Skip to content

v7.166.4

Choose a tag to compare

View all tags
@github-actions github-actions released this 05 Sep 23:28
v7.166.4
e58e383
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What

  • One executable PR-review report contract (issue #2494): the swarm-pr-review skill now publishes the exact mapping from its display verdicts (APPROVE, APPROVE_WITH_NOTES, REQUEST_CHANGES, BLOCK) to the complete_pr_workflow machine report_verdict enum (APPROVE, REQUEST_CHANGES, INCOMPLETE) in both the Merge Recommendation Table and the Final Output ## Verdict section, and advertises INCOMPLETE as the machine-only terminal for PARTIAL/NO_COVERAGE coverage (issue #2383). Previously the advertised values APPROVE_WITH_NOTES/BLOCK were rejected by the registered tool schema with no published bridge, and INCOMPLETE was never advertised. The mapping is a vocabulary bridge only — severity/action-to-verdict policy remains owned by #2491. A frozen acceptance-check battery binds every advertised example to the registered schema (check-ac4).
  • Partial-coverage wording reconciled: Hard Rules 14/15, the Phase 8 critic rule, and the COVERAGE GATE are now scoped to the truthful N-of-6 terminal settlement, so a literal reader is no longer barred from the PARTIAL/NO_COVERAGE exits the runtime requires. Structured discovery receipts stay primary; legacy transcript rows remain labeled deprecated compatibility.
  • Missing-capability honesty before activation: pre-phase-briefing (canonical + Claude mirror) gains a capability gate — probe the session's actual tool list before the first dispatch; when the controller's lane tools are absent (Claude Code/Codex/ZCode), run the same lanes as native parallel subagents (or sequential passes), preserving the partition, settlement gate, and per-lane provenance, and never fabricate batch_id/lane receipts. deep-dive, deep-research, and council (canonicals + mirrors) move their controller-availability check from after-first-failure to before-first-dispatch. The codebase-review-swarm Claude/Codex adapters qualify retrieve_lane_output with a controller-absent alternative, and gate-attribution declares itself controller-only.
  • Prefixed-agent rendered-surface pin: a new durable test (tests/unit/agents/prefixed-architect-pr-review-guidance.test.ts) renders multi-swarm prefixed architects (local_architect, mega_architect) and asserts the MODE: PR_REVIEW guidance, the bundled swarm-pr-review skill reference, and the controller tools in the rendered YOUR TOOLS / Available Tools blocks — closing the rendered-contract capture gap for prefixed agent names.

Why

Issue #2494 (Workstream F2): every shipped first-class skill must execute on its declared hosts or refuse honestly with an available alternative before activating an unreachable workflow. The display/machine verdict split (source of #1965/#2131 residues) could leave a compliant Profile A session unable to complete the workflow through the registered tool, and four shipped skills still activated controller-only dispatch unconditionally on hosts without the controller.

Upgrade notes

No runtime or schema changes; all edits are skill/adapter instruction text plus one new test. The swarm-pr-review entry skill stays at its ratcheted 2024-line baseline. Downstream consumers: #2585/#2586 can now consume the published vocabulary mapping; #2491 remains the owner of severity/action-to-verdict policy tables.

7.166.4 (2026-09-05)

Bug Fixes

  • skills: close PR-review findings on host-executability guardrails (2f4175e)
  • skills: make first-class skills executable across supported hosts (c262842)
Assets 2
Loading