R76.H2 Contract — colibri-verification SKILL.md behavioral contract

Task: R76.H2. Artifact under contract: .agents/skills/colibri-verification/SKILL.md (canon) + .claude/skills/colibri-verification/SKILL.md (mirror, byte-identical to canon after this task).

Purpose (unchanged)

The skill provides verification checklists for every stage of a Colibri execution session: pre-execution readiness, execution-time safety, post-execution completion, and per-phase acceptance criteria for GATHER / ANALYZE / PLAN / EXECUTE / VERIFY / CLOSE. It is referenced from every T3 executor chain and from colibri-pm / colibri-executor / colibri-tier1-chains.

The behavioral purpose is not modified by R76.H2. Only the tool names that the checklists call out are rewritten to the Phase 0 shipped 14-tool surface.

Allowed tool set (Phase 0, 14 tools)

Every tool citation on the live path of the rewritten SKILL.md must come from this set — or must drop to a narrative / file-read / git / npm reference:

Family Tool Purpose
β (Task Pipeline) task_create Create a task
β task_list List tasks
β task_get Read one task
β task_update Update fields (including status, which routes via internal state machine)
β task_next_actions Unblocked-task query
ζ (Decision Trail) audit_session_start Start an audit session (includes optional context)
ζ thought_record Record reasoning / plan / decision / reflection (single Phase 0 reasoning-capture tool)
ζ thought_record_list Query recorded thoughts
ζ audit_verify_chain Verify hash chain integrity
η (Proof Store) merkle_finalize Finalize the Merkle tree; seals the session
η merkle_root Retrieve the root hash (Phase 0 attestation primitive)
ε (Skill Registry) skill_list List registered skills
system server_ping Liveness check
system server_health Health/status check

Forbidden tool families (all struck or relabelled HERITAGE)

No live-path reference to any of the following is allowed outside an explicit > HERITAGE: block:

  • roadmap_* (entire family) — not in Phase 0; roadmap is a narrative doc
  • unified_* (entire family: unified_init, unified_set_project, unified_vitals, unified_backup, unified_metrics, etc.)
  • context_* (entire family: context_create, context_verify, context_ensure, context_diff)
  • memory_* (entire family: memory_pack, memory_bundle, memory_verify, memory_gc, memory_smart_store, memory_tier_migrate, memory_cross_session, memory_semantic_search, memory_map)
  • gsd_* (entire family: gsd_plan, gsd_verify_comprehensive, gsd_workflow_create, gsd_agent_pool_create, gsd_agent_spawn)
  • task_eisenhower, task_deps, task_depends_on, task_link_roadmap, task_create_batch, task_transition, task_delete — donor-era task-edge tools (Phase 0 β is exactly 5 β tools)
  • thought_plan, thought_decide — collapsed into thought_record in Phase 0
  • audit_session_set_context — merged into audit_session_start
  • audit_session_end — not in Phase 0; session seals at merkle_finalize
  • merkle_attest — not in Phase 0; attestation = merkle_root hash
  • analysis_* (incl. analysis_recommend, analysis_rag_index, analysis_rag_search, rag_gc)
  • ams_session_resume
  • server_info, server_shutdown — never shipped in Phase 0

