kimi-k2.6 main agent stream interrupted mid-generation (user did not press Esc) #132

Bug Summary

The kimi-k2.6 (Moonshot API) model stream drops or aborts mid-generation, causing the harness to display "× interrupted (esc)" even though the user did not press Escape. This is a model/API-side bug, not a harness timeout.

Evidence from harness.trajectory.jsonl

The exact failure chain (lines 410–418):

Line 410: turn 20, gpt-5.5  → network error: TlsInitializationFailed (gave up after 6 attempts)
Line 411: turn 21, gpt-5.5  → task about lawplain issue #91, ok=true
Line 413: turn 22, gpt-5.5  → task about recently-viewed tab, ok=false
Line 414: turn_error → network error: TlsInitializationFailed (gave up after 6 attempts)
Line 415: turn 23, gpt-5.5  → same task, ok=false
Line 416: turn_error → network error: TlsInitializationFailed (gave up after 6 attempts)
Line 417: turn 24, kimi-k2.6 → task about trajectory report, ms=182632 (3 min) ok=false
Line 418: turn_error → Interrupted

Key insight: After two GPT-5.5 TLS failures, the harness fell back to kimi-k2.6. kimi-k2.6 ran for 182,632ms (3 minutes) executing tool calls (read_file, codedb, bash) and then the Moonshot API stream dropped, which the harness interpreted as Interrupted.

Evidence from .graff/sessions/adding-recently-viewed-documents.session.json

• msg[5] (assistant): [response interrupted by user] — after the user asked to add Recents as a separate tab
• msg[59] (assistant): [response interrupted by user] — after the user complained about the error

After each API-side drop, the harness UI duplicated the user's message 3-4 times (auto-retry/duplicate injection).

Evidence from harness.trace.jsonl

• First line: {"t":3,"ev":"session","detail":"kimi-k2.6"} — confirms the trace is for the correct model
• The trace file is corrupted with null bytes (37KB binary), which likely happened when the Moonshot API stream dropped at line 418

Root Cause Analysis

This is a Moonshot API (kimi-k2.6) streaming bug, not a harness timeout:

1. Model/API stream drops after extended generation: kimi-k2.6 ran for 3+ minutes with multiple tool calls, then the SSE stream from Moonshot abruptly closed without a finish reason.
2. Harness interprets the drop as interruption: The harness sees the stream die and records turn_error: "Interrupted", but the root cause is the API dropping the connection.
3. Preceding OpenAI TLS failures are unrelated: The GPT-5.5 TlsInitializationFailed errors are a separate issue (see GPT-5.5 subagent calls fail with ms=0 / resp_bytes=0 — model crashes immediately #131). The harness falling back to kimi after those failures is correct behavior; the subsequent kimi stream drop is independent.
4. Trace corruption is a side effect: When the Moonshot stream drops, the harness is in the middle of writing trace events, causing null bytes to be written to harness.trace.jsonl.

Impact

• User loses 3+ minutes of model work (tool calls, reasoning, partial edits)
• The harness auto-saves [response interrupted by user] into the session JSON, which pollutes context on resume
• Message duplication after each API drop wastes tokens and confuses the user
• The corrupted trace file makes post-mortem debugging harder

Suggested Fixes (for harness side)

1. Distinguish API drops from user interruption: When the stream drops without an explicit stop signal, check if the last chunk was a valid tool call or mid-sentence. If so, mark it as API_ERROR not Interrupted.
2. Retry on API drop: If the stream drops after tool calls, automatically retry the generation with a continue prompt rather than marking [response interrupted by user].
3. Fix trace file atomic writes: Ensure trace is flushed/closed before recording an error, so null bytes don't corrupt the file.

Suggested Fixes (for Moonshot/kimi side — requires upstream report)

1. Report to Moonshot: kimi-k2.6 appears to drop SSE streams after ~3 minutes of generation with tool calls. This may be a server-side timeout or a bug in their streaming infrastructure.
2. Workaround: Use shorter generation batches or reduce context size when using kimi-k2.6 for complex ultracode workflows.

Environment

• Harness version: v0.0.14-2-g11f341f-dirty (from trajectory line 1)
• Session model: kimi-k2.6 (Moonshot API)
• Previous attempts: gpt-5.5 (failed 2× with TLS errors — see GPT-5.5 subagent calls fail with ms=0 / resp_bytes=0 — model crashes immediately #131)
• Session file: .graff/sessions/adding-recently-viewed-documents.session.json
• Trace file: harness.trace.jsonl (corrupted, 37KB with null bytes)
• Trajectory file: harness.trajectory.jsonl (281KB, intact)