---
name: rw-reflect
worker-type: hook
sidecar-path: _reflections/
blocking: true
requires: []
capabilities: [no-web-access]
eval-signals: [coverage-improved]
trigger: on-case-close
---

# rw-reflect: Post-Case Reflection Worker

## Input

- Completed case: Q-NNN (answered or deferred) or D-NNN (proposed)
- All artifacts produced during the case lifecycle (discovery logs, corpus files, notes, comparisons, findings, critiques)

## Output

- `{domain}/_reflections/YYYY-MM-DD-{slug}.md`

## Responsibilities

1. **Compare planned vs actual**: Count artifacts planned (from Q-NNN sub-questions and evidence needed) vs artifacts actually produced. Report specific numbers: sources discovered, sources captured, notes written, comparisons written, findings produced, open questions promoted.
2. **Identify surprises**: Document plan-vs-actual deltas. What took longer than expected? What was easier? What failed? What unexpected results appeared? Reference specific file paths — no vague assessments like "went well" or "was challenging."
3. **Extract lessons**: Concrete, actionable lessons. Each lesson must state: what happened, why it matters, and what to do differently next time. Categorise as: Process, Planning, Accuracy, Infrastructure, Quality, or Design.
4. **Promote unanswered threads**: Review `synthesis/open-questions/` for threads that emerged during this case. List each as an OQ-NNN candidate with priority and suggested next steps. These feed back into Phase 0 (Question Framing) for future cases.
5. **Quantitative metrics table**: Include a table with columns: Metric, Planned, Actual. Cover at minimum: sources discovered, sources captured, notes written, comparisons written, findings produced, open questions promoted, self-correction iterations triggered, challenge catch rate (if rw-challenge ran).

## Constraints

- **Read-only** on all research artifacts. This worker reads and summarises — it never modifies research outputs.
- Reference specific file paths and counts. Every claim in the reflection must point to a concrete artifact.
- No vague assessments. "Quality was good" is not acceptable. "3/4 findings have high confidence, 1/4 medium (F-007)" is acceptable.

## Job-Hunter Reflection Focus Areas

Beyond generic process metrics, note findings relevant to the UK SaaS context:

- **Regulatory surprises**: Any ICO guidance, case law, or GDPR interpretation that was more restrictive or permissive than expected
- **Platform instability signals**: Any evidence of platform API changes, new bot-detection, or ToS updates during the research
- **UK data quality issues**: Any gaps or inconsistencies found in Home Office CSV, ONS data, or salary survey data
- **Scope creep flags**: Any research question that expanded significantly beyond its original scope — flag for future budget calibration

## Self-Check (Level 1 Self-Correction)

Before completing, verify:
- [ ] Quantitative metrics table is present with Planned vs Actual columns
- [ ] Every surprise references a specific file path or artifact
- [ ] Every lesson has a category and an actionable recommendation
- [ ] Unanswered threads section lists OQ-NNN candidates (or states "none")
- [ ] Reflection file name follows `YYYY-MM-DD-{slug}.md` convention

Max 2 self-correction iterations. If self-check still fails after 2 retries, emit `status-recommendation: blocked` with a description of what failed.

## Reflection Format

See `references/reflection-format.md` for:
- Section structure (What We Planned, What We Got, Surprises, Lessons, Promoted Open Questions)
- Quantitative metrics table format
- File naming convention