chore(main): release 7.181.3

[github-actions]

Instruction selection and caching with paired outcome evidence (#2672)

What changed

• The architect knowledge-injector's context cache key now includes a
  payload-input fingerprint covering every input the cached instruction text
  embeds: the curator briefing, rejected lessons, the run-memory summary,
  recent escalations, and the latest curator drift report (each read once per
  invocation and shared with the assembly path). Previously a changed
  briefing or drift report was re-served stale while the conversational
  context was unchanged; now any instruction-set change invalidates the
  cache. Unrelated event-file churn does not invalidate (escalation content
  is hashed, not file stamps).
• New paired cached-vs-uncached evaluation:
  /swarm memory evaluate --instruction-pairing runs deterministic offline
  tasks through the real injector hook on a cold cache (regeneration
  reference) and a warm cache replaying identical context, reporting per pair
  the quality outcome (which directive labels survived selection), per-arm
  latency, cache reads, uncached cost, and the rendered prefix length at the
  host-renderable carrier boundary. The durable report lands at
  .swarm/memory/instruction-pairing-report.json. Negative results are
  retained rows (negative_result: true when caching did not improve the
  paired quality outcome), and the report deliberately contains no
  percentage or savings fields — every quantity names its measurement
  denominator. The report's instruction_set_digest is the handle a
  HarnessOpt lineage record ([Workstream F] PR 11 of 21: Ship the governed HarnessOpt capstone with held-out validation #2503, the governed held-out comparison owner)
  can reference; this command feeds that harness evidence without
  duplicating it.
• New measured reachability dispositions for the six bundled skills named by
  the issue (ci-failure-batching, gate-attribution,
  merge-queue-readiness, skill-edit-validation,
  worktree-retry-cleanup, parallel-work-check) in
  src/config/bundled-skill-dispositions.ts: all six are reachable with
  verified consumer references (including the Claude-side commit-pr and
  editing-skills adapters, which previous closure scans did not visit).
• Six named consumer-control tests plus guards: a skill with no references
  and no retirement fails the control (a missing literal search hit can
  never delete a skill), and retirement requires full inventory parity. The
  bundled-skill runtime-closure test now scans all consumer trees
  (.opencode, .claude, .agents, src), closing the scan-root gap.
• Docs: the instruction cache key and invalidation inputs, the pairing
  measurement denominators, and the six-skill disposition table
  (docs/configuration.md, docs/skills.md, docs/commands.md).

Why

Instruction-selection and cache behavior shipped without paired outcome
evidence, the instruction cache could serve a superseded instruction set
after its inputs changed on disk, and bundled-skill consumption was asserted
only by inventory lists and a closure scan rooted at .opencode/skills
that missed two consumer trees. Issue #2672 (Workstream E PR 09 of 09)
required all three to be measurable and guarded.

How to use it

Nothing to configure. Run /swarm memory evaluate --instruction-pairing for
the paired report; the injector cache now invalidates automatically when any
embedded instruction input changes. Skill dispositions are verified by the
consumer controls on every test run — retiring a skill remains a deliberate,
full-parity act.

Caveats

• On the offline deterministic corpus both pairing arms run the same
  selection algorithm, so quality_outcome is identical by design — the
  honest negative result the issue contract asks to retain; latency and cost
  deltas carry the efficiency signal in absolute milliseconds.
• The fingerprint adds bounded reads (briefing/rejected/run-memory/
  escalations/drift) to the cache-hit path; the paired report quantifies the
  remaining cache benefit rather than assuming it.

---

fix(agents): validate final enabled swarm model selections and fallbacks during preflight

Issue: #2680

What

• The model preflight now validates the final effective model selection of every enabled role — legacy unprefixed and multi-swarm prefixed (swarms.<id>.agents.<role>.model) — by reusing the shipped exact-name resolution helpers, instead of only reading top-level agents overrides plus a blanket DEFAULT_MODELS pass that missed per-swarm overrides and warned about roles that cannot dispatch.
• Disabled roles and feature-gated optional roles (council.*, docs_design, designer while their flags are off) are never collected, so preflight warnings carry no consumer-default noise; primary (architect) roles classify as host-controlled and stay silent — OpenCode's UI owns their selection.
• New distinct bounded outcome classes with actionable diagnostics: missing-selection (enabled role with no final selection — catalog-independent) alongside the existing provider/model unresolved classes; explicit fallback_models entries are validated as their own recorded class and are never synthesized.
• Dispatch admission now covers every registered swarm agent, not just critics: a Task dispatch whose final selection is positively unresolved is denied with SWARM_AGENT_MODEL_UNRESOLVED (critics keep the PLAN_CRITIC_MODEL_UNRESOLVED identity); a role with no final selection is denied with SWARM_AGENT_MODEL_MISSING_SELECTION. Primary agents are exempt (short-circuit before any catalog lookup) and an unreachable catalog never denies a dispatch — a catalog warning alone is not dispatch-failure evidence.
• /swarm doctor reports failing selections with configured/enabled/resolved/fallback distinctions and per-class operator actions (bounded to 20 rows).

Why

A per-swarm model override pointing at a missing provider or model was invisible to every preflight surface: the startup warning stayed silent, /swarm doctor missed it, and the first observed failure was a permanent dispatch error with no typed preflight class. The #2271 preflight was built for top-level-config-only deployments and never followed multi-swarm generation.

Notes

• collectConfiguredAgentModels remains exported unchanged for compatibility; runModelPreflight accepts an optional options object carrying the live generated-name registry.
• The deferred catalog fetch stays bounded (2 s timeout, 30 s TTL cache) and fail-open; no network call was added to plugin startup.
• Issue [Workstream H] PR 11 of 15: Refuse unregistered lane agents at dispatch time (bare canonical roles on multi-swarm hosts) #2614's unregistered-lane dispatch refusal is untouched.
• Documentation: new "Model preflight" section in docs/configuration.md covering configured vs enabled vs resolved vs fallback, prefixed names, optional-role gating, and operator actions.

🤖 I have created a release beep boop

7.181.3 (2026-09-15)

Bug Fixes

• agents: validate final enabled swarm model selections and fallbacks during preflight (#2680) (64a7f64)
• agents: validate final enabled swarm model selections during preflight (#2680) (0f75c2d)
• ci: register the pairing-report stream and canonicalize test tmpdirs (#2672) (4bea76b)
• knowledge: pair instruction-selection evidence and close cache/consumer gaps (#2672) (8dbab19)
• memory: resolve PR #2781 review findings PRR-001..016 (5039c3c)
• memory: satisfy G2 cache-invalidation scan and #2526 system-role ratchet (#2672) (40afb4f)
• review: close PR #2782 review findings (sanitize sinks, class alignment, cache epoch) (616bd9a)
• tests: complete mocked bindings for the widened memory-barrel graph (#2672) (0c4d07d)
• tests: make consumer-control path check POSIX-portable (#2672 final-critic F2) (5bb5338)

---

This PR was generated with Release Please. See documentation.

[github-actions]

Drift check report

Found 2 drift finding(s): 0 error, 0 warning, 2 notice.

required-check-contract (2)

• 🔵 notice scripts/required-check-contract.json: [RULESET_DIVERGENCE] intended-required context "drift" is not yet required by the captured ruleset
• 🔵 notice scripts/required-check-contract.json: [RULESET_DIVERGENCE] intended-required context "drift" is not present for every expected event in captured external workflow evidence

[github-actions]

🤖 Created releases:

• v7.181.3

🌻