses above the configured harness match floor.
 
 The result: you always know what skills, agents, and MCP servers are available
 for your current task, and which harness fits when you choose your own model.
@@ -235,7 +236,7 @@ ones are flagged. New ones self-ingest.
     ---
 
     Current main is **v1.0.21** — MIT, CI-matrixed (Ubuntu 3.12 plus Windows/macOS 3.11/3.12),
-    4,584 test inventory. Adds enterprise OpenTelemetry-ready telemetry and
+    4,606 test inventory. Adds enterprise OpenTelemetry-ready telemetry and
     ships console scripts including `ctx-init`,
     `ctx-monitor` (local dashboard with graph + wiki + load/unload for
     skills, agents, and MCP servers, plus Harness Setup for user-owned LLMs),

docs/knowledge-graph.md

@@ -215,9 +215,11 @@ The graph backs these recommendation paths:
   those records, the same recommender falls back to the index file.
 - Harness recommendations are a separate path for custom/API/local
   model onboarding (`ctx-init --model-mode custom ...`),
-  `ctx-harness-install`, and LoopFlow/agent-loop adapter calls that declare a
-  user-owned/API/local model. They use the same graph filtered to `harness`
-  nodes and the higher harness match floor from `config.json`.
+  `ctx-harness-install`, and LoopFlow/agent-loop adapter calls that pass
+  explicit user-owned/API/local model consent. Model/provider fields can
+  improve ranking, but the LoopFlow adapter does not treat them as consent by
+  themselves. Harness paths use the same graph filtered to `harness` nodes and
+  the higher harness match floor from `config.json`.
 - Repository scans still start from stack detections, then turn that profile
   into the same tag/query bundle used by the execution recommender. If a
   shipped graph is unavailable, scan output falls back to the legacy installed
@@ -273,6 +275,10 @@ Pack directories live inside the installed wiki:
   wiki-packs/overlay-<id>/
 ```
 
+Pack manifests require every checksum value to be a lowercase 64-character
+SHA-256 hex digest. Non-hex placeholders are rejected before merged graph/wiki
+readers can trust the pack contents.
+
 Overlay packs are the normal local-update path. Full rebuilds are still needed
 when you intentionally refresh the release artifact, change global scoring
 configuration, rebuild communities, or compact a long chain of overlays into a
@@ -434,6 +440,9 @@ JSON, graph-pack node/edge metadata, markdown members, and merged wiki-pack
 markdown must not contain host-user paths; markdown members are capped at
 10 MiB; local generated markdown such as `catalog.md`, `converted-index.md`,
 `log.md`, and `versions-catalog.md` must stay out of the full archive.
+When a shipped root artifact has sibling promotion metadata, validation also
+checks the `current` snapshot against the actual artifact size and SHA-256 so
+stale promotion records cannot describe a different file.
 
 ## Current artifact record
 
@@ -448,7 +457,7 @@ the wiki content and the checked-in graph build configuration.
 
 ## Pre-ship gates
 
-Two advisory gates run before the tarball is repackaged. Both produce
+Three advisory gates run before the tarball is repackaged. All three produce
 review reports and never auto-modify the inventory.
 
 - **`ctx-dedup-check`** — flags entity pairs (skill ↔ skill, skill ↔

docs/plans/ctx-agent-reviewer-workbench.md

@@ -0,0 +1,324 @@
+# ctx Agent-Reviewer Workbench
+
+This plan adapts the Claude Science actor-critic pattern to ctx review work.
+The source pattern is a coordinator agent that delegates to specialists, while
+separate reviewer agents check evidence, traceability, calculations, and
+artifact correctness:
+<https://www.anthropic.com/news/claude-science-ai-workbench>.
+
+This is an internal execution plan. It is intentionally kept under
+`docs/plans/`, which `mkdocs.yml` excludes from public docs.
+
+## Goal
+
+Review ctx with paired agents that behave like production review teams:
+
+1. An actor agent discovers, tests, or proposes the review artifact.
