retrolisted

This skill runs once per session, on request (`/retro`), over every task the session completed — not after each `tasqx_complete_task`. Until 2026-09-16 a PostToolUse hook nudged a retro per completion; each run cost ~56k tokens, 14 completions in 36 h cost ~800k, and a measurement (memory doc 01a0a754-e116) found the docs those retros wrote were never read back. Decided in claude-config #675: one run, one agent, one bundle per task that earned it, and nothing written on a task that went straight through. Its evidence is an evidence bundle the main session assembles from tasqx and the executors' reports (Step 0); the session transcript is not read. Its outputs are an annotation on each reviewed task, and optionally a memory entry, a hookify recommendation, or a skill candidate. A skill is promoted only on the second sighting of the same procedure — retros overrate their own novelty. ## Step 0: Pick the tasks, then delegate to a fresh agent on an evidence bundle **Which tasks.** With a ref (`/retro 123`), that task. With none, every task whose `tasqx_complete_task` this session made; if unsure, `tasqx list status:done` newest first and take the ones completed since the session began. **Which of those earn a bundle.** Only a task whose work had rework (a reverted edit, a re-run command, a wrong assumption), a discovery written down nowhere, a blocked or refused step, or an executor that came back wrong. The main session judges this from what it watched. A task that went strai