Team Coordination

Contract

Independent analysis before shared conclusions: teammates who see each other's findings are contaminated reviewers. Applies only when agentTeams: true and CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS is set; commands/review.md Path A applies it, using the tables in references/paired-review-protocol.md verbatim. Returns consolidated P1/P2/P3 findings, each carrying a confidence and a disposition from the controlled set consensus|validated|refined|kept|unchallenged. The only permitted skip is the documented per-facet fallback to single-reviewer Path B; a paired-protocol failure never blocks a review.

When to Spawn Teams

Task Sizing

5-6 tasks per teammate, each completable without coordination and carrying its own acceptance criteria. Overlapping file sets go to the same teammate.

Adversarial Review Protocol

Each facet runs as a paired reviewer: a skeptic ("assume the diff is broken until proven otherwise") and a verifier ("assume correct; look for missed edge cases and unenforced invariants"), dispatched in parallel with no view of each other. The default 6-facet topology is 12 invocations: 5 Agent(...) pairs plus 2 Skill(holdout-validation) lens calls.

1. Independent analysis. Both variants return P1/P2/P3 findings with file:line citations. No challenge information in this phase.
2. Share. The lead indexes findings by facet. No LLM call. Holdout findings are indexed for consensus matching only.
3. Challenge (disposition-only). Each variant labels the other's findings AGREE, DISAGREE: <reason>, or REFINE: priority= category= WITHOUT re-reading the diff. Holdout-validation is excluded by design: it is objective claim-verification, so DISAGREE has no meaning there.
4. Synthesize. The lead applies the consolidation table mechanically: both raised independently (same facet, file, lines ±2, priority ±1) → HIGH consensus; AGREE → HIGH validated; REFINE → MEDIUM refined; DISAGREE → LOW kept; timeout/error → MEDIUM unchallenged; both DISAGREE → dropped and journaled under ## Dropped after challenge. A System One answer is not a challenger answer and never counts toward a drop. The lead never adjudicates.
5. Emit. Finding tables carry a trailing _(confidence · disposition)_ suffix; the FLOW_REVIEW_CYCLE marker carries the 7-field rows per references/finding-ledger-parser.md.

Cost: about 23 LLM calls per review (12 + 10 challenge + 1 consolidation), roughly 3.8× the single-session baseline; opt-in.

Bias rules: no reading of another reviewer's findings before producing your own (anchoring); HIGH-confidence-on-everything is a smell (groupthink); no third-agent judge (central-judge bias).

Fallback Semantics

Per facet, never blocking the review: agentTeams: false → Path B; env var unset → stderr WARN and Path B; one variant fails to spawn or times out (timeouts.teammateTimeout, default 300s) → use the responding variant, mark its findings unchallenged; both fail → re-dispatch that facet on Path B; challenge round fails → include findings as unchallenged. The full matrix is in references/paired-review-protocol.md.

Teammate Health

Timeout: collect partial results and mark the facet "incomplete review", else "not reviewed". Crash: log, fall back to single-session for that facet, note it in the summary. Idle: nudge-idle-teammate.sh sends "check task list for unclaimed tasks" after 60s (stderr, exit 0).

Implementation Team Protocol

The lead creates the dependency-ordered task list, assigns non-overlapping file groups, monitors via TaskList, and runs final verification after all teammates complete. Teammates claim, implement, commit, and mark tasks complete.

Single-Session Fallback

When teams are disabled or spawning fails: sequential Agent dispatch (parallel where independent), main-thread coordination, no adversarial protocol. The calling command owns the fallback; this skill supplies team-specific knowledge only.