Retro

Turn one agent run into evidence that can improve future runs. Reconstruct only what matters, identify supported improvement candidates, and preserve the result for a separate decision. Do not improve the workflow during the retrospective itself.

Write the report and response in the user's active language.

Establish the run

Analyze the run the user identifies, or the current run when none is named. Prefer primary evidence you can access directly: conversation and tool history, session logs, commits and diffs, test or validation output, produced artifacts, and the exact installed instructions or skills involved.

Do not ask the user to retell evidence that is already accessible. When history was compacted, logs are unavailable, or a claim cannot be checked, name the gap and reduce confidence instead of reconstructing missing events from plausibility. Record the repository revision, worktree, session identifier, or artifact provenance only when available and useful.

Keep secrets, credentials, personal data, and unrelated source content out of the report. Include the minimum sanitized fact or excerpt needed to make each finding understandable on its own. Refer to large logs or artifacts by path and relevant location only as supplementary provenance, never as the report's sole evidence.

Reconstruct the useful story

Describe the original objective, actual outcome, and only the turning points that changed cost, quality, direction, or confidence. Capture both friction and behavior that should be preserved. Do not produce a turn-by-turn transcript or judge success only by whether the final code passed.

For each material friction, separate: