jwilger Factory Review
Value: Communication -- factory mode concentrates the human's attention on decisions that matter. This skill ensures the review interface is clear, concise, and actionable so the human can provide meaningful oversight without re-reading every line of code.
Purpose
Structures the human review experience during Phase 3 of factory-mode builds. Produces audit trail summaries, PR digests, retrospective synthesis, quality trend reports, and an autonomy tuning interface. The goal is high-signal review in minimal time.
Practices
Audit Trail Summary
Read from .factory/audit-trail/ and present a concise build summary:
- Slices completed: Count and list (by slice ID and title)
- Rework rate: Percentage of slices that required rework cycles
- Gate failure distribution: Count of failures by gate type (tdd, review, ci, mutation)
- Escalations pending: Count and brief description of each unresolved escalation
Format as a dashboard-style summary. Lead with the numbers, follow with details only if the human asks.
PR Digest
For each merged PR in the build, produce a one-paragraph summary:
- What changed: Feature or fix description in plain language
- Which slice: The vertical slice ID this PR implements
- Which pair: The two engineers who built it
- Gate results: Pass/fail for each quality gate (tdd, review, ci, mutation)
- Rework count: Number of rework cycles before all gates passed
Keep each digest to 3-5 sentences. The human should be able to scan all PR digests in under a minute for a typical build session.
Retrospective Synthesis
Aggregate findings from team retrospectives (stored in
.factory/audit-trail/retrospectives/) and surface patterns:
- Recurring rework causes: What kinds of issues triggered the most rework?
- Pair effectiveness: Which pairings produced the fewest rework cycles?
- Domain hotspots: Which areas of the domain model generated the most discussion or revision?
- Process friction: Any team-identified impediments or improvement suggestions
Present as bullet points grouped by theme. Do not editorialize -- report what the team said.
Quality Trend Tracking
Read from .factory/audit-trail/metrics/ and present trends:
| Metric | Current | Previous | Trend |
|---|---|---|---|
| Mutation score | % | % | up/down/stable |
| Rework rate | % | % | up/down/stable |
| Cycle time per slice | duration | duration | up/down/stable |
| Gate failure rate by type | counts | counts | up/down/stable |
"Previous" means the last completed build session. If no previous session exists, omit the comparison column.
Tuning Interface
Accept adjustments to .factory/config.yaml during review. For each proposed
change:
- Validate: Check that the change is consistent (e.g., cannot enable
auto-merge at
conservativeautonomy level; cannot disable a gate that another setting depends on) - Explain implications: What will change in behavior if this setting is modified? Be specific.
- Apply or reject: If valid, apply the change. If invalid, explain why and suggest the nearest valid alternative.
Never apply config changes silently. Always confirm with the human before
writing to .factory/config.yaml.
Escalation Review
Present each pending escalation with full context:
- Which gate: The quality gate that triggered the escalation
- Rework attempts: How many rework cycles were attempted
- What was tried: Brief summary of each rework attempt
- Current state: What the code looks like now (diff or description)
- Recommendation: The team's suggested resolution (if any)
The human decides: resolve, override, or send back for more rework.
Enforcement Note
Advisory in all modes. Summaries and trend reports are self-enforced formatting disciplines.
Hard constraints:
- No silent config changes:
[H]
Constraints
- "No editorializing": Report what the team said, not what you think about what the team said. "The team discussed X and decided Y" is reporting. "The team wisely decided Y" is editorializing. "The team discussed X but missed the important point about Z" is editorializing. If you have a concern, present it separately as your own analysis, clearly labeled.
- Config validation: "Never apply config changes silently" means every config change goes through explicit user confirmation, even if the user previously said "go ahead and configure things." Each individual change is confirmed. Blanket permission does not exist for config changes.
Verification
After completing a factory review session, verify:
- Audit trail summary was presented with current build metrics
- Every merged PR has a digest (no PRs skipped)
- Retrospective synthesis covers all team retrospectives from the session
- Quality trends include comparison to previous session (if available)
- All pending escalations were presented with full context
- Any config changes were validated before being applied
- Human confirmed or overrode each escalation
If any criterion is not met, revisit the relevant practice.
Dependencies
This skill requires the pipeline skill for factory mode infrastructure.
It integrates with:
- pipeline: Reads audit trail files and metrics produced by the pipeline orchestrator during Phase 2
- ensemble-team: Retrospective synthesis reads team retro output; pair effectiveness data comes from pairing history
- ci-integration: Gate failure data includes CI_RESULT packets from the ci-integration skill
Missing a dependency? Install with:
npx skills add jwilger/agent-skills --skill pipeline