Finding Disposition

Reviewers use plan-review and adversarial-review; they receive nothing from this file. Every dispatch runs through subagent-dispatch.

Every finding gets one written disposition, and the record is committed. Silence is not a disposition; "unlikely" is not a reason. Rejecting via an unwritten invariant requires writing that invariant into the entity page or gotchas.md as part of the rejection.

Read first

• docs/wiki/gotchas.md and the entity ## Behavior cases the round covers.

Brief round — before Red (plan-adversary)

1. Subject (exactly one): .handoff/<slug>-plan.md ([complex]/batched cycle) or the verbatim todo line plus $ARGUMENTS instruction (simple cycle).
2. Send only: subject, entity slug(s), case IDs, test command, and branch name — never your opinion or defense.
3. Verify report: If no findings and no Checked: line, re-dispatch once demanding it.

4. Dispose (Applied is default):

   Disposition	When	Action
   Applied	Valid, and fixable via plan edit, sharper brief, or added case ID	On a planned cycle, edit .handoff/<slug>-plan.md. On a simple cycle, pass the refinement in the developer's Inputs: brief (and if you update docs/wiki/todos.md, stage it in the docs(<slug>): plan review commit of step 6 below, so the tree is clean before the developer runs).
   Escalated	The spec is wrong, ambiguous, or missing	human-checkpoint recommending /project:interview. Stop the cycle.
   Rejected	Misreads the brief, or outside this cycle's cases	One sentence. Out-of-scope work worth doing becomes an ordinary todo line.

   A contested blocker is a human-checkpoint, never a unilateral rejection. Optionally dispatch triage when a finding hinges on code you have not read.

5. Re-plan only for a structural blocker (wrong decomposition or impossible order), once, with findings attached. A second disagreement → human-checkpoint.

6. Commit the record in the cycle's work log entry (docs/wiki/log.md, plus docs/wiki/todos.md if edited) as docs(<slug>): plan review before dispatching developer (/project:work step 4a):

   - Plan review: 3 findings — 2 applied, 1 rejected
     - P1 blocker spec-fidelity — Applied: B3 had no step; plan step 4 now covers it.
     - P3 note scope — Rejected: retry policy belongs to entity `billing`, not this cycle.

Diff round — after the cases land (adversary)

1. Scope: the whole cycle, once. <sha-before-cycle>..HEAD, where <sha-before-cycle> is the commit just before the cycle's first case commit (the docs(<slug>): plan review commit when step 4a ran). Never one review per case. Check git diff --stat <range> -- . ':(exclude)docs/wiki/log.md'; empty → skip and say so.
2. Send only: commit range, entity slug(s), case IDs, test command — plus previous round findings on re-review (step 9). Never send .handoff/, conductor reasoning, or intent prose.
3. Check independence: The adversary diffs with docs/wiki/log.md excluded; flag any report citing .handoff/ or log.md. HEAD and tracked files must be unchanged.
4. Triage (optional): Dispatch triage with the range, slug, case IDs, and verbatim findings. Never downgrade severity solely because a failure was not reproduced live.
5. Dispose (Filed is default):
   • Filed: Add a todo line (docs/wiki/todos.md § Tags, tagged [adversary]): critical → ## Now (P0 — next), major → ## Next (P1), minor → ## Later (P2). nit is never filed.
   • Fixed: Only a human-approved critical/major. Dispatch developer for a failing test + fix commit: fix(<slug>): <what> — adversary F<n> (include Wiki-Update: none (adversary fix) footer if the entity page needs no change).
   • Rejected: One sentence citing the invariant, misread, or scope boundary.
6. Human gate: Group all critical and major findings into one human-checkpoint (claim, failure scenario, files touched, recommendation). If declined or unreachable in fast mode → file at P0/P1 and highlight in the cycle report. A human instruction such as "fix all the findings" is the approval, at that scope.

7. Commit the round record (stage docs/wiki/todos.md plus any gotcha/ADR from triage; do not stage log.md in a round a re-review follows — /project:work records the tally in step 8's work entry, and /project:adversary stages its adversary entry in the final round's commit):

   git add docs/wiki/todos.md   # plus any gotcha or ADR the triage produced
   git commit --allow-empty -m "$(cat <<'MSG'
   docs(<slug>): adversary round 1 — 3 findings, 1 fixed, 1 filed, 1 rejected, 2 nits
   
   F1 critical correctness — Fixed in a1b2c3d (human approved): empty token rejected before lookup.
   F2 major durability — Filed P1 (human declined immediate fix): unbounded retry on 5xx.
   F3 minor test-integrity — Rejected: covered by tests/integration/test_auth.py.
   Nits: 2 — not filed.
   MSG
   )"

   --allow-empty handles rounds where every finding was rejected or Fixed in its own commit. git log --grep="adversary round" is the audit trail.

8. P0 saturation: Count open P0 items (docs/wiki/todos.md § P0 saturation threshold). At or above P0_MAX → human-checkpoint with the count, [adversary] share, three oldest entries, and a recommendation.
9. Re-review only if something was Fixed, over the fix commits only, attaching prior findings and their disposition lines. Stop after 3 rounds maximum: at round 3, gate/file any open critical/major items and list open minor/nit items in the round commit as unfiled (reason: round cap).

Record the yield — every round

Record exact counts per severity and disposition in the round commit and the cycle's log entry (if run via dispatch-handoff, name the external harness and model so its counts are not attributed to the role's pinned model). When open [adversary] todos reach FINDINGS_MAX (docs/wiki/todos.md § Filed-findings backlog), note it in the cycle report.

Anti-patterns

• Leaking author context into either reviewer's brief.
• Fixing minor findings inline or fixing a critical/major without human approval.
• Filing brief findings for later when you are about to implement their subject.
• Re-reviewing the whole diff or opening a fourth round.