retro

Sprint 1.3 fan-out + Sprint 3 composable skill (v8.16+). Sub-reflectors merge via dedup-by-title.

When to invoke

• After ship completes (PASS) or after audit FAIL/WARN (capture lessons either way)
• Cycle is in ship or audit phase, transitioning to learn

When NOT to invoke

• Cycle aborted before producing any artifacts
• Pure-documentation cycles (no behavioral changes to reflect on)

Workflow

Step	Action	Exit criteria
1	Read build-report.md, audit-report.md, state.json	Inputs loaded
2	Dispatch 3 sub-reflectors in parallel	3 worker artifacts
3	Aggregator dedups ## Lesson: blocks across workers	<workspace>/retrospective-report.md produced
4	Optionally write to .evolve/instincts/lessons/<slug>.yaml	Lessons persisted

The three sub-reflectors

Sub-reflector	Focus	Output style
retro-instinct	Reusable patterns across cycles	## Lesson: <pattern title>
retro-gene	Gene-pool updates based on cycle outcome	## Lesson: <gene insight>
retro-failure	Failed approaches with reproduction steps	## Lesson: <failure pattern>

Report format

retrospective-report.md carries deduplicated ## Lesson: blocks. Optionally one or more YAML files in .evolve/instincts/lessons/.

Phase facts

Fact	Value
Phase	retrospective (control archetype, optional)
Persona	agents/evolve-retrospective.md
Profile	.evolve/profiles/retrospective.json — CLI codex-tmux, tier deep, fan-out ×3
Inputs	audit-report.md · build-report.md
Artifact	retrospective-report.md (cycle workspace)

Output contract

retrospective-report.md — no required sections (classified by the phase runner).

Composition

Invoked by:

• /evo:retro
• loop macro after /evo:ship

Fan-out prompts live in .evolve/profiles/retrospective.json:parallel_subtasks (count projected into Phase facts above).

Reference

• .evolve/profiles/retrospective.json
• legacy/scripts/dispatch/aggregator.sh (phase=learn → lessons mode with dedup)
• agents/evolve-retrospective.md
• skills/loop/phase6-learn.md (legacy detailed workflow)