cogni-knowledge

Phase-4.5 answer-claim distiller for the inverted pipeline. Reads a per-question claim bundle — each existing type:question node grouped with the pre-extracted claims of the sources that answer it — and selects, per question, the claims that actually ANSWER it, writing a raw-text answer-records file the knowledge-distill orchestrator feeds to question-store.py answer-merge. Pure proposal — never writes wiki pages, never builds JSON/YAML, never computes slugs or decides claim-dedup. The constrained per-question sibling of concept-distiller.

Phase-4 claim extractor for the inverted pipeline. Reads one cached source body plus the calling sub-question refs, identifies 5-20 verifiable factual claims grounded in the body, and returns a JSON array of {id, text, excerpt_quote, excerpt_position, sub_question_refs, extracted_at} entries the calling source-ingester writes into the wiki page's pre_extracted_claims frontmatter. Pure extraction — never writes files.

Phase-4.5 distiller for the inverted pipeline. Reads the run's source-claim bundle plus an index of the distilled pages already on the wiki, clusters recurring facts into a small set of concept / entity / person proposals, and writes a raw-text concept-records file the knowledge-distill orchestrator feeds to concept-store.py. Pure proposal — never writes wiki pages, never builds JSON/YAML, never computes slugs or decides claim-dedup.

Phase-4.5 summary re-narrator for the inverted pipeline. For each distilled page (concept / entity / person) that gained claims this run, rewrites the

Phase-7 concepts-outline lead-in narrator for the inverted pipeline (concept-map auto-refresh). For each theme group in wiki/concepts/index.md, (re)writes the per-theme lead-in paragraph the ENGINE owns — framing prose (why the theme matters + what to read first), never a bullet dump. Reads a bundle (per theme: heading + current machine lead-in + that theme's concept titles and one-line summaries), writes a raw-text records file the knowledge-finalize concepts auto-refresh sub-step stages or applies. Pure proposal — never writes wiki pages, never builds JSON/YAML, never touches a human (non-sentineled) lead-in, and never lists, reorders, or invents concept bullets (the deterministic renderer owns those).

Phase 5 of the inverted pipeline. Reads <project>/.metadata/plan.json + <project>/.metadata/ingest-manifest.json + the populated cogni-wiki, dispatches a wiki-composer pass (plus ONE bounded fail-soft zero-network coverage-gated re-dispatch when a sub-question has ingested evidence the draft left uncited), and lands <project>/output/draft-vN.md + <project>/.metadata/citation-manifest.json. Supports --source wiki to compose a report grounded only in the bound wiki + fetch-cache with no web crawl (default web unchanged; local/hybrid staged). Use this skill whenever the user says 'compose the draft', 'write the report from the wiki', 'wiki-only report', 'compose from the wiki only', 'no web crawl report', 'phase 5 of the knowledge pipeline', 'knowledge compose', 'draft v1', or 'run the writer'. After compose, knowledge-verify will run the zero-network claim alignment.

Phase 2 of the inverted pipeline. Reads plan.json, dispatches one source-curator agent per sub-question to WebSearch + score candidate sources AND fetch each survivor's body via WebFetch into the shared fetch-cache, then merges the per-sub-question batches into candidates.json via candidate-store.py. Each candidate carries a fetch sub-object. Use this skill whenever the user says 'curate sources for the eu-ai-act plan', 'discover candidates for project X', 'run the curators on plan.json', 'knowledge curate', 'phase 2 of the knowledge pipeline'. After curate, run knowledge-fetch (cobrowse reconcile; a near no-op unless --cobrowse).

Render an HTML dashboard for a bound cogni-knowledge base — runs the vendored wiki-dashboard scripts (render_dashboard.py + build_graph.py) natively against the bound wiki and writes a knowledge-overlay.md sidecar that lists deposited research projects and the latest lint-audit claim_drift count. Use this skill whenever the user says 'show the dashboard for my <slug> base', 'knowledge dashboard', 'visualize my eu-ai-act knowledge base', 'render the knowledge base as HTML', 'knowledge-dashboard <slug>'. The sidecar makes the binding's contribution visible alongside the wiki's own dashboard.

Phase 4.5 of the inverted pipeline (between ingest and compose). Distills the run's source claims into recurring type:concept / type:entity / type:person pages, creating-or-merging them under a lock with claim-level dedup so the bound wiki compounds across runs (distilled pages get enriched, not duplicated). Also synthesizes a citable answer_claims: surface onto each type:question node from its findings' claims (Step 6.1). An optional cross-lingual pass merges DE↔EN twin claims on mixed-language bases (auto-skips otherwise). Fail-soft and optional: a distill failure never blocks compose. Use this skill whenever the user says 'distill the concepts', 'build the concept web', 'phase 4.5', 'knowledge distill', 'extract entities and concepts', 'answer the question nodes', or 'dedupe claims'. After distill, knowledge-compose reads the distilled pages as framing context.

Phase 3 of the inverted pipeline. Builds fetch-manifest.json from the fetch results the curators already produced in Phase 2 (bodies are already fetched during curate). Cobrowse is OPT-IN (--cobrowse): the skill walks the user through enabling the Claude-in-Chrome extension, then dispatches source-fetcher sequentially to recover WebFetch misses AND to additively top up thin primary-tier sources (reading the fuller browser body beyond WebFetch's cap, superseding only if strictly longer). Use this skill whenever the user says 'fetch candidates for project X', 'build the fetch manifest', 'phase 3 of the knowledge pipeline', 'recover the failed sources', 'deepen the primary sources', 'knowledge fetch'. After fetch, run knowledge-ingest to deposit per-URL wiki pages.