Review Learnings

Interactively review pending learnings in the backlog. For each learning, you can Approve, Reject, Edit, Skip, or Quit.

Batch Operations

• --approve-all-high: Auto-approve all high-confidence pending learnings (with confirmation)
• --reject-all-low: Auto-reject all low-confidence pending learnings (with confirmation)

Preflight Checks

1. Check if backlog file exists: !test -f .claude/learnings/backlog.yaml && echo "EXISTS" || echo "NOT_EXISTS"

Context

Parse $ARGUMENTS for batch operation flags:

• --approve-all-high: Batch approve high confidence learnings
• --reject-all-low: Batch reject low confidence learnings

If neither flag is present, run interactive review.

Based on preflight:

• If EXISTS: Proceed with review flow
• If NOT_EXISTS: Show helpful message about empty backlog

Task Instructions

1. Handle Missing Backlog

If backlog doesn't exist (NOT_EXISTS from preflight):

## Review Learnings

No backlog found at .claude/learnings/backlog.yaml

The learning backlog is empty. You can:
- Add a learning manually: /m42-signs:add
- Extract from a session: /m42-signs:extract <session-id>

2. Read and Parse Backlog

If backlog exists:

1. Read .claude/learnings/backlog.yaml
2. Parse the YAML structure
3. Filter learnings to only status: pending
4. If no pending learnings, show message and exit

3. Handle Batch Operation: --approve-all-high

If --approve-all-high flag is present:

1. Count all pending learnings with confidence: high
2. Display count: "Found N high-confidence pending learnings"
3. If count == 0: "No high-confidence pending learnings to approve"
4. If count > 0:
   • Show confirmation: "About to approve N high-confidence learnings. Continue? [Y/N]"
   • Use AskUserQuestion with Yes/No options
   • On Yes: Update status from pending to approved for all matching learnings
   • On No: "Batch approval cancelled"
5. Save updated backlog.yaml
6. Show summary: "Approved N learnings"

4. Handle Batch Operation: --reject-all-low

If --reject-all-low flag is present:

1. Count all pending learnings with confidence: low
2. Display count: "Found N low-confidence pending learnings"
3. If count == 0: "No low-confidence pending learnings to reject"
4. If count > 0:
   • Show confirmation: "About to reject N low-confidence learnings. Continue? [Y/N]"
   • Use AskUserQuestion with Yes/No options
   • On Yes: Update status from pending to rejected for all matching learnings
   • On No: "Batch rejection cancelled"
5. Save updated backlog.yaml
6. Show summary: "Rejected N learnings"

5. Interactive Review Flow

For each pending learning (1 of N):

5.1 Display Learning Details

Show the learning in this format:

## Learning: <id> (<current>/<total>)

**Title**: <title>
**Confidence**: <confidence>
**Target**: <target>

### Problem
<problem description - multi-line>

### Solution
<solution description - multi-line>

### Source
- Tool: <source.tool>
- Command: <source.command>
- Error: <source.error>

---
Actions: [A]pprove | [R]eject | [E]dit | [S]kip | [Q]uit

5.2 Prompt for Action

Use AskUserQuestion with these options:

• Approve: Mark learning as approved (ready to apply)
• Reject: Mark learning as rejected (will not apply)
• Edit: Enter edit mode to modify fields
• Skip: Leave as pending, move to next
• Quit: Stop reviewing, save progress

5.3 Process Action

On Approve:

1. Update the learning's status from pending to approved
2. Save backlog.yaml immediately
3. Show: "✓ Learning approved"
4. Move to next pending learning

On Reject:

1. Update the learning's status from pending to rejected
2. Save backlog.yaml immediately
3. Show: "✗ Learning rejected"
4. Move to next pending learning

On Edit:

1. Enter edit mode (see section 6)
2. After edit completes, return to this learning's action prompt

On Skip:

1. Keep status as pending
2. Show: "→ Skipped"
3. Move to next pending learning

On Quit:

1. Show review summary (approved/rejected/skipped counts)
2. Save any pending changes
3. Exit review

6. Edit Mode

When user selects Edit:

6.1 Show Editable Fields

Present current values and ask which field to edit:

## Edit Learning: <id>

Current values:
- Title: <current title>
- Problem: <current problem>
- Solution: <current solution>
- Target: <current target>

Which field would you like to edit?

Use AskUserQuestion with options:

• Edit Title
• Edit Problem
• Edit Solution
• Edit Target
• Done Editing

6.2 Collect New Value

For the selected field:

• Title: "Enter new title:"
• Problem: "Enter new problem description:"
• Solution: "Enter new solution description:"
• Target: "Enter new target path:" (suggest available CLAUDE.md files)

6.3 Validate After Edit

After any edit:

1. Run validation: "${CLAUDE_PLUGIN_ROOT}/scripts/validate-backlog.sh" .claude/learnings/backlog.yaml
2. If validation fails:
   • Show validation errors
   • Revert the change
   • Prompt to try again or cancel edit
3. If validation passes:
   • Save the change
   • Consider confidence adjustment (see 6.4)

6.4 Confidence Adjustment

If the learning was from automated extraction (source.tool != "manual"):

• If title or solution is significantly changed (>50% different), show: "This learning was auto-extracted. Significant edits may affect confidence."
• Offer to adjust confidence: Keep current | Set to medium | Set to low

6.5 Return to Review

After edit mode completes ("Done Editing"):

1. Save all changes to backlog.yaml
2. Re-display the learning with updated values
3. Return to action prompt (Approve/Reject/Edit/Skip/Quit)

7. Review Complete

When all pending learnings have been processed:

## Review Complete

| Action | Count |
|--------|-------|
| Approved | N |
| Rejected | N |
| Skipped | N |

Next steps:
- Run `/m42-signs:apply` to write approved learnings to CLAUDE.md files
- Run `/m42-signs:status` to see current backlog status

8. Handle Edge Cases

• No pending learnings: "No pending learnings to review"
• Single learning: Skip (N/M) counter, just show "Learning: <id>"
• Invalid YAML: "Error parsing backlog.yaml - check format"
• File write error: "Failed to save backlog.yaml - check permissions"

Success Criteria

• Pending learnings are displayed with title, problem, solution, target, and confidence
• Interactive flow allows Approve, Reject, Edit, Skip, and Quit actions
• Status updates work correctly: pending → approved or pending → rejected
• Edit mode allows modifying title, problem, solution, and target
• Edits are validated before saving (using validate-backlog.sh)
• Batch operations (--approve-all-high, --reject-all-low) work with confirmation
• Changes are saved after each decision
• Review summary shows counts at completion