v2.5.5 - the four registered findings, fixed

An override now advances the task instead of demanding the PASS it waives; the handoff checker no longer accepts a negated precedence claim; the documented handoff template passes the documented checker; and both installers actually install, create their target directory, support full and light modes, and finish with a smoke check. CI runs the shell installer for real, because this machine has no bash. 102 engine tests, 8 acceptance, 8 handoff. Gate: 15 steps.

v2.5.4 - the release artifact is checked, and an independent verifier broke it first

An independent verifier (tianshu) attacked the new artifact check and found five defects: the build packed the verifier's own runtime database into the release, entry names and case variants were not scanned, text in a .png was skipped entirely, and a UTF-16-encoded address was fully present yet invisible to the byte scan. All five are closed and pinned by self-tests; the verifier's own adversarial packages are now the regression corpus. The boundary is written down: base64, percent-encoding and HTML entities are not detected. Gate: 15 steps, 89 engine + 8 acceptance + 8 handoff tests.

v2.5.2 - conformance audited, artifact built by script

Conformance matrix (40 requirements: 37 ENFORCED, 1 DOCUMENTED ONLY, 2 OUT OF SCOPE), machine-checked by scripts/check-conformance.mjs; a reproducible release builder that sets install.sh to 100755 and refuses a CRLF artifact; the last MIT statement removed; every shipped file LF. Gate: 14 steps, 89 engine + 8 acceptance + 8 handoff tests.

v2.5.1 - the specification and the runtime agree

Two release blockers, found by verifying the specification against the runtime

Both were the class of defect this project exists to remove: a check that verified what it thought it
verified. Reproduced first, then fixed, then pinned by a test.

• The frontmatter did not parse, and the gate said it did. description was a plain YAML scalar
  containing ": ", which YAML reads as a nested mapping - GitHub rendered mapping values are not
  allowed in this context. The gate looked for name:/description: with a regex, so it reported
  PASS on a document no parser accepts. Now the description is a block scalar, the gate parses the
  frontmatter (scripts/check-frontmatter.mjs), and the self-test mutates it back into the broken
  form and requires rejection.
• The L2 evidence floor was documented, not enforced. completeStage only tested whether the
  evidence list was non-empty, so L2 + a single E3 self-report reached done while the standard
  promised E1 or E2. The floor is a table now (L1 any grade, L2 E1/E2, L3 E1) with tests for the
  refusal and both ways through.

Fidelity fixes

• Approval gates the external step (externalAction.at, default implement), not the whole task,
  and the scope must equal the target - a substring test let "not production" unlock production.
• The domain reviewer must be domain-qualified and now fails closed instead of falling back to any
  third party.
• Roster lookups use Object.hasOwn; a route named constructor no longer resolves.
• install.sh shipped with CRLF, so bash -n failed on the published ZIP. Installers are LF, the
  gate checks line endings, and CI runs bash -n on a fresh checkout.

Verification

13-step gate (it injects eleven known faults and must reject every one), 87 engine tests, 8 acceptance,
8 handoff tests, and a conformance benchmark whose table the gate re-derives on every run.

v1.1.0 — Configurable Activation

What's New in v1.1.0

You now choose when this skill engages:

• global — engages for every task (simplest)
• keyword — engages only when task text matches your keywords (default, token-saving)
• manual — engages only when you explicitly invoke it (most economical)

Keyword mode

• Built-in bilingual trigger words (English + Chinese)
• Exclude keywords act as a veto (e.g. casual chat never triggers)
• Case-insensitive matching, any/all match modes
• New script: node scripts/detect-trigger.mjs --text "<task>" --config config/agents.example.yaml
  • Exit 0 = ENGAGED, exit 1 = NOT_ENGAGED

Files changed

• config/agents.example.yaml — activation block + bilingual keyword list
• scripts/detect-trigger.mjs — new trigger detector
• scripts/validate-config.mjs — validates activation block
• SKILL.md, README.md, CHANGELOG.md — activation docs

License

MIT.

v1.0.0 — Multi-Agent Orchestration Skill

What's Inside

Multi-Agent Orchestration Skill — coordinate 3+ AI agents with capability-based routing, single-writer ownership, independent verification, risk-aware execution, and token-efficient workflows.

• 🧠 Capability discovery: auto-detect agent tools, permissions, models, and cost tiers
• 🔗 Role routing: coordinator, implementer, verifier, environment specialist, domain reviewer
• 🔒 Single-writer ownership per file/resource — no parallel overwrites
• ⚠️ Risk-aware pipelines: routine / important / critical with evidence gates
• ✅ Independent verification + explicit user approval for deploy/publish/send
• 💰 Token efficiency: compact packets, model tiering, parallel read-only checks, caching

Files

• SKILL.md — portable protocol entry
• references/ — protocol, roles, token policy, agent manifest schema
• config/agents.example.yaml — risk/cost/approval configuration
• scripts/validate-config.mjs — local config validation

License

MIT — free to use, modify, and redistribute.