Audit — R77.C .claude/skills/ mirror resync (3 drifted skills)

Round: R77 (pages-finish) Task: R77.C Tier: T3 executor Branch: feature/r77-c-mirror-resync Worktree: .worktrees/claude/r77-c-mirror-resync Base: origin/main @ ea5055e0 Date: 2026-04-18

1. Purpose

R76 Wave 1 (PR #153) resynced 9 drifted .claude/skills/colibri-* mirrors from .agents/skills/ canon. Subsequent R76 waves (H2 colibri-verification rewrite, H3 HERITAGE tagging, Phase B seal) edited canon SKILL.md files but the mirror was not re-pushed. This audit inventories the residual drift across 3 skills and confirms scope before the contract/packet/implement steps.

2. Scope

Three skills are in scope for R77.C. All three exist in both .agents/skills/ (canon) and .claude/skills/ (mirror) and have byte-divergent SKILL.md files:

colibri-mcp-server — 6 files (SKILL.md + agents/openai.yaml + references/{errors,tools,workflow}.md + scripts/setup-claude.ps1)
colibri-tier1-chains — 2 files (SKILL.md + agents/openai.yaml)
colibri-growth-strategy — 2 files (SKILL.md + agents/openai.yaml)

No other Colibri skill is in scope for R77.C.

3. File inventory — canon vs mirror

File trees are structurally identical between canon and mirror in every case; only the SKILL.md body diverges. There are zero mirror-only UX files to preserve.

3.1. colibri-mcp-server

Canon .agents/skills/colibri-mcp-server/ (6 files):

SKILL.md (182 lines)
agents/openai.yaml
references/errors.md
references/tools.md
references/workflow.md
scripts/setup-claude.ps1

Mirror .claude/skills/colibri-mcp-server/ (6 files):

SKILL.md (176 lines) — DIFFERS
agents/openai.yaml — identical
references/errors.md — identical
references/tools.md — identical
references/workflow.md — identical
scripts/setup-claude.ps1 — identical

Mirror-only files: none.

3.2. colibri-tier1-chains

Canon .agents/skills/colibri-tier1-chains/ (2 files):

SKILL.md (262 lines)
agents/openai.yaml

Mirror .claude/skills/colibri-tier1-chains/ (2 files):

SKILL.md (256 lines) — DIFFERS
agents/openai.yaml — identical

Mirror-only files: none.

3.3. colibri-growth-strategy

Canon .agents/skills/colibri-growth-strategy/ (2 files):

SKILL.md (221 lines)
agents/openai.yaml

Mirror .claude/skills/colibri-growth-strategy/ (2 files):

SKILL.md (217 lines) — DIFFERS
agents/openai.yaml — identical

Mirror-only files: none.

4. Diff character

All three divergences are canon-ahead-of-mirror drifts consisting of post-Phase-0 reality-stamp additions authored during R76 (H2/H3 waves). The mirror contains an older revision of the same SKILL.md body.

4.1. colibri-mcp-server (9 diff lines, canon +6 lines)

Canon adds three blockquote blocks not present in the mirror:

A “Post-Phase-0 reality stamp (2026-04-19, R76.H3)” blockquote after the top-of-file banner, pointing at ADR-004 for the authoritative 14-tool surface.
A “Phase 0 note” blockquote calling out donor-era tool families (gsd_*, ams_autonomous_*, roadmap_*, workflow_*, analysis_rag_*, observe_*, watcher_*, unified_*, plus memory_pack / memory_bundle) as out-of-scope for Phase 0 per ADR-004/ADR-005.
A second “Phase 0 note” blockquote clarifying that gsd_* / ams_* / roadmap_* / analysis_rag_* / unified_* / workflow_* families plus skill_get / skill_reload (Phase 1) and agent_spawn* (Phase 1.5) are not shipped.

Signal: purely doc-level HERITAGE tagging. No behavioral change.

4.2. colibri-tier1-chains (25 diff lines, canon +6 lines)

Canon adds two new blockquote blocks and inline ~~name~~ <sup>(Phase 0: deferred…)</sup> strike-through annotations on the following inline tool references within existing paragraphs:

context_create, analysis_rag_search, task_eisenhower, analysis_recommend, gsd_plan, task_depends_on, gsd_validate, gsd_agent_spawn.

Two new blockquotes:

A “Post-Phase-0 reality stamp (2026-04-19, R76.H3)” after the top banner.
A “Phase 0 note” before the bullet list of donor families (unified_*, gsd_*, roadmap_*, analysis_*, memory_*, context_*, observe_*).

Signal: HERITAGE tagging + inline strike-through of donor names; no behavioral change.

4.3. colibri-growth-strategy (25 diff lines, canon +4 lines)

Canon adds:

A “Post-Phase-0 reality stamp (2026-04-19, R76.H3)” blockquote adjacent to the existing “Phase 0 note” block.
Inline ~~name~~ <sup>…</sup> strike-throughs on task_create_batch, task_link_roadmap, task_deps, unified_set_project, roadmap_nodes.
A HERITAGE do-not-invoke tag on the colibri-roadmaps-tasks reference in the “Related skills” section, plus an explanatory blockquote.

Signal: same HERITAGE tagging + inline strike-through pattern as tier1-chains; no behavioral change.

5. Preservation requirement

Per CLAUDE.md §9.2 the mirror must not be hand-edited. Since none of the three mirror skills contain UX-specific files beyond what canon provides, this resync is a pure canon-wins overwrite — no mirror-side content needs to survive the operation.

6. Verification command (pre-resync)

Runs at the repo root of the worktree:

for skill in colibri-mcp-server colibri-tier1-chains colibri-growth-strategy; do
  diff -qr ".agents/skills/$skill" ".claude/skills/$skill"
done

Expected output (confirmed at audit time):

Files .agents/skills/colibri-mcp-server/SKILL.md and .claude/skills/colibri-mcp-server/SKILL.md differ
Files .agents/skills/colibri-tier1-chains/SKILL.md and .claude/skills/colibri-tier1-chains/SKILL.md differ
Files .agents/skills/colibri-growth-strategy/SKILL.md and .claude/skills/colibri-growth-strategy/SKILL.md differ

All three are in scope. None is trivially clean.

7. Out of scope

Any edit to canon (.agents/skills/*). Canon is the source of truth and is untouched by this task.
Any edit to other .claude/skills/colibri-* skills not listed in §2.
Any edit to src/ code. R77 ships no src/ code per the round manifest.
Any tool-surface change. Zero MCP tools touched.
Mirror-policy rewrite (CLAUDE.md §9.2 stands as-is).

8. Next step

Proceed to contract (docs/contracts/r77-c-mirror-resync-contract.md) to state the canon-wins resync policy and declare the per-file operations authorized.