knowledge-compose

name	knowledge-compose
description	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.
allowed-tools	Read, Write, Bash, Task

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) + prior wiki/syntheses/*.md. Since the distillation interphase (knowledge-distill), it also reads the distilled wiki/{concepts,entities}/*.md pages (topic-matched, lazily) — these serve both as narrative framing and as citable cross-source evidence: when ≥2 sources converge on a fact the distilled page already captures, the composer cites the distilled page itself via its dcl-NNN claim id, so the convergence carries epistemic weight rather than a row of source markers. Distilled pages carry distilled_claims: (not pre_extracted_claims:), and a distilled-page citation is scored by the verifier against that claim's text. Distillation stays optional and fail-soft: when it hasn't run, the composer simply has no distilled pages to draw on and composes from sources + syntheses alone.

The composer also reads the type: question nodes at wiki/questions/*.md (topic-matched, lazily) — first-class wiki pages each recording one research question the base has already explored, with ## Findings [[links]] to the sources that answered it. These serve both as narrative framing and as a citable cross-source answer surface: a question node may carry an answer_claims: block (acl-NNN ids, synthesized by knowledge-distill), and when its backlinks[] list ≥2 distinct sources the composer cites the node directly via its acl-NNN claim — one citation carrying "N sources agree on the answer" — exactly mirroring the distilled-page rule just above. A single-source answer, or a question node with no answer_claims: block yet, stays framing-only: the composer reads it for orientation but cites the backing source page, never the node (an inline citation to a claim-less node would score unsupported). A question-node citation is scored by the verifier against that answer claim's text.

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):

{
  "schema_version": "0.1.1",
  "draft_version": 1,
  "citations": [
    {"id": "cit-001", "draft_position": "02:03", "draft_sentence": "Article 6 classifies a system as high-risk…<sup>[1](https://…)</sup>.", "wiki_slug": "eu-ai-act-article-6", "claim_id": "clm-001", "url": "https://…"}
  ]
}

Read ${CLAUDE_PLUGIN_ROOT}/references/inverted-pipeline.md §"Phase 5 — knowledge-compose" and references/claim-at-ingest.md once to anchor on the contract.

When to run

• web mode (default): <project>/.metadata/ingest-manifest.json exists with non-empty ingested[] (Phase 4 has run) AND either no draft yet OR the user explicitly wants a new draft version.
• --source wiki mode: the bound wiki has wiki/sources/*.md to compose from (Phase 4 ran for an earlier project on this base, or the wiki was populated by another path). No ingest-manifest.json is required — the wiki itself is the evidence base.
• User explicitly invokes /cogni-knowledge:knowledge-compose.

Never run when

• No <project>/.metadata/plan.json — offer knowledge-plan first.
• web mode only: no <project>/.metadata/ingest-manifest.json or ingested[] empty — offer knowledge-ingest first. (Under --source wiki this is not a block — the wiki/sources/ glob is the evidence base; see the --source wiki gate below.)
• --source wiki mode only: the bound wiki has no wiki/sources/*.md pages — there is nothing to compose from. Offer knowledge-ingest (to web-ingest sources) or point the operator at a populated base.
• No binding.json at the resolved knowledge root — offer knowledge-setup first.
• binding.wiki_path does not resolve to a directory containing .cogni-wiki/config.json — the binding is stale.

Parameters

Parameter	Required	Description
--knowledge-slug	Yes	Slug of the bound knowledge base.
--project-path	Yes	Absolute path to the project directory.
--knowledge-root	No	Override the default knowledge-base directory.
--source	No	Evidence source mode. web (default when omitted) is the standard inverted-pipeline path — compose from the web-ingested ingest-manifest.json sources. wiki is the wiki-only rung: compose a full structured report grounded only in the bound wiki (wiki/sources/*.md + wiki/syntheses/*.md + distilled pages) and the .cogni-knowledge/fetch-cache/, with no web crawl — the preserved research-report --source wiki capability. local/hybrid are accepted but staged (treated as wiki until implemented), mirroring the staged apa/mla/harvard citation formats. Omitting the flag is byte-identical to the prior behavior.
--target-words	No	Soft target word count. Default reads target_words from plan.json if present, else 2000. A soft upper budget under standard density (never a floor — a tight, fully-grounded draft is the better outcome), a ceiling under executive. It drives no expansion by word count under either density — Step 5.5 expands on a coverage deficit. Under executive it additionally bounds the expansion: the actuator only fires below this ceiling and the composer stops deepening at it.
--no-expand	No	Skip the Step 5.5 bounded coverage-gated expansion. Default: OFF (expansion may run under standard density when a sub-question has uncited ingested evidence, and under executive density for a zero-cited sub-question while headroom remains below the target_words ceiling). Pass to keep the single composer pass even when a coverage deficit exists. Mirrors finalize's --no-reviewer/--no-contradictor.
--no-contradiction-surfacing	No	Skip threading the ingest-time recency-survivor annotations into the composer. Default: OFF — when the project's .metadata/contradiction-ingest.json carries ≥1 recency-resolved contradiction, its path is threaded to the composer (Step 3.5) so the composer can build the recency-survivor map. Pass to suppress the surfacing entirely (dispatch byte-identical to the pre-surfacing form). Mirrors --no-expand / finalize's --no-contradictor.
--contradiction-act	No	Mode-C opt-in (default OFF). Gates whether the composer acts on the recency-survivor surfacing — i.e. actually prefers the more-recent survivor claim over a superseded loser for high-severity contradictions. Default OFF: the central contradiction-ingest.json is still threaded for the composer to read (observability), but Phase 2 never prefers a survivor (dispatch byte-identical to the pre-acting form). Pass to enable acting; only takes effect when a non-empty CONTRADICTION_INGEST_PATH also resolves (Step 3.5). Independent of --no-contradiction-surfacing (which suppresses the path — and therefore acting — entirely).
--prose-density	No	Override plan.json::prose_density for this draft: standard (soft upper budget, cite/ground every claim) or executive (BLUF + Pyramid ceiling, one citation per claim). Default reads plan.json, else standard.
--tone	No	Override plan.json::tone for this draft (see references/writing-tones.md). Default reads plan.json, else objective.
--citation-format	No	Override plan.json::citation_format: ieee/chicago (wired) or apa/mla/harvard (staged). Default reads plan.json, else ieee.
--draft-version	No	Force a specific draft version N. Default: max(existing output/draft-v*.md) + 1, or 1.
--dry-run	No	Print the resolved inputs (WIKI_ROOT, DRAFT_VERSION, RESUME_FROM_OUTLINE) without dispatching the composer.

