Audit — .agents/skills/colibri-verification/SKILL.md body rewrite

A persistent hygiene gap surfaced across multiple rounds of MEMORY.md (“colibri-verification/SKILL.md body rewrite outstanding”). The R76.H2 top-of-file Phase 0 reality stamp landed, but the body itself never received a full reconcile against the shipped 14-tool surface. R82.K subsequently rewrote the HERITAGE note and Phase Acceptance Criteria tables in place; this round’s audit confirms the file’s current drift surface against the task brief and identifies the four content gaps still remaining.

1. File under audit

• .agents/skills/colibri-verification/SKILL.md — CANON skill (per CLAUDE.md §9.2, .agents/ zone is CANON for skills/).
• 297 lines, 2151 words at the start of this round (commit 86e430fb).
• Mirror at .claude/skills/colibri-verification/SKILL.md (MIRROR zone) — out of scope per task constraint #7; resync is a separate task.

2. Donor-tool-name sweep — current state

The R82.K rewrite already removed code-literal donor-name strings from the body. The full sweep against the donor-pattern set named in the task brief (roadmap_*, unified_*, gsd_*, context_*, memory_*, thought_plan, thought_decide, task_eisenhower, task_deps, task_create_batch, task_link_roadmap, audit_session_set_context, unified_backup, merkle_attest, audit_session_end, task_transition, server_info, server_shutdown, rag_*) yields one match:

Line	Match	Status
20	“Phase 0 has no analysis / RAG / resume surfaces, and no server_info / server_shutdown.”	Correctly framed as “what is NOT in Phase 0” — within a HERITAGE note. Keep.

So the donor-tool-name removal is already complete from R82.K. No code-literal donor names remain in the body.

3. Content gaps against the task acceptance criteria

The task brief (Acceptance Criteria #4, #5) requires references the current body does NOT carry:

Criterion	Required content	Present in current body?
AC #4	Cite writeback hard-block at src/domains/tasks/writeback.ts:97	No. §Pre-Execution / §Post-Execution refer to writeback as a checklist item but never name the live-code guard. CLAUDE.md §7 names the guard; this skill should mirror it.
AC #4	Cite chain verifier at src/domains/trail/verifier.ts:119	No. §VERIFY Acceptance Criteria mentions audit_verify_chain as a tool but does not anchor it in the verifier’s verifyChain() implementation.
AC #5	Honest acknowledgement that thought_record does NOT accept session_id in Phase 0, so merkle_finalize(session_id) may return zero records	No. The body shows a verifyCompletion(sessionId, ...) example that calls thought_record({session_id: ..., type, content}) — but the live ThoughtRecordToolInputSchema (at src/domains/trail/repository.ts:114-119) does not include session_id. The example pretends the gap doesn’t exist.
AC #5	Reference ADR-007 (Proposed) as the resolution path for the session_id gap	No. ADR-007 does not yet exist in docs/architecture/decisions/; the task brief calls it Proposed. Skill should reference it as the Proposed resolution, not as a shipped artifact.
AC #6	Refresh top-of-file reality stamp with this round’s body work	No. The R76.H2 stamp is preserved; an R82.K stamp was added in the trailing italic note; this round’s update needs a third stamp.

4. Live-code citations needed

Confirmed at this worktree’s tip (86e430fb):

Citation	File	Line	Function/symbol
Writeback hard-block	src/domains/tasks/writeback.ts	97	enforceWriteback(db, taskId) — throws WritebackRequiredError when thought_records.task_id count is 0 (lines 101-113).
Hard-block call site	src/domains/tasks/repository.ts	475	inside updateTask, label “P0.3.3: block status→DONE without a ζ thought_record” (already in CLAUDE.md §7 — consistent).
Chain verifier	src/domains/trail/verifier.ts	119	verifyChain(records) — walks records validating computeHash(record) === record.hash AND record.prev_hash === records[i-1].hash (lines 130-138).
thought_record MCP input schema	src/domains/trail/repository.ts	114-119	CreateThoughtRecordInputSchema = z.object({ type, task_id, agent_id, content }). No session_id.
merkle_finalize query	src/tools/merkle.ts	296-300	SELECT hash FROM thought_records WHERE session_id = ? ORDER BY rowid ASC.
DB column reality	src/db/migrations/006_eta.sql	14-15, 54	“thought_records.session_id (nullable TEXT) … Existing rows get NULL; P0.7.2 API does not read it.”