+2. A critic agent independently checks the artifact against code, tests, docs,
+   and user-visible behavior.
+3. The CTO orchestrator merges only evidence-backed findings into the canonical
+   trackers.
+4. Fixes land as small commits. No rushed PRs. No long no-mistakes runs until
+   the fix batch is ready.
+
+## Non-Goals
+
+- Do not create a PR during this review setup phase.
+- Do not run no-mistakes as the default review loop.
+- Do not replace `qa/bug_smoke_status.csv` or `qa/feature_status.csv`.
+- Do not let specialist notes become new canonical trackers.
+- Do not accept "tests passed" as evidence unless the behavior under review is
+  directly demonstrated.
+
+## CTX Review Workbench Team
+
+### CTO Orchestrator
+
+Owns scope, task graph, conflicts, and merge criteria.
+
+Inputs:
+
+- `qa/feature_status.csv`.
+- `qa/bug_smoke_status.csv`.
+- CI/no-mistakes evidence.
+- Repo tree.
+
+Output:
+
+- One canonical review board with owner, status, evidence path, and blocker
+  state, derived from the canonical CSVs. Do not create a third tracker file.
+
+### Feature Mapper + Feature Reviewer
+
+Mapper discovers every product capability: CLI, MCP server, dashboard,
+telemetry, graph/wiki, skills, agents, harnesses, and Loopflow adapter.
+
+Reviewer checks every discovered feature has a user story, expected behavior,
+test evidence, and tracker row.
+
+### Runtime/API/MCP Agent + Contract Reviewer
+
+Agent reviews `src/ctx/api.py`, `src/ctx/mcp_server`, generic adapters, tool
+router, and lifecycle flows.
+
+Reviewer checks API schemas, backwards compatibility, permissions, error
+behavior, and trace propagation.
+
+### Telemetry/Enterprise Agent + Privacy Reviewer
+
+Agent reviews OTel-style traces/metrics/logs, token usage KPIs, exporters,
+retention, and local defaults.
+
+Reviewer checks redaction, hashed error handling, salt boundaries, exporter
+config safety, and enterprise runbook gaps.
+
+### Dashboard/UX Agent + UX Critic
+
+Agent walks dashboard pages and user flows.
+
+Critic checks broken links, misleading states, graph/LLM-wiki/dashboard health,
+visual regressions, and empty/error states.
+
+### Graph/Wiki/Recommendation Agent + Evidence Reviewer
+
+Agent reviews graph artifacts, wiki packs, skill/agent/MCP recommendations, and
+subgroup selection behavior.
+
+Reviewer checks recommendations are explainable, reproducible,
+permission-aware, and backed by tests.
+
+### QA/Test Gate Agent + Gate Critic
+
+Agent runs local-fast, focused pytest, CI classifier, no-test policy,
+no-mistakes, and PR checks.
+
+Critic checks bottlenecks, flaky tests, overbroad gates, missing fixtures, and
+whether evidence is real or just "tests passed."
+
+### Security/Supply Chain Agent + Red-Team Reviewer
+
+Agent scans install scripts, sync/HF paths, tokens/secrets, LFS/artifacts, and
+external tool ingestion.
+
+Reviewer tries adversarial cases: malicious skill/MCP metadata, path traversal,
+unsafe exporters, and poisoned DSL/tool manifests.
+
+### Docs/Runbook Agent + Docs Reviewer
+
+Agent validates docs against code behavior.
+
+Reviewer checks every doc claim has command/test/source evidence and flags stale
+prose or unsupported marketing language.
+
+### Loopflow/DSL Integration Agent + DSL Author Reviewer
+
+Agent designs the Loopflow adapter surface: permissions, harnesses, skills,
+MCPs, agents, recommendations, and agent loops.
+
+Reviewer evaluates from the Loopflow owner's view: "Would I trust ctx as a
+recommendation/runtime layer without losing control of my DSL?"