Workflow

0. Pre-flight

Required plugins. Probe only cogni-wiki (clean-break — no cogni-research, no cogni-claims):

probe_plugin() {
  local plugin="$1" skill="$2"
  test -f "${CLAUDE_PLUGIN_ROOT}/../${plugin}/skills/${skill}/SKILL.md" && return 0
  for d in "${CLAUDE_PLUGIN_ROOT}/../../${plugin}/"*/skills/"${skill}"/SKILL.md; do
    [ -f "$d" ] && return 0
  done
  return 1
}
probe_plugin cogni-wiki wiki-setup && WIKI_OK=yes || WIKI_OK=no

If WIKI_OK=no, abort with the standard missing-plugin message.

Binding + wiki root. Resolve knowledge_root (same logic as knowledge-ingest). Read the binding:

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/knowledge-binding.py read \
    --knowledge-root <knowledge_root>

On success: false → abort, offer knowledge-setup.

Parse data.binding.wiki_path as WIKI_ROOT. Confirm <WIKI_ROOT>/.cogni-wiki/config.json exists; abort otherwise. Confirm <WIKI_ROOT>/wiki/ exists.

Source mode. Resolve SOURCE_MODE from --source (default web when the flag is absent — byte-identical to the prior behavior). local/hybrid are staged: when either is passed, print staged: --source=<mode> treated as wiki (not yet implemented) here at pre-flight (before any dispatch), then proceed as wiki. SOURCE_MODE decides how <project_path>/.metadata/ingest-manifest.json is obtained below — it does not change anything downstream of Step 0, because both modes hand the composer a normal manifest.

Project manifests. <project_path>/.metadata/plan.json is always required — abort with "run knowledge-plan first" otherwise. How the ingest manifest is obtained is source-mode-dependent, but the result is the same: a populated <project_path>/.metadata/ingest-manifest.json the composer reads in Step 4.

• SOURCE_MODE=web (default): confirm <project_path>/.metadata/ingest-manifest.json exists; read it and, if ingested[] is empty, abort with "no ingested sources to compose from — re-run knowledge-ingest". This is the unchanged web path — the manifest was produced by knowledge-ingest from a web crawl.

• SOURCE_MODE=wiki: synthesize the ingest manifest from the bound wiki (no web crawl) by mapping each wiki/sources/*.md page to the current plan's sub-questions via the shared wiki-grounding primitive:

  python3 ${CLAUDE_PLUGIN_ROOT}/scripts/wiki-source-manifest.py build \
      --wiki-root "<WIKI_ROOT>" \
      --plan "<project_path>/.metadata/plan.json" \
      --out "<project_path>/.metadata/ingest-manifest.json"
  

  Parse the envelope: on success: true read data.ingested_count. If it is 0, abort with "no wiki sources cover this plan's sub-questions — ingest more sources (knowledge-ingest), broaden the plan, or check the bound wiki". On success: false, surface error and abort. The synthesized manifest carries source_mode: "wiki" and the same ingested[] shape ({url, slug, title, publisher, summary, claims_extracted, sub_question_refs[]}) the web path produces, so Steps 1–7 run identically — the composer reads it in Step 4 with no awareness it was wiki-sourced. INGESTED_SOURCES for the dry-run/summary is data.ingested_count.

Phase timing accumulators. Capture PHASE_START=$(date -u +%FT%TZ) (the phase wall-clock the Step 8 ledger records) and initialise MAX_DURATION_MS=0 (the slowest single composer dispatch, the orchestrator-measured Option B serial-tail signal) now, at the top of the run. They are run-level — set once here so a run that also re-dispatches the composer in Step 5.5 folds across both dispatches. The composer agents (wiki-composer) have no Bash tool to self-report duration_ms, so this phase measures each dispatched Task's wall-clock itself (Option B); see references/run-metrics-wiring.md.

1. Resolve draft version N

KNOWLEDGE_SCRIPTS="${CLAUDE_PLUGIN_ROOT}/scripts" \
PROJECT_PATH="<project_path>" \
python3 -c '
import os, re
from pathlib import Path
out = Path(os.environ["PROJECT_PATH"]) / "output"
existing = sorted(int(m.group(1)) for p in out.glob("draft-v*.md")
                  for m in [re.match(r"draft-v(\d+)\.md$", p.name)] if m)
print(existing[-1] + 1 if existing else 1)
'

If --draft-version is passed, use that integer instead and overwrite any existing file with the same N (the user asked for it explicitly).

2. Outline-recovery check

Check for <project_path>/.metadata/writer-outline-v{N}.json:

• Exists (and --draft-version is the same N): a prior dispatch wrote the outline before crashing. Pass RESUME_FROM_OUTLINE=true to the composer so Phase 1 is skipped. Surface this in the dry-run summary and the final summary — the operator should know a recovery happened.
• Absent: the composer runs both phases. Pass RESUME_FROM_OUTLINE=false (or omit; the agent treats unset as false).

If --dry-run, print the resolved inputs:

WIKI_ROOT=<wiki_root>
PROJECT_PATH=<project_path>
DRAFT_VERSION=<N>
SOURCE_MODE=<web|wiki>
RESUME_FROM_OUTLINE=<true|false>
TARGET_WORDS=<resolved>
PROSE_DENSITY=<resolved>
TONE=<resolved>
CITATION_FORMAT=<resolved>
OUTPUT_LANGUAGE=<resolved>
INGESTED_SOURCES=<web: count from ingest-manifest.json; wiki: count of wiki/sources/*.md>
CONTRADICTION_INGEST_PATH=<resolve Step 3.5 early for the dry-run: path|empty>
CONTRADICTION_ACT=<resolve Step 3.5 early for the dry-run: 1 (mode-C acting ON, high-severity only)|empty (observe-only)>

(For --dry-run, resolve Step 3.5's contradiction-gate vars before printing so the operator can confirm whether mode-C acting will fire before a live run — the acting-vs-observing branch is exactly the high-stakes one the new --contradiction-act flag introduces.)

and stop.

