kimi-k2.6 stream drops mid-reasoning (3rd occurrence, 160K context, 76% window) #133

Bug Summary

kimi-k2.6 (Moonshot API) stream drops mid-generation again, showing "× interrupted (esc)" even though the user did not press Escape. This is a recurring occurrence of the Moonshot API bug (related to #132).

Screenshot evidence

• Model was in the middle of its reasoning/thought process (✓ Thought visible)
• Status bar showed: [kimi-k2.6 · cwd /Users/limyuxi/lawbook · 160331/209k tok (76%) · ⚡158961 cached]
• The model had not yet emitted its final response — it died during internal reasoning

Evidence from harness.trajectory.jsonl

Latest interruption at the end of the current session:

{"kind":"turn_error","parent":55,"t":9425434,"detail":"Interrupted"}

This is turn 55 of the current session, interrupted after ~2.6 hours of session runtime.

The session started with harness version 0.0.187 (newer than the v0.0.14-2-g11f341f-dirty from the previous occurrence in #132).

Preceding turns in the same session show gpt-5.5 successfully completed its workflow (turns with ok: true), then the harness switched to kimi-k2.6 for the user's follow-up message, which was then interrupted.

Pattern confirmation

This is the third confirmed occurrence of the same Moonshot API stream drop:

1. First: msg[5] in .graff/sessions/adding-recently-viewed-documents.session.json — mid-response
2. Second: msg[59] in same session — mid-response
3. Third: this current session, turn 55 — mid-reasoning/thought (most severe, as it aborts before any output)

Context at time of death

• Model: kimi-k2.6
• Context tokens: 160,331 / 209,000 (76%)
• Cached tokens: 158,961
• Session runtime: ~2.6 hours
• Harness version: 0.0.187

The high context load (160K tokens, 76% of context window) may be a contributing factor to the Moonshot API dropping the stream.

Impact

• The model dies during reasoning, producing zero output for the user
• The user has to re-prompt with the same request
• Token cost is wasted — 160K context + reasoning time is burned with no result
• UX trust is severely degraded because the model appears "unreliable"

Suggested fixes (harness side)

1. Retry on interruption: If turn_error: "Interrupted" occurs with a kimi-k2.6 model AND no user stop signal was received, automatically retry the generation with continue or a reduced context window
2. Distinguish API drops from Esc: Add telemetry to confirm whether the interruption was user-initiated or API-initiated. If API-initiated, do not save [response interrupted by user] into the session
3. Context-window aware routing: When context exceeds 70% of the model's window (e.g. 150K+ for kimi-k2.6), warn the user or switch to a model with a larger context window
4. Recover partial reasoning: If the stream drops mid-reasoning, attempt to recover the partial thought output rather than discarding everything

Suggested fixes (upstream — Moonshot)

1. Report again: Moonshot's kimi-k2.6 consistently drops SSE streams after extended generation with high context. This is reproducible.
2. Possible workaround: Moonshot may have a server-side timeout on SSE connections. The harness should send periodic keep-alive pings or break long requests into smaller chunks.

Environment

• Harness version: 0.0.187
• Model: kimi-k2.6
• Session runtime: ~2.6 hours
• Context: 160,331 / 209,000 tokens (76%)
• Occurrence: 3rd confirmed instance