+
+## Operating Loop
+
+Each pair produces:
+
+- finding
+- evidence
+- repro
+- risk
+- fix recommendation
+- reviewer verdict
+
+Then CTO merges them into:
+
+- canonical tracker updates
+- prioritized fix batches
+- local-fast/no-mistakes validation plan
+- PR-ready changelog
+
+## Operating Principles
+
+- Keep one coordinator. The CTO owns scope, conflicts, ordering, and commit
+  boundaries.
+- Keep agents paired. Every actor output gets an independent critic pass before
+  it becomes a finding.
+- Keep evidence portable. Findings must include file paths, commands, observed
+  output, screenshots when visual, or generated artifacts when behavior is not
+  visible in logs.
+- Keep trackers canonical. Feature coverage updates go to
+  `qa/feature_status.csv`; bug, smoke, UX, and garbage-code findings go to
+  `qa/bug_smoke_status.csv`.
+- Keep fixes separate. Review findings first, fix batches second.
+- Keep validation cheap until ready. Use focused tests and `local-fast`; reserve
+  no-mistakes and PR CI for the final integrated batch.
+
+## Canonical Tracker Hygiene
+
+Before any agent or CTO step writes tracker rows, validate the canonical CSVs:
+
+- Parse `qa/feature_status.csv` as CSV and require unique non-empty
+  `feature_id` values.
+- Parse `qa/bug_smoke_status.csv` as CSV and require unique non-empty
+  `finding_id` values.
+- Check proposed new row keys against both files before writing; collisions
+  must become `needs-cto-review`, not silent overwrites.
+- Check stale status/evidence markers before writing: historical phrases such
+  as `pending no-mistakes`, bare `TODO`, obsolete blocker text, or retest claims
+  without matching `last_verified_at` and evidence.
+- If a stale row is found, preserve the old text in review notes, add current
+  evidence, and update status only after the critic verifies the behavior.
+- Treat the "canonical review board" as a derived view over these CSVs with
+  owner, status, evidence path, and blocker state. Do not create a separate
+  board file unless the user explicitly approves it.
+
+## CTO Orchestrator
+
+The CTO orchestrator owns the shared review board and makes final calls.
+
+Inputs:
+
+- Current branch and diff.
+- `qa/feature_status.csv`.
+- `qa/bug_smoke_status.csv`.
+- `docs/qa/feature-user-story-status.csv`.
+- `docs/qa/dashboard-user-story-status.csv`.
+- Local-fast timing evidence in `.gate/local-fast.json` when available.
+- CI/no-mistakes evidence only when intentionally run.
+
+Outputs:
+
+- Updated canonical trackers.
+- A prioritized fix queue.
+- Small commit boundaries.
+- A final "ready for big PR" report.
+
+Decision rules:
+
+- If actor and critic disagree, the finding stays `needs-cto-review`.
+- If the critic cannot reproduce, the finding stays `needs-evidence`.
+- If a behavior needs product judgment, the finding is `human-decision`.
+- If a fix would touch more than 5 files, split it into phases before coding.
+
+## Evidence Contract
+
+Every pair returns this shape:
+
+```yaml
+pair: string
+surface: string
+status: pass | finding | blocked | human-decision
+finding:
+  summary: string
+  category: feature-gap | bug | ux | security | docs | gate | architecture | none
+evidence:
+  files: []
+  commands: []
+  artifacts: []
+  observed_output: []
+risk:
+  severity: critical | high | medium | low | info
+  blast_radius: string
+repro:
+  steps: []
+fix_recommendation:
+  scope: string
+  suggested_files: []
+  validation: []
+reviewer_verdict:
+  status: accepted | needs-evidence | duplicate | human-decision | invalid
+  notes: string
+tracker_update:
+  file: qa/feature_status.csv | qa/bug_smoke_status.csv | none
+  row_key: string
+```
+
+The CTO may reject any result without concrete evidence.
+
+## Detailed Pair Contracts
+
+The canonical team roster is `CTX Review Workbench Team` above. This section
+adds execution contracts only, so agents do not maintain a second divergent
+team list.