3. Resolve composer knobs (TARGET_WORDS, PROSE_DENSITY, TONE, CITATION_FORMAT)

Each knob resolves --<flag> if passed, else the matching plan.json field, else the hard default. They were already resolved + validated in knowledge-plan Step 0.5 and written to plan.json, so this is a read-with-override (the --flag lets an operator re-compose a single draft with a different knob without re-planning):

Knob	--flag	plan.json field	Default
TARGET_WORDS	--target-words	target_words	2000
PROSE_DENSITY	--prose-density	prose_density	standard
TONE	--tone	tone	objective
CITATION_FORMAT	--citation-format	citation_format	ieee

TARGET_WORDS is a soft target — a soft upper budget under standard (never a floor), a ceiling under executive. The composer itself is single-pass per dispatch; a coverage deficit (a sub-question whose ingested evidence the draft left uncited) may trigger ONE bounded expansion re-dispatch in Step 5.5 (below) under both densities — under standard for any thin-or-zero-cited deficit section, under executive only for a zero-cited deficit sub-question and only while the draft is below the TARGET_WORDS ceiling (the actuator's pre-check and the composer's in-pass stop both enforce the ceiling). Word count itself drives no re-dispatch under either density — it only bounds the executive one. CITATION_FORMAT is now live: ieee/chicago render end-to-end (the composer differs only in the reference-list string); apa/mla/harvard are accepted but render as numbered until the author-date follow-up lands (references/citation-formats.md).

3.5. Resolve the recency-survivor surfacing path (CONTRADICTION_INGEST_PATH) + the mode-C acting gate (CONTRADICTION_ACT)

knowledge-ingest Step 4.6 may have written <project_path>/.metadata/contradiction-ingest.json — the ingest-time contradiction tripwire, where each contradiction finding optionally carries a zero-network resolution {survivor_claim_id, strategy: "recency", rationale} recency-survivor suggestion (the more-recent side of a contradicting claim pair) plus a severity ∈ {high, medium, low}. When that file holds ≥1 resolved contradiction, thread its path so the composer can build the recency-survivor map. Acting on that map — preferring the survivor over the superseded loser — is mode C: gated behind the explicit --contradiction-act opt-in AND restricted to high-severity contradictions (the classification decision for the shipped slice). By default the path is threaded for observability but the composer never prefers a survivor. This is pure surfacing — it changes only which of an existing claim pair the composer cites, never the citation-manifest schema, the verifier, or any wiki page.

Resolve CONTRADICTION_INGEST_PATH — fail-soft, where any miss leaves it empty and the dispatch is byte-identical to the pre-surfacing form:

• If --no-contradiction-surfacing was passed → leave empty (omit the param).
• Else if <project_path>/.metadata/contradiction-ingest.json does not exist → leave empty.
• Else read it and check the canonical file's top-level resolution_coverage.resolved (the merged contradiction-ingest.json is a canonical artifact, not a {success,data,error} script envelope — there is no data. wrapper): when it is a positive integer → set CONTRADICTION_INGEST_PATH to that file path; when it is 0, absent, or the file is unreadable / malformed JSON → leave empty.

CONTRADICTION_INGEST_PATH=""
ci="<project_path>/.metadata/contradiction-ingest.json"
# (skip this block entirely when --no-contradiction-surfacing was passed)
if [ -f "$ci" ]; then
  resolved=$(CI="$ci" python3 -c "import json,os; d=json.load(open(os.environ['CI'])); print(int((d.get('resolution_coverage') or {}).get('resolved', 0) or 0))" 2>/dev/null || echo 0)
  [ "${resolved:-0}" -gt 0 ] 2>/dev/null && CONTRADICTION_INGEST_PATH="$ci"
fi

# Mode C (acting) is a separate, explicit opt-in. Skip the CONTRADICTION_ACT="1"
# line entirely when --contradiction-act was NOT passed (the default) — exactly as
# the block above is skipped when --no-contradiction-surfacing was passed. When
# --contradiction-act WAS passed, the line sets CONTRADICTION_ACT="1" only if a
# non-empty CONTRADICTION_INGEST_PATH also resolved.
CONTRADICTION_ACT=""
# (run the next line ONLY when --contradiction-act was passed:)
[ -n "$CONTRADICTION_INGEST_PATH" ] && CONTRADICTION_ACT="1"

Pass CONTRADICTION_INGEST_PATH to the composer (Step 4, and the Step 5.5 expansion re-dispatch) only when it resolved to a non-empty path — omit the parameter entirely otherwise, so a project with no resolved contradictions dispatches exactly as before. Mode C (acting) is gated separately: pass CONTRADICTION_ACT=1 to the composer (both dispatches) only when --contradiction-act was passed AND CONTRADICTION_INGEST_PATH is non-empty — omit it otherwise, so the default run threads the path for the composer to read but never acts on it (the survivor preference is suppressed unless the maintainer explicitly opts in, and even then only for high-severity contradictions). The high-severity restriction itself is enforced composer-side, not threaded as a parameter: with CONTRADICTION_ACT=1 the wiki-composer builds its recency-survivor acting map from severity == "high" central-JSON findings only (${CLAUDE_PLUGIN_ROOT}/agents/wiki-composer.md Phase 0 step 7), so the orchestrator threads only the boolean opt-in — the severity scope lives where the severity-bearing contradiction-ingest.json is read. The frontmatter contradiction_resolutions: block carries no severity field and so never drives acting.

This central path is the fallback layer: knowledge-ingest Step 4.6.4 also persists each resolution durably onto the participating pages' contradiction_resolutions: frontmatter (mode-B), and the composer prefers a page's own frontmatter block over this central file when present (fallback hierarchy frontmatter-resident → central → none, resolved entirely composer-side in Phase 0 step 7.1). No orchestrator change is needed — the resolution here, and the resolution_coverage.resolved gate, are unchanged; the frontmatter preference is invisible to this skill.

4. Dispatch wiki-composer (single Task call)

Stamp START_MS=$(python3 -c 'import time; print(int(time.time()*1000))') immediately before the dispatch (the Option B per-dispatch timing — see Step 0 + references/run-metrics-wiring.md). Dispatch via the Task tool (matches the upstream knowledge-ingest / knowledge-fetch agent-dispatch convention):

Task(wiki-composer,
     PROJECT_PATH=<project_path>,
     WIKI_ROOT=<wiki_root>,
     DRAFT_VERSION=<N>,
     TARGET_WORDS=<resolved>,
     PROSE_DENSITY=<resolved, default standard>,
     TONE=<resolved, default objective>,
     CITATION_FORMAT=<resolved, default ieee>,
     OUTPUT_LANGUAGE=<plan.json::output_language, default en>,
     CONTRADICTION_INGEST_PATH=<resolved in Step 3.5; OMIT this param when empty>,
     CONTRADICTION_ACT=<resolved in Step 3.5; OMIT this param when empty>,
     RESUME_FROM_OUTLINE=<true|false>)

OUTPUT_LANGUAGE is read from <project_path>/.metadata/plan.json (output_language, default en — the same value knowledge-finalize reads for its reference heading). It controls the draft body, section headings, and the reference-section heading. PROSE_DENSITY / TONE / CITATION_FORMAT (Step 3) shape the draft's structural discipline, rhetorical register, and citation rendering respectively — all single-pass (the composer never loops on any of them). The agent derives the plan and ingest-manifest paths from PROJECT_PATH (fixed .metadata/plan.json and .metadata/ingest-manifest.json). wiki-composer lives at ${CLAUDE_PLUGIN_ROOT}/agents/wiki-composer.md — dispatched via Task, not Skill. Single-pass, no fan-out, no per-section sharding — the agent reads the wiki itself and writes both output files atomically.

