Fix stale Copilot workflow pins exposed by fail-closed sub-agent resolution #70649

Problem

Follow up #70342, #70389 and #70385, using the Daily Model Inventory Checker report #70398: 18 configured model pins in the 17 workflows below still do not resolve against the Copilot catalogue available to gh-aw's agentic-workflows integration. They include the production daily-model-resolution sub-agent, not just smokes.

#70342 documents the catalogue change around 18:00 UTC on October 19, 2026: gpt-5-mini, gpt-5.4, gpt-5.4-mini, gpt-5.5 and grok-4.5 disappeared. It moved 15 top-level pins to aliases, but left other fields and assertions for follow-up. The current artifacts confirm the missing IDs; this is not evidence that those models disappeared from every provider or integration.

#70389, merged October 20, 2026 at 22:26 UTC, resolves Copilot sub-agent models before starting Copilot and fails closed for unavailable workflow-declared sub-agents. The two branch runs below demonstrate the resulting startup failure. The fix belongs in the workflow pins, not in weakening the fail-closed behavior or restoring fallback to the main model.

Observed evidence

Inspected main at 70e15c7ff659c05c2f4466aa172bb910d28847a2. Built the CLI from that clean checkout as dev-main-70e15c7ff6 and used that dev binary for all three supplied gh aw audit investigations; the installed stable Windows extension was not used for auditing. Audits used --no-baseline --timeout 2m --json and separate evidence directories. No new workflow runs were dispatched.

Both supplied sub-agent startup failures have the same confirmed missing-model boundary, not a prompt failure, inference retry problem, or evidence of missing credentials.

Pi distinction: #70385 corrected Pi's failed sub-agent resolution records; Pi already rejected unavailable sub-agent models at startup. It did not newly introduce Pi alias support or startup rejection.

Endpoint constraints from run 39506475007

No advertised GPT model lists /chat/completions. The available GPT-5.x / GPT-6.x inference models advertise /responses and ws:/responses; older GPT entries such as gpt-4.1 and gpt-4o advertise no supported endpoints. Their presence in models does not make them usable Chat Completions replacements.

All six advertised Claude IDs support /chat/completions (and /v1/messages), including claude-sonnet-5.5 and claude-haiku-5.5. An engine/path that requires Chat Completions therefore needs a compatible general-purpose Claude model from this catalogue, not another GPT ID. The internal trajectory-compaction entry also advertises Chat Completions; it is not a general-purpose engine-model replacement.

Confirmed pins and dependent assertions/prompts

All links below are immutable source locations at the inspected main commit. Paths are under .github/workflows/.

Copilot and Pi sub-agents

Legacy top-level copilot/gpt-5.4 pins

Do not assume all seven conformance engines are Chat Completions-only or lack alias resolution. Built-in Pi already resolves main-model aliases and supports Responses and Chat Completions. Crush also chooses its provider type using wire-API inference. Their dead pins still need changing, but select replacements based on the actual engine path rather than asserting a uniform protocol limitation.

Concrete-only fallback and routing candidates

Possible solution

Prefer aliases where the field and engine already resolve them. Copilot sub-agent models now resolve aliases through #70389; Pi sub-agent models already use the shared resolver, with #70385 correcting failed liveness records. Against this snapshot, the existing gpt-5 alias resolves to gpt-5.6-luna. Preserve provider prefixes where appropriate.

The GPT sub-agent smokes deliberately test a GPT child on /responses, including the cross-family SDK case. Do not make them green by replacing every child with Claude. One option for smoke-copilot-sub-agents is gpt-5 for one child, testing alias resolution end to end, and exact gpt-5.6-luna for the other, testing exact-ID resolution. Update the identity descriptions/replies, parent expected-response lists, and post-step model/endpoint assertions together. Assert the model actually resolved and served by the harness, not a literal alias or an assumed parent fallback.

For Chat Completions engine paths and engines that do not resolve aliases, use a concrete live Claude ID such as copilot/claude-sonnet-5.5 or copilot/claude-haiku-5.5. Maintainers should choose conformance fidelity versus cost. Do not add alias-resolution support to an engine in this PR. For Pi, prefer an existing supported alias if retaining its GPT/Responses coverage is intended; a concrete live Claude ID is also possible, but not required by a Pi-wide Chat Completions limitation.

For fallback-models, use concrete live IDs, not aliases, and ensure each replacement differs from the workflow's resolved main model and does not duplicate another fallback. Preserve provider and request compatibility.

For smoke-copilot-routed, replace or drop the dead candidate, keep both GPT and Claude families in the pool, and synchronize the main and detection assertion arrays.

