This records recovery acceptance for the interruption. Scope follows its
gap table: four sources’ integration requirements/missing-field/incremental tests and test data,
an independent oh-my-pi adapter decision, five real_verify examples and local read-only native checks.
| Item |
Value |
| Date |
2026-09-25 recovery; interrupted 2026-09-24 |
| Environment |
Windows 11 x64; rustc 1.98.0; Node v24.21.0 / npm 12.0.2 |
| Revision |
Uncommitted worktree: accepted M2-A changes plus recovery additions |
| Requirements |
Adapters; validation V07/V12/V17; execution plan M2 |
Commands and results
| # |
Command and cwd |
Exit |
Results |
| 1 |
cargo test –locked -p llm-usage-core; desktop/src-tauri |
0 |
249 passed/zero failed: interruption baseline 143 plus 106, comprising 103 integration tests across five sources and three omp inline tests |
| 2 |
cargo clippy –locked -p llm-usage-core –all-targets – -D warnings |
0 |
Passed the broader all-targets check; four existing/new test lint issues were fixed as listed below |
| 3 |
npm run verify; repository root |
0 |
Markdown lint: 89 files/zero issues; assets/scripts/Svelte/fmt/clippy passed; core 249/app 2 = 251 Rust tests; Vite build passed |
| 4 |
cargo run -p llm-usage-core –example real_verify_pi – ~/.pi/agent build/desktop-usage-validation/real-check-pi |
0 |
Local read-only native check; permitted aggregates below |
| 5 |
cargo run -p llm-usage-core –example real_verify_omp – ~/.omp/agent build/desktop-usage-validation/real-check-omp |
0 |
Same check for omp |
| 6 |
real_verify_claude / real_verify_gemini / real_verify_qwen |
0 |
Each discovered zero roots; no_data/not_found results below |
Independent oh-my-pi adapter decision
Created adapters/omp.rs with 931 lines/three inline tests. It shares pub(crate) pi parsing helpers
(parse_usage/map_cost/event construction), rather than the entire adapter. Native reads of
58 local session files established these differences:
- First line is type:“title”, v=1; the session header follows. Detection reads at most four
lines: a first line other than title/session is UnknownFormat; title without session is Pending.
- Assistants carry omp-specific floating-point duration/ttft milliseconds, rounded for storage.
- model_change writes a combined model=“provider/model” field, consistent in three redacted
native samples. omp uses this field first and separate modelId/provider fields as fallback;
the separate-field shape was not observed locally. Native samples exposed this implementation
error during recovery.
- For sub-agents, filenames other than
<ts>_<uuid>.jsonl use the nearest <ts>_<uuid> ancestor
directory as parent, taking the portion after its first underscore. Nested sub-agents do not
inherit a more distant ancestor’s identity.
Implementation and tests
| Source |
Native samples |
Synthetic sample count |
contract |
gaps |
incremental |
| pi |
session-error-zero-usage: one error call, all zero |
7 |
2 |
8 |
8 |
| oh-my-pi |
session-glm-reasoning: reasoningTokens=54; session-k3-cache-abort: seven calls including aborted; subagent-community-research: five sub_agent calls |
9 |
4 |
12 |
8 |
| Claude |
None; local no_data |
8 |
2 |
11 |
8 |
| Gemini |
None; local not_found |
8 |
3 |
11 |
5 |
| Qwen |
None; local not_found |
8 |
2 |
11 |
8 |
- Ignored build/desktop-usage-validation/tools/extract-pi-omp.mjs extracts/redacts native samples.
Numbers/booleans/null remain; strings default to REDACTED, enum keys are allowlisted, and
IDs map stably to anon-N. Deduplicated leakage checks left only enum values. Each sample
directory has manually calculated _expectations.md, independently checked record-by-record
with jq and compared against test expectations.
- input_total is derived input+cacheRead+cacheWrite through usage_map.rs map_pi_family;
total_tokens is derived input_total+output; source_total preserves reported totalTokens.
Recovery fixed two expectation documents that confused derived input_total with raw input sums.
- omp gaps tests cover four auxiliary usage formats, assistants without usage (count calls,
leave tokens unknown), copied fork entries, nested sub-agents, error/aborted, cost>0 as
estimated (zero omitted), four first-line detection outcomes, rejection of unknown versions/
formats and one diagnostic for an unknown entry type. These describe this historical stage.
