feat(discovery): discover omp session transcripts in the pi adapter

[mdc2122]

Summary

omp (Oh My Pi) writes pi-compatible version:3 JSONL session transcripts, but the pi adapter could not see them. Two gaps:

1. Store root: omp writes to ~/.omp/agent/sessions/ (it honors PI_CODING_AGENT_DIR, defaulting to ~/.omp/agent), which was not among the adapter's roots.
2. Header layout: omp prepends a fixed-width {"type":"title",...,"pad":"…"} record as line 1 (rewritten in place when the title generates), so the {"type":"session","version":3,"id":…,"cwd":…} entry is line 2. classify() read line 1 only and required type:"session" + cwd, so every omp file classified as null.

Changes

• storeSpecs(): add ~/.omp/agent/sessions as a default nested root. An env redirect via PI_CODING_AGENT_DIR dedupes against it through the existing realpath map.
• classify(): scan the first head lines for the type:"session" entry instead of requiring it on line 1.
• enumerate(): descend one extra level inside nested roots — omp writes subagent transcripts at <escaped-cwd>/<session-id>/<Name>.jsonl.
• read(): accept model_change.model as a fallback for modelId (omp writes model).

Verification

Against a live omp store (~500 .jsonl across ~55 project dirs, including ~250 nested subagent transcripts): all files classify with a real cwd, and read() parses message/toolCall/toolResult events and the model unchanged — the transcript body format is identical to pi's. New fixture test/fixtures/omp-session.jsonl covers the title-record layout and model field; the enumerate test covers the omp root and the depth-3 subagent layout. pnpm run check passes (662 tests).

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge, with the documented omp layouts covered without weakening existing transcript validation.

_{Reviews (1) · Last reviewed commit: "feat(discovery): discover omp session tr..."}

[kunchenguid]

Speaking as Kun's firstmate.

Reviewed head 82448d1290245b545b5aa3271d5a7267ad4141d3 (diff + fixture + prior OMP thread).

VISION verdict (per rule):

1. Evidence is the only currency — aligns. Inspected test/fixtures/omp-session.jsonl: title line then type:session + cwd, real user/assistant/toolResult turns, model_change.model. classify()/read() keep quote-bearing turns; no summary/checkpoint path.
2. The human owns the weights — aligns (discovery only; no write-path change).
3. Nothing the model says is taken on faith — aligns. Pinned fixture + adapter tests (classify past title, model fallback, enumerate omp root + depth-3 subagent). Not a README-only row.
4. The budget is the constraint — n/a (discovery).
5. It reads what you already have — aligns. Local ~/.omp/agent/sessions, fail-soft via existing listDirs/listFiles, no upload/key. Same evidence class as pi. More harnesses welcome; folding omp into the pi adapter (identical version:3 JSONL) should not destabilise other adapters. Matches prior OMP ready-for-pr (Features: Add support for OMP "Ohmypi" #53) and feat(discovery): add OMP session discovery #21 triage (default omp discovery was not held as a captain default-behavior call).
6. Failure is loud and named — aligns. Non-session heads still classify null and are skipped; missing store stays empty like other nested roots.
7. Scope — aligns (adapter discovery, not a writer/provider/CI).

contract-class: restore — OMP support is already product-accepted (#53 ready-for-pr); this closes the pi-adapter gaps (store root, title-record line 2, nested subagent depth, model field) so unconfigured discovery matches that promise for pi-compatible omp transcripts. Not treating as new-default (same stance as #21).

Security: clean. Local read-only discovery; no secrets, no egress, no permission changes.

CI: first-time fork workflows approved. CI (ubuntu+macos) and Guard green on this SHA. Require no-mistakes is red — PR body has no no-mistakes attestation / signature (opened run 35051548667 failed). That check is blocking.

Still waiting on you, not on the captain: run git push no-mistakes so the body attestation matches head 82448d12. Optionally add Closes #53 in that body rewrite. Branch is ahead 1 / behind 0 of main — no rebase needed. I will not merge until no-mistakes is green on this head.