Repository navigation
Progress lines on stderr #10
Description
Parent
What to build
Progress on stderr. While a 10–40 minute run is going, thirdshift condenses each session's stream-json output into short stderr lines (skill invoked, tool calls, commits, pushes) and prints its own deterministic steps (worktree created, pushing, checking the PR, merging the Base branch, waiting for CI, starting a Repair, cleaning up). stdout stays reserved for the PR URL.
Acceptance criteria
- A fake session emitting stream-json events produces one short, readable stderr line per notable event
- thirdshift's own steps each print a stderr line
- stdout still contains only the PR URL
- Unknown or malformed stream events don't crash the run; they're skipped
Activity
JacobStephens2 commentedon Sep 25, 2026
Finding from the first real run of the walking skeleton (
vacation#3→ JacobStephens2/vacation#13):A single implement session's stream contained 4
resultevents, not one. The agent ended a turn while its/thirdshift:code-reviewStandards and Spec sub-agents ran in the background ("I'll wait for both before addressing findings"). Their completions then resumed it. The skeleton was unaffected because it waits forclaudeto exit, not for aresultevent. The progress lines should handle this too:resultevent doesn't mean the session finished. Only the process exiting does, so don't print a "done" line on the firstresult.total_cost_usdis cumulative across those events, so report the value from the lastresult, not a sum. For scale, this run took about 5.5 minutes of session time, 34 turns and $1.82 as reported. On a claude.ai subscription that figure is the API-price equivalent, not a charge (apiKeySource: "none"in the init event).system/initevent per resumption, so the "session started" line should come from the first one only.Merge origin/main into issue-10
Merge origin/main into issue-6