repro: BYOK sub-agent with sonnet model fails with Responses API 400

[Copilot]

All 5 pr-processor sub-agents in PR Sous Chef run 28359605525 returned sub_agent_error because claude-sonnet-4.6 does not support the OpenAI Responses API wire format that AWF injects into every agent container.

Failure chain

• AWF docker-compose sets COPILOT_PROVIDER_WIRE_API=responses unconditionally → all Copilot CLI requests use Responses API format
• PR Sous Chef's main agent (gpt-5-mini) overrides the sub-agent model via the task tool's optional model parameter, passing "sonnet" despite the pr-processor agent definition specifying claude-haiku-4.5
• AWF API proxy resolves sonnet → claude-sonnet-4.6 (Anthropic)
• Anthropic rejects: 400 unsupported_api_for_model: model claude-sonnet-4.6 does not support Responses API

Repro workflow (repro-byok-sonnet-subagent.md)

• workflow_dispatch, bare Copilot engine with explicit BYOK config:

  engine:
    id: copilot
    model: gpt-5-mini
    env:
      COPILOT_PROVIDER_BASE_URL: https://api.githubcopilot.com
      COPILOT_PROVIDER_BEARER_TOKEN: ${{ secrets.GH_AW_GITHUB_TOKEN }}
      COPILOT_PROVIDER_WIRE_API: responses

• Main agent calls dummy-sonnet-sub (defined with model: claude-sonnet-4.6) via task with model: "sonnet" override
• Expected result: 400 model claude-sonnet-4.6 does not support Responses API

[pelikhan]

/scout suggest a resolution

[github-actions]

🔭 Recon complete! Scout has charted the territory. Map ready! 🗺️

[github-actions]

🔍 Scout Research Report

Triggered by @github-actions[bot]

Executive Summary

The 400 model claude-sonnet-4.6 does not support Responses API failure is not a Claude bug — it is a known wire-format mismatch when a Responses-API-only client (COPILOT_PROVIDER_WIRE_API=responses) is pointed at a non-OpenAI model. Anthropic models are served only via /v1/messages and an OpenAI chat/completions compat surface, never the OpenAI Responses API. The ecosystem fix is to route wire-API per model: OpenAI family → responses, everything else → chat/completions. This repro is correct.

Click to expand detailed findings

Research Findings

Anthropic does not implement the Responses API

Claude's only OpenAI-compatible surface is /v1/chat/completions; the canonical native endpoint is /v1/messages. There is no Responses-API support, so any responses-wire request for claude-* returns 400 unsupported_api_for_model. Caveats also noted (no prompt caching, JSON via Structured Outputs) ([Claude OpenAI SDK compat docs]((platform.claude.com/redacted)

Identical bug confirmed in LiteLLM — and fixed

unable to call Claude models via Responses API on GitHub Copilot was filed as BerriAI/litellm#20103. Fix in PR #19650: non-OpenAI Copilot models automatically fall back to chat/completions. Workaround until merged: hit /v1/chat/completions directly for Claude. This validates that wire-API is model-dependent, not global.

Copilot routes wire-API per model

gpt-5.1-codex-max etc. require responses; the toolchain tracks a per-model route table (opencode#5866). Forcing one global responses for all models is the defect being reproduced.

Recommendations

• The repro is sound; expect 400 unsupported_api_for_model. Keep it as a regression guard.
• Real fix: make the AWF proxy / wire-API selection per-model — responses only for OpenAI family, chat/completions for claude-*. Mirrors litellm PR [daily regulatory] Regulatory Report - 2026-03-04 #19650.
• Stop-gap: when a sub-agent override resolves to an Anthropic model, unset/override COPILOT_PROVIDER_WIRE_API to chat-completions, or restrict task-tool model overrides to Responses-capable models.

Key Sources

• (platform.claude.com/redacted)
• [Bug]: GitHub Copilot: unable to call Claude models via Responses API BerriAI/litellm#20103
• [BUG]: GitHub Copilot Responses API routing for GPT-5+ anomalyco/opencode#5866

Suggested Next Steps

1. File the per-model wire-API routing fix in the API proxy.
2. Add a guard rejecting responses wire for claude-* before dispatch.

