| Item |
Value |
| Date |
2026-09-24 |
| Environment |
Windows 11 Pro 26200 x64; Node v24.21.0; read-only extraction, no agent-source changes |
| Code revision |
Uncommitted working tree |
| Design references |
implementation-readiness.md, real-data validation during implementation; execution.md M0 item 5 |
Scope and method
- Search only known candidate directories, not entire disks. Versions come from local
packages/install metadata; agent entrypoints were not started.
- Retained fields: event kind, schema/version hints, stable anon-N IDs, time, model, token
values and inclusion relations. Bodies become constants. SQLite queried read-only from
staged copies, then copies removed.
- Ignored build/desktop-usage-validation/ holds
fixtures/<source>/, each with manually
checkable _expectations.md, agent-inventory.md and tools/ extraction/leak scanners.
- tools/leak-check.py scanned all 42 outputs for UUID/path/email/body/username leaks:
zero. Two redaction defects were found/fixed before regenerating all samples.
Observed local products
| Product |
Version and reference |
Local format |
Sample |
| Codex, VS Code extension host |
CLI 0.155.0-alpha.16.3 from rollout session_meta; openai.chatgpt extension 26.917.62051 |
rollout JSONL token_usage_record per call, token_count cumulative, turn_context; 6 SQLite files |
extracted; 3 sessions, up to 49 calls/1,457 lines |
| New Kimi Code |
Desktop 1.0.3; wire protocol_version=1.5 |
sessions/<wd>/session_*/agents/*/wire.jsonl; camelCase usage.record inputOther/output/inputCacheRead/inputCacheCreation; epoch ms |
extracted; main/subagent comparison |
| ZCode |
3.14.3, EXE 3.14.3.7762; protocol client 0.16.9/schemaVersion 1 |
model-io JSONL AI SDK camelCase and Anthropic snake_case with opposite input definitions; db.sqlite model_usage/turn_usage |
extracted; last 8 requests, daily log, two tables |
| Copilot CLI |
1.0.73 from events.jsonl copilotVersion |
events.jsonl has no per-call tokens; session-store.db assistant_usage_events has per-turn fields |
extracted |
| Kilo Code CLI |
pnpm 7.4.21; session.version observed 7.4.15–7.4.20 |
~/.local/share/kilo/kilo.db, 1.6 GB, not ~/.config/kilo; OpenCode-derived SQLite, message.data.tokens |
extracted schema/redacted rows |
| VS Code |
code –version: 1.139.0 |
copilot-chat session-store.db no token columns, recheck M5; kilocode.kilo-code extension 7.7.9 |
schema_recorded |
| pi |
Scoop 0.87.1 |
Empty sessions |
no_data |
| WorkBuddy |
version file 37.10.3-24 |
Three directories empty or version-only |
no_data, remnants |
| Claude Code |
No CLI install metadata found |
~/.claude lacks projects; sessions/backups empty |
no_data |
| Gemini CLI |
No install metadata found |
~/.gemini lacks chats |
no_data |
| oh-my-pi |
Scoop manifest 18.2.11 |
~/.omp/agent/{agent.db,history.db} exists |
exists_only; this stage permits existence checks only |
| qwen/opencode/mimocode/openclaw/hermes/zoo/dsh |
— |
Candidate paths absent |
not_found |
Observed field meanings for M1/M2 parsers
- Codex: total=input+output, cached is an input subset, reasoning an output subset (98/98).
Per-call sum equals final cumulative snapshot without compaction. Compaction resets
cumulative snapshots: 32.06M sum differs from 31.59M snapshot. usage has no model;
attribution requires turn_context.
- Kimi wire: four exclusive fields, no total. event.usage echoes usage.record; choose one
to prevent double counting. subagent.completed.usage equals per-field child wire sums.
- ZCode: response.usage.inputTokens includes cache reads; anthropic.usage.input_tokens
excludes them. turn_usage equals model_usage sums for 16 request rounds.
- Copilot: input=uncached+read+write; request_multiplier=27.0 is a paid multiplier.
- Kilo: input/output/reasoning/cache.read/cache.write are all exclusive, unlike the above.
Codex threads.tokens_used also agrees with rollout per-call totals across stores.
Failures and unexecuted checks
| Item |
Status |
Cause/next step |
| Positive cache writes |
Copilot only |
Codex/Kimi/ZCode samples all zero; labeled synthetic cases in M2/M3 |
| Mid-session model changes |
Codex only |
Not observed in other samples |
| Reasoning |
Kimi/ZCode samples lack values |
Keep unknown, no zero filling |
| Active Kimi wire file |
Point-in-time snapshot |
Grew 394→515 lines during sampling; expectations state snapshot time |
| .omp DB contents, WSL/container instances, F1 products |
Not inspected |
Stage restrictions/out of scope |
Validation files
Ignored build/desktop-usage-validation/: agent-inventory.md; fixtures/ for Codex, Kimi Code,
ZCode, Copilot CLI and Kilo; tools/extract-jsonl.mjs, sqlite-probe.py, kilo-sample.py and
leak-check.py. Review redaction before importing samples; CI uses reviewed samples only.
