feat(model): add built-in kimi-for-coding provider

[yan-ad]

Summary

Adds Kimi For Coding as a built-in, drop-in provider. Kimi For Coding is a separate product from Moonshot's general API and speaks the Anthropic Messages protocol at https://api.kimi.com/coding/v1. The previous 401 was caused by crabcode routing the Moonshot provider through the OpenAI-compatible path against the wrong endpoint.

This change registers a static provider extension so users only need to add an auth.json key — matching OpenCode's behavior:

"kimi-for-coding": { "type": "api", "key": "sk-kimi-..." }

No provider URL or custom-provider config required. Run /refreshmodels after adding the key.

Details

• src/model/extensions/kimicode.rs: static provider with npm: @ai-sdk/anthropic, api: https://api.kimi.com/coding/v1, and models kimi-for-coding, kimi-for-coding-highspeed, k3, k3-256k. Routes through the Anthropic transport (base URL normalized to .../coding, /v1/messages appended).
• Registered in src/model/extensions/mod.rs PERSISTENT_EXTENSIONS.
• Declares header: [("User-Agent", "KimiCLI/1.5")] so requests carry the UA Kimi whitelists (avoids the 429 seen otherwise).

User-Agent handling (refactor)

The Kimi User-Agent requirement is not hard-coded in the generic Anthropic provider. Instead:

• The catalog Provider gained a header field (src/model/extensions/mod.rs, src/llm/client.rs).
• src/aisdk/providers/anthropic.rs builds request headers generically and applies any provider-declared extra_headers.
• src/llm/client.rs (create_anthropic_provider) threads the catalog provider's declared headers into the Anthropic client.

This keeps vendor-specific knowledge (the Kimi UA) entirely inside the kimicode extension and the generic Anthropic provider unaware of any vendor.

Verification

• cargo build clean.
• 12 related tests pass, including build_anthropic_request_headers_applies_extra_headers and create_anthropic_provider_applies_provider_headers.

🤖 Generated with Claude Code

[Blankeos]

Hi @yan-ad have you tested this with your kimi sub? I don't have one so I can't test.

[yan-ad]

@Blankeos not yet, still waiting the whishlist approved tho. I just find the implementation inside the opencode, will test immediately if the sub succeed / approved

[Blankeos]

I bought a $19 moderato Kimi sub just now and it works I think... even without this PR. idk what happened haha.

We can close this PR I think.

EDIT: Scratch that.. I noticed a different problem.. {"error":{"type":"invalid_request_error","message":"an assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'. The following tool_call_ids did not have response messages: edit:18"},"type":"error"}

[Blankeos]

Alright fixed the problem in 6322075

I think we can close this now.

[Blankeos]

Btw I also fixed prompt caching for this... Noticed my sub burned to 20% in the 5 hr limit with Kimi k3 low w/ just 3 prompts. That was weird. I think that was definitely bad prompt caching. I fixed them here 0e10b43

I can confirm prompts are getting cached when I check cc --emit-logs. But yeah the sub still burns kinda quick. But I'm only burning like 0.8% every prompt with Kimi K3. The $19/m plan is kinda bad I think lol.