retro

name	retro
version	1.0.0
standalone	true
description	Periodic cross-agent learning consolidation — review what worked, what didn't, propagate insights
uses	["core/memory"]
requires	{"tools":["curl"],"env":[]}
security	{"risk_level":"read","capabilities":["memory:read","memory:write"],"requires_approval":false}

Retro — Cross-Agent Learning Consolidation

Periodic retrospective that reviews agent learnings, consolidates patterns, and propagates high-value insights across agents. Run weekly or after significant milestones.

Standalone mode: Works without agent-coordinator. Scans .os/agents/*/memory/ directly instead of using the coordinator API.

When to Use

• Weekly scheduled consolidation
• After a sprint or milestone completion
• When learning propagation thresholds are met
• When you notice agents repeating the same mistakes

Steps

1. Gather Agent Learnings

Scan each agent's L1 memory for recent learnings:

# Framework mode
curl -s "${AC_COORDINATOR_URL:-http://localhost:9889}/memory/recent?days=7"

# Standalone mode
find .os/agents/*/memory/ -name "*.md" -newer .os/retro-last-run 2>/dev/null

2. Categorize Findings

Group learnings into:

• Patterns: Recurring approaches that worked (promote to L2)
• Anti-patterns: Mistakes that keep happening (add to skill references)
• Process gaps: Where the workflow broke down (create issues)
• Tool insights: Better ways to use tools (update SKILL.md files)

3. Propagate High-Value Learnings

For each pattern with applications >= 2 and successes >= 1:

# Framework mode — trigger propagation
curl -s -X POST "${AC_COORDINATOR_URL:-http://localhost:9889}/memory/${AGENT_ID}/consolidate"

For standalone mode, manually copy high-value learnings from .os/agents/{name}/memory/ to .os/memory/shared/.

4. Update Skill References

If a learning reveals a common mistake or pattern:

1. Identify the relevant skill
2. Add the pattern to that skill's references/ directory
3. Example: review misses → add to review/pr-review/references/common-misses.md

5. Create Issues for Process Gaps

If the retro reveals systematic problems:

1. Create a GitHub issue describing the gap
2. Tag it with the relevant category
3. Prioritize based on frequency and impact

6. Record Retro Summary

Write a summary to .os/retro/YYYY-MM-DD.md:

# Retro — {date}

## What Worked
- {pattern}: {evidence}

## What Didn't
- {anti-pattern}: {evidence}

## Actions
- [ ] {action item}

## Learnings Propagated
- {learning} → L2 shared memory

Touch .os/retro-last-run to mark the timestamp.

Completion States

• DONE: Retro complete, learnings propagated, summary written
• DONE_WITH_CONCERNS: Retro complete but some propagation failed
• BLOCKED: Cannot access agent memories
• NEEDS_CONTEXT: Need clarification on which time period to review