task-registry

Task Registry

One registry.jsonl at the repo root. Every task by every agent, sub-agent, or worker job gets entries. Git-tracked — committed together with the work it describes.

Why

Git captures what changed. It loses work with no diff: audits, verification runs, RCA findings, decisions, abandoned approaches. The registry captures task state and outcomes so the next agent reads one file instead of reconstructing history.

Rules

1. Session start: if registry.jsonl exists, read the last ~30 lines before the first action. Active entries from other agents = work in flight; don't duplicate or collide with it.
2. Task start: append an active line before beginning work.
3. Task end: append a done / failed / abandoned line. Always include outcome — one sentence, the finding or result. Failed/abandoned entries state the reason; they are as valuable as done ones.
4. Progress lines (long tasks): append a progress line at objective milestones — a commit made, a phase completed, a blocking finding, a decision taken. Never on a timer. Each line must state new information; repeating the last line is noise.
5. No-diff work still registers: audits, RCA, decisions, verification runs get entries with "commits":[] and the finding in outcome.
6. Commit linkage: completion lines list the commit SHAs of the work. Commit the registry change in the same commit as the work it describes.
7. Append only: never edit or delete existing lines. Corrections are new lines referencing the same id.
8. One file: no per-workstream splits. Rotate to registry-YYYY-MM.jsonl only if reads become slow.

Line schema

One JSON object per line. Fixed keys, no nesting:

{"id":"T-20260806-1231-idx-audit","ts":"2026-08-06T12:31:00Z","agent":"sub-idx","task":"fpd covering index audit","status":"active"}
{"id":"T-20260806-1231-idx-audit","ts":"2026-08-06T13:02:00Z","status":"done","commits":["a3f9c2e"],"outcome":"idx unused (50 scans in window vs 197M on neighbor); drop decision deferred pending 1 day of worker traffic"}