panda850819 Code Review
The delta beyond an unstructured single-pass review is scope provenance, risk-adaptive lanes, evidence gates, bounded correction, and cold escalation.
1. Bind scope
- Read
AGENTS.mdorCLAUDE.mdwhen present. - Resolve the base from the PR, upstream, or merge-base; print the base and changed-file list. Never silently review the whole repository.
- Read the issue, brief, or request that defines intent. If none exists, report
INTENT GAPand infer only from the diff. - Check branch state and uncommitted changes; use history only to explain scope.
2. Choose the risk lane
- low: local, reversible, with no trust boundary, persistent data, concurrency, public interface, generated contract, or production operation.
- medium: shared behavior, multiple files, compatibility, or concurrency.
- high: auth, secrets, money, permissions, migrations, destructive writes, production infrastructure, or unfamiliar costly behavior.
Choose the lane from the diff and intent. Promote when the first pass reveals a risk trigger, candidate finding, coverage gap, or scope drift.
Low-risk fast path
Only in the low lane with explicit scope and intent: trace the changed path,
match acceptance and branches to tests, and self-refute the likeliest defect. An
INTENT GAP promotes rather than entering this fast path. If no finding, gap, or
drift remains: Do not load review learnings; enumerate no lenses,
print no empty scope-drift or cold-review fields, and return:
Review scope: <base>..<head> | <n> files | risk: low
No actionable findings.
Coverage: <verified check>
Self-refute: <assumption and observed result>
3. Escalated review
Read lib/learning-recall.md and apply relevant repo learnings. Map the diff to
security, data integrity, concurrency, architecture, or operations lenses.
Medium uses only triggered lenses. High uses every relevant lens plus a
cold review.
Trace changed inputs through code, callers, contracts, tests, and failure
handling. A finding survives only with severity P0–P3, a tight file/line
range, a trigger, mechanism and user-visible consequence, and a correction
direction. Exclude style,
unreachable speculation, and pre-existing defects; review does not edit code.
Use an isolated read-only reviewer when the lane is high, the diff exceeds
roughly 5K tokens, or a load-bearing conclusion remains disputed. Give it only
the bound diff, intent, and applicable repository contract, without the author's
analysis or proposed verdict. Any fresh context that satisfies those semantics is
valid; the transport and model family do not define the review. If no isolated
context is available, record Cold review: unavailable — no isolation capability.
A second pass in this same context is not a cold review. Merge findings by
mechanism; disagreement becomes NEEDS TRACE, not a vote.
Match acceptance and branches to tests, run the narrowest available checks, and
self-refute the highest-risk assumption. Report COVERAGE GAP only for unproved
concrete behavior and SCOPE DRIFT only for changes outside intent. Conclude with
findings, No actionable findings., or BLOCKED.
Output and completion
Review scope: <base>..<head> | <n> files | risk: <lane>
Findings
- [P1] <title> — <file:line>
Trigger: <input/state>
Mechanism: <why it fails and impact>
Direction: <correction>
Coverage: <verified checks or concrete gaps>
Scope drift: <none or entries>
Cold review: <not earned | completed | unavailable>
Self-refute: <assumption and result>
Done when findings are evidence-backed and the report names applicable coverage, scope drift, cold review, and self-refutation.
Anti-patterns
- Fixed three-pass fan-out for a small reversible diff.
- Skipping scope provenance or risk promotion because the diff looks small.
- Treating a repository-wide scan as evidence about the changed path.
- Auto-fixing findings and reviewing the rewrite as independent proof.
- Reporting hypothetical security language without an attacker-controlled path.
- Calling an unavailable cold review clean; record the gap.