1. Most turns: hunting for a cause that probably isn’t in the unpushed commits. The test passed locally and on develop. I spent about a dozen calls reading the grammar-check chain (check_core_a_block_over_a_grammar, run_grammar, Workshop::run) to explain an empty findings list, and never found how it could happen. The CI-log fetch that showed [] came late.

2. Lookups the brief could have handed me:
   • The brief points at docs/briefs/qa-input.txt, but this lane’s input is docs/briefs/async/qa-input.txt. I only found it by listing directories.
   • The input held only the tail of suite.log, so I had to run gh run view --log-failed to see the actual panic.
   • The .claude/agents/qa.md pointer was fine.
   • I found the test file, the check path and the shared workshop by grep.
3. Checks:
   • Useful: the single failing test (passed), the dsl_grammar module on a cold cache (passed), the arm64 history on the two earlier develop runs, and the diff review.
   • Wasted: the cold-cache run of dsl grammar filtered across the whole it binary. It filled the disk and gave unrelated failures, and I then had to clean up the scratch cache. I also read parser/mod.rs and check_core_expr diffs that had nothing to do with the failing test.
   • I ran no fmt, clippy or pre-push, because nothing changed.
4. One change: have the driver put the failing test’s panic output, not just the last 200 lines, in qa-input.txt. Add a rule to qa.md: if a failing test passes locally and on the last develop CI, and the diff doesn’t touch its path, report it as a flake after at most 10 turns. Don’t run cold-cache runs on a disk that is almost full.

This site is open source. Improve this page.