Kimi (moonshotai) stops agent loop after first tool call - reasoning_content not handled in multi-step #36316

Summary

When using a Kimi model (e.g. cf/@cf/moonshotai/kimi-k2.6 / kimi-k2.7-code) through an OpenAI-compatible provider, the agent loop stops after the first tool call instead of continuing to the next step. A model like MiniMax M3 on the same provider works fine (read -> edit -> done).

Environment

• OpenCode: 1.17.18 (latest)
• Provider: OpenAI-compatible (@ai-sdk/openai-compatible), baseURL http://100.95.250.13:20128/v1
• Models affected: cf/@cf/moonshotai/kimi-k2.6, cf/@cf/moonshotai/kimi-k2.7-code (also nvidia/moonshotai/kimi-k2.6, cline-pass/kimi-k2.6)
• Agent: build

Reproduction (agent loop)

Task: "read alvo.py, find the bug and fix by adding the missing import".

1. Model calls glob (or read).
2. OpenCode returns the tool result.
3. Model returns finish_reason: "stop" with no follow-up tool call -> agent ends without editing.

Observed step_finish after the tool result: reason:"stop". The model is capable of the next step (see below).

Key evidence — raw provider API works

Calling the same provider/model directly (curl) with the same multi-turn tool sequence returns a correct continuation:

• After glob result only: finish_reason:"tool_calls", model calls read.
• After glob + read results: finish_reason:"tool_calls", model calls edit_file with correct args.

So the endpoint/model is fine; the issue is in how OpenCode drives the multi-step loop for this model.

Suspected cause

Kimi returns its reasoning in a separate reasoning_content field (not interleaved in the streamed content). OpenCode appears to not continue the loop correctly when the model emits reasoning_content after a tool result, finalizing the turn early (stop) instead of executing the next tool call.

Config flags tested (no effect)

• reasoning: true / false
• interleaved: true / false
• attachment + modalities present / absent (minimal config with only name/limit/tool_call)
• Multiple provider endpoints (cf/, nvidia/, cline-pass/)

All still stop after the first tool call. MiniMax (no separate reasoning_content behavior) works in every configuration.

Expected behavior

After a tool result, the agent should continue calling tools (read -> edit) until the task is complete, as it does for MiniMax.

Minimal raw repro (works, proves endpoint OK)

POST /v1/chat/completions
{
"model": "cf/@cf/moonshotai/kimi-k2.6",
"messages": [
{"role":"system","content":"You are a coding agent."},
{"role":"user","content":"Read alvo.py and fix the missing import."},
{"role":"assistant","tool_calls":[{"id":"call_1","type":"function","function":{"name":"glob","arguments":"{"pattern":"**/alvo.py"}"}}]},
{"role":"tool","tool_call_id":"call_1","content":"/sandbox/alvo.py"}
],
"tools":[{"type":"function","function":{"name":"read","description":"Read","parameters":{"type":"object","properties":{"filePath":{"type":"string"}},"required":["filePath"]}}}],
"tool_choice":"auto"
}

Response: finish_reason:"tool_calls" -> read tool call (correct).