workflow-retro

原文涵盖的主题

• Workflow retro
• 1. Collect
• 2. Find causes
• 3. Decide with the running log
• 4. Converge
• 5. Apply and record
• Running log format

Turn what agents logged while working (ledger.sh, global CLAUDE.md) into fewer round trips, fewer errors and checks sized to the change, without growing the instructions. The session owner runs it; workers only hand over their ledger. A retro whose only output is deleting or merging rules is a good retro; zero edits is a fine outcome. Ledgers: the run's dir (bash /ledger.sh dir ), or this session's file (ledger.sh path), plus your own. Metrics: per stage wall time (orchestrated: orca orchestration task-list --run --json --full (the concise view has no timestamps), createdat → completedat), fix rounds, review findings by severity, every [check] line, and the profile's "Retro" metric sources (e.g. hook logs). The running log (path in the profile's "Retro") and the last context-size row. 1. Group entries by root cause, not symptom. The same cause in two workers' ledgers is the strongest signal of the run. 2. Attribute each to the stage or file that could have prevented it. A Build spec-gap is a Specify or spec-template fix; a review finding is an escaped against the check that should have caught it. 3. Scope against checks, per PR: diff size and area against what ran ([check] lines, g

这里只展示节选，请以完整原文和当前要求为准。