Kimi K3 works as primary but OpenCode Go rejects subagent requests with 400 #37552

Activity

1. github-actions

   assigned

   MrMushrooooom

   on Jul 17, 2026
2. github-actions commented on Jul 17, 2026

   

   github-actionsBot

   on Jul 17, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of existing issues. Please check:

   • Bug: opencode-go/kimi-k2.6 fails in child sessions with "image.png" error #24454: Kimi K2.6 also fails in child (subagent) sessions via OpenCode Go — same symptom of subagent HTTP error while primary works fine
3. mihneaptu commented on Jul 17, 2026

   

   mihneaptu

   on Jul 17, 2026

   Author

   More actions

   Additional reproduction: native Build fails after multi-turn tool history

   I reproduced the same error with no Fusion agents or custom agent files active. The live config contained only:

   {
     "$schema": "https://opencode.ai/config.json",
     "model": "opencode-go/kimi-k3",
     "small_model": "opencode-go/kimi-k3"
   }

   In a native Build session using Kimi K3 max, the model completed several reasoning/tool-call turns successfully and then failed on a later continuation:

   HTTP 400 invalid_request_error
   Error from provider (Console Go): Upstream request failed
   isRetryable: false
   endpoint: https://opencode.ai/zen/go/v1/chat/completions
   

   The session had 12 stored messages with successful read, glob, question, and todowrite calls before the failure. Entering continue in that same conversation produced the same error again (same native Build session, at 20:53:38 and 20:57:04 in the log).

   Control test immediately afterward:

   opencode run --variant max "Do not inspect or modify files. Reply with exactly: FRESH_KIMI_MAX_OK"
   FRESH_KIMI_MAX_OK
   > build · kimi-k3
   

   Updated conclusion: this is not Fusion-specific, not strictly subagent-only, and not a general Kimi K3 or max outage. Fresh primary sessions succeed, while a longer native Build conversation can fail repeatedly when its accumulated reasoning/tool-call history is replayed. This points to a possible multi-turn reasoning/tool-history serialization incompatibility.

   Evidence screenshots:

   
4. mihneaptu commented on Jul 18, 2026

   

   mihneaptu

   on Jul 18, 2026

   Author

   More actions

   Update after controlled A/B tests and replaying the affected session through patched source:

   1. temperature is one independent 400 path

   Using identical fresh Kimi K3 max requests:

   • with temperature: 0.2 -> HTTP 400 Upstream request failed
   • with the temperature field omitted -> completed successfully

   The same omission also let the Fusion Sidekick and Design child agents complete live tool-call tasks. Max thinking remains enabled; only the unsupported sampling field is removed.

   Metadata fix: anomalyco/models.dev#3328
   Fusion prompt fix: mihneaptu/opencode-fusion#28

   2. Long-session history is a separate 400 path

   The broken session contained this stored sequence:

   step-start -> empty metadata-free reasoning -> step-start -> reasoning -> text

   Replay turned the empty boundary into an assistant message with no usable content. The patch skips that empty reasoning part. The focused regression suite passes (37/37), and replaying the exact previously broken Kimi K3 max session continued for 26 model steps and finished normally.

   OpenCode fix: #37624

   One correction to avoid overclaiming: the public Console code loads the production provider routing table from deployment secrets. It does not prove that OpenCode Go routes Kimi K3 through OpenRouter. What the A/B test proves is that the upstream selected by OpenCode Go rejects the temperature field.
5. rekram1-node commented on Jul 18, 2026

   

   rekram1-node

   on Jul 18, 2026

   Contributor

   More actions

   Fixed upstream in models.dev: anomalyco/models.dev#3331

   Kimi K3 locks sampling server-side (temperature fixed at 1.0; other values → HTTP 400). The catalog had advertised temperature = true, so clients sent the field and OpenCode Go rejected the request.

   #3331 marks the base model temperature = false so clients omit it. After catalog refresh, subagent calls that were injecting temperature should stop 400ing.

   Related incomplete fix attempt: anomalyco/models.dev#3328 (closed as superseded).
6. rekram1-node

   closed this as completedon Jul 18, 2026