+
+| Canonical pair | First local review focus | Primary tracker target |
+| --- | --- | --- |
+| QA/Test Gate Agent + Gate Critic | Run `scripts/no_mistakes_run.sh fast`, focused pytest, CI classifier checks, no-test policy checks, and local gate timing summaries. Identify slow, flaky, redundant, or missing local gates. Separate local CPU bottlenecks from GitHub-hosted runner delays and no-mistakes agent overhead. | `qa/bug_smoke_status.csv` for gate findings |
+| Feature Mapper + Feature Reviewer | Discover user-visible features, map each to current tests/docs, and reject vague or duplicate rows. | `qa/feature_status.csv`; missing-story findings in `qa/bug_smoke_status.csv` |
+| Runtime/API/MCP Agent + Contract Reviewer | Exercise API and MCP boundaries from a user perspective; check schema stability, error payloads, backwards compatibility, permissions, and trace propagation. | `qa/bug_smoke_status.csv` for contract gaps |
+| Telemetry/Enterprise Agent + Privacy Reviewer | Verify traces, metrics, logs, events, token usage KPIs, exporters, retention, local defaults, redaction, hashed errors, salt boundaries, and exporter opt-in safety. | `qa/bug_smoke_status.csv` for telemetry/privacy gaps |
+| Dashboard/UX Agent + UX Critic | Walk dashboard pages and capture browser evidence for visual or interaction findings. Reject visual UX findings without screenshot or rendered artifact evidence. | `qa/bug_smoke_status.csv` for UX findings |
+| Graph/Wiki/Recommendation Agent + Evidence Reviewer | Verify graph artifacts, wiki packs, recommendations, subgroup selection, explainability, permission awareness, and LFS hydration expectations. | `qa/bug_smoke_status.csv` for graph/recommendation findings |
+| Security/Supply Chain Agent + Red-Team Reviewer | Review install scripts, Hugging Face sync, Git LFS artifacts, external imports, MCP ingestion, exporters, generated artifacts, and adversarial skill/MCP/DSL inputs. | `qa/bug_smoke_status.csv` for security findings |
+| Docs/Runbook Agent + Docs Reviewer | Compare docs to code behavior and reject unsupported claims, stale prose, or instructions without command/test/source evidence. | `qa/bug_smoke_status.csv` for docs drift |
+| Loopflow/DSL Integration Agent + DSL Author Reviewer | Check permissions, harnesses, skills, MCPs, agents, recommendations, and agent loops from the DSL author's perspective. | `qa/bug_smoke_status.csv` for adapter UX/permission findings |
+
+## First Wave
+
+Run these pairs first because local-fast is the review bottleneck reducer and
+tracker hygiene must be clean before broader feature mapping:
+
+1. QA/Test Gate Agent + Gate Critic.
+2. Feature Mapper + Feature Reviewer.
+3. Runtime/API/MCP Agent + Contract Reviewer.
+4. Telemetry/Enterprise Agent + Privacy Reviewer.
+5. Loopflow/DSL Integration Agent + DSL Author Reviewer.
+
+First-wave success criteria:
+
+- Local-fast lanes are run first and `.gate/local-fast.json` is inspected.
+- Canonical tracker IDs are checked for collisions before any row update.
+- Stale tracker status/evidence text is identified before any row update.
+- Every reviewed surface has a row or explicit deferral in canonical trackers.
+- Every finding has repro evidence.
+- Every fix candidate has a scoped validation command.
+- No PR is opened.
+- No no-mistakes run is started.
+- At least one commit records the review plan and tracker updates.
+
+## Subagent Prompt Template
+
+Use this template for each actor:
+
+```text
+You are the ACTOR for <pair name>. Work only in ctx. Inspect the named surface.
+Return findings using docs/plans/ctx-agent-reviewer-workbench.md Evidence
+Contract. Do not fix code. Do not create trackers. Use existing canonical
+trackers only. Include exact files, commands, and observed behavior.
+```
+
+Use this template for each critic:
+
+```text
+You are the CRITIC for <pair name>. Review the actor output. Verify against
+code, docs, tests, and artifacts. Reject weak evidence. Mark each item as
+accepted, needs-evidence, duplicate, human-decision, or invalid. Do not fix
+code. Do not create trackers.
