On October 19, gpt-5-mini, gpt-5.4, gpt-5.4-mini, gpt-5.5 and grok-4.5 were removed from the Copilot catalogue. Since #70389, Copilot refuses to start when a workflow-declared sub-agent model doesn't resolve. Run 39506475007's catalogue (awf-reflect.json) shows 25 dead pins across the repository's workflows. This PR fixes the 18 from the issue and 6 same-kind pins that a maintainer added to scope. It leaves one: video-analyzer's copilot/gemini-3.7-flash, which is the owner's decision.

Alias definitions, the resolver, the harnesses, AWF configuration and the pin checker are unchanged.

How replacements were chosen

• Aliases where the field and engine resolve them: Copilot and Pi sub-agents, and top-level model: on Pi.
• Concrete live IDs for fallback-models, routing candidates, routed-smoke sub-agents, and engines that don't resolve aliases.
• copilot/claude-haiku-5.5 for every engine path that uses Chat Completions, since no GPT model in this catalogue advertises it. I picked haiku over sonnet so these workflows stay within their existing max-ai-credits limits (5 and 20).
• GPT sub-agent coverage on /responses is kept in every sub-agent smoke.

Sub-agents

Notes:

• run-analyzer moves from GPT to Claude. No small GPT model remains in the catalogue, and the built-in gpt-5-mini alias still finds no match.
• Routed smokes use exact IDs. The shared routing assertions match requests to sub-agents by a fixed model name, and their C1 check requires each sub-agent model to differ from the routing candidates. In smoke-pi-routed the only candidate is gpt-5.6-luna, which is why its sub-agent uses gpt-5.3-codex.
• smoke-pi-sub-agents deliberately pins exact IDs, so it stays exact.

Sub-agent assertions

• smoke-copilot-sub-agents: the post-step now reads each sub-agent's model.resolution record. It requires status resolved for the alias child and exact for the others, then checks the served model and endpoint against the resolved model rather than the literal alias:

  const expected = [
    ["haiku-whoami", "claude-haiku-4.5", "exact", "/chat/completions"],
    ["gpt5-whoami", "gpt-5", "resolved", "/responses"],
    ["luna-whoami", "gpt-5.6-luna", "exact", "/responses"],
  ];

• smoke-copilot-sdk-routed-claude: resolutionRoles: ["mini-whoami"], so a successful resolution of the sub-agent is now required.
• Descriptions, replies and the parent's expected-response lists are updated to match.

Top-level copilot/gpt-5.4 pins

Fallbacks and routing

Each fallback differs from the workflow's main model and from the other fallback.

smoke-copilot-routed: dropped gpt-5.4-mini, so allowed-models is [gpt-5.6-luna, claude-haiku-4.5] (GPT and Claude). Both allowedModels assertion arrays match it.

Regression test (repository-pins-reflection-regression)

• Fixture: actions/setup/js/test_data/awf_reflect/run-39506475007-awf-reflect.json, a byte-for-byte copy of the run's awf-reflect.json. SHA-256 is 3d9fa7af7e07e93100bb614e45eb85abafbd16a3bbb2cb62a687a04cb5fff9b3; it is checked for credentials and its source is recorded in that folder's README.
• Scans the repository's real workflows and asserts on the report's classifications, not the CLI exit code:
  • each of the 24 fixed workflow/field pairs is exact or alias;
  • the 21 affected workflows have no unresolvable pins;
  • the repository's unresolvable pins are exactly {video-analyzer.md, model, copilot/gemini-3.7-flash}.
• Other tests updated because they read the changed workflow text: smoke_copilot_sub_agents.test.cjs, pi_subagent_extension.test.cjs, goose_runtime_test.go, prompts_test.go.

Verification still needed (maintainer dispatch on this branch)

None of these smoke runs has happened on the branch yet.

Known failures on main

• Go lint: three modernize (embedlit) findings in claude_engine.go, codex_engine.go and comment_memory.go. This PR doesn't touch those files.
• smoke-copilot: a separate shell/URL-denial failure, out of scope here.

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