Managing review campaigns

Managing Review Campaigns

Overview

A reviewer invocation may run deeply, but the campaign around it must end. Rounds and total external reviewer invocations are hard-capped; the lead closes the campaign from recorded code and test evidence.

REQUIRED SUB-SKILL: Use razorback:receiving-code-review to verify findings before accepting them. Apply the blocker semantics from razorback:using-razorback.

Start Once

Before the first review invocation, emit this immutable setup:

REVIEW CAMPAIGN
scope: <problem class and change range>
workflow: ordinary | pre-merge | convergence
participants: <lead and selected reviewers>
required_reviewers: <names or none>
evidence_target: lead-only | fresh-session | external-reviewed | cross-model-reviewed
severity_floor: <default medium>
discovery_scopes: <named scopes>
external_invocation_budget: <integer>
max_rounds: <1-3>
round: 0/<max>
external_invocations: 0/<budget>

Setup fields are immutable; round and external_invocations counters only increase. Participants and budgets cannot grow mid-campaign. Each external CLI call counts as one external invocation, including separate general and security passes. There are no per-invocation turn/spend caps; that does not relax the campaign budget. Persist the setup and counters through compaction and continuation.

Availability and Evidence

The lead always participates. Select others only when available, policy-allowed, and budgeted.

Available evidence	Label	Rule
Lead only	lead-only	Ordinary review may complete; explicit external or cross-model requirements cannot.
Fresh isolated same-model session	fresh-session	Adds context independence; never reported as cross-model.
Lead plus external reviewer	external-reviewed	Adds independent evidence within the fixed campaign.
Lead plus a real different-model reviewer	cross-model-reviewed	Satisfies an explicit cross-model requirement.

A reviewer is explicit when named by the user, required by repo instructions, or recorded in the approved plan. An explicit reviewer that is unavailable before completing its required discovery scope blocks. A required reviewer obligation is satisfied once usable evidence covers every declared required discovery scope. Failure of an optional confirmation after that point does not retroactively block; count the call when dispatched, do not retry or replace it, and let the lead confirm. An unavailable optional participant is omitted and recorded before setup. An optional participant lost mid-campaign is not retried or replaced; use the strongest evidence label already earned and continue within the original budget.

Bounded Rounds

Round 1 — discovery

• Review the declared scope broadly; each selected reviewer runs at most once per discovery scope.
• Verify and deduplicate findings, then the lead assigns critical, high, medium, or low.
• Freeze the accepted finding set after triage.

Round 2 — scoped confirmation

• Mark every accepted finding addressed, not addressed, contested, or deferred.
• Inspect only the fix diff for new breakage. Observations outside the fix diff are recorded and cannot extend the campaign. A new medium/low observation may be fixed or deferred but cannot authorize another external sweep.
• A deferred finding at or above the severity floor remains open. A finding stops counting as open only when addressed, dismissed from evidence, or recorded outside the approved campaign scope.
• The lead confirms by default. At most one predeclared external confirmer may make one targeted invocation.

Round 3 — exceptional targeted confirmation

Enter only when the lead verifies a new critical/high regression introduced inside the fix diff. Target only that regression and its fix; broad discovery is forbidden. Use at most one confirmer, or let the lead confirm when none is available. Reviewer disagreement is not a trigger. Push back once on the evidence, record any dispute, and stop after Round 3 regardless of outcome.

Extra reviewers never add rounds. They add Round 1 evidence and consume the predeclared invocation budget.

Workflow Profiles

Workflow	Discovery	Confirmation	Default external budget	Max rounds
Ordinary lead-only	Lead sweep	Lead	0	2
Standalone external	One selected reviewer	Lead	1	2
Pre-merge	General plus security once	Lead	2	2
Cross-model convergence	Selected reviewers once	At most one targeted external confirmer	selected reviewers + 1	3
Exceptional regression	None	One targeted confirmer	+1 only when predeclared	1

Callers may choose a stricter profile, never a looser one.

Close on Evidence

Canonical severity comes from skills/codex-cli/schemas/review-output.schema.json: critical, high, medium, or low. For every accepted finding, record its classification and canonical severity; file:line or symbol evidence; red-to-green test, existing covering test, or inspection-only; and the fix, dismissal, dispute, or deferral reason. Green tests on an uncovered path are inspection-only. Majority vote may raise confidence but cannot override code evidence, scope, or the campaign budget.

Terminal states are clean when nothing above the floor remains open in scope, capped when the maximum round ends or no permitted action remains while an above-floor finding is open, and blocked when a required discovery obligation cannot be satisfied or an unresolved critical/high finding meets the blocker taxonomy. An exhausted external budget stops external dispatch but does not prevent an already-permitted lead-only confirmation. Emit this block for every terminal state:

REVIEW CAMPAIGN STATUS
state: clean | capped | blocked
evidence: lead-only | fresh-session | external-reviewed | cross-model-reviewed
round: 2/2
external_invocations: 3/3
open_critical_high: 0
open_medium_low: 2
open_above_floor: 1
campaign_closed: yes

clean, capped, and blocked are all terminal. Goal runners must treat campaign_closed: yes as terminal, even with remaining findings.

Observed Rationalizations

Excuse	Reality
“One more pair or review will probably be clean.”	The budget is a hard stop; close as capped.
“Inflate severity so this qualifies for Round 3.”	The lead verifies severity from code evidence; reviewer labels cannot buy a round.
“The majority vote says it is fixed.”	Model agreement is evidence, not closure authority.
“Reset the counters after compaction or continuation.”	Setup and counters are immutable campaign state.

Red Flags

• Dispatching beyond the declared external budget
• Reopening broad discovery in Round 2 or Round 3
• Adding rounds because more reviewers are available
• Letting a reviewer assign the campaign's canonical severity
• Resetting setup or counters after context loss
• Continuing after campaign_closed: yes; it is terminal

Any red flag means stop dispatching and emit the terminal status.