Every session cut into work segments by the repository it was actually about. 250 segments · 21522 turns · 1925 commits · 2026-06-29 → 2026-10-19
How a segment is made. A session is not one topic. Each record is scored for which repository it concerns — a file path under
/projects/<repo> counts most, a bare repo name counts less — and the session is cut wherever that changes and
stays changed for three records. A commit joins a segment when the repository matches and the time falls inside it (
in), or within six hours of its edge (
near).
590 of 1,925 commits (30%) attach to nothing, and the reason is almost entirely a gap in the transcripts rather than a failure of the join:
| why | commits | where |
|---|
| inside the window, but no session was running | 509 | ams-demo-0 175 · resolve-questions 99 · knia-analysis 70 · knia-dag 70 · question-ledger 55 |
| committed after the transcripts end | 29 | visus (post 2026-10-13) |
| repository never appears in any session | 23 | acci-vlm |
| committed before the transcripts begin | 16 | CarAccident (from 2026-05) |
| the 2026-10-14 handover snapshot itself | 13 | one per archived repository |
The corpus covers 53 days. July and August hold
477 commits against 9 days of session coverage; September holds 1,165 across 28. So the work on ams-demo-0, resolve-questions, knia-analysis and knia-dag — all of it July and August — happened in sessions that were never captured.
Those four repositories account for 414 of the 509.Nothing is redacted. Usernames, paths, plate numbers and KB material appear as they were written. Five AWS access-key
ids appear; no secret keys, database URLs or tokens do. In-house only.