Main agent K3 requests routed to 256K kimi-for-coding backend after subagent/AgentSwarm returns

Version

0.40.1

Platform

Linux x86_64

Model

Main agent: kimi-code/k3 (status bar shows K3)

What happened

In two separate sessions, the main agent's llm.request wire records always show modelAlias: "kimi-code/k3" and model: "k3", but the request was rejected by the backend with:

400 Invalid request: Your request exceeded model token limit: 262144 (requested: ...)

At the same time, my Kimi Code console's model-usage records show these main-agent requests were actually served as k2.7 code (kimi-for-coding), which has a 256K context limit.

Trigger pattern observed

In both cases, the error happened immediately after a subagent using kimi-code/kimi-for-coding returned:

1. Session session_76a8b5ca-4efe-466b-8e4e-514576e44de8

   • Turn 31 step 5 called Agent (description: "修复 K1+R2 review 问题", model kimi-code/kimi-for-coding).
   • Step 5 ended at 2026-09-02 10:25:03 UTC with ~520K input tokens.
   • Step 6 at 2026-09-02 10:25:05 UTC failed with:

     model token limit: 262144 (requested: 501524)
     

2. Session session_01a1ce7b-441e-4d4b-9b52-38cf39cd40f4

   • Turn 4 step 3 called AgentSwarm (description: "anima LoRA NSFW 调研打标", subagent model kimi-code/kimi-for-coding).
   • Step 3 ended at 2026-09-03 03:28:39 UTC.
   • Step 4 at 2026-09-03 03:28:40 UTC failed with:

     model token limit: 262144 (requested: 322518)
     

Expected behavior

When the main agent is bound to kimi-code/k3, its requests should always be routed to the K3 (1M context) backend, regardless of whether Agent/AgentSwarm subagents using other models were invoked earlier in the same session.

Additional context

• The client-side wire log always records model: "k3" for the failing main-agent requests, so this does not look like the client changed the requested model.
• It looks like either the managed:kimi-code provider backend is routing the session to a k2.7 code deployment after a subagent uses it, or some session/provider-level state is leaking from the subagent model back to the main agent.

Start by tracing the llm.request path and the Agent and AgentSwarm model/provider handling, comparing main-agent state before and after a kimi-code/kimi-for-coding subagent returns. The issue is done when main-agent kimi-code/k3 requests consistently reach the K3 backend after either subagent flow, with regression coverage for both observed patterns.