workflow-team-code-review

name	workflow-team-code-review
description	Perform final Smelter multi-perspective code review before human check on implementation workflows.
version	0.55
type	workflow
consumes	full task change + e2e artifacts
produces	team_review.md
default_pattern	B
default_agents	["advocate","critic","arbitrator"]
supports_patterns	["B"]
team_template	{"B":{"agents":["advocate","critic","arbitrator"],"consensus_threshold":0.95,"max_rounds":5,"aggregator":"arbitrator"}}
result_types	["pass","fail"]
min_verification_rounds	2
verification_rounds	[{"n":1,"focus":"omission","prompt_template":"templates/verification/round-1-omission.md"},{"n":2,"focus":"contradiction","prompt_template":"templates/verification/round-2-contradiction.md"}]
gate	{"postcondition":[{"file_exists":"team_review.md"},{"consensus_reached":true},{"severity_classified":true}]}

workflow-team-code-review

Overview

The final team review. Re-examines the whole implementation from 3 independent perspectives (advocate / critic / arbitrator). This is distinct from workflow-agent-review:

• workflow-agent-review = diff-scoped, dual adversarial (code + security), runs immediately after coding
• workflow-team-code-review = whole-task scope, 3-role consensus, runs after E2E — last technical gate before human-check

Core principle: The last technical gate before human-check must see the whole task (diff + artifacts + prior reviews) from 3 independent perspectives and reach 95% consensus.

Violating the letter of this rule is violating the spirit of this rule.

Announce at start: "I'm using workflow-team-code-review to run 95% consensus + 2-round review on the complete task."

The Iron Law

NO HUMAN-CHECK HANDOFF WITHOUT 95% CONSENSUS AND 2/2 ROUNDS AT pass

Both must hold: consensus_reached === true AND completed_rounds === 2 && all rounds result === pass.

Pattern B consensus (95%)

advocate:   advocate for implementation merits and correct decisions
critic:     call out defects, missed cases, and latent bugs
arbitrator: synthesize both sides, compute score, render final judgment

Rounds repeat until agreement ≥ 95% (max 5).

Review scope

1. Re-verify that prior workflow-agent-review findings have been resolved
2. Look for new defects
3. Assess security, performance, and maintainability
4. Alignment of implementation vs. plan

Severity classification & routing

Severity	Criterion	Routing
CRITICAL	data loss, security vulnerability, service outage	workflow-tasker (redesign)
HIGH	major bugs, missed edges	workflow-tasker
MEDIUM	minor bugs, improvements	workflow-coding
LOW	style, naming	record in ## Risks, pass

Output

team_review.md:

• ## Advocate Report
• ## Critic Report
• ## Arbitrator Synthesis
• ## Final Verdict (severity)
• ## Consensus Rounds

Red Flags - STOP

Thought	Reality
"workflow-agent-review passed, this is rubber-stamp"	Team review sees the WHOLE task + E2E artifacts. Agent review saw only the diff. New findings here are common.
"95% is close to 90%"	Threshold is 95%. 90% is not pass. Run another round.
"Advocate and critic have stopped disagreeing early — force consensus"	95% computed by arbitrator, not forced.
"All findings are MEDIUM, route to coding and call it done"	MEDIUM routes to coding. But the review itself still needs 2-round pass before the handoff.
"Prior agent-review findings are auto-resolved by now"	Re-verify that each finding has been addressed in the diff. Missing resolutions are HIGH.

Rationalization Prevention

Excuse	Reality
"Nothing will change in round 2"	Round 2 targets contradiction. Absence is what the round reveals.
"Skip round 2 — contradictions are too rare"	Contradictions in review evidence are costly. Round 2 catches them.
"I'm tired"	Iron Law applies.

Consensus not reached

5 rounds exceeded → enter section 11-7 Stall Cascade Level 2. Do NOT force-pass.

Multi-Pass Verification (2-Round Enforcement)

This skill runs 2 mandatory rounds before declaring pass. Each round has a distinct focus:

Round	Focus	Question
1	Omission	Are any diff paths, prior review findings, or plan items unaddressed?
2	Contradiction	Are there conflicts between advocate/critic findings, or with earlier review artifacts?

Agent assignment per round

• Pattern A: Prefer different agent types across rounds. If reusing the same type, reset prompt context per round (fresh perspective).
• Pattern B: Each round runs with 95% consensus (3 × N agents × consensus rounds).
• Pattern D: Lead orchestrates 2 rounds, assigning different viewpoint sub-agents.

State recording

All rounds recorded in state.json.team_runtime.workflow-team-code-review.rounds[]. Skill-level pass is declared only when completed_rounds === 2 && all rounds result === pass.

Failure handling

• Any round fail → skill-level fail with cause: verification_failed, evidence: {round, focus, findings[]}
• Producer-chain routes to the skill's consumed-artifact producer based on severity (workflow-tasker or workflow-coding)
• On re-entry after upstream fix, both rounds re-run (not just failed ones)

Anti-evasion enforcement

1. Do not inject prior round conclusions into current round prompts (bias prevention)
2. Critic Watchdog blocks "already verified" skip statements
3. Declaring pass with completed_rounds < 2 is blocked by hook

Terminal State — Required Next Skill

REQUIRED NEXT SKILL on pass: workflow-human-check

On fail: route per severity:

• CRITICAL / HIGH → workflow-tasker
• MEDIUM → workflow-coding
• LOW → record in ## Risks, proceed to human-check

On consensus exceeded (5 rounds): Stall Cascade Level 2.

Do NOT:

• Declare the task complete here — only workflow-human-check does that
• Stop after pass, report "all reviews passed", or ask "shall I finalize?"
• Offer A/B/continue choices — workflow-human-check is the next skill and it's the ONLY allowed halting point

Evidence Integrity (Mechanical Enforcement)

Every fail verdict MUST cite at least one anchor in the strict form:

Evidence: path/to/file.ext:LINE[-LINE] "verbatim quote substring"

Rules:

• The path must exist on disk at this session's cwd.
• The quoted substring must appear on the cited line (or within the range).
• Quote is a substring match after whitespace normalization — not a regex.
• A PostToolUse hook (scripts/review-evidence-verifier.mjs) blocks writes that violate these rules.
• Do not paraphrase — quote the line verbatim.
• If you cannot produce a verified anchor, the symptom is not grounded; do not emit a fail verdict.