Parse the return envelope:

• ok: true → fold this dispatch's wall-clock into the accumulator: MAX_DURATION_MS=$(python3 -c "import time; print(max($MAX_DURATION_MS, int(time.time()*1000) - $START_MS))") (fail-soft — an unset START_MS contributes 0, never aborts), then continue to Step 4.5 (build the manifest).
• ok: false, error: "no_ingested_sources" → re-emit the abort message and stop (shouldn't happen if Step 0 ran, but defence-in-depth).
• ok: false, error: "write_failed" → surface the reason; do not retry blindly. The composer already retried once internally. Direct the user to inspect output token-budget conditions or re-run.
• ok: false, error: "outline_write_failed" → surface; no recovery in this slice (Phase 1 couldn't even land the outline).

4.5 Build citation-manifest.json from the composer's records

The composer wrote a raw-text citation-records file (<project_path>/.metadata/citation-records-v<N>.txt), never JSON — so a draft_sentence containing a straight " (routine in German/FR/IT/ES/PL prose) can't break the manifest. Serialize and self-check the manifest with citation-store.py build. Pass each path as a quoted literal CLI arg — a quoted string is one space-/apostrophe-safe argv element, so quoting alone is sufficient. Do not wrap this in a command-prefix env-var form (RECORDS_PATH=… python3 … --records "$RECORDS_PATH"): the shell expands "$RECORDS_PATH" against the current environment — where the var is still unset — before applying the prefix assignment and exec'ing, so --records would receive an empty string and the build aborts resolving the cwd as the records path.

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/citation-store.py build \
    --records "<project_path>/.metadata/citation-records-v<N>.txt" \
    --draft "<project_path>/output/draft-v<N>.md" \
    --out "<project_path>/.metadata/citation-manifest.json" \
    --draft-version <N> \
    --ingest-manifest "<project_path>/.metadata/ingest-manifest.json"

This is unchanged in --source wiki mode: the synthesized manifest from Step 0 carries each source's real sources: URL in ingested[].url, so the --ingest-manifest URL gate (every inline citation URL must be a known source URL) runs exactly as it does on the web path — pass the same --ingest-manifest "<project_path>/.metadata/ingest-manifest.json". The wiki-only rung gets the same integrity checks as the web path, not fewer.

citation-store.py build parses the records, json.dumps the manifest (ensure_ascii=False — escaping owned by the serializer, never the LLM), asserts every draft_sentence is a verbatim substring of the draft, asserts every inline citation URL is a known ingested-source URL (the --ingest-manifest gate; the composer must copy each cited page's real sources: URL, never reconstruct it from the slug), and round-trips the file it wrote (json.loads + count). Parse the envelope:

• success: true → capture data.citations_count (the authoritative count) and data.claim_kinds (the per-kind breakdown — {distilled, source, answer, null, other}, keyed by claim_id prefix; distilled is the dcl-NNN cross-source-convergence count and answer is the acl-NNN question-node answer-citation count — both surfaced in Step 6 + Step 7) and continue to Step 5.
• success: false, error: "write_failed" → surface error + data (e.g. failed_check: "sentence_not_in_draft" with the offending ids; or failed_check: "url_not_in_sources" with the offending urls — an inline citation URL the composer slug-derived instead of copying the cited page's sources: value) verbatim and stop — do not auto-retry. A sentence the composer claims to have written verbatim is not in the draft it just wrote, the manifest did not round-trip, or a cited URL is not a real ingested source (re-compose).
• success: false, error: "records_not_found" / "draft_not_found" → surface and stop; the composer's write did not land (re-run the composer).

Reconcile the count. Compare the composer's returned citations (its own tally) against data.citations_count (the authoritative manifest length). If they differ, surface a ⚠ citation count mismatch: composer claimed <X>, manifest built <Y> line — a smaller manifest count points at a phantom/truncated records write (records the composer thought it wrote but didn't land), which would otherwise sail through as a silently-undersized manifest. Do not hard-fail (the composer's count is an LLM tally and may be slightly off), but the operator must see the discrepancy.

5. Verify outputs on disk

One Python subprocess validates all three artefacts (draft non-empty + carries a [[sources/ wikilink; citation-manifest parses with schema_version ∈ {"0.1.0", "0.1.1"} and a list-typed citations[] each carrying id / draft_sentence / wiki_slug / claim_id and a draft_sentence that is a verbatim NFC substring of the draft; outline file is on disk). An empty citations[] is NOT a hard fail — it emits a stderr WARN line and surfaces in the final summary per the edge-case section below (zero claims is an upstream-data symptom, not a composer bug). On any structural failure, the subprocess exits non-zero with the assertion message; surface verbatim in the summary and stop — do not auto-retry. Paths go via env vars so spaces / apostrophes in project paths can't break the Python literal:

DRAFT_PATH="<project_path>/output/draft-v<N>.md" \
MANIFEST_PATH="<project_path>/.metadata/citation-manifest.json" \
OUTLINE_PATH="<project_path>/.metadata/writer-outline-v<N>.json" \
python3 -c '
import json, os, sys, unicodedata
from pathlib import Path
draft    = Path(os.environ["DRAFT_PATH"])
manifest = Path(os.environ["MANIFEST_PATH"])
outline  = Path(os.environ["OUTLINE_PATH"])
assert draft.exists() and draft.stat().st_size > 0, f"draft missing or empty: {draft}"
dtext = draft.read_text(encoding="utf-8")
assert "[[sources/" in dtext, "draft contains no [[sources/...]] wikilink"
nfc_draft = unicodedata.normalize("NFC", dtext)
m = json.loads(manifest.read_text(encoding="utf-8"))
schema = m.get("schema_version")
assert schema in ("0.1.0", "0.1.1"), "bad schema: " + repr(schema)
cites = m.get("citations", [])
assert isinstance(cites, list), "citations must be a list, got " + type(cites).__name__
for c in cites:
    assert "id" in c and "draft_sentence" in c and "wiki_slug" in c and "claim_id" in c, c
    # Authoritative gate: citation-store.py already builds + self-checks the
    # manifest, but re-assert here that every draft_sentence is a verbatim (NFC)
    # substring of the draft so a future regression cannot ship a stale surface.
    assert unicodedata.normalize("NFC", c["draft_sentence"]) in nfc_draft, "draft_sentence not in draft: " + repr(c.get("id"))
assert outline.exists(), f"outline missing: {outline}"
if not cites:
    print("WARN: citations[] empty — every cited statement will fail verification", file=sys.stderr)
print(len(cites))
'

The trailing print(len(cites)) is captured for the final summary's citations count; the stderr WARN line is captured separately so the summary can surface the ⚠ Zero citations line documented in the edge-case section below.