+```
+
+Use this template for the CTO:
+
+```text
+Merge actor and critic outputs. Update only canonical trackers. Split accepted
+fixes into small commits. Keep no-mistakes and PRs out of scope until the user
+approves the final integrated validation phase.
+```
+
+## Commit Policy
+
+- Commit review infrastructure separately from bug fixes.
+- Commit tracker updates separately from code fixes.
+- Commit each fix batch by surface.
+- Do not open PRs until the user asks.
+- Do not run no-mistakes until the user asks or the final integrated batch is
+  ready.
+
+## Fast Validation
+
+Use these checks for review-plan-only commits:
+
+```bash
+python -m mkdocs build --strict
+scripts/no_mistakes_run.sh fast --lane cheap --lane docs
+```
+
+Use broader local validation only when tracker updates or code fixes justify it.

docs/qa/feature-user-story-status.csv

@@ -47,7 +47,7 @@ CLI-045,CLI,Backup mirror CLI,"python -m backup_mirror create, snapshot-if-chang
 CLI-046,CLI,Generic harness planner evaluator and contract modes,"ctx run --planner, ctx run --evaluator, ctx run --contract","src/ctx/cli/run.py, src/tests/test_harness_planner.py, src/tests/test_harness_evaluator.py, src/tests/test_harness_contract.py",As a model-agnostic harness user I can run planning evaluation and contract review modes,ctx.cli.run exposes --planner --evaluator and --contract modes while preserving session and runtime semantics,Provider fixture or local test provider available,"python -m pytest src/tests/test_harness_planner.py src/tests/test_harness_evaluator.py src/tests/test_harness_contract.py -q",Tested Pass,harness planner evaluator and contract tests cover CLI parser runtime dispatch and persisted outputs.,,,,,2026-06-29,Split from broad CLI-001 so deep ctx.cli.run modes are explicit.
 CLI-047,CLI,Resume MCP restoration safety,ctx resume --restore-session-mcp and credential_env,src/ctx/cli/run.py and src/tests/test_harness_cli_run.py,As a returning harness user I can resume without restoring executable MCP metadata unless I opt in,ctx resume skips recorded MCP command and env metadata by default restores credential_env only with explicit --restore-session-mcp and emits previous trace correlation when available,Recorded session fixture,python -m pytest src/tests/test_harness_cli_run.py -q,Retested Pass,PASS: test_harness_cli_run.py covers default skip restore opt-in credential_env behavior failed resume telemetry and previous_trace_id correlation.,AUDIT-018,Early ctx resume failures and resumed sessions lacked reliable previous trace correlation telemetry.,Fixed,PASS: .venv/bin/pytest src/tests/test_harness_cli_run.py -q --tb=short -> 73 passed.,2026-07-05,Tracks resume security and trace correlation semantics inside broad CLI-001.
 CLI-048,CLI,Harness update and uninstall controls,ctx-harness-install --update and ctx-harness-install --uninstall,src/harness_install.py and src/tests/test_harness_install.py,As a harness user I can dry-run update or uninstall an installed harness safely,Update and uninstall flows detect installed harness state show planned changes and avoid destructive surprises,Installed harness fixture,python -m pytest src/tests/test_harness_install.py -q,Tested Pass,test_harness_install.py covers update dry-run uninstall safety and missing-record behavior.,,,,,2026-06-29,Split from CLI-026 so install recommend update and uninstall semantics are separately tracked.
-CLI-049,CLI,LoopFlow and agent-loop adapter,python -m ctx.adapters.loopflow,src/ctx/adapters/loopflow.py and src/tests/test_loopflow_adapter.py,As a DSL runner or custom agent-loop author I can ask ctx for scoped planning capabilities before each loop plan,Adapter emits a JSON contract with permission-gated skills agents MCPs ctx MCP server metadata and gated harness recommendations only when a user-owned/API/local model is declared,Graph/wiki available or recommendation functions mocked,python -m pytest src/tests/test_loopflow_adapter.py -q,Tested Pass,test_loopflow_adapter.py covers loop-file parsing permission filtering MCP tool metadata harness gating shell-quoted harness install commands CLI JSON output and empty permission fail-closed behavior.,,,,,2026-06-30,Tracks the module entry point documented in docs/harness/attaching-to-hosts.md.
