kimicode

Overview ¶

Package kimicode implements an event-level adapter for Kimi Code (MoonshotAI/kimi-code, MIT). Kimi Code keeps one append-only wire log per AGENT per session:

<home>/sessions/<workspace>/<session>/agents/<agent>/wire.jsonl

The log is a mixed record stream (prompts, messages, loop events, config). Exactly two record types matter here: `llm.request`, written before every provider call, and `usage.record`, written after one returns.

THE MODEL TRAP. `usage.record` carries a `model` field and it is NOT a model id: kimi records usage as `usage.record(request.modelAlias, ...)`, so the field holds the CONFIG ALIAS the profile was bound under — observed here as the literal `__kimi_env_model__`, the synthetic alias for an environment-configured profile. The real model id appears only on the `llm.request` record that preceded the call, which carries BOTH (`model`: `gemma4:31b-cloud`, `modelAlias`: `__kimi_env_model__`). Every usage record is therefore resolved against the nearest PRIOR request in the same file, and the carry-forward survives a tail read in the checkpoint. Taking the record's own field would file every session on this machine under one unpriceable pseudo-model.

THE DOUBLE-COUNT TRAP. The same token counts appear TWICE in the file: once as `usage.record`, and once inside the `context.append_loop_event` whose `event.type` is `step.end` (which also carries the provider `messageId`). They are one API response reported by two writers, not two calls. Only `usage.record` is read, because it is the complete stream — a compaction or title call records usage with `usageScope: "session"` and no loop step at all — and because counting both would exactly double every loop turn.

SCOPE IS NOT CUMULATIVE. `usageScope` is `"turn"` or `"session"` and both are per-request DELTAS: kimi's own replayer folds every record with `addUsage(byScope[scope], rec.usage)` and `addUsage(byModel[model], rec.usage)`, i.e. it SUMS them. Every record is counted once, whatever its scope; filtering by scope would drop the compaction and summarisation calls that only ever arrive as `"session"`.

PRIVACY. The wire log is full of content — system prompts, user input, message parts, tool arguments, tool schemas. None of it is decoded: the line struct names only counters, model ids, times and scopes, so encoding/json discards every other key as it parses and the content never becomes a value in this process. No raw audit payload is built either, so `privacy.no_raw` is satisfied by construction rather than by a switch.