🔭 Intelligence gathered by Scout · 56.6 AIC · ⌖ 4.34 AIC · ⊞ 5.8K · ◷
_{Comment /scout to run again}

[pelikhan]

/scout generate a bug report for copilot cli

[github-actions]

🔭 Recon complete! Scout has charted the territory. Map ready! 🗺️

[github-actions]

🔍 Scout Research Report

Triggered by @github-actions[bot]

Executive Summary

Research confirms the PR's root cause. Anthropic/Claude models do not support OpenAI's Responses API wire format — they use the Anthropic Messages API (with optional OpenAI Chat Completions compatibility). Forcing COPILOT_PROVIDER_WIRE_API=responses against a model that resolves to claude-sonnet-4.6 will legitimately yield 400 unsupported_api_for_model. The repro is valid; the deeper fix is wire-format/provider-type selection per model rather than an unconditional responses injection.

Click to expand detailed findings

Research Findings

Anthropic does not natively support the Responses API

• OpenAI exposes two formats: Chat Completions (legacy, broad) and Responses (OpenAI-specific). Anthropic offers Messages API + an OpenAI Chat Completions compatibility shim — but not Responses. ([Claude Platform — OpenAI SDK compatibility]((platform.claude.com/redacted)
• The ecosystem treats Responses→Anthropic as needing an explicit adapter, not native support (e.g. open LiteLLM request to add Responses support to the Anthropic adapter). So a 400 unsupported_api_for_model for claude-sonnet-4.6 under responses is expected, not a flaky error. (LiteLLM #16215)

Copilot CLI BYOK uses provider type, not a uniform wire API

• Official GitHub docs: Copilot CLI BYOK supports three provider types — openai (Chat Completions), azure, and anthropic — each with its own native protocol. Claude should run under COPILOT_PROVIDER_TYPE=anthropic, which speaks Messages, not Responses. (GitHub Docs — BYOK, [Changelog 2026-04-07]((github.blog/redacted)
• This matches the failure chain: AWF sets COPILOT_PROVIDER_WIRE_API=responses unconditionally, the task model override swaps haiku→sonnet, and the proxy routes a Claude model down the OpenAI Responses path → 400.

Recommendations

• Keep the repro — it isolates a real bug: a single global responses wire setting cannot serve sub-agents that resolve to Anthropic models.
• Fix should make wire API / provider type follow the resolved model: Claude → Messages, gpt-5-mini → Responses. Don't inject responses unconditionally.
• Guard the task model override so an override that crosses providers (sonnet) also flips the wire format, or reject incompatible combos early with a clear error instead of a raw 400.

Key Sources

• GitHub Docs — Using your own LLM models in Copilot CLI: https://docs.github.com/en/copilot/how-tos/copilot-cli/customize-copilot/use-byok-models
• Claude Platform OpenAI SDK compatibility: (platform.claude.com/redacted)
• LiteLLM Responses→Anthropic adapter request: [Feature]: Add OpenAI Responses API support to Anthropic experimental adapter BerriAI/litellm#16215

Suggested Next Steps

1. Confirm the failing run reproduces with the dispatch workflow.
2. Make AWF select wire-api per resolved model rather than globally.
3. Add a precondition error for incompatible model+wire combinations.

🔭 Intelligence gathered by Scout · 62.6 AIC · ⌖ 4.29 AIC · ⊞ 6.5K · ◷
_{Comment /scout to run again}

[github-actions]

🤖 PR Triage — §28376613466

Field	Value
Category	chore (investigation/repro)
Risk	🟡 Medium
Priority	🟢 Low
Score	28 / 100
Action	⏸️ defer
Age	3.4h

Score breakdown: Impact 15 + Urgency 8 + Quality 5

Rationale: Standalone reproduction workflow for the BYOK + sonnet sub-agent 400 unsupported_api_for_model failure. Large addition (1660+/0−, 2 files) but repro/test artifacts only. Draft with no CI. Valuable investigation context but not a fix — defer until the root cause is addressed and a fix PR is opened. This PR documents the failure chain clearly.

Generated by 🔧 PR Triage Agent · 93.2 AIC · ⌖ 14.6 AIC · ⊞ 5.4K · ◷

[github-actions]

🔍 PR Triage — §28395315609

Field	Value
Category	chore
Risk	medium
Score	23 / 100
Action	defer
Batch	—

Score breakdown: impact 10 + urgency 5 + quality 8

Carried over (8.5h). Repro workflow for BYOK+claude-sonnet-4.6 Responses API 400 failure. Investigation/documentation artifact — not a fix. Labels correct. Deferred until root cause addressed in a separate fix PR.

Generated by 🔧 PR Triage Agent · 99.1 AIC · ⌖ 11.6 AIC · ⊞ 5.4K · ◷