knowledge-compose

Knowledge Compose

Phase 5 of the inverted pipeline. Reads the per-project plan.json + ingest-manifest.json + the populated wiki at <binding.wiki_path>/wiki/, dispatches wiki-composer once, and verifies the output files land on disk.

The default --source web path composes from this run's web-ingested sources (the ingest-manifest.json Phase 4 produced). --source wiki is the wiki-only rung — a full structured, verified report grounded only in the already-populated wiki (wiki/sources/*.md + wiki/syntheses/*.md + distilled pages) and the .cogni-knowledge/fetch-cache/, with no web crawl: it preserves the retired research-report --source wiki capability so that capability does not vanish when cogni-research is archived.

The wiki-only mode does not change the composer or the verify/finalize tail at all. Instead, Step 0 synthesizes an ingest-manifest.json from the bound wiki's wiki/sources/*.md pages — mapping each source to the current plan's sub-questions via the shared wiki-grounding discovery primitive (scripts/wiki-source-manifest.py) — so the dispatched wiki-composer reads a normal manifest and runs byte-for-byte identically whether that manifest came from a web crawl or from the wiki. The composer does not branch on source mode and needs no new input. (See --source in Parameters and the Step 0 pre-flight.)

The composer reads wiki/index.md + selected wiki/sources/*.md (lazily, claims under pre_extracted_claims:) + prior wiki/syntheses/*.md. It also reads two optional knowledge-distill surfaces, topic-matched and lazily: the distilled wiki/{concepts,entities}/*.md pages (claims under distilled_claims:, ids dcl-NNN) and the type: question nodes at wiki/questions/*.md — first-class pages each recording one research question the base has already explored, with ## Findings [[links]] to the sources that answered it (answers under an optional answer_claims: block, ids acl-NNN).

Both surfaces serve both as narrative framing and as citable cross-source evidence, under one shared rule: when ≥2 distinct sources converge (≥2 sources agreeing on a fact the distilled page captures; ≥2 distinct sources in a question node's backlinks[]), the composer cites the distilled page or question node directly via its dcl-NNN/acl-NNN claim id, so the convergence carries epistemic weight rather than a row of stacked source markers. Below that bar — a single-source answer, or a page/node carrying no claim block yet — the surface stays framing-only: the composer reads it for orientation but cites the backing source page (an inline citation to a claim-less node would score unsupported). The verifier scores each such citation against that claim's text. Both surfaces are optional and fail-soft: with neither present the composer composes from sources + syntheses alone.

The composer then writes:

• <project>/output/draft-v{N}.md — the draft, with clickable numbered [N] inline citations (wikilinks confined to the reference list).
• <project>/.metadata/citation-records-v{N}.txt — one raw-text record per citation (the composer writes this; it never hand-builds JSON). This skill then runs citation-store.py build to serialize and validate <project>/.metadata/citation-manifest.json (schema 0.1.1, one {id, draft_position, draft_sentence, wiki_slug, claim_id, url} entry per citation). Escaping is owned by json.dumps, never the LLM — a straight " in a draft_sentence would otherwise break a hand-built manifest's json.loads and kill the verify phase.

A writer-outline-v{N}.json is persisted by the composer's Phase 1 before any draft Write attempt — this is the outline-recovery contract. If the composer crashes between outlining and drafting, re-running this skill detects the leftover outline and re-dispatches the composer with RESUME_FROM_OUTLINE=true so only Phase 2 runs.

citation-manifest.json shape (consumed by the wiki-verifier):