| name | adversarial-review |
| description | The pre-merge review playbook — reproduce, don't adopt. Use for any branch review; defines the minimum attack set, the verdict format, and the honesty requirements for what was and wasn't reproduced. |
Adversarial review playbook
Your job is to try to make the change fail, not to confirm it works. The author's report is a list of claims; reproduce every claim you rely on.
Minimum attack set
- Merge reality. Fetch, then
merge-tree against the LIVE tip (it moves during long reviews — re-check before verdict). Check file overlap with recent history for semantic conflicts, not just textual ones.
- Fresh-binary gates. Delete the test executables, rebuild, confirm exit code AND new mtimes, then run the suites yourself. Beware false greens: a test runner whose filter matches nothing can exit 0 — verify the suite ran by its test counts.
- Red-green. Reproduce the red state for new tests (revert the change, keep the tests, observe the failure). A test that was never red is a claim, not a proof. Check for tautologies — a control assertion that stays green in the red build.
- Hostile inputs. Anything parsing external or untrusted bytes gets crafted-input attacks — overflowed length fields, truncation at every structural boundary, declared-size bombs, wrong-type fields — run under a watchdog. A hang is a finding, not a timeout.
- Threading and lifetime. For concurrency changes: who owns destruction; what happens on the cancel, shutdown, and error paths; can a slot or lock leak or double-release; does any worker ever wait on work scheduled to its own pool; does anything publish state outside the lock its reader holds.
- Stale-context sweep. Grep for comments and docs the diff falsifies. A correct fix sitting next to a contradicting comment is a future bug — require the comment fixed in the same change.
- Excuse verification. When the author skipped something ("no test target exists", "not reproducible"), verify the excuse against the build system or the repro before accepting it.
- Deletions review. Every removed line justified; no dead code left behind; no defense removed without an equivalent replacement.
- Evidence audit. Claims must be tiered verified/inferred/assumed; measurements must exist as artifacts — spot-check that cited artifacts exist and say what the report says they say.
Verdict
MERGE-READY or NOT-READY, with numbered findings tagged [BLOCKER] / [SHOULD-FIX] / [NIT], each with file:line and a minimal fix. State explicitly: what you reproduced vs reviewed statically, and what you deliberately skipped and why. If your verdict depends on something you did not reproduce, say so in the verdict line itself.
