collect_agent_files now head-scans every agent jsonl per call — multiplies the whole-corpus walk cost

[coryking]

is_conversion_artifact detection calls read_provenance (file open + ≤8-line head read) on every agent transcript, on every collect_agent_files call. session_sources calls collect_agent_files per session per search; activity and grep paths pay the same. On a corpus with hundreds of sessions × many agents this multiplies the already-tracked whole-corpus walk cost (see the perf note in search.py session_sources docstring).

No caching exists on this path (load_transcript's mtime cache doesn't cover it). Breadcrumbs: a (path, mtime)-keyed module cache, or folding the probe into a single discovery pass. Surfaced by code review of PR #30 (efficiency angle + angle C, finding 6).

[Necmttn]

This sounds like a single discovery/index pass rather than a per-call provenance probe.

Cache path, size, mtime, first-line hash, provenance classification, and parse status for every agent JSONL. session_sources, activity, and grep can reuse that index and invalidate only changed files, instead of multiplying head reads across each session search.

---

Generated with ax.