Copilot sub-agents fail with 400 when their model needs a different API than the main model (Claude under GPT, GPT under Claude) #66280

Summary

The Copilot engine runs inside AWF with a BYOK provider pointed at the API proxy. In that mode,
Copilot CLI sends every request in the session through one wire API, responses or
completions, and that includes sub-agents. Copilot's API doesn't serve Claude models on
/responses, and it doesn't serve GPT-5 family models such as gpt-5.4-mini on
/chat/completions. So an inline or imported sub-agent whose model: is from the other family
fails with an upstream 400 on every call:

• a Claude sub-agent under a GPT-5 main model;
• a GPT-5 sub-agent under a Claude main model.

This happens with and without model routing, in the default CLI mode. The inline sub-agents
reference uses model: claude-haiku-4.5 as its example, and that fails under any GPT-5 main
model.

#66234 makes this more visible under model routing. Declared sub-agent models are now admitted
by the request policy, but the main model is chosen per run. A workflow with a fixed-model
sub-agent therefore works on some routed arms and fails on others, and the author can't tell
which in advance.

Evidence

The test repository runs gh-aw main 3fd44f8dc1 (compiled with --gh-aw-ref) and AWF 0.28.35
images on GitHub-hosted runners.

• The workflow has two inline sub-agents: file-summarizer, with model: claude-haiku-4.5, and
  quick-checker, with a GPT model.
• The task tells the agent to delegate one question to each and not to answer them itself.

Upstream error messages:

• 400 model claude-haiku-4.5 does not support Responses API.
• 400 model "gpt-5.4-mini" is not accessible via the /chat/completions endpoint

These are recorded in each run's artifacts:

• the Copilot session events, as subagent.failed with the sub-agent name and model;
• the API proxy's upstream-errors.jsonl;
• for routed runs, model-routing.jsonl.

They aren't policy rejections: the request policy admitted these models (no 403
model_policy_violation).

The harness logs show where the session's wire API came from:

• GPT main, not routed: COPILOT_PROVIDER_WIRE_API already set to responses — skipping auto-configure.
• Routed: inference routing: mode=awf-routed model=gpt-5.6-luna effort=high wire_api=responses,
  and ... model=claude-sonnet-5 effort=medium wire_api=completions for the Claude arm.
• Claude main, not routed: no wire API is set, and Copilot CLI uses completions for the session.

Cause

• With AWF, the agent always runs Copilot CLI in BYOK mode: COPILOT_PROVIDER_BASE_URL points
  at the API proxy, for credential isolation. A BYOK provider has a single wire API, set by
  COPILOT_PROVIDER_WIRE_API or Copilot CLI's default.
• Three places set it from the main model only:
  • AWF's copilot-credential-env.ts sets responses whenever COPILOT_MODEL is GPT-5 family;
  • gh-aw's applyCopilotWireAPI in copilot_harness.cjs does the same when it's unset;
  • under routing, applyCopilotRoutingSelection sets it from the selected arm's endpoint.
• Sub-agents spawned by the task tool reuse the session's provider, so their requests use the
  main model's wire API, whatever their own model needs.
• gh-aw already decides the wire API from the model name: from the catalogue's wire_api in
  models.json when present, otherwise from the GPT-5 family rule in applyCopilotWireAPI. So
  the mismatch can be detected at compile time.

Proposed plan

1. Give each model its own wire API.
   • Configure the agent's BYOK provider so that each model uses the wire API it needs.
   • One way: two providers pointing at the same API proxy, one for responses and one for
     completions, with each model (main, sub-agent, and routed candidate) mapped to the
     provider that matches its catalogue wire_api.
   • The SDK driver path already builds a multi-provider config from AWF's reflect data
     (resolveMultiProviderFromReflect). Check whether CLI mode can use the same mechanism, and
     whether Copilot CLI supports a per-model provider in CLI mode.
   • If Copilot CLI can't do this, raise it with the Copilot CLI team. Use steps 2 and 3 in the
     meantime.
2. Warn at compile time. When a sub-agent's model needs a different wire API from the main
   model, emit a compiler warning naming the sub-agent and both models.
   • Under routing, compare against every routing candidate and warn when any candidate differs.
     That routed arm would break the sub-agent.
   • Use the catalogue's wire_api, falling back to the GPT-5 family rule the harness already
     uses.
3. Document the constraint in the inline sub-agents reference until step 1 lands.
   • Its claude-haiku-4.5 example fails under any GPT-5 main model.
   • State that a sub-agent's model must use the same API family as the main model, or (under
     routing) as every allowed model.
4. Tests.
   • Harness: a session with a GPT main model and a Claude sub-agent model resolves a provider
     with the right wire API for each.
   • Compiler: warnings for a Claude sub-agent under a GPT model, for a GPT sub-agent under a
     Claude model, and for a routed workflow whose candidates span both families.
   • The existing smoke test "Smoke Copilot Sub Agents" covers one direction in SDK mode only.
     Add a CLI-mode case.

Open questions

• SDK mode. "Smoke Copilot Sub Agents" passes daily with copilot-sdk: true, a
  gpt-5.3-codex main model, and a claude-haiku-4.5 sub-agent, so SDK mode may already handle
  this. We couldn't test SDK mode in our repository: the SDK driver exited with No GitHub OAuth token or Copilot HMAC key provided, because the workflow authenticates with github.token and
  copilot-requests: write and has no Copilot token. Is SDK mode the intended path for
  mixed-family sub-agents, and does it work under routing?
• Ownership of the default. AWF sets the GPT-5 default in the agent environment, and gh-aw
  sets it again in the harness. Whichever fix is chosen should leave one owner for the wire API
  decision.

Notes

• Related: Model routing: let sub-agents use their own models by emitting routing.candidateModels #66231 / Separate routing candidates from sub-agent model policy #66234 (sub-agent models in the request policy), [aw-failures] [aw-fix] P1: Smoke Copilot Sub Agents — subagents pinned to claude-haiku-4.5 hard-fail with CAPI 400 (model does not support R [Content truncated due to length] #46326 (same error in July,
  expired without a fix).