[Epic] Support OpenAI Responses API for GPT-5 / Codex reasoning models

[HXYerror]

Background

GitHub Copilot's upstream now exposes both /chat/completions and /responses. Codex-family models (gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.3-codex) and o-pro models (o1-pro, o3-pro) are /responses-only on the upstream — they cannot be reached via /chat/completions at all. Other reasoning models (gpt-5, gpt-5.1, gpt-5.2, o1, o3, o4-mini) work on /chat/completions but in a degraded mode: there is no way to round-trip reasoning.encrypted_content, so multi-turn reasoning quality and cost both regress.

copilot-api 0.7.0 is a strict Chat Completions–only proxy. A full audit confirms:

• No /v1/responses Hono route (src/server.ts:17–31)
• No upstream ${copilotBaseUrl}/responses call site (src/services/copilot/)
• Zero matches for reasoning, reasoning_content, reasoning_effort, encrypted_content, output_text, response_id across the entire repo
• Anthropic adapter explicitly suppresses thinking block output (src/routes/messages/non-stream-translation.ts:305)

Impact on users

Use case	Today
gpt-5-codex / 5.1-codex / 5.3-codex / -codex-max	❌ unreachable
o1-pro, o3-pro	❌ unreachable
gpt-5 / o-series multi-turn reasoning	⚠️ encrypted_content lost, re-thinks every turn
Codex CLI (local_shell, apply_patch, previous_response_id)	❌ impossible
Claude Code → Copilot reasoning models with thinking visibility	❌ adapter drops it

Sub-issues (suggested implementation order)

• Add /v1/responses route scaffolding #2 — Add /v1/responses route scaffolding
• Add upstream Responses API service client #3 — Add upstream Responses API service client
• Model-to-endpoint routing: chat-completions vs responses #14 — Model-to-endpoint routing (chat vs responses)
• Preserve encrypted_content for multi-turn reasoning #15 — Preserve encrypted_content for multi-turn reasoning
• Reasoning types & reasoning_effort passthrough in chat-completions path #16 — Reasoning types & reasoning_effort passthrough in chat-completions
• Anthropic /v1/messages -> Responses API adapter (Claude Code on Codex) #17 — Anthropic /v1/messages → Responses API adapter (Claude Code on Codex)
• Emit Anthropic thinking blocks from reasoning channel #18 — Emit Anthropic thinking blocks from reasoning channel
• Streaming reasoning event translation (SSE) #19 — Streaming reasoning event translation (SSE)
• Add Copilot-Vision-Request header for image inputs in Responses path #20 — Copilot-Vision-Request header for image inputs in Responses path
• previous_response_id stateful chain support #21 — previous_response_id stateful chain support
• Tests for Responses path & reasoning fidelity #22 — Tests for Responses path & reasoning fidelity

Dependency graph

       #2 (route) ──┐
                    ├─► #14 (routing) ─► #17 (anthropic→responses) ─► #18 (thinking) ─► #22 (tests)
       #3 (client) ─┤                  │                              │
                    │                  └─► #19 (streaming) ───────────┘
                    ├─► #15 (encrypted_content)
                    ├─► #20 (vision header)
                    └─► #21 (previous_response_id)

  #16 (chat-completions reasoning_effort) — independent, can land anytime

References

• OpenAI Responses API: https://platform.openai.com/docs/api-reference/responses
• OpenAI Reasoning guide: https://platform.openai.com/docs/guides/reasoning
• litellm reference impl: litellm/llms/github_copilot/responses/transformation.py (GithubCopilotResponsesAPIConfig(OpenAIResponsesAPIConfig))
• litellm Anthropic→Responses adapter: litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py
• Relevant litellm PRs: #17130, #19650, #22370, #22448, #25278