| Item |
Value |
| Date |
2026-09-24 |
| Environment |
Windows 11 Pro 26200 x64; Rust 1.98.0; Node v24.21.0 |
| Code revision |
Uncommitted working tree |
| Design references |
execution.md M2; adapters.md delivery requirements/samples; architecture.md incremental strategy; validation.md V07/V12/V17 |
Commands and results
| # |
Command (working directory) |
Exit |
Result |
| 1 |
npm run verify (root) |
0 |
lint/assets/svelte-check/fmt/clippy -D warnings/cargo test/vite build pass |
| 2 |
cargo test –locked –workspace (desktop/src-tauri) |
0 |
135 passed / 0 failed, M1’s 72 plus new tests |
| 3 |
cargo run -p llm-usage-core –example real_verify_codex – ~/.codex build/desktop-usage-validation/real-check |
0 |
Read-only local checks; permitted-field statistics below |
Implemented changes
- adapters/framework.rs: SourceAdapter discover→detect→scan→capability; scan flow includes
generation rescans, skipping unchanged files, rejecting unrecognized data and restricted
diagnostics. Reconcile per-call totals with final snapshots and carried compaction records.
- adapters/jsonl.rs: all V07 boundaries: no cursor advance for partial lines, UTF-8 across
chunks, controlled errors above 8 MiB per line, BOM, isolated bad rows without content,
generation redetection for same-length replacement/truncation/renaming.
- adapters/codex.rs: Codex 0.155.0-alpha.16.3. Per-call token_usage_record becomes model_call
with resp:{response_id}, falling back to seq:{session}:line. token_count cumulative formats
remain separate. Deduplicate/reconcile compaction carry records; turn_context determines
models, otherwise unknown; codex_vscode maps to host_application=vscode; subagent categories
mapped. source_instances stores capabilities.
- ingest.rs repeated-final fix: same key/content differing only in time is Keep, avoiding
false conflicts; real same-revision differences still retain conflict diagnostics.
Tests: 63 new cases, all real temporary files/SQLite
- codex_contract.rs: end-to-end expectations for three M0 real samples: one call, 49 calls,
multimodel compaction; capabilities/discovery/instances/storage.
- codex_gaps_synthetic.rs: clearly labeled synthetic samples: repeated final, subagent,
positive cache writes, usage-free tool/user messages, missing response_id fallback;
V17 rejects unknown versions/formats.
- codex_incremental_v12.rs: appends, partial lines, truncation/same-length replacement/rename,
segmented reads under scan limits, no new records on rescan, retain existing conflicting final.
- jsonl_v07.rs: each V07 boundary; review_regressions.rs: prior-failure regressions.
Read-only local checks: ~/.codex, 265 rollout files
| Metric |
Observed result |
| Fully scanned files |
53; another 6 degraded for empty files/abnormal metadata |
| Rejected |
212 other-version files: 0.154.0-alpha.6.2×89, 0.155.0-alpha.16×18 and 12+ versions including 0.146/147/149/150/151/153; per-version samples required |
| Lines/records/events |
23,059 lines → 2,620 model-call events |
| Snapshot comparison |
matched 48 / mismatch 5 / no_snapshot 0; final snapshot exceeds per-call total by 2.9M–31.5M tokens, consistent with source retention truncation or cross-file sessions; reconcile_mismatch retained, no silent addition |
| Rescan |
rescan_added=0; no duplicates |
| UTC totals |
calls=2620; input_total 268,545,932; cache_read 257,653,888, about 95.9%; output 1,271,994; total 269,817,926; 3 models |
Diagnostics: unsupported_version 212, snapshot_out_of_order 76, unexpected_session_meta 6,
reconcile_mismatch 5, unknown_record_type 4, usage_shape_deviation 1. No bodies/paths leaked.
examples/real_verify_codex.rs restricts fields; temporary database is in ignored
build/desktop-usage-validation/real-check/.
Completion after interruption
The M2-A subagent stopped at its quota. The main session repaired/validated its implementation:
capability cost assertion, repeated-final Keep logic, same-length replacement sample (original
corrupted cli_version and conflicted with rejection rules; replaced by whole-line swapping),
six Clippy issues, formatting and –force-local for bundle-report tar with Windows drive colons.
Outstanding work at this stage
| Item |
Status |
Next step |
| Other Codex 0.146–0.155 versions |
Partial; M2-D verifies 0.153.0/0.154.0-alpha.6.1/6.2; 0.139–0.151 lack per-call format and need dedicated implementation; unlisted versions use latest_fallback attempts |
Verify token_count/last_token_usage identity before extending |
| OTel intake |
Not implemented |
M5 |
| Actual GUI sources page |
Not implemented |
M6 |
Validation files
- desktop/src-tauri/crates/core/src/adapters/: framework/jsonl/codex; tests/: codex_*,
jsonl_v07, review_regressions and fixtures/codex/.
- Ignored build/desktop-usage-validation/real-check/ contains real-check.sqlite.