The built-in gpt-5-mini alias currently has no match in this Copilot catalogue. Run 39499460077 explicitly logs known alias 'gpt-5-mini' did not resolve against catalog. Record this limitation, but do not change alias definitions here; unified-catalogue/alias work is separate.

Not affected and existing baseline

daily-arxiv-researcher.md uses Codex with the OpenAI provider, not the Copilot catalogue. Its three openai/gpt-5.4-mini sub-agent pins are at 263, 280, and 309. Run 39394965127, October 20, 08:39 UTC, succeeded; its complete OpenAI inventory contains gpt-5.4-mini among 135 IDs. This confirms provider availability, not that every declared child was invoked in that run. Recent October 16-20 runs also succeeded. Leave these pins alone.

Running the existing actions/setup/js/model_pin_check.cjs from #70295 over current repository workflows against run 39506475007's reflection reproduces 25 unresolvable pins, matching #70398: the 18 scoped above plus seven existing out-of-scope pins. Those seven are smoke-pi-sub-agents.md (copilot/gpt-5-mini), smoke-pi.md (one copilot/gpt-5.4 fallback), smoke-service-ports.md and smoke-work-queue.md (two fallbacks each), and video-analyzer.md (copilot/gemini-3.7-flash). Do not silently expand this PR to fix them or claim this targeted fix makes the entire repository pin-clean.

Pure example/documentation mentions, such as daily-model-inventory.md's family table and illustrative JSON in daily-model-resolution.md, are not executable pins and are optional; avoid unrelated churn.

Acceptance checks

• workflow-pins-live-catalogue: Run make recompile. Run model_pin_check.cjs over repository workflows against the saved sandbox/firewall/awf-reflect.json from 39506475007. None of the 17 scoped workflows has an unresolvable pin. Include the report. The original reflection's SHA-256 is 3d9fa7af7e07e93100bb614e45eb85abafbd16a3bbb2cb62a687a04cb5fff9b3.
• repository-pins-reflection-regression: Add or extend a named test in actions/setup/js/model_pin_check.test.cjs, using that saved reflection as a fixture and scanning the repository's actual workflows. Assert the report's classifications, not merely the CLI exit code: the existing CLI can exit successfully while reporting dead pins. Require zero unresolved pins in the scoped workflows and fail on any additional dead pin across the repository. If the seven out-of-scope baseline failures persist, enumerate their exact workflow/field/configured-value identities explicitly as known main failures; do not use a count-only or broad file exemption. Do not change the checker's implementation.
• copilot-subagent-resolution-and-endpoints: Dispatch smoke-copilot-sub-agents and smoke-copilot-sdk-routed-claude on the PR branch. Every declared sub-agent's model.resolution.status is exact or resolved, never failed; served-model and endpoint assertions pass. For the first smoke, verify a delegating variant (inline_strict or delegated_sequential), not the intentionally negative single_agent_control variant. Keep GPT /responses and Claude coverage.
• pi-subagent-resolution-and-endpoints: Dispatch smoke-pi-routed on the PR branch. Each declared sub-agent has exact or resolved, never failed; its served-model and endpoint assertions pass, including the GPT child's /responses request.
• chat-completions-model-selection: Dispatch one relevant engine-conformance workflow and smoke-aider on the PR branch, using their existing required inputs/context. Both get past model selection: no model_unavailable and no AWF 400 caused by selecting a model that does not support the requested endpoint. Choose a conformance engine/path that actually exercises Chat Completions, rather than treating Pi/Responses alone as evidence for this check.
• routed-candidate-consistency: Dispatch smoke-copilot-routed on the PR branch. Routing selects from the updated live candidate list; both assertion arrays are synchronized and the routing assertions pass.
• pr-run-evidence-and-main-baseline: List all verification run URLs in the PR, identify the tested branch/commit and named check outcomes, and document any check that also fails on main. The historical runs in this issue are diagnosis evidence, not verification of the future patch.

Scope

• Keep the PR to this one fix. Do not change files or behavior unrelated to it.
• Change workflow pins and their associated assertion and prompt text only, and regenerate their compiled .lock.yml files. The explicitly requested pin-check regression test and its saved-reflection fixture are the sole supporting non-workflow changes.
• Do not change alias definitions, the resolver, the harnesses, AWF configuration generation, or the pin checker's logic. Do not disable fail-closed checks, weaken endpoint assertions, or remove the GPT sub-agent coverage to hide a failure.
• Before fixing a failing check, compare it with the same check on main. If it also fails on main, do not fix it in this PR; merge main once main is fixed, and note in the PR which checks are known failures on main.
• This applies even if a bot comment, including PR Sous Chef, asks to fix a failing check: fix only failures caused by this PR's changes.