M3’s first adapter, Kilo, and M4’s three products, ZCode/Kimi Code/Kimi Work, are implemented.
The user directed remaining locally absent M3 tools, Cline/OpenCode/MiMo/Zoo/DSH/OpenClaw/Hermes,
to be implemented from documents/source with native acceptance deferred.
| Item |
Value |
| Date |
2026-09-25 |
| Environment |
Windows 11 x64; rustc 1.98.0; Node v24.21.0 |
| Code revision |
Uncommitted working tree after M6, plus these changes |
| Design references |
adapters.md A11/A12/A13/A21; architecture.md#database source SQLite rules; V07/V12/V17/V30 |
| Execution |
Three parallel subagents implemented four adapters using samples extracted before the earlier interrupted session; main session integrated scanner registration, Clippy/format fixes and full checks |
Commands and results after integration
| # |
Command (directory) |
Exit |
Result |
| 1 |
cargo test -p llm-usage-core (desktop/src-tauri) |
0 |
373 passed / 0 failed, M6 baseline 285 plus net 88 |
| 2 |
cargo clippy –workspace / cargo fmt –check |
0 |
Core clean; app main.rs test-module-order warning was an existing M0 issue |
| 3 |
npm run verify (root) |
0 |
49 test binaries pass; lint clean; Vite build unchanged |
| 4 |
APPDATA=temporary cargo run -p llm-usage-m0 – –headless |
0 |
All seven instances succeeded, 27,311 events; rescan adds only 3 from active ZCode file |
Adapter findings: real_verify outputs and sample expectations contain details
Kilo Code CLI: M3, A11
- Inspected ~/.local/share/kilo/kilo.db: active 519 MB database plus 30 MB WAL, 25 tables.
message.data.tokens exists on all 13,342 assistant rows, five exclusive buckets. Session
token columns compare totals only: 275/276 matched; semantics vary by version, never
mapped as events. session.cost actually contains model JSON and is ignored. New
session_message core tables have zero rows and are only recorded as present.
- Read-only SQLite; Busy/lock/CANTOPEN may use a staged Online Backup, bounded to 2s/2GiB
and removed afterward. Native direct reads did not require staging; synthetic tests cover Busy.
- 7.4.8/7.4.9 samples verified. Observed but unregistered 7.3.42–7.7.12 remain latest_fallback.
- Native check: 255 sessions, 13,342 calls, primary 11,745 plus sub_agent 1,597. Total
1,645,598,767 matches independent Python sums field by field. Repeats add nothing;
254/255 snapshots match, one unsynchronized source snapshot remains a visible diagnostic.
ZCode: M4, A21
- Separate input definitions: primary AI SDK camelCase five fields, inputTokens includes
cache reads; Anthropic snake_case comparison input_tokens excludes them. When both views
exist, check in+cr+cw==inputTokens and out==outputTokens per row; mismatches get dual_caliber_mismatch.
- Version comes from request.headers[“x-zcode-app-version”]; 3.14.3 verified.
- Read-only db.sqlite comparison: active 418/438 match, with in-flight/cancelled mismatches
visible; static M0 sample 16/16.
- Native check: three files/eight events, stable repeated reads. ~/.zcode/v2 and desktop
%APPDATA%/zcode storage not connected pending inspection.
Kimi Code: M4, A12; Kimi Work: M4, A13
- Shared adapters/kimi_wire.rs handles metadata/usage.record/echo deduplication/subagent
comparison/epoch-ms checks; independent directories/registries for protocols 1.5 and 1.4.
- Work uses
conv-* or ctitle-* under daimon, lacks usage.record.agentId and uses bare model
IDs. Two same-ms cross-file pairs from parallel swarms require session:agent in event keys;
this fixed first-version conflicts that lost two events.
- All observed times are epoch milliseconds. No second-based sample; out-of-range values
get diagnostics/skipping, without guessed ×1000 conversion.
- Native values match independent jq sums by field and stay stable on rescan:
- Code: 13 files, 965 calls, primary 586 + auxiliary 5 + sub_agent 374; total 116,428,813.
Ten files match, two echo_subset interrupted steps lack echoes and remain visible; one
subagent.completed snapshot also equals child totals.
- Work: 69 files, 1,337 calls, 833+8+496; total 125,225,933; all 69 comparisons match.
Integrated source totals: actual headless collection
| Agent |
Calls |
total_tokens |
| kilo-code |
13,342 |
1,645,598,767 |
| oh-my-pi |
8,767 |
1,026,791,695 |
| codex |
2,597 |
299,747,341 |
| kimi-work |
1,337 |
125,225,933 |
| kimi-code |
965 |
116,428,813 |
| zcode |
266 plus 3 from active file |
100,651,285 |
| pi |
37 |
2,864,419 |
Claude/Gemini/Qwen had no local data; existing not_found/no_data status retained.
Outstanding work at this stage
| Item |
Status |
Next step |
| Kilo new session_message tables, zero rows |
Not collected |
Inspect dedicated implementation when Kilo changes |
| ZCode v2/desktop sessions |
Not collected |
Inspect |
| Kimi usage.record lacks stable ID; provider/per-call latency omitted |
unavailable |
Connect after upstream correlation keys exist |
| Work positive cache writes/subagent.completed |
No native sample |
Synthetic paths covered; obtain native sample later |
| Work official default layout/environment |
No documentation found |
Add manual roots after installation moves |
| Remaining M3 Cline/OpenCode/MiMo/Zoo/DSH/OpenClaw/Hermes |
Not implemented |
User-directed document/source implementation, native acceptance later |
Validation files
- adapters/{kilo,zcode,kimi-code,kimi-work}/ and adapters/kimi_wire.rs; product mappings
in Kilo/ZCode common.rs and kimi_wire.rs, removed from usage_map.rs.
tests/{kilo,zcode,kimi_code,kimi_work}_{contract,gaps_synthetic,incremental_v12}.rs;
tests/fixtures/{kilo,zcode,kimi-code,kimi-work}/: real redacted and synthetic samples,
each with _expectations.md.- real_verify_{kilo,zcode,kimi_code,kimi_work}.rs; ignored
build/desktop-usage-validation/tools/extract-kimi.mjs.
- src/scanner.rs registers ten adapters.
