orion v0.7.2

August 28, 2026

A cost report is now posted to the ticket when Orion transitions it to Done, and printed to the watch/collect console the moment the merge is announced — both rendered by the same code path, so the two cannot disagree. It totals every agent run the ticket caused (the implementation run, each fix-loop re-entry, and runs that died on a timeout, breaker or quota), broken down per actor: runs, turns, input, output, cache creation and cache reads reported separately, wall time, and the estimated USD the runner itself reported per session. Runs that failed are marked, and a run whose usage never arrived is stated rather than dropped — the report then says its totals are a floor. Re-running collect on a closed ticket never posts it twice.

A passing verification command (go test, make test, etc.) no longer counts toward the identical-repeat loop breaker. Re-running the tests is the normal edit-test cycle and exactly what the unverified-edits breaker demands after every edit; counting it as a loop made the two breakers trip each other. A failing verify still counts (OR-124).
A claude -p subprocess that exits cleanly without ever emitting its own stream result line (killed externally mid-run, a crash after a partial flush) is now reported as a failed run with a clear reason, instead of reading as a silent success with no cost recorded and an empty session id (OR-127).
Every gh call on the orion watch/orion collect hot path (PR status, merge, prune, and the fix loop's CI-log fetch) is now bounded by a 45s timeout, so a hung gh process can no longer block the watcher indefinitely with nothing on the console to explain why (OR-128).
orion collect prints a line before searching Jira for tickets awaiting CI, so the first network call of every watch tick has visible feedback rather than possible silence (OR-128).