+CLI-049,CLI,LoopFlow and agent-loop adapter,python -m ctx.adapters.loopflow,src/ctx/adapters/loopflow.py and src/tests/test_loopflow_adapter.py,As a DSL runner or custom agent-loop author I can ask ctx for scoped planning capabilities before each loop plan,Adapter emits a JSON contract with permission-gated skills agents MCPs scoped ctx MCP server metadata and gated harness recommendations only with explicit user-owned model consent,Graph/wiki available or recommendation functions mocked,python -m pytest src/tests/test_loopflow_adapter.py -q,Tested Pass,test_loopflow_adapter.py covers loop-file parsing permission filtering scoped MCP tool metadata harness gating shell-quoted harness install commands CLI JSON output and empty permission fail-closed behavior.,,,,,2026-06-30,Tracks the module entry point documented in docs/harness/attaching-to-hosts.md.
 HOOK-001,Hooks,Backup PostToolUse runtime hook,hooks/backup_on_change.py,hooks/backup_on_change.py and src/tests/test_quality_hook.py,As a Claude Code user I can snapshot tracked config changes from a PostToolUse hook,The hook reads stdin JSON detects tracked files runs snapshot-if-changed and always exits 0 so edits are not blocked,Temp Claude home and tracked config fixture needed,python -m pytest src/tests/test_quality_hook.py -q,Tested Pass,test_quality_hook.py covers root backup hook stdin parsing tracked-file dispatch snapshot reason shaping and nonblocking exit even when snapshot returns nonzero.,,,,,2026-06-30,Validated in Phase 2 runtime hook behavior slice without increasing the test inventory count.
 HOOK-002,Hooks,Quality Stop runtime hook,hooks/quality_on_session_end.py,hooks/quality_on_session_end.py and src/tests/test_quality_hook.py,As a Claude Code user I can refresh touched skill quality on session stop without blocking shutdown,The hook reads event state recomputes bounded touched slugs advances state only after successful recompute writes audit session rows and always exits 0,Temp Claude home and skill event fixture needed,python -m pytest src/tests/test_quality_hook.py -q,Tested Pass,test_quality_hook.py covers touched slug selection state cutoff max slug cap missing file behavior and failed recompute exit zero without state advancement.,,,,,2026-07-05,Validated after checkpoint hardening for recompute failures.
 PYAPI-001,Python API,Public Python API,ctx.api and ctx top-level re-exports,src/ctx/api.py and src/ctx/__init__.py,As a harness author I can import ctx from Python without depending on internal modules,recommend_bundle recommend_related graph_query wiki_search wiki_get list_all_entities default_wiki_dir and CtxCoreToolbox keep stable signatures safe defaults and top-level re-exports,Runtime wiki or synthetic fixture available,pytest src/tests/test_public_api.py -q,Tested Pass,test_public_api.py covers __all__ exports top-level re-exports signatures happy paths empty paths and telemetry privacy payloads.,,,,,2026-07-02,Added during canonical feature inventory gap pass and updated when related recommendations became public API.
@@ -139,7 +139,7 @@ SEC-002,Distribution,No-mistakes development gate,no-mistakes axi run,imported-s
 MAINT-001,Maintainer,Backup snapshot audit,scripts/audit_backup.py,scripts/audit_backup.py,As a maintainer I can inspect backup snapshots for missing expected files,Script summarizes backup contents while accepting --help without snapshot access and flags missing graph/wiki/report artifacts without mutating the repo,Backup fixture available,python scripts/audit_backup.py --help; run script against representative backup fixture,Retested Pass,PASS: --help printed usage without touching snapshot files; temp backup fixture with BOM manifest printed entries and OK coverage; focused pytest passed.,CTX-AZ-008; CTX-QA-034,Backup audit previously rejected UTF-8 BOM manifests and treated --help as a snapshot path,Fixed by utf-8-sig manifest reads and snapshot-free help parsing in PR #210,PASS: /tmp/ctx-verify-venv/bin/python scripts/audit_backup.py --help; focused pytest 62 passed 1 skipped; no-mistakes run 01KWFXQTHJ83VNZ9ZJGMTAJPWC checks-passed.,2026-07-02,Read-only reporting; --help is snapshot-free and does not repair backup gaps.
