Auto Mode
When --yes or -y : Auto-confirm task decomposition, skip interactive validation, use defaults.
Team Review
Usage
$team-review "src/auth/**/*.ts" $team-review -c 2 --full "src/components" $team-review -y --dimensions=sec,cor "src/api" $team-review --continue "RV-auth-review-2026-03-08" $team-review -q "src/utils" $team-review --fix "src/auth/login.ts"
Flags:
-
-y, --yes : Skip all confirmations (auto mode)
-
-c, --concurrency N : Max concurrent agents within each wave (default: 3)
-
--continue : Resume existing session
-
--full : Enable scan + review + fix pipeline
-
--fix : Fix-only mode (skip scan/review)
-
-q, --quick : Quick scan only
-
--dimensions=sec,cor,prf,mnt : Custom dimensions (security, correctness, performance, maintainability)
Output Directory: .workflow/.csv-wave/{session-id}/
Core Output: tasks.csv (master state) + results.csv (final) + discoveries.ndjson (shared exploration) + context.md (human-readable report)
Overview
Orchestrate multi-agent code review with three specialized roles: scanner (toolchain + LLM semantic scan), reviewer (deep analysis with root cause enrichment), and fixer (automated fixes with rollback-on-failure). Supports 4-dimension analysis: security (SEC), correctness (COR), performance (PRF), maintainability (MNT).
Execution Model: Hybrid — CSV wave pipeline (primary) + individual agent spawn (secondary)
┌─────────────────────────────────────────────────────────────────────────┐ │ Team Review WORKFLOW │ ├─────────────────────────────────────────────────────────────────────────┤ │ │ │ Phase 0: Pre-Wave Interactive │ │ ├─ Parse arguments and detect pipeline mode │ │ ├─ Validate target path and resolve file patterns │ │ └─ Output: refined requirements for decomposition │ │ │ │ Phase 1: Requirement → CSV + Classification │ │ ├─ Generate task breakdown based on pipeline mode │ │ ├─ Create scan/review/fix tasks with dependencies │ │ ├─ Classify tasks: csv-wave (scanner, reviewer) | interactive (fixer)│ │ ├─ Compute dependency waves (topological sort → depth grouping) │ │ ├─ Generate tasks.csv with wave + exec_mode columns │ │ └─ User validates task breakdown (skip if -y) │ │ │ │ Phase 2: Wave Execution Engine (Extended) │ │ ├─ For each wave (1..N): │ │ │ ├─ Execute pre-wave interactive tasks (if any) │ │ │ ├─ Build wave CSV (filter csv-wave tasks for this wave) │ │ │ ├─ Inject previous findings into prev_context column │ │ │ ├─ spawn_agents_on_csv(wave CSV) │ │ │ ├─ Execute post-wave interactive tasks (if any) │ │ │ ├─ Merge all results into master tasks.csv │ │ │ └─ Check: any failed? → skip dependents │ │ └─ discoveries.ndjson shared across all modes (append-only) │ │ │ │ Phase 3: Post-Wave Interactive │ │ ├─ Generate final review report and fix summary │ │ └─ Final aggregation / report │ │ │ │ Phase 4: Results Aggregation │ │ ├─ Export final results.csv │ │ ├─ Generate context.md with all findings │ │ ├─ Display summary: completed/failed/skipped per wave │ │ └─ Offer: view results | retry failed | done │ │ │ └─────────────────────────────────────────────────────────────────────────┘
Task Classification Rules
Each task is classified by exec_mode :
exec_mode Mechanism Criteria
csv-wave
spawn_agents_on_csv
One-shot, structured I/O, no multi-round interaction
interactive
spawn_agent /wait /send_input /close_agent
Multi-round, clarification, inline utility
Classification Decision:
Task Property Classification
Scanner task (toolchain + LLM scan) csv-wave
Reviewer task (deep analysis) csv-wave
Fixer task (code modification with rollback) interactive
CSV Schema
tasks.csv (Master State)
id,title,description,deps,context_from,exec_mode,dimension,target,wave,status,findings,error 1,Scan codebase,Run toolchain + LLM scan on target files,,,"csv-wave","sec,cor,prf,mnt","src/**/*.ts",1,pending,"","" 2,Review findings,Deep analysis with root cause enrichment,1,1,"csv-wave","sec,cor,prf,mnt","scan-results.json",2,pending,"","" 3,Fix issues,Apply fixes with rollback-on-failure,2,2,"interactive","","review-report.json",3,pending,"",""
Columns:
Column Phase Description
id
Input Unique task identifier (string)
title
Input Short task title
description
Input Detailed task description
deps
Input Semicolon-separated dependency task IDs
context_from
Input Semicolon-separated task IDs whose findings this task needs
exec_mode
Input csv-wave or interactive
dimension
Input Review dimensions (sec,cor,prf,mnt)
target
Input Target path or pattern
wave
Computed Wave number (computed by topological sort, 1-based)
status
Output pending → completed / failed / skipped
findings
Output Key discoveries or implementation notes (max 500 chars)
error
Output Error message if failed (empty if success)
Per-Wave CSV (Temporary)
Each wave generates a temporary wave-{N}.csv with extra prev_context column (csv-wave tasks only).
Agent Registry (Interactive Agents)
Agent Role File Pattern Responsibility Position
fixer agents/fixer.md 2.3 Apply fixes with rollback-on-failure post-wave
COMPACT PROTECTION: Agent files are execution documents. When context compression occurs, you MUST immediately Read the corresponding agent.md to reload.
Output Artifacts
File Purpose Lifecycle
tasks.csv
Master state — all tasks with status/findings Updated after each wave
wave-{N}.csv
Per-wave input (temporary, csv-wave tasks only) Created before wave, deleted after
results.csv
Final export of all task results Created in Phase 4
discoveries.ndjson
Shared exploration board (all agents, both modes) Append-only, carries across waves
context.md
Human-readable execution report Created in Phase 4
interactive/fixer-result.json
Results from fixer task Created per interactive task
agents/registry.json
Active interactive agent tracking Updated on spawn/close
Session Structure
.workflow/.csv-wave/{session-id}/ ├── tasks.csv # Master state (all tasks, both modes) ├── results.csv # Final results export ├── discoveries.ndjson # Shared discovery