workflow-agent-review

name	workflow-agent-review
description	Perform focused Smelter code and security review after implementation before E2E validation.
version	0.55
type	workflow
consumes	src/** diff
produces	agent-review.md, "## Risks" updated
default_pattern	B
default_agents	["code-reviewer","security-reviewer"]
supports_patterns	["A","B"]
team_template	{"B":{"agents":["code-reviewer","security-reviewer"],"mode":"dual_adversarial","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":"agent-review.md"},{"contains_verdict":"pass|fail"}]}

workflow-agent-review

Overview

Immediately after implementation, a Pattern B Dual Adversarial review:

• code-reviewer: quality, readability, bugs, logic errors
• security-reviewer: vulnerabilities, data exposure, permissions

The two agents review in parallel and independently; arbitrator then merges results.

Core principle: Implementation is unverified until two independent reviewers (code + security) both pass through 2 rounds (omission, contradiction).

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

Announce at start: "I'm using workflow-agent-review to run Pattern B dual adversarial review + 2-round verification on the diff."

The Iron Law

NO E2E HANDOFF WITHOUT DUAL REVIEW AT pass AND 2/2 ROUNDS AT pass

Two independent enforcements must both hold:

1. Both code-reviewer and security-reviewer return pass (Pattern B consensus)
2. Both verification rounds return pass

Declaring pass with fewer than two recorded rounds or only one reviewer is invalid. Record evidence in agent-review.md.

Demotion condition (Pattern B → A)

Only when tasker declares security_surface: false does code-reviewer run alone (Pattern A). Default remains B. The demotion is a tasker-time decision; workflow-agent-review cannot demote itself.

Output

agent-review.md:

• ## Code Review — code-reviewer findings
• ## Security Review — security-reviewer findings
• ## Risks Added — updates to the ## Risks section (LOW/MEDIUM/HIGH/CRITICAL)
• ## Verdict — pass / fail

Risks section

## Risks
- [LOW] <style / naming improvement suggestion>
- [MEDIUM] <possible minor bug>
- [HIGH] <major logic flaw>
- [CRITICAL] <data loss / security vulnerability>

Red Flags - STOP

Thought	Reality
"The diff is small, skip security-reviewer"	Pattern B demotion is a tasker decision, not a reviewer decision. Run both unless security_surface: false was declared.
"Code-reviewer passed, ship it"	Pattern B needs BOTH reviewers. One is not consensus.
"Only round 1 is needed for small diffs"	Run both rounds; downstream E2E relies on the artifact.
"Reviewer disagrees with me, I'm right"	Push back with evidence, not confidence. If you cannot verify, say so and ask.
"Findings are just style, ignore them"	LOW findings go in ## Risks. Ignoring is not the same as acknowledging.
"We already reviewed this in the previous session"	Session-local. Re-verify. Critic Watchdog blocks "already verified" skips.

Rationalization Prevention

Excuse	Reality
"The reviewer doesn't understand the context"	Provide missing context, don't dismiss findings.
"Great point!" / "You're absolutely right!"	Forbidden response. Technical acknowledgment or reasoned pushback only.
"Let me implement that now" (before verification)	Verify the reviewer's finding is technically correct for THIS codebase first. Then implement.
"I'll bundle feedback for later"	Fix Critical immediately. Fix Important before proceeding to workflow-e2e. Minor goes in ## Risks.

Forbidden Responses

When receiving review feedback, NEVER say:

• "You're absolutely right!" — performative; forbidden per feedback_precision memory
• "Great point!" / "Excellent feedback!" — performative
• "Let me implement that now" — before technical verification

INSTEAD:

• Restate the technical requirement
• Ask clarifying questions if unclear
• Push back with technical reasoning if incorrect
• Just start working on verified fixes (actions > words)

Fail routing

• fail → workflow-coding
• On security cause detection → workflow-coding with security guidance attached via active_feedback

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, error branches, or security surfaces unreviewed?
2	Contradiction	Are there inconsistencies between code review and security review findings, or with the plan?

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.

Artifact recording

All rounds are recorded in agent-review.md. Skill-level pass is declared only when both rounds are present and both results are pass.

Failure handling

• Any round fail → skill-level fail with cause: verification_failed, evidence: {round, focus, findings[]}
• Producer-chain routes to workflow-coding (the consumed artifact's producer)
• 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. Do not use "already verified" as a reason to skip a fresh round
3. Do not declare pass without two recorded rounds

Terminal State — Required Next Skill

REQUIRED NEXT SKILL on pass: workflow-e2e

On fail: route to workflow-coding with active_feedback.

Do NOT:

• Skip to workflow-team-code-review — workflow-e2e must run first on interface-changing surfaces
• Stop after pass, report "review complete", or ask "shall I run E2E?"
• Offer A/B/continue choices — Iron Law #1 forbids pausing at non-human-check stages

If the surface is in exempt_if_surface: [style, typography, typo, dialogue], the mode's configuration may route directly to workflow-team-code-review. Otherwise workflow-e2e is mandatory.

Evidence Integrity

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.
• Record only anchors you verified against files in the current checkout.
• 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.