Behavioral contract of the rewritten skill

  1. Quick Reference table (line 13-25): identical content; no rewrite.
  2. Pre-Execution Readiness Checklist:
    • Roadmap Validation subsection: renamed to Task Context and rewritten to reference docs/guides/implementation/task-breakdown.md (file read, not a tool call), the colibri-docs-check skill for doc integrity, and task_get for task lookup.
    • Task Validation: rewritten to use task_list + task_get only. unified_set_project row dropped (no project-switching in Phase 0). task_deps row rewritten to “dependencies noted in task description field”.
    • Environment Validation: unified_init row dropped (auto-init on MCP handshake). unified_vitals replaced with server_health. Worktree + branch rows kept unchanged (they reference git, not MCP tools).
    • Audit Session: audit_session_set_context row rewritten to “context arg passed inline to audit_session_start”.
  3. Execution Safety Checklist:
    • Dry-Run First: kept (it is a pattern, not a tool).
    • Worktree Isolation: kept (git-only).
    • Incremental Progress: context_create row dropped. Other rows (task_update, thought_record, error handling) kept.
    • Rollback Preparation: unified_backup / context_create row dropped and replaced with “use git reset + worktree discard; seal prior state via thought_record before the destructive step”.
  4. Post-Execution Completion Checklist:
    • Colibri Writeback: kept (task_update).
    • Thought Record: kept (thought_record).
    • Memory Pack: entire subsection dropped. No memory surface in Phase 0.
    • Merkle Proof: merkle_attest row replaced with “attestation hash retrieved via merkle_root”.
    • Git Checkpoint: kept (git-only).
  5. Phase Acceptance Criteria tables:
    • GATHER: roadmap_list / roadmap_nodes rows dropped. context_create row replaced with thought_record. Other rows kept.
    • ANALYZE: task_eisenhower, task_deps, roadmap_prerequisites_check, analysis_recommend, gsd_plan rows dropped. thought_plan and thought_decide rows collapsed into thought_record.
    • PLAN: task_create_batch, task_link_roadmap, task_deps/task_depends_on, gsd_workflow_create, gsd_agent_pool_create rows cleaned. Only task_create and thought_record remain in PLAN table with narrative cells.
    • EXECUTE: context_create row replaced with thought_record. Other rows kept.
    • VERIFY: roadmap_validate, gsd_verify_comprehensive, context_verify rows dropped. Replaced with audit_verify_chain row and the npm run build && npm run lint && npm test gate as a narrative row (per CLAUDE.md §5).
    • CLOSE: memory_pack row dropped. audit_session_end mention reworded. All rows otherwise kept and cite only the 14-tool set.
  6. Verification Tools Quick Reference (JS example): rewritten to the CLAUDE.md §7 6-call chain: 
    audit_session_start(...)
// ... work ...
audit_verify_chain(session_id)
thought_record(session_id, thought_type: "reflection", ...)
merkle_finalize(session_id)
merkle_root(session_id)

    With a comment noting the ordering rule (final thought_record MUST precede merkle_finalize).

  7. Common Verification Failures table: Roadmap validation fails, Context verify fails, Memory pack empty rows dropped. Replaced with Phase 0-shaped rows: “Audit chain verify fails”, “thought_record ordering wrong”, “Merkle root missing”.
  8. Integration with Tier-1 Chains section: kept — it only references skill names, not tools.
  9. See Also section: kept. colibri-tier1-chains, colibri-pm, colibri-audit-proof all exist. data/tool-contracts/ reference replaced with docs/reference/mcp-tools-phase-0.md (the live Phase 0 tool contract doc).

Frontmatter contract

  • name: colibri-verification — unchanged.
  • description: — kept quoted (YAML safety per Wave H memory lesson on gray-matter; contains colons in Greek: η (eta) segment). Wording lightly tightened to drop donor tool hints but keep η + π Greek-letter labels.

Reality-stamp blockquote at line 9

Replaced with a cleaner R76.H2 stamp:

Phase 0 reality stamp (R76 — 2026-04-18). This checklist skill runs on the Phase 0 shipped 14-tool surface: task_create, task_list, task_get, task_update, task_next_actions, audit_session_start, thought_record, thought_record_list, audit_verify_chain, merkle_finalize, merkle_root, skill_list, server_ping, server_health. Donor-era tool families (roadmap_*, unified_*, context_*, memory_*, gsd_*, task_eisenhower/deps/link_roadmap, thought_plan/decide, audit_session_set_context/end, merkle_attest) are either deferred to later phases or collapsed into the 14-tool set. See ADR-004 R75 Wave H amendment and CLAUDE.md §7 for the canonical verification chain.

Mirror resync rule

After canon is clean, the mirror at .claude/skills/colibri-verification/SKILL.md is overwritten from canon byte-for-byte. This discharges one of the Wave H carry-over drift items for this one skill (other skills remain drifted and are out of scope for R76.H2).

Acceptance gates

  1. rg -n "gsd_|unified_|roadmap_|memory_pack|context_|thought_plan|thought_decide|audit_session_set_context|merkle_attest|task_eisenhower|task_deps|task_link_roadmap|task_create_batch|task_depends_on|audit_session_end|task_transition|server_info|server_shutdown" .agents/skills/colibri-verification/SKILL.md returns zero matches.
  2. diff .agents/skills/colibri-verification/SKILL.md .claude/skills/colibri-verification/SKILL.md is empty.
  3. 4 chain docs exist: audit / contract / packet / verification under docs/…/r76-h2-verification-rewrite-*.md.
  4. npm run build && npm run lint && npm test all green.

Out-of-scope confirmations

  • No edit to any other skill under .agents/skills/ or .claude/skills/.
  • No rename or move.
  • No src/ touch.
  • No new tools invented.
  • No docs-sync to Obsidian vault.
  • No new frontmatter fields.
  • No new dependencies.

Back to top

Colibri — documentation-first MCP runtime. Apache 2.0 + Commons Clause.

This site uses Just the Docs, a documentation theme for Jekyll.