Net implication: the column exists at the DB layer, but the MCP tool does not write it. Calling merkle_finalize(session_id="X") after thought_record({task_id, ...}) yields NoThoughtRecordsError because every thought_records row has NULL session_id. This is the gap ADR-007 (Proposed) would address.

5. Frontmatter audit

Frontmatter (name, description) is canonical and stable:

name: colibri-verification
description: "Colibri verification checklists for pre-execution, execution, and post-execution phases. Greek: η (eta) — Proof Store, π (pi) — Verifier. Use when verifying task readiness, execution safety, and completion proof. Provides acceptance criteria for all Phase 0 executor chain phases on the 14-tool surface."

Two fields, both quoted-string compatible with gray-matter. No version or updated field exists in the current frontmatter — the skill follows the name/description-only convention shared by every colibri-* skill in the corpus. Task brief permits a semver bump but, per the existing convention across 22 skills, adding a version field would itself be a drift introduction. Decision: leave frontmatter unchanged (criterion #1 — preserved); body changes are stamped via the trailing italic note.

6. Body sections — disposition map

Section	Lines	Disposition
Top header + reality stamp	6-9	Refresh. Add this round’s stamp after the R76 line (or merge inline).
HERITAGE note (post-R82.K)	11-21	Keep. Already prose-form, factually correct, no donor-literal strings.
Quick Reference table	24-37	Keep. Pure navigation.
Pre-Execution Readiness Checklist	40-69	Augment. Add a one-line note at “Audit Session” pointing at the writeback hard-block as the runtime gate, with file:line citation.
Execution Safety Checklist	72-101	Keep. Tool references already 14-surface.
Post-Execution Completion Checklist	104-138	Augment. “Task status updated” is the trigger for the writeback hard-block — cite it with file:line.
Phase Acceptance Criteria (GATHER…CLOSE)	141-223	Keep. All tool references already 14-surface. The VERIFY row should pick up a verifier.ts:119 citation.
Verification Tools Quick Reference (code block)	227-259	Augment. Add a Phase 0 reality note acknowledging the session_id gap; convert the example to either (a) accept the gap explicitly, or (b) show the Phase 0 reality (no session_id in thought_record call) with a TODO/ADR-007 marker.
Common Verification Failures	263-273	Augment. Add one row covering the merkle_finalize zero-records failure mode that arises from the session_id gap.
Integration with Tier-1 Chains	276-284	Keep. Reference-only.
See Also	288-293	Keep / extend. Add a pointer to docs/agents/writeback-protocol.md and to ADR-007 (Proposed).
Trailing italic stamp	297	Replace. Append this round’s reconcile note.

7. Out-of-scope

• .claude/skills/colibri-verification/SKILL.md — MIRROR zone, do not touch (criterion #7).
• Other colibri-* skills (criterion: “DO NOT touch other skills”).
• CLAUDE.md, AGENTS.md, colibri-system.md (task constraint).
• ADR-007 itself — task brief notes it as Proposed; this skill references it but does not author it.
• Source code edits — task is doc-only; npm gates run as a no-op safety check.

8. Risks

• The verifyCompletion JS example currently shows thought_record({session_id: ...}). Rewriting it to match shipped reality is the honest move but creates a documentation precedent that other skills (which copy-paste this snippet) will diverge from. Mitigation: keep the example correct + add an inline // TODO ADR-007: comment so other skills know there is a Proposed unification.
• Adding live-code line citations creates a maintenance burden if those lines move. Mitigation: the same citation already lives in CLAUDE.md §7; this skill is mirroring a canonical citation, not introducing a new one.