fix: route anthropic-messages models to /v1/messages instead of /chat/completions

[gitawego]

Problem

The vision delegate hard-codes the OpenAI-compatible endpoint POST {baseUrl}/chat/completions for every vision model (lib/delegate.ts → callVisionModel), ignoring the model's registered api type.

For models registered as api: "anthropic-messages" — e.g. minimax-cn/MiniMax-M3 with baseUrl: https://api.minimaxi.com/anthropic — the resulting URL https://api.minimaxi.com/anthropic/chat/completions doesn't exist. The request fails with:

Vision tool error: Vision model returned 404: 404 page not found

Reproduced live:

URL	Result
api.minimaxi.com/anthropic/chat/completions (what the delegate builds)	404 page not found
api.minimaxi.com/anthropic/v1/messages (correct Anthropic Messages route)	200

This also means any Anthropic-Messages-family vision model (MiniMax, Claude, Qwen Anthropic-compat endpoints, …) can never be used as the vision model, regardless of config correctness.

Fix

Make the request API-aware:

• api: "anthropic-messages" → POST {baseUrl}/v1/messages with the Messages schema:
  • image passed as a source.base64 content block (media_type + data)
  • systemPrompt as a top-level system field (not a messages entry)
  • no temperature / reasoning_effort (not part of the Messages schema)
  • response parsed from the content[] array — first text block, with a thinking-block fallback for reasoning-only replies
• everything else → unchanged OpenAI /chat/completions path (data-URL image, system message, reasoning_effort for reasoning models)

Tests

6 new tests in tests/delegate.test.ts covering: endpoint URL, Anthropic body shape, system-field placement, content[] response parsing (text + thinking fallback), error surfacing, and reasoning_effort omission.

tests/delegate.test.ts: 39 passed
Full suite (serialized): 357/359 passed (2 pre-existing timing flakes on slow filesystems: disk-LRU mtime race + wall-clock batch assertions — unrelated to this change, pass in CI/isolated runs)

Also verified end-to-end against the real minimax-cn/MiniMax-M3 endpoint (200 OK with a live request to .../anthropic/v1/messages).

Notes

• Retry/fallback behavior is unchanged; the error message for a 4xx (like this 404) still surfaces the status + body excerpt.
• openai-responses / openai-codex-responses models still use the /chat/completions compat path — this PR deliberately scopes to the Anthropic-Messages fix; the OpenAI-responses shape can be a follow-up if needed.