ever-works/orgs/code-review

Code Review

A pre-merge structural audit run in paranoid mode. The operating assumption: passing tests mean the tests pass, nothing more. The job is to find the bugs that survive CI and detonate in production.

Steps

1. Diff against the mainline, whole and cold. Read the full diff before

judging any hunk — half the real bugs live in the interaction between two hunks that are each fine alone.

1. Pass one — intent. What does the branch claim to do? Check the diff

against the plan or task it implements. Flag scope drift both ways: changes the plan never asked for, and promised behavior that is missing.

1. Pass two — the adversarial checklist. Hunt specifically for:

• N+1 queries and missing indexes on new query paths
• Race conditions, stale reads, and check-then-act windows
• Broken invariants — state that two places update and one forgets
• Trust-boundary violations: unvalidated external input, IDs accepted

without ownership checks, model or user output executed as instructions

• Injection and escaping bugs in queries, shells, templates, and paths
• Retry logic that duplicates side effects, and conditional side effects

hiding in helpers

• Error paths that swallow, mislabel, or double-report failures
• Tests that assert the happy path while the named failure mode goes

untested

1. Interrogate the tests. For each behavior the plan promises, find the

test that would fail if it broke. No such test is a finding even when the code is correct today.

1. Write findings, not fixes. Each finding: location, the failure

scenario as a concrete "given X, then Y breaks", severity, and the required change. Style preferences are not findings; keep them out or mark them explicitly as optional.

1. Verdict. Approve, or return a numbered findings list. Approval with

unresolved critical findings does not exist.

Quality bar

• Every finding has a concrete failure scenario, not an aesthetic objection.
• The review says what was checked and found clean, so silence is

information rather than ambiguity.

• Re-review after fixes covers the new diff, not just the patched lines.

Output format

A review verdict: Scope Check, numbered Findings (location, scenario, severity, required change), Checked-Clean list, and APPROVED or CHANGES REQUIRED as the final line.