feat(proxy): serve the Anthropic Messages API

[OnlyTerp]

Summary

Makes Local Studio a native ANTHROPIC_BASE_URL target — the integration llama.cpp (Jan 2026), Ollama, and LM Studio all shipped this year because Claude Code and other Anthropic-wire-protocol agents became the default way people use local models for real work.

• POST /v1/messages — translates Anthropic requests to OpenAI internally (llama.cpp's approach) and forwards to the running backend, so it works for every engine. Handles: system (string or blocks), text/image(base64+url)/tool_use/tool_result content blocks, tools + tool_choice (auto/any/tool), stop_sequences, temperature/top_p.
• Streaming — full Anthropic SSE event sequence (message_start → content_block_start/delta/stop → message_delta → message_stop) translated from OpenAI chunks by a pure state machine, including input_json_delta for streamed tool calls and text→tool block transitions. 15s ping keepalive (same Cloudflare-prefill rationale as the OpenAI path).
• POST /v1/messages/count_tokens — flattens the request and tokenizes upstream, with a fallback chain (/tokenize → TabbyAPI's /v1/token/encode).
• Observability for free: requests through /v1/messages are recorded by the existing inference accounting — per-model requests, tokens, TTFT show up in the usage page. None of the other local Anthropic endpoints have this.
• Errors are Anthropic-shaped ({type:"error",error:{...}}), and the model-not-running guard behaves exactly like the OpenAI route (proxy never launches models).

Live validation — RTX 5090, TabbyAPI/ExLlamaV3 backend (branch of #168)

$ ANTHROPIC_BASE_URL=http://127.0.0.1:8188 ANTHROPIC_AUTH_TOKEN=local-studio \
    claude -p "Reply with exactly: LOCAL STUDIO OK" --model llama32-1b-exl3
"LOCAL STUDIO OK"        # real Claude Code v2.1.179, end to end

• Non-streaming /v1/messages: correct message shape, stop_reason mapping
• Streaming: verified full event sequence; message_delta carried real usage ({"input_tokens":39,"output_tokens":19})
• count_tokens: {"input_tokens":27} via the TabbyAPI fallback
• Usage page after the Claude Code turn: llama32-1b-exl3 | requests: 3 | tokens: 27171 | avg TTFT 931ms

Test plan

• 10 unit tests: request translation (system/blocks/tools/tool_choice/images/tool_result splitting), response translation (text + tool_use + stop reasons), stream state machine (text stream, text→tool transition, usage propagation), SSE line parsing, prompt flattening
• bun run typecheck, node scripts/validate-shared-contracts.mjs
• Full pre-push check:quality gate
• Live Claude Code session against the 5090 (above)

Note: fresh-clone npm ci (without --legacy-peer-deps) currently fails on main — the frontend lockfile looks out of sync with the pinned deps; npm install resolves. Happy to PR the regenerated lockfile separately if wanted.

[0xSero]

Solid translation layer — clean state machine, real tests, and the live Claude Code validation is convincing. One blocker: main inlined services/inference-client into http/local-fetch.ts (3c146e9), so this compiles against a stale import path — GitHub reports it mergeable but typecheck fails post-merge; please rebase and update the import. Two asks with the rebase: pass through reasoning_content as Anthropic thinking blocks (or explicitly note the drop) — local reasoning models otherwise lose thinking output that the OpenAI route preserves via reasoning.ts — and consider mirroring the OpenAI route's strict-model 404 for unmatched models instead of forwarding upstream. With those this is ready.