Local read-only native checks: 2026-09-25 permitted aggregates
pi: ~/.pi/agent, pi 0.87.1
| Metric |
Result |
| Completely scanned files |
Two; diagnostics=0 |
| Lines/records/events |
130 lines produced 37 model-call events |
| Repeat scan |
rescan_added=0 |
| UTC totals |
calls=37; input_total 2,805,788; cache_read 2,592,768; output 58,631; total 2,864,419 |
The interruption’s empty pi sessions/no_data result was measured on the morning of 2026-09-24.
Local sessions began at 16:37 that day; the first redacted sample came from that session.
Two files existed at recovery. The invariant holds: 2,805,788+58,631=2,864,419.
oh-my-pi: ~/.omp/agent, omp 18.2.7
| Metric |
Result |
| Completely scanned files |
59: 17 primary/42 sub-agent files; diagnostics=0 |
| Lines/records/events |
28,004 lines produced 8,767 model-call events |
| Categories |
primary=6,334; sub_agent=2,433; auxiliary=0, consistent with native reads finding no usage in the four auxiliary formats |
| Repeat scan |
rescan_added=0 |
| UTC totals |
calls=8,767; input_total 1,021,931,700; cache_read 1,005,572,583, about 98.4%; output 4,859,995; total 1,026,791,695 |
The invariant holds: 1,021,931,700+4,859,995=1,026,791,695.
Claude Code, Gemini CLI and Qwen Code
| Source |
Local discovery repeated 2026-09-25 |
Result |
| Claude |
~/.claude exists with backups/ide/sessions/skills, but no projects/; real_verify_claude discovered_roots=0 |
no_data |
| Gemini |
~/.gemini absent; real_verify_gemini found zero roots |
not_found |
| Qwen |
~/.qwen absent; real_verify_qwen found zero roots |
not_found |
These three adapters/tests rely on documentation/fixed-source references listed in their
headers. Extract native samples under m0-agent-fixtures redaction rules when local data appears.
Observed fork behavior: pi and omp
The original capability text claimed identical copied four-field keys/content would select Keep.
In native checks, copied entries carry the fork file’s session_id/parent_session_id from its
header. They therefore share an existing key but differ in content, producing conflict and
retaining the first scanned record. Usage is not counted twice, source-session ownership is
retained and distinct-session counts do not increase; the result is conflict rather than Keep.
pi.rs/omp.rs dedup.fork_copies descriptions and synthetic-fork-inherited/_expectations.md now
describe the observed behavior.
Recovery fixes
| Location |
Problem |
Correction |
| tests/qwen_incremental_v12.rs |
Four writes lacked trailing newline; incomplete last lines were not consumed, causing seven of eight tests to fail |
Added \n endings; eight of eight passed |
| omp.rs model_change |
Parsed pi’s separate fields; native omp uses a combined field |
Prefer model=“provider/model”, retain separate-field fallback; updated capability/header |
| omp synthetic-detect-supported |
Derived input_total expectation was incorrectly 110, equal to totalTokens |
Changed to 100, from 100+0+0; synchronized test/_expectations.md |
| omp k3/subagent _expectations.md |
Stored input_total described as raw-input sum |
Corrected to 202,560/78,310 and explained derivation |
| Two pi sample _expectations.md files |
markdownlint MD056/MD033 |
Added missing table columns and code spans around angle brackets |
| models_v05/retention_v14/claude_contract/qwen_incremental |
Four all-targets Clippy issues: numeric grouping, needless borrow, type_complexity, useless vec! |
Small local changes preserving semantics |
Remaining work outside this recovery
| Item |
Status |
Follow-up |
| Version-specific legacy Codex samples |
Partial, 2026-09-25; M2-D: 0.153.0/0.154.0-alpha.6.1/6.2 verified; 0.139–0.151 lack per-call records, requiring a separately verified implementation |
Follow M2-D conditions |
| Adapter directories/unknown-version compatibility, execution.md#m2-layout |
Complete, 2026-09-25; M2-D |
— |
| OMP_PROFILE named-profile paths |
Unverified |
Default ~/.omp only; limitation recorded in omp.rs |
| Claude/Gemini/Qwen native samples |
No local data |
Extract/redact/recheck when data appears |
| OTel telemetry |
Not implemented at this stage |
M5 |
Verification artifacts
- Adapters: adapters/{pi,omp,claude,gemini,qwen,usage_map}.rs; tests:
tests/{pi,omp,claude,gemini,qwen}_*.rs, tests/common/mod.rs and tests/fixtures/{pi,omp,claude,gemini,qwen}/.
- Examples: real_verify_{codex,pi,omp,claude,gemini,qwen}.rs.
- Ignored build/desktop-usage-validation/real-check-{pi,omp,claude,gemini,qwen}/ contains temporary
verification databases; tools/extract-pi-omp.mjs is the ignored extraction tool.
