feat: OpenAI tool calling + array/object content fix

[nayrosk]

Summary

Ports two upstream PRs from joesobo/claude-max-api-proxy onto this fork, merged so both fixes coexist.

• joesobo#3 — handle array/object content in messagesToPrompt (fixes the [object Object] bug when clients send OpenAI multimodal-style content: [{type:"text",...}]).
• joesobo#2 — full OpenAI-compatible function/tool calling (tools, tool_choice, tool_calls, role=tool, finish_reason=tool_calls).

Encoder (src/adapter/openai-to-cli.ts)

• Injects a Tool-Use Protocol system section when the request carries tools: tool list, JSON-Schema params, emission rules, tool_choice policy.
• tool_choice → auto / none / required / specific function mapped to explicit policy text.
• Assistant tool_calls replayed as <tool_call>…</tool_call>; role=tool rendered as <tool_result id="…">…</tool_result> for multi-turn re-hydration.
• contentToString from PR#3 retained and now called on every role branch (system/user/assistant) so array/multimodal content keeps working alongside the tool flow.

Decoder (src/adapter/cli-to-openai.ts)

• extractToolCalls scans Claude's text for <tool_call> blocks, parses inner JSON, returns OpenAI tool_calls + cleaned text. Robust to missing id/arguments, malformed JSON, and unterminated blocks.
• cliResultToOpenai / cliToOpenaiChunk accept extractTools flag; finish_reason flips to "tool_calls" when any call is parsed.

Routes (src/server/routes.ts)

• Passes hasTools through to the decoder.
• Streaming buffers content while tools are active (partial JSON is unsafe to stream piece-by-piece) and emits the parsed tool_calls in the final chunk.
• Per-request timing logs ([Timing rid=…]) + opt-in DEBUG=1 prompt dump for diagnosing new agent frameworks.

Types (src/types/openai.ts)

Full alignment with OpenAI Chat Completions v1 spec:

• Discriminated OpenAIChatMessage union (system | user | assistant | tool).
• OpenAITool, OpenAIToolChoice, OpenAIToolCall, OpenAIToolCallDelta, OpenAIFinishReason (incl. "tool_calls").
• Nullable assistant content, OpenAIToolMessage.tool_call_id.
• OpenAIContentPart retained from PR#3; system/user/assistant content widened to string | OpenAIContentPart[] (assistant also | null).

Tests

14 new cases in src/adapter/tool-calling.test.ts:

• Encoder: hasTools flag, protocol injection, tool_choice variants (auto/none/required/specific), multi-turn tool role rendering.
• Decoder: single block, multiple blocks in order, missing id/args, custom id, generated id, malformed JSON, missing name, unterminated block, plain-text passthrough.

✔ 14 pass / 0 fail  (npm test)
✔ tsc clean         (npm run build)

Test plan

• npm install && npm run build — clean, no type errors.
• npm test — 14/14 pass.
• PR#3 content-parts path still exercised (array content reaches all message branches via contentToString).
• Manual curl end-to-end against a running proxy with a real OpenAI-tool-calling client (optional).

Notes

• Commit is GPG-signed.
• package-lock.json picked up a name re-sync (upstream rename from claude-code-cli-provider → claude-max-api-proxy) during npm install.