Human Review Handoff

Build the final reviewer packet that a human reviewer actually reads.

When to use and inputs

After finding-synthesizer has produced the ranked finding set — always for medium/high-risk PRs, and for low-risk PRs when findings exist. Consumes synthesized findings from finding-synthesizer and the evidence pack from pr-evidence-builder.

Steps

1. Build the reviewer packet using the packet structure below.

Packet structure

The packet contains these sections (see the example output for exact formatting):

• TL;DR — 1–2 sentence summary: what changed, risk lane, top concern if any
• Risk — LOW / MEDIUM / HIGH with contributing factors and AI-assisted flag
• Verification Status — table: verifier name | status | notable findings
• Findings — ordered by severity × confidence; each entry includes title, file:line, why it matters, evidence type/source, and suggested action (fix | verify | discuss)
• Unresolved Assumptions — things the review couldn't determine; questions for the human
• Recommended Review Focus — specific files/hunks where human attention is most needed and why
• Metadata — reviewer mode, reviewer model family, authoring model family (if detected), wall-clock time, context isolation

Example output (abbreviated)

## PR Review Packet

### TL;DR
Adds retry logic with exponential backoff to `api/client.py`; risk is MEDIUM due to
unbounded retry count and missing circuit breaker. One unresolved assumption about
upstream timeout configuration requires human input.

### Risk: MEDIUM
Contributing factors:
- Modifies external service call path (high-impact surface)
- No new tests for the retry-exhaustion path
AI-assisted: yes (GitHub Copilot detected in commit metadata)

### Verification Status
| Verifier          | Status  | Notable Findings                        |
|-------------------|---------|-----------------------------------------|
| Static analysis   | PASS    | No new lint errors                      |
| Test coverage     | WARN    | retry-exhaustion branch uncovered (0%)  |
| Dependency audit  | PASS    | No new vulnerable packages              |

### Findings (2 items)
1. **Uncovered retry-exhaustion branch** — `api/client.py:114`
   Why it matters: If all retries fail, the caller receives None instead of
   an exception, silently dropping the request.
   Evidence: Coverage report (pr-evidence-builder)
   Suggested action: fix

2. **Missing circuit breaker on retry loop** — `api/client.py:87`
   Why it matters: Under sustained upstream failure, every request retries 5 times
   with backoff, multiplying latency and connection pool pressure.
   Evidence: Code inspection (finding-synthesizer)
   Suggested action: discuss

### Unresolved Assumptions
- What is the upstream service's timeout and rate limit for this endpoint?

### Recommended Review Focus
- `api/client.py` lines 87–120: retry loop and the uncovered exhaustion branch
  are the highest-risk hunks; verify behaviour under sustained upstream failure
  and concurrent request load.

### Metadata
- reviewer mode: fresh_eyes
- reviewer model family: claude-3
- authoring model family: unknown
- wall-clock time: 47 s
- context isolation: yes

---
*This is an evidence-based review aid, not an approval. The tile produces
findings and questions. Humans produce decisions.*

2. Add the review-boundaries disclaimer. Every packet must include: "This is an evidence-based review aid, not an approval. The tile produces findings and questions. Humans produce decisions."

3. Mark escalation areas. If any human-escalation triggers are present, surface specific questions the human reviewer should answer.

4. Validate the packet before delivering it. Run through this checklist:

   • TL;DR is present and ≤ 2 sentences
   • Risk rating is one of LOW / MEDIUM / HIGH with contributing factors listed
   • Verification Status table is populated (or explicitly marked "no verifiers run")
   • Findings are ordered by severity × confidence and each has a suggested action
   • Unresolved Assumptions section is present (may be empty — explicitly state "None")
   • Recommended Review Focus names specific files or hunks
   • Review-boundaries disclaimer is included verbatim
   • Metadata block is complete

   If any item is missing, regenerate or patch that section before handing off.

Related skills

• finding-synthesizer — produces the ranked finding set consumed by this skill
• pr-evidence-builder — produces the evidence pack consumed by this skill