llmwiki

Update the persistent knowledge wiki after source documents change. Writes wiki pages or review candidates according to project policy. Extract concepts from new/changed sources, generate wiki pages, resolve interlinks, and rebuild the index. Requires an LLM provider with credentials.

Describe one declared workflow action, including its EFFECTIVE permission per surface = min(profile request, local grant, surface cap). The `mcp` surface is hard-capped at staged-write. Read-only.

Export the wiki as an Open Knowledge Format (OKF) v0.1 bundle. `out` defaults to dist/exports/okf and must stay inside the project.

Retrieve bounded evidence from this project knowledge base for research, document questions, or agent task context. Use it when you will reason over the evidence yourself. Build a token-budgeted pack for `prompt` over the compiled wiki: primary pages, semantic chunks, graph neighbors, citations, warnings, and suggested next actions. Returns the same v1 JSON envelope as `llmwiki context --json`. Read-only; no provider credentials required. Use this to PREPARE evidence; use `query_wiki` to GENERATE a grounded natural-language answer.

Import an OKF bundle from a path INSIDE the project as review candidates. Always staging-only — imported knowledge requires human review/approval before it enters the wiki. Use dryRun to preview.

Add a user-requested document to this knowledge base: fetch a trusted URL or copy a trusted local file into sources/. Writes project files. Returns the saved filename, character count, and whether content was truncated to fit the size limit.

Run rule-based quality checks (broken wikilinks, orphans, duplicates, empty pages, broken citations). Returns structured diagnostics. No LLM call.

List the workflow actions declared by the active profile (id, label, workflow, operation). Read-only; takes no lock and mutates nothing.

Get a model-generated answer grounded in this knowledge base. For source evidence to reason over yourself, use get_context_pack instead. Selects relevant pages, loads them, and returns a grounded answer with citations. Set save=true to request publication as a wiki page; citation or profile refusals return the answer with publicationRefusal and no saved slug. Set debug=true to include the selected chunks and their scores. Requires an LLM provider.

Read a single wiki page by slug. Searches concepts/ first, then queries/. Returns the parsed frontmatter and body. No LLM call required.

Run the wiki quality eval harness. fast suite checks health and citation coverage without LLM calls. full suite also LLM-judges a sample of citations (requires an LLM provider). Set record: true to append results to eval history (default false — read-only).

Execute a declared workflow action. MCP actions are HARD-CAPPED at staged-write: they CANNOT perform trusted writes or satisfy human gates, regardless of what a profile or local config requests. A denied, unknown, or invalid action returns a clean error result rather than crashing.

Select pages relevant to a question and return their full content. Uses semantic embeddings when available, falling back to LLM-based selection over the wiki index. Requires an LLM provider.

Verify a hash-pinned artifact ref against the active profile's declared artifact types. Returns manifest metadata (artifactType, slug, sha256, bytes, contentKind, writtenAt) plus a health verdict — NEVER the artifact body. Read-only; no write or store-wide list tool is exposed over MCP.

Start here to inspect this knowledge base before retrieval or updating it. No provider credentials required. Summarize page count, source count, last compile time, pending source changes, and freshness-derived page health. stalePages lists concept slugs whose source changed or partially disappeared since last compile. orphanedPages lists concept slugs whose every owning source was deleted OR that are frontmatter-flagged orphaned (superset of prior behavior). stateStatus reports state.json readability (ok | missing | corrupt) so corrupt state is never silent. Each list (stalePages, orphanedPages, pendingChanges) is capped at 100 entries for response size; the corresponding *Count fields (staleCount, orphanedCount, pendingChangesCount) give the true totals. Read-only — never modifies the workspace.

Report the status of one workflow run (by runId) or of all runs, classified against the active profile (current / historical / needs-adaptation / blocked-by-config). Read-only and fail-visible: an unavailable store surfaces a problem row, never an empty result. Takes no lock and mutates nothing.