 MAINT-002,Maintainer,Dashboard graph index builder,scripts/build_dashboard_graph_index.py,scripts/build_dashboard_graph_index.py,As a dashboard user I get fast graph search and neighborhoods,Script builds the dashboard SQLite/index artifacts from graph exports for fast monitor reads,Graph fixture available,Run script against small graph fixture and query resulting index,Retested Pass,FAIL: BOM-prefixed graph.json from PowerShell Set-Content -Encoding UTF8 raised JSONDecodeError.,CTX-AZ-008,Dashboard graph index builder read graph JSON with utf-8 and rejected common Windows BOM input,Fixed by reading graph JSON with utf-8-sig,PASS: temp graph fixture built SQLite index with 2 nodes and neighbor mcp-server:github; regression pytest passed.,2026-06-23,
 MAINT-003,Maintainer,CI change classifier,scripts/ci_classifier.py,scripts/ci_classifier.py and .github/workflows/test.yml,As a contributor I get the right CI gates for the files I changed,Classifier marks docs graph source package browser similarity telemetry and CI workflow changes conservatively,CI classifier tests available,python -m pytest src/tests/test_ci_classifier.py -q,Tested Pass,CI classifier suite covers path classes workflow contract behavior and ci_changed output use.,,,,,2026-07-06,Validated by focused pytest in this tracker slice including CI workflow classification.
-MAINT-004,Maintainer,No-test policy gate,scripts/ci_no_test_policy.py,scripts/ci_no_test_policy.py and .github/workflows/test.yml,As a reviewer I can reject risky code or workflow changes without test updates,Policy fails contract changes including release sync and graph artifact maintainer scripts that omit matching tests while only allowing metadata-only releases when diff hunks prove version or stats-only edits,CI policy tests available,python -m pytest src/tests/test_ci_classifier.py src/tests/test_ci_preflight.py -q,Tested Pass,CI policy and preflight suites cover contract files release metadata exceptions and rejection when diff context is missing.,,,,,2026-07-05,Validated after preflight started passing real per-file diffs to the policy helper and after maintainer script contract coverage expanded.
+MAINT-004,Maintainer,No-test policy gate,scripts/ci_no_test_policy.py,scripts/ci_no_test_policy.py and .github/workflows/test.yml,As a reviewer I can reject risky code or workflow changes without test updates,Policy fails contract changes including release sync and graph artifact maintainer scripts that omit matching tests while allowing proven metadata-only releases or maintainer-applied no-tests-needed labels for genuinely untestable changes,CI policy tests available,python -m pytest src/tests/test_ci_classifier.py src/tests/test_ci_preflight.py -q,Tested Pass,CI policy and preflight suites cover contract files release metadata exceptions no-tests-needed label exemptions and rejection when diff context is missing.,,,,,2026-07-05,Validated after preflight started passing real per-file diffs to the policy helper and after maintainer script contract coverage expanded.
 MAINT-005,Maintainer,Required CI aggregate,scripts/ci_required.py,scripts/ci_required.py and .github/workflows/test.yml,As a branch maintainer I can trust the required CI summary,Aggregate rejects failed missing or wrongly skipped required jobs including skipped full pytest matrix on CI workflow PRs while allowing documented conditional skips,CI required tests available,python -m pytest src/tests/test_ci_classifier.py -q,Tested Pass,CI classifier suite covers required job dependency success failure conditional skip cases and ci-changed PR full-matrix skip rejection.,,,,,2026-07-06,Validated by focused pytest in this tracker slice including CI workflow PR matrix skip rejection.
