Vouch

Field-level diff between two claim revisions or two page revisions. new_id is optional for a superseded claim: resolves to superseded_by.

Read-only reviewer briefing: pending proposals oldest-first, recent decisions, stale claims, and followups due. since: window spec — a duration ("7d", "12h"), an ISO date, or "all".

Per-artifact effectiveness: did surfacing this claim change the outcome? Read-only. For every artifact that entered a context pack, reports how the sessions it appeared in ended (good: confirm/approve, bad: contradict/supersede/archive/redact/reject), an associational lift against the corpus baseline, and a 95% Wilson interval. Verdicts are power-gated: `useful` / `harmful` only when the interval clears the baseline and the sample meets min_samples, otherwise `unverified` / `insufficient`. Writes nothing and never expires an artifact — it ranks what a reviewer should look at. window: durati…

Model identity, per-kind counts, query cache stats.

Run retrieval eval over a JSONL queries file.

Rank entities by evidence density on a topic (read-only).

Expire stale pending proposals (dry-run unless apply=True).

Explain why each candidate for a query ranked where it did. Read-only introspection over the retrieval pipeline: per candidate it returns the lexical and semantic rank, the RRF contribution, a row per stage (fusion, scope/status filters, recency, pages_first, rerank, strategy, limit, and the optional budget/citation gates) with the rank and score delta that stage caused, and the gate that kept or dropped it. max_chars / require_citations opt into explaining kb.context's budget and citation gates. project/agent set the viewer context for scope filtering, the same as kb.search — nothing is exp…

Return the full markdown body of a named skill / slash command. Errors with ``permission_denied`` when ``mcp.publish_skills`` is ``false``, and ``not_found`` when the name isn't in the catalogue.

Render the provenance DAG (or one session's subgraph) as dot/mermaid.

Forward impact: dependents, and breakage if op (archive/contradict/supersede).

Drop and rebuild state.db from the durable files.

List all claims, optionally filtered by status.

List approved goals, oldest first. Defaults to the open ones.

List pages, optionally filtered by kind and frontmatter. type: page kind (built-in or config-declared, e.g. "followup"). meta: frontmatter equality, e.g. {"followup_status": "open"}. meta_before / meta_after: inclusive bounds — numbers or ISO dates, e.g. meta_before={"due_at": "2026-07-10"} for followups due by then.

List proposals awaiting human review.

List all relations; if node_id is given, only edges touching it.

List captured sessions in the review pipeline: open buffers awaiting a summary, and filed summary proposals awaiting review. Read-only. Each row: session_id, stage ("buffer" | "pending"), proposal_id, kind, title, summarized, observations, last_activity.

Enumerate every Claude Code skill / slash command visible to vouch. Scans, in priority order: 1. ``<kb_root>/.claude/skills/<name>/SKILL.md`` — project-local skills 2. ``<kb_root>/.claude/commands/<name>.md`` — project-local commands 3. ``~/.claude/skills/<name>/SKILL.md`` — user-global skills 4. ``~/.claude/commands/<name>.md`` — user-global commands Project entries override user ones with the same name. Returns ``[{name, description, scope, kind, path}]``. Returns an empty list when ``mcp.publish_skills`` is ``false`` in config.yaml.

Return graph neighbors of a claim, page, entity, or source.

Propose a new claim. Becomes durable only after `kb_approve`.

Propose hard-deleting a durable artifact (claim/page/entity/relation). Files a PENDING delete request that a *different* reviewer approves via kb.approve. Refused if the target is still referenced by another artifact, unless cascade=True, which records the referrer edits (pages and claims lose their pointer, relations are deleted) in the same proposal so the reviewer approves the whole set as one decision.

Propose an in-flight objective — what this project is trying to do now. Files a PENDING goal a human approves via kb.approve; the approved goal then re-surfaces in the session-start digest so a fresh session picks up the intent, not just the facts. Status moves after approval go through kb.set_goal_status, never through a second proposal.

Propose a theme synthesis page from a detected cluster. Routes through the review gate — appears in kb.list_pending. Pass a cluster from kb.detect_themes directly.

Rebuild the prov_edges cache from durable files; returns edge count.

Return a claim with its citation list.

Return an evidence record — the span/quote a claim cites.

Return a page (title, body, claim ids).

Return a registered source's metadata (not its content).

Register a source. Evidence intake is NOT gated (registering raw evidence is harmless and de-duplicates by content hash).

Re-encode every artifact under the current embedding adapter.

Mass-reject pending edges the auto-extractor filed (issue #224). Scope to one page's edges with `page_id`, or omit it to clear every pending auto-extracted edge across the KB.

Search the KB. backend: None (default) defers to retrieval.backend in config.yaml; "auto"/"hybrid" fuse embedding + fts5 via RRF (substring fallback under "auto"), or pin "embedding", "fts5", or "substring". The response retrieval block reports the backend that actually served the query. project/agent: optional viewer context for scope filtering.

Render a captured session's full transcript from its raw agent JSONL. Read-only. Locates the raw Claude Code / Codex file on disk and normalizes it into message blocks (text, thinking, tool_use with paired results). ``agent`` restricts the search ("claude" | "codex"); omit to try both. Degrades to compact capture observations when the raw file is unavailable, and to a bare unavailable result when no KB resolves at all.

Move an approved goal to open / done / abandoned / blocked. Appends the transition to the audit log — this is the only write path for goal status.

Re-hash every source and report any drift.

Observability: pending by agent, review rates, citation coverage. days: decision window in days; 0 means all-time.

Return KB artifact counts and health summary.

Summarize a captured session into PENDING page proposals. Reads the host-neutral observation buffer for `session_id` and files either one mechanical rollup page (small sessions) or several LLM-drafted topical `session` pages (large sessions). `mode` is "auto" | "split" | "mechanical". Long-running when it splits (the LLM call is synchronous). Never approves.

Answer a query from the review-gated KB, with inline `[id]` citations, an explicit gaps block, and a synthesis_confidence grade. Unlike `kb_context` (a ranked list), this returns prose where every sentence is traceable to a source. Deterministic by default (approved claims only); `llm=True` drafts the answer with the deployment-configured LLM (compile.llm_cmd) grounded in pages and approved claims — citations are still verified mechanically, and the call is synchronous.

Shortest typed-edge path between two artifacts (or found=false).

Advisory triage scoring over the pending-review queue. Attaches `_meta.vouch_triage` (recommendation/score/signals/rationale) to each pending proposal's view. Read-only — never approves, rejects, or otherwise decides; a human still calls `kb_approve` / `kb_reject`. Opt-in: disabled unless `triage.enabled: true` is set in config.yaml.

Poll confidence-gated context volunteered for an active session.

Backward provenance for a claim: cites, session, supersedes, approval. depth: how many hops of provenance to expand (default 3).

Strip references to claims that no longer exist, KB-wide. Covers durable pages and pending page proposals: the frontmatter claim list and inline `[claim: …]` markers both lose ids that resolve to no claim file. Audited as one bulk event. Use after claims were archived, redacted, or bulk-cleared and pages still point at them. Args: dry_run: If True, report what would be stripped without writing. Returns: Per-page and per-proposal dead ids, total dropped count.