Anthropic /v1/messages → Responses API adapter

[HXYerror]

Part of #1. Depends on #4, #5.

Goal

Let Claude Code (and other Anthropic-shape clients) hit `/v1/messages` with an Anthropic body and have copilot-api dispatch all the way to upstream Copilot's `/responses` endpoint when the target model is Codex/o-pro/etc. Today `/v1/messages` only routes to `/chat/completions` (`src/routes/messages/anthropic-to-openai.ts` and friends).

Current state

The Anthropic adapter chain is exclusively chat-completions:

• `src/routes/messages/non-stream-translation.ts` — Anthropic → OpenAI Chat Completions
• `src/routes/messages/stream-translation.ts` — same, streaming
• No translation layer exists for Anthropic ↔ Responses

`AnthropicMessagesPayload.thinking?` (`src/routes/messages/anthropic-types.ts:21–24`) is parsed but silently dropped.

Tasks

• Add `src/routes/messages/anthropic-to-responses.ts` (request side):
  • Map Anthropic `messages[]` → Responses `input[]`
  • Map Anthropic `tools[]` → Responses tools format (different shape: tools live at top level, not inside messages)
  • Map Anthropic `thinking: { type: "enabled", budget_tokens }` → Responses `reasoning: { effort: "minimal"|"low"|"medium"|"high" }` using budget tiers (litellm's tiers: ≥10k=high, ≥5k=medium, ≥2k=low, else minimal — mirror these)
  • Preserve any prior `thinking` blocks in assistant messages as Responses `reasoning` items with `encrypted_content` from the Anthropic `signature`/`data` field (Preserve encrypted_content for multi-turn reasoning #6)
• Add response side: Responses `output[]` → Anthropic content blocks
  • `reasoning` items → Anthropic `thinking` blocks (see Emit Anthropic thinking blocks from reasoning channel #9 for emission)
  • `message` items → Anthropic `text` blocks
  • `function_call` items → Anthropic `tool_use` blocks
• Wire the model-mode classifier (Model-to-endpoint routing (chat vs responses) #5) into `/v1/messages` so requests targeting `responses`-mode models go through this new adapter instead of the chat-completions path
• Add streaming variant (`anthropic-to-responses-stream.ts`) — see Streaming reasoning events translation #10

Acceptance criteria

• `POST /v1/messages` with `model: "gpt-5.3-codex"` succeeds end-to-end (Claude Code can drive a Codex session)
• Multi-turn calls preserve reasoning (covered by Preserve encrypted_content for multi-turn reasoning #6's acceptance test)
• Existing chat-completions path for non-Responses models is unchanged

Reference impl

litellm: `litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py` — `LiteLLMAnthropicToResponsesAPIAdapter`

[HXYerror]

Implemented in commit c5fd7ab (initial implementation) with security hardening in 4f2c00a and 4c3b10d.

What was built:

• src/routes/messages/anthropic-to-responses.ts: Translates Anthropic /v1/messages payload → Responses API (budget tier mapping, thinking→reasoning, tool_use→function_call, signature→encrypted_content)
• src/routes/messages/handler.ts: Routes codex/o-pro models via Responses API path in handleCompletion
• Adaptive thinking output_config.effort forwarded correctly with allowlist validation
• Media type allowlist prevents data-URI injection
• Prototype pollution guard strips dangerous keys from upstream tool arguments
• Runtime narrowing guard replaces unsafe cast; failed status → 500 error

176 tests passing, lint clean.