 MAINT-006,Maintainer,CI preflight gate,scripts/ci_preflight.py,scripts/ci_preflight.py scripts/ci_classifier.py and scripts/local_fast_gate.py,As a contributor I can run the same preflight logic locally that CI uses,"Preflight classifies changes then emits source static gates (ruff format/check, ruff lint, mypy, pip check), public docs tracker including bug-smoke, feature, dashboard, and toolbox coverage, strict build, graph, browser, package, and similarity gates required; local-fast groups those same checks into isolated parallel worktree lanes before the serial gate",Repo checkout with git available,Run scripts/ci_preflight.py against representative changed files; run scripts/no_mistakes_run.sh fast before PR validation,Tested Pass,"select_checks and workflow parity coverage confirm the public docs tracker includes bug-smoke, feature, dashboard, and toolbox checks before strict MkDocs build; local_fast_gate grouping coverage confirms graph hydration stays with graph validation and source changes split across independent lanes.",,,,,2026-07-06,Validated through focused preflight tests and the first real local-fast run which exposed missing tracker coverage before the final serial gate.
 MAINT-007,Maintainer,Clean host contract script,scripts/clean_host_contract.py,scripts/clean_host_contract.py and .github/workflows/clean-host-contract.yml,As a new installer I can smoke ctx on a clean host,Script builds a clean venv installs ctx and runs init scan harness and monitor checks,Network and clean temp venv available,GitHub clean-host contract or local script smoke,Tested Pass,Clean host contract job passed inside PR #195 required checks.,,,,,2026-06-23,Workflow evidence covers the script path.

docs/telemetry.md

@@ -137,9 +137,12 @@ The shipped config keeps telemetry local:
 `local_redacted` removes or hashes raw input fields such as `query`, `prompt`,
 `tool_input`, `stdout`, `stderr`, repo names, secrets, `paths`, dotted or
 hyphenated path keys such as `ctx.repo.path` and `repo-path`, and keys that
-normalize to `_path` or `_paths` suffixes. The only accepted modes are
-`local_redacted`, `disabled`, `off`, and `none`; unknown modes fail closed
-instead of emitting raw fields.
+normalize to `_path` or `_paths` suffixes. It also scans arbitrary string
+payload values for local host paths such as `/Users/...`, `/home/...`, `~/...`,
+and Windows drive paths, replacing each match with
+`[path_hash:sha256:...]`. The only accepted modes are `local_redacted`,
+`disabled`, `off`, and `none`; unknown modes fail closed instead of emitting
+raw fields.
 
 Runtime lifecycle records apply the same local-redacted posture to their
 evidence-bearing strings. Top-level `reason`, `evidence`, `command`, `summary`,
@@ -261,6 +264,11 @@ advance the checkpoint past malformed pending records; retrying after repair may
 re-export already delivered events, so downstream collectors should deduplicate
 by `event_id`.
 
+Manual and continuous exporters re-sanitize legacy local records at the export
+boundary before writing `local_jsonl` or sending OTLP, so older spool entries
+that predate current redaction rules do not leak raw query, path, repo, stdout,
+stderr, token, or secret payload values.
+
 ## Retention
 
 Retention is explicit operator action, never a background deletion. Plan first:
@@ -307,6 +315,13 @@ overrides live at `~/.claude/skill-system-config.json`.
 `OTEL_EXPORTER_OTLP_ENDPOINT` is also supported and automatically appends
 `/v1/logs`.
 
+Continuous export drains the pending spool before advancing the checkpoint, not
+just the event or metric that triggered the write. If all pending well-formed
+records export successfully and no malformed pending records remain, the
+checkpoint advances to the last exported id. If malformed pending records or a
+checkpoint anomaly are present, ctx records `status: degraded` and leaves the
+checkpoint in place so an operator can repair or intentionally replay.
+
 ## Collector Examples
 
 Use an OpenTelemetry Collector between ctx and the observability backend. That