5.5 Bounded coverage-gated expansion (capped at ONE, fail-soft — executive is ceiling-bounded + zero-cited-only)

A draft is complete when every sub-question is grounded in the evidence the wiki actually holds — not when it reaches a word count. target_words is a soft upper budget under standard (never a floor) and a ceiling under executive, so a tight draft that grounds every sub-question is the better outcome and must NOT trigger a second LLM call. This step fires ONE bounded, zero-network expansion pass only when a sub-question has ingested evidence the draft never cited — a real coverage deficit with evidence on hand to close it. It cannot pull new evidence (cogni-knowledge is zero-network); it only deepens the named sections from claims already on the wiki, so it fixes "a thin section left grounded evidence uncited", not "wiki too sparse" (the latter routes to more ingestion via knowledge-curate/-fetch). The load-bearing safety net is the accept check below: an expansion that adds no new grounded citation is discarded, so the system can never ship padding.

standard vs executive firing. Under standard density the actuator fires on any thin-or-zero-cited section that maps to a coverage deficit, treating target_words as a soft upper budget. Under executive density it fires conservatively, within the BLUF/Minto-Pyramid ceiling: only the zero-cited deficit sub-questions are eligible (never a thin-but-already-cited section — executive caps length, so it must not pad a section that already has its citation), and only when the draft has headroom below target_words (the executive ceiling). Both the pre-expansion ceiling guard below and the composer's in-pass executive ceiling stop (wiki-composer EXPANSION_MODE) keep the expanded draft from breaching the ceiling, and the one-citation-per-claim executive discipline applies to the expansion pass unchanged (PROSE_DENSITY is threaded into the re-dispatch).

Skip this step entirely (proceed to Step 6 with draft-vN) when ANY of:

• --no-expand was passed → log expansion skipped: --no-expand.
• PROSE_DENSITY is neither standard nor executive (no coverage actuator for any other density) → log expansion skipped: density=<PROSE_DENSITY>.
• This dispatch is already an expansion round (defence-in-depth against a manual re-entry) → log expansion skipped: already an expansion round.
• executive ceiling pre-check. When PROSE_DENSITY == executive, compute BODY_WORDS for <project_path>/output/draft-v<N>.md via the canonical _knowledge_lib.body_word_count(draft, lang) helper (reference list excluded — the same body-word surface Step 7's over-ceiling warning uses) and, if BODY_WORDS >= TARGET_WORDS, skip with expansion skipped: executive ceiling already met (<BODY_WORDS>/<TARGET_WORDS> body words) — the draft is already at or over the ceiling, so deepening it would breach the executive contract. Compute it with the same env-var python3 -c fail-soft pattern the summary uses (never interpolate the path into the literal); an unreadable draft degrades to no-skip (the in-pass composer stop is then the backstop). This pre-check does not run under standard density (no ceiling there).

Compute the coverage deficit (deterministic). Word count plays no role in deciding to expand. Read plan.json + ingest-manifest.json + citation-manifest.json and call _knowledge_lib.coverage_report (the single canonical coverage surface, unit-tested in tests/test_knowledge_lib.sh) to get, per sub-question, the ingested source slugs available / cited / uncited, plus uncited_evidence_sq_ids (the sub-questions with ≥1 uncited ingested source — a coverage deficit WITH evidence to close it). Paths via env vars:

COVERAGE=$(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/compose-coverage.py" coverage-deficit \
    --plan "<project_path>/.metadata/plan.json" \
    --ingest "<project_path>/.metadata/ingest-manifest.json" \
    --citation "<project_path>/.metadata/citation-manifest.json")

compose-coverage.py coverage-deficit calls _knowledge_lib.coverage_report and prints one JSON object — {"uncited_evidence_sq_ids": [...], "zero_cited_sq_ids": [...]} (zero_cited_sq_ids = the sub-questions with no cited ingested source) — captured verbatim into $COVERAGE. It prints nothing and exits non-zero on any error, so the empty-output skip below is the fail-soft path.

If the snippet errors or returns empty (e.g. an unreadable manifest), treat it as no deficit and skip expansion with expansion skipped: coverage report unavailable — the whole step is fail-soft, so a missing measurement never blocks the deposit. If uncited_evidence_sq_ids is empty, skip with expansion skipped: coverage met (every sub-question's ingested evidence is cited) — a short-but-fully-grounded draft is the intended outcome, not a deficit.

Gate on ceiling_hit. If the composer's returned ceiling_hit == true, skip with expansion skipped: at single-call ceiling — raise coverage via more ingestion (knowledge-curate/-fetch) — re-rolling the composer cannot fit more in one call; the fix is more wiki coverage, not a re-roll.

The coverage gate is deliberately independent of the wiki-reviewer advisory Word-Count Gate (finalize Step 10.7): that gate is now a brevity-neutral backstop (it flags only a likely-truncated draft at < 0.50 of budget, never a short-but-complete one), while this actuator fires on a coverage deficit regardless of word count. The two serve different roles and neither tracks the other.

Select EXPAND_SECTIONS from the just-written outline <project_path>/.metadata/writer-outline-v<N>.json — the topical sections (excluding the References section, covers_sub_questions: []) that cover ≥1 sq in uncited_evidence_sq_ids. Under standard density a section qualifies when it is thin (drafted_words < budget × 0.9) or covers a zero-cited sq (a sq in zero_cited_sq_ids). Under executive density the thin path is dropped — only a section that covers a zero-cited sq qualifies (executive caps length, so a section that already carries its citation must never be padded toward a thin-budget threshold; deepen only the thinnest, zero-cited sub-questions). This is conservative by design — the brevity-first intent prefers under-firing (no padding) over over-firing. Paths via env vars (COVERAGE is the JSON the snippet above printed; PROSE_DENSITY is the resolved density):

EXPAND_SECTIONS=$(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/compose-coverage.py" expand-sections \
    --outline "<project_path>/.metadata/writer-outline-v<N>.json" \
    --coverage-json "$COVERAGE" \
    --density "<PROSE_DENSITY>")

compose-coverage.py expand-sections reads the outline + the $COVERAGE JSON and prints the bare comma-list of qualifying section indices. Its selector branches on density == "executive" (only a zero-cited section qualifies — executive caps length, so it never deepens a thin-but-already-cited section) vs standard (a section qualifies when it is thin, drafted_words < budget × 0.9, or covers a zero-cited sq); a section whose covers_sub_questions doesn't intersect uncited_evidence_sq_ids is left alone. The gate fires iff EXPAND_SECTIONS is non-empty. If it is empty, skip with expansion skipped: no thin/zero-cited section maps to the uncited evidence (under executive: expansion skipped: no zero-cited section maps to the uncited evidence) — deepening a section already at budget (or, under executive, one that already carries its citation) would only pad.

Stamp START_MS=$(python3 -c 'import time; print(int(time.time()*1000))') immediately before the re-dispatch (same Option B timing as Step 4). Re-dispatch the composer ONCE at N+1 in expansion mode (same knob values as Step 4). Its purpose is to deepen the named sections from the specific not-yet-cited wiki evidence for their sub-questions, not to close a word count:

Task(wiki-composer,
     PROJECT_PATH=<project_path>,
     WIKI_ROOT=<wiki_root>,
     DRAFT_VERSION=<N+1>,
     EXPANSION_MODE=true,
     BASELINE_DRAFT_VERSION=<N>,
     EXPAND_SECTIONS=<comma-list>,
     TARGET_WORDS=<resolved>,
     PROSE_DENSITY=<resolved>,
     TONE=<resolved>,
     CITATION_FORMAT=<resolved>,
     OUTPUT_LANGUAGE=<resolved>,
     CONTRADICTION_INGEST_PATH=<same as Step 4; OMIT when empty>,
     CONTRADICTION_ACT=<same as Step 4; OMIT when empty>)

Snapshot the canonical manifest before the re-dispatch. A successful N+1 build overwrites citation-manifest.json (to describe v<N+1>) before Step 5 runs, so a copy of the current (vN) manifest is the only way a failed expansion can restore consistent vN state — the same discipline knowledge-verify uses before a revise round (.citation-manifest.pre-r<round>.json):

cp "<project_path>/.metadata/citation-manifest.json" \
   "<project_path>/.metadata/.citation-manifest.pre-expand.json"

On ok: true, fold this dispatch's wall-clock into the accumulator: MAX_DURATION_MS=$(python3 -c "import time; print(max($MAX_DURATION_MS, int(time.time()*1000) - $START_MS))") (fail-soft, same as Step 4), then re-run Step 4.5 (citation-store.py build … --draft-version <N+1> with the same --ingest-manifest gate, writing the canonical citation-manifest.json from citation-records-v<N+1>.txt) and Step 5 (the on-disk verifier) against v<N+1>. Accept check (load-bearing) — keep v<N+1> only when both pass AND the expansion added at least one grounded citation: data.citations_count from the v<N+1> build must exceed the data.citations_count Step 4.5 captured for vN (the one authoritative citation count — len(citation-manifest::citations), never an LLM tally). An expansion that grew the prose but added no new citation is padding → treated as a failure below; words alone never survive. On success, v<N+1> becomes the canonical latest draft — set N := N+1 so Steps 6/7 report on it, then drop the now-stale snapshot:

rm -f "<project_path>/.metadata/.citation-manifest.pre-expand.json"

Cap = 1; fail-soft (the orchestrator must leave vN AND its matching manifest as the canonical state on any failure): if the expansion dispatch returns ok: false, OR its Step-4.5 manifest build is rejected, OR its Step-5 on-disk verify fails, OR it added no new grounded citation (citations_count<N+1> ≤ citations_count<N>) — remove the partial v<N+1> artifacts (output/draft-v<N+1>.md, .metadata/citation-records-v<N+1>.txt, .metadata/writer-outline-v<N+1>.json) so the latest-draft resolver lands back on vN, and restore the snapshot so the canonical citation-manifest.json describes vN again:

mv "<project_path>/.metadata/.citation-manifest.pre-expand.json" \
   "<project_path>/.metadata/citation-manifest.json"

then log ⚠ expansion failed — kept draft-vN (manifest restored) (or ⚠ expansion added no new citation — kept draft-vN (manifest restored) on the no-citation branch) and proceed to Step 6 with vN. The restore is load-bearing precisely in the build-OK-but-verify-fail (and no-citation) window: a successful N+1 build has already overwritten citation-manifest.json to describe the about-to-be-removed draft-v<N+1>, so removing the artifacts alone would leave a stale manifest pointing at a deleted draft (its draft_sentences no longer verbatim substrings of vN), breaking the downstream knowledge-verify/knowledge-finalize read. (After a rejected N+1 build the manifest is still the vN one — citation-store.py build writes only on success — but the unconditional restore is correct there too: it simply moves the snapshot back over an identical file.)

6. Append wiki/log.md

Append one summary line (Bash >> append; wiki/log.md is append-only by cogni-wiki convention):

DATE_STAMP=$(date -u +%F)
TOPIC=<topic from plan.json>
N_WORDS=<words from composer return>
N_CITES=<the script-derived count from Step 5's print(len(cites)) — NOT the composer return's "citations" field>
N_DCL=<data.claim_kinds.distilled from Step 4.5, default 0>
N_ACL=<data.claim_kinds.answer from Step 4.5, default 0>
LOG_PATH=$(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/control-path.py" log --wiki-root "${WIKI_ROOT}")
echo "## [${DATE_STAMP}] compose | project=${TOPIC} draft=v${N} words=${N_WORDS} citations=${N_CITES} dcl=${N_DCL} acl=${N_ACL}" >> "${LOG_PATH}"

The dcl=<n> suffix is the cross-run record of the distilled-citation rate, and acl=<n> the question-node answer-citation rate — the cross-source-convergence loops firing (or not) show up directly in wiki/log.md.

Note on the compose prefix: cogni-wiki's log-format enum (per cogni-wiki/CLAUDE.md §"Key Conventions") does not yet list compose, but readers count unknown prefixes in their catch-all bucket without crashing — compose is additive and safe.

7. Final summary

Print ≤ 11 lines:

• Project: <topic> at <project_path>
• Wiki: <WIKI_ROOT>
• Source: <SOURCE_MODE> (wiki = composed only from the bound wiki + fetch-cache via a synthesized manifest, no web crawl, <INGESTED_SOURCES> wiki sources mapped to the plan; web = <INGESTED_SOURCES> web-ingested sources). The staged --source local|hybrid notice, if any, was already printed at pre-flight.
• Draft: output/draft-v<N>.md (<N_WORDS> words across <N_SECTIONS> sections)
• Citations: <N_CITES> (authoritative count = len(citation-manifest.json::citations), from Step 5)
• Sources (cited vs ingested): print Sources: <X> of <Y> ingested cited (<Z> compounding on wiki), computed fail-soft from the two on-disk manifests — X = distinct ingested sources cited ({c.wiki_slug for c in citation-manifest::citations} ∩ {s.slug for s in ingest-manifest::ingested}, the same field pair the canonical _knowledge_lib.coverage_report intersects), Y = len(ingest-manifest::ingested), Z = Y − X. Compute via the env-var python3 -c pattern (never interpolate paths into the literal), e.g. C="$PROJECT_PATH/.metadata/citation-manifest.json" I="$PROJECT_PATH/.metadata/ingest-manifest.json" python3 -c 'import json,os;c=json.load(open(os.environ["C"]));g=json.load(open(os.environ["I"]));cit={x.get("wiki_slug") for x in c.get("citations",[])};ing={s.get("slug") for s in g.get("ingested",[]) if s.get("slug")};X=len(cit&ing);Y=len(ing);print(f"Sources: {X} of {Y} ingested cited ({Y-X} compounding on wiki)")' 2>/dev/null || echo "Sources: (signal unavailable)". The <Z> uncited-but-ingested pages are the deliberate read-before-web investment — they compound for future knowledge-curate runs (references/differentiation-thesis.md §"The compounding loop"), not waste.
• Distilled citations: <N_DCL> of <N_CITES> (dcl-NNN cross-source convergence cited directly, from Step 4.5's data.claim_kinds.distilled) — 0 on a base with no distilled pages is expected; 0 on a base with distilled pages whose claims show ≥2 backlinks is the inert-loop symptom the operator should notice (the cross-source-convergence evidence is never load-bearing).
• Answer citations: <N_ACL> of <N_CITES> (acl-NNN question-node answers cited directly, from Step 4.5's data.claim_kinds.answer) — 0 on a base whose question nodes carry no answer_claims: is expected; 0 on a base with answer_claims: whose claims show ≥2 backlinks is the inert symptom the operator should notice (same posture as the distilled-citation rate above).
• Contradiction acting (mode C): print the line Contradiction acting: mode-C ON (high-severity recency-survivor preference enabled) only when CONTRADICTION_ACT=1 was threaded (the --contradiction-act opt-in fired with a resolved path). Omit it entirely in the default observability-only run (CONTRADICTION_ACT unset) — the line's presence is itself the audit signal that acting was enabled, paralleling the dcl/acl rates above.
• Outline: .metadata/writer-outline-v<N>.json (outline-recovery anchor; recovery used: <RESUME_FROM_OUTLINE>)
• Expansion (standard or executive density): one of coverage-expansion ran (vN-1 → vN, deepened <sections>) / expansion skipped: <reason> / ⚠ expansion failed — kept draft-vN (manifest restored) / ⚠ expansion added no new citation — kept draft-vN (manifest restored) — from Step 5.5; omit the line only when Step 5.5 did not run at all (a density with no coverage actuator, i.e. neither standard nor executive).
• Cost: $X.XXX (from composer return; accumulate the expansion dispatch's cost_estimate when it ran)
• Next: knowledge-verify will run zero-network claim alignment by reading the citation manifest + each cited page's claim block — pre_extracted_claims[] on a source/synthesis page, distilled_claims[] on a cited distilled page, or answer_claims[] on a cited question node.

Surface a density-aware summary line — but do not auto-retry. The two densities now read oppositely:

• Under PROSE_DENSITY=standard: target_words is a soft upper budget, so a draft under it is the intended brevity-first outcome — emit no under-budget word warning at all. Instead surface a coverage line coverage: <#grounded>/<#sub-questions> sub-questions grounded (grounded = sub-questions NOT in coverage_report's uncited_evidence_sq_ids; reuse the COVERAGE JSON from Step 5.5, or recompute it cheaply when Step 5.5 was skipped via --no-expand). A coverage: line below full is the honest coverage signal — and, when uncited evidence remained, Step 5.5 already attempted to deepen it.
• Under PROSE_DENSITY=executive: target_words is a ceiling, so compute BODY_WORDS for <project_path>/output/draft-v<N>.md with the _knowledge_lib.body_word_count helper (the wiki-reviewer-aligned surface, reference list excluded — body words, not the composer's total words, so the ~1.1k-word bibliography never triggers a false warning) and, if BODY_WORDS is over TARGET_WORDS, print ⚠ Over ceiling (BODY_WORDS/TARGET_WORDS). Under-ceiling is the correct executive outcome — no warning.

Then, under either density, surface a per-sub-question source-coverage breakdown so the operator sees an evidence-breadth gap (e.g. a sub-question whose ingested first-party sources went uncited) before finalize — executive density legitimately caps length, never breadth. The two aggregate lines above (Sources: and the coverage:/ceiling line) hide which sub-questions are thin, and on an executive run Step 5.5's COVERAGE may not be in this scope (it is computed inside the Step-5.5 subshell, and Step 5.5 can short-circuit before computing it — e.g. the executive ceiling pre-check), so compute _knowledge_lib.coverage_report here density-independently. Re-declare the three manifest paths inline (they live only inside the Step-5.5 subshell, not in this scope, exactly as the Sources: line re-declares its C/I) and print, per sub-question with ≥1 ingested (available) source, a sq-NN: <cited>/<available> ingested sources cited line — omitting any sub-question with no ingested evidence (zero available is not a gap, mirroring coverage_report's own exclusion from uncited_evidence_sq_ids). The per_sq[sq].available/.cited values are source-slug lists, so display their len(). Fail-soft via the env-var python3 -c pattern (never interpolate paths into the literal): on any compute error omit the breakdown silently, never blocking the summary.

