conv-main._expectations.md (REAL, anonymized local extraction)
Source: Kimi Work’s embedded kimi-code home,
…/Kimi/share/daimon-share/daimon/runtime/kimi-code/home/sessions/<wd>/conv-<hexid>/agents/main/wire.jsonl.
The original has 363 lines, wire protocol_version=1.4 and host daimon. state.json
createdBy=daimon-kernel-adapter also identifies the product. Extracted on 2026-09-25:
tools/extract-kimi.mjs retains 115 lines of selected fields, with bodies REDACTED and IDs anon-N.
Structural expectations
- First line: metadata with protocol_version=“1.4”.
- 38 turn-scoped usage.record entries, with no agentId: the 1.4 main wire omits agentId; Agent identity comes from agents/main. All 38 step.end repeats contain usage.
- Model k28-agent-preview ×38; llm.request provider=kimi.
Expected usage values (tool output, independent source jq sums and roundtrip agree)
| n | inputOther | output | inputCacheRead | inputCacheCreation |
|---|---|---|---|---|
| 38 | 47,574 | 18,658 | 1,310,720 | 0 |
Repeat and record sums match exactly (38=38); this session has no interrupted steps.
Measured differences from Kimi Code 1.5 (separate native samples)
- Directory:
conv-<hexid>rather thansession_<uuid>; this session has no subagent directory. - usage.record omits agentId, which 1.5 main records include. Models use bare IDs rather than alias/model combinations. A shared kernel name does not merge product verification.
Expected import (manually calculated)
38 primary events: input_uncached=47,574, cache_read=1,310,720, cache_write=0 (reported), output_total=18,658, input_total=1,358,294 (derived), total_tokens=1,376,952 (derived).
76 anonymized IDs: anon-1…anon-76.