Four products were locally absent in the 2026-09-25 inventory, with candidate paths such
as ~/.cline missing and all four real_verify examples discovering zero roots. The user
directed implementation against fixed source/official documentation. Synthetic tests pass;
native-data acceptance was deferred until installation/discovery and real_verify checks.
Adapters and samples explicitly identify document/source references.
| Item |
Value |
| Date |
2026-09-25 |
| Environment |
Windows 11 x64; rustc 1.98.0 |
| Code revision |
Uncommitted tree after M3/M4 plus this work |
| Design references |
adapters.md A03/A08/A24/A09 fixed source; V03/V07/V12/V17/V30 |
| Execution |
Three parallel subagents stopped at API limits, 2026-09-25 21:33; each adapter’s five files compiled. Main session added tests/examples, fixed inline failures and integrated |
Implementation and references
Cline: A03, fixed source dcf8c3c
- Usage comes from text JSON on type=say with say in api_req_started/deleted_api_reqs/
subagent_usage. tokensIn/tokensOut/cacheWrites/cacheReads/cost individually optional;
four exclusive buckets derive input_total=in+cw+cr. Compaction estimates excluded.
Task ui_messages.json is a whole-file array under the task directory.
- Inline tests exposed missing documented request key and missing negative/nonfinite
map_cost rejection; main session fixed both consistently with parse rules.
- Three format tests: synthetic end-to-end three calls/input 475/total 530 and stable
repeats; documentation-based capability label; partially usable boundary scenarios.
DSH: A08, fixed token-meter README 46a7f68
- Final replaces streaming samples of the same attempt. retry-started ends replacement
scope and starts another billable attempt. Attempt/step boundaries finalize the latest
sample. occurred_at uses observation time because README lacks per-event timestamps.
pressure/contextBreaks estimates excluded.
- Four tests: two attempts/input 140/total 158 including streaming→final/retry; replacement
scenario three attempts/input 195/total 208; non-usage vocabulary creates no requests;
capability clearly says synthetic assumptions.
- Persistence path/row serialization undocumented; proposed JSONL shape remains a labeled
synthetic assumption.
Hermes: A24, fixed source ef70b36
- session_model_usage is cumulative per model/route/task, not request-level. Map to source
intervals/UsageObservation; api_call_count does not become model_call. Cross-day first_seen/
last_seen never assigns the entire total to the last day, following V03. Auxiliary task
and main-session cumulative values are mutually exclusive.
- Eight tests from subagent, with two type/assertion fixes during integration: basic
cumulative, cross-day, exclusive auxiliary totals, v20 backfill/compaction inheritance,
capability evidence_level=doc-level.
- projection.json rebuilds SQLite; schema DDL accompanies that extracted test data.
OpenClaw: A09, official documentation
- Documents then established only ~/.openclaw/agents/id/agent/openclaw-agent.sqlite and
legacy sessions archive locations, without table/entry schema. Runtime DB/archive
return unknown_format with pending-native-sample reasons. No table/field guesses or
invented zeros. Archives treated as migration input following doctor –fix.
- Three tests: runtime rejection/stable repeats, archives excluded even with placeholder
1200+300 values, accurate capabilities. layout.json rebuilds temporary SQLite/JSON/JSONL layout.
Commands and results after integration
| # |
Command (directory) |
Exit |
Result |
| 1 |
cargo test -p llm-usage-core |
0 |
Cline 3, DSH 4, OpenClaw 3, Hermes 8 and additional inline tests pass; verify includes 53 test binaries |
| 2 |
cargo run –example real_verify_{cline,dsh,hermes,openclaw} |
0 |
All four discovered_roots=0; local data directories absent |
| 3 |
npm run verify |
0 |
Lint/Clippy/fmt clean; frontend build unchanged |
| 4 |
APPDATA=temporary cargo run -p llm-usage-m0 – –headless |
0 |
14 registered adapters; 7 instances/27,158 events, same order of magnitude as prior batch with active-file changes |
Problems and fixes during work
| Location |
Problem |
Fix |
| scanner.rs |
Registry overwritten twice with stale external editor/tool buffers; app discovered zero versus probe seven roots |
Restore integration and assert with script; discover_probe.rs prints per-adapter roots using app context |
| Cline inline tests |
request absent from known keys; map_cost accepted negatives |
Add key; reject nonfinite/negative cost as parse does |
| hermes_contract |
Fourth tuple field incorrectly nonnullable; evidence_level assertion too strict |
Nullable type and prefix assertion |
OpenCode / MiMo Code / Zoo Code: Second batch, 2026-09-26, A17/A14/A19
- OpenCode, 0027387: part.data.type=step-finish has input, output, reasoning,
cache.read/write and cost. Source projector/publish-llm-event explicitly establishes
inputTokens=in+cr+cw and total=sum of five exclusive fields. session.tokens_* compares
totals, matching here; message.data turn totals excluded to avoid duplicates. Exact
call time/model-switch order needs event layer; this stage uses observed_at.
- MiMo, 456678b: related layout with distinguishing message.agent_id, no session
cumulative columns, MIMOCODE_HOME/data. Shared ancestry does not establish compatibility;
both adapters reject the other’s format.
- Zoo, f780647: Roo-related whole ui_messages.json, LIFO start/finish pairing with
finish replacing start and unmatched finish discarded. condense_context.cost contributes
upstream totalCost as auxiliary work with unknown tokens. tokensIn includes cache,
unlike the OpenCode family.
- adapters/opencode_family.rs shares step-finish scan for OpenCode/MiMo with injected product
identity. Sixteen format tests plus twelve module tests pass.
- Cross-product collision fixed: Kilo held matching opencode-rc.db; initial discovery
misidentified it as OpenCode and matched 14 comparison events. Restrict discovery to
opencode/mimocode directories or direct MIMOCODE_HOME/data children. After repair,
not_found agrees with inventory; negative tests and capability collision warning retained.
Outstanding work at this stage
| Item |
Status |
Next step |
| Native data for seven products |
Deferred; locally not_found |
After installation, real_verify/redacted samples; verify each session.version before known_version |
Validation files
- adapters/{cline,dsh,hermes,openclaw}/.
tests/{cline,dsh,openclaw,opencode,mimo_code,zoo}_contract.rs, tests/hermes_contract.rs.- tests/fixtures/{cline,dsh,hermes,openclaw}/: synthetic data and expectations.
- real_verify_{cline,dsh,hermes,openclaw,opencode,mimo_code,zoo}.rs and discover_probe.rs.