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/compose-coverage.py" per-sq-coverage \
    --plan "$PROJECT_PATH/.metadata/plan.json" \
    --ingest "$PROJECT_PATH/.metadata/ingest-manifest.json" \
    --citation "$PROJECT_PATH/.metadata/citation-manifest.json" 2>/dev/null || true

compose-coverage.py per-sq-coverage calls _knowledge_lib.coverage_report and prints, per sub-question with ≥1 ingested (available) source, a sq-NN: <cited>/<available> ingested sources cited line under a header (omitting any sub-question with no ingested evidence); it prints nothing on a compute error, so the breakdown is silently skipped fail-soft.

The advisory wiki-reviewer (finalize Step 10.7) independently re-scores the draft — under standard it only flags a likely-truncated draft (< 0.50 of budget), never a short-but-complete one; under executive it caps on excess. The compose-time line is a fast heads-up, not a gate.

8. Record run metrics (phase-exit ledger)

Persist this phase's timing + cost to <project_path>/.metadata/run-metrics.json so the run leaves a durable per-phase ledger (read by knowledge-resume / knowledge-dashboard / a perf study). Capture PHASE_START=$(date -u +%FT%TZ) and init MAX_DURATION_MS=0 at the top of this skill's run (Step 0); the composer dispatches (Step 4, and Step 5.5 when it runs) fold their wall-clock into MAX_DURATION_MS; then at exit:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/run-metrics.py" record \
    --project-path "<project_path>" --phase compose \
    --started-at "$PHASE_START" --ended-at "$(date -u +%FT%TZ)" \
    --agent-count <1 composer, +1 if Step 5.5 expansion ran> \
    --cost-usd <composer cost_estimate.estimated_usd (+ expansion when it ran)> \
    --max-agent-duration-ms <MAX_DURATION_MS — the slowest single composer dispatch (Option B orchestrator-measured)>

Fail-soft — a record failure never blocks the phase. Full contract: ${CLAUDE_PLUGIN_ROOT}/references/run-metrics-wiring.md.

Edge cases

• Outline recovery in action. The outline file exists from a prior crashed run. The composer skips Phase 1 (saves model time and avoids re-deriving the section plan), runs Phase 2 fresh, and writes the draft + citation manifest. The outline's drafted_words placeholders get filled by the resume pass. Surface "RESUME_FROM_OUTLINE=true (outline recovery)" in the summary so the operator sees what happened.
• Re-run with same N. The user explicitly passes --draft-version <N> against an existing draft. The composer overwrites draft-v<N>.md and citation-manifest.json (and re-writes the outline — Phase 1 runs unless writer-outline-v<N>.json is present and RESUME_FROM_OUTLINE=true was inferred). No automatic backup — the user asked for it.
• Empty ingested[] after a re-ingest cleanup. Step 0 aborts with the "no ingested sources" message; do not dispatch the composer against an empty manifest.
• Citation manifest empty. If the composer returns ok: true but citations[] == 0 (every cited page had zero claims — no source pre_extracted_claims:, no distilled distilled_claims:, and no question-node answer_claims: — unusual but possible if the claim-extractor failed across the board), surface as ⚠ Zero citations — every cited statement will fail verification. Do not block — that's an upstream-data issue, not a composer bug.
• Plan changed between ingest and compose. Step 1.2 of the composer aligns covers_sub_questions from ingest-manifest.json (resolved sources carry sub_question_refs[]), so a sub-question added to plan.json after knowledge-ingest ran will have no sources mapped to it. The introduction and conclusion still list it (synthesis sections list all plan.json sub-question ids), but a topical section for that sub-question won't have evidence. Surface in the summary as ⚠ Sub-question <id> has no ingested sources.

Out of scope

• Does NOT verify citations — Phase 6 (knowledge-verify).
• Does NOT deposit the draft into the wiki as wiki/syntheses/<slug>.md — Phase 7 (knowledge-finalize).
• Does NOT modify binding.json — Phase 7 appends the project entry. knowledge-finalize records report_source: wiki for every inverted-pipeline deposit regardless of compose --source (the deposit always lands in the wiki), so the new mode changes nothing here.
• Does NOT re-run any earlier phase. In particular --source wiki does not trigger knowledge-curate/-fetch/-ingest — it synthesizes its manifest from the already-populated wiki (Step 0), which is the whole point of the wiki-only rung (a report with no new web crawl).
• Does NOT implement the local or hybrid source modes yet — they are accepted-but-staged (treated as wiki, with a pre-flight notice); a real local (fetch-cache-only) / hybrid (wiki + bounded top-up curate on uncovered sub-questions) path is a deferred follow-up.
• Does NOT run an unbounded expansion loop or story arcs — the composer is single-pass per dispatch. This skill runs ONE bounded, fail-soft, zero-network coverage-gated expansion (Step 5.5) only when a sub-question has uncited ingested evidence AND the expansion adds a grounded citation. Under standard density it fires on any thin-or-zero-cited deficit section; under executive density (prose_density: executive, BLUF + Pyramid ceiling) it fires conservatively — only for a zero-cited deficit sub-question, only with headroom below the target_words ceiling, and never breaching that ceiling (the pre-expansion guard + the composer's in-pass executive stop). The expansion re-elaborates existing wiki claims only — it never fetches new evidence (that is knowledge-curate/-fetch's job), and never pads toward a word count.

Output

• <project_path>/output/draft-v<N>.md
• <project_path>/.metadata/citation-records-v<N>.txt (composer's raw-text records; input to citation-store.py build)
• <project_path>/.metadata/citation-manifest.json (schema 0.1.1; built by citation-store.py build)
• <project_path>/.metadata/writer-outline-v<N>.json (outline-recovery anchor)
• One new ## [YYYY-MM-DD] compose | … line in <WIKI_ROOT>/wiki/log.md.

References

• ${CLAUDE_PLUGIN_ROOT}/references/inverted-pipeline.md — Phase 5 contract
• ${CLAUDE_PLUGIN_ROOT}/references/claim-at-ingest.md — claim shape on the wiki page
• ${CLAUDE_PLUGIN_ROOT}/references/writing-tones.md — the TONE catalog threaded to the composer
• ${CLAUDE_PLUGIN_ROOT}/references/citation-formats.md — the CITATION_FORMAT menu (ieee/chicago wired)
• ${CLAUDE_PLUGIN_ROOT}/references/absorption-roadmap.md — remaining deferrals (story arcs, author-date citation rendering, expansion loops)
• ${CLAUDE_PLUGIN_ROOT}/agents/wiki-composer.md — dispatched agent
• ${CLAUDE_PLUGIN_ROOT}/scripts/citation-store.py --help — builds + self-checks citation-manifest.json
• ${CLAUDE_PLUGIN_ROOT}/scripts/knowledge-binding.py --help