lsm / neokai / 28490817416

82%

push

github

fix(agent): enable SDK native auto-compact for Kimi to stop prompt-too-long (#2218)

Kimi sessions overflowed with terminal "Prompt is too long" (~7.7% of sessions
vs ~0.6–4.2% for codex/glm/anthropic) because NeoKai disabled the SDK's native
auto-compact for Kimi (the shouldUseNeoKaiCompactFallback special-case) and
replaced it with an async post-turn /compact fallback. That fallback fires
after turns/events, so it cannot prevent within-turn context growth or
resume overflow — the SDK pre-flights against its 200k belief and, with
auto-compact off, emits a 0-token terminal result without ever calling Kimi.

Verified empirically against SDK 0.3.x: the SDK resolves every non-Anthropic,
non-[1m] model ID (codex gpt-5.5 AND kimi-for-coding) to a 200k window, and
clamps every override to it — settings.autoCompactWindow,
CLAUDE_CODE_AUTO_COMPACT_WINDOW, /v1/models (never queried), and
message_start.usage.model_context_window (inert) all leave maxTokens=200000.
The codex bridge does NOT raise this: with auto-compact OFF a >200k prompt is
blocked before the bridge sees it; with auto-compact ON (codex's config) the
same prompt passes through. So the sole codex-vs-Kimi difference is
auto-compact ON vs OFF.

Drop the Kimi special-case so Kimi uses SDK native auto-compact (armed at
200k − 33k = 167k, safely below Kimi's real 262k window). Kimi now matches
codex: prompts pass through instead of hard-blocking at 200k, and the SDK
compacts between turns at its threshold.

11447 of 16495 branches covered (69.4%)

Branch coverage included in aggregate %.

97764 of 115951 relevant lines covered (84.31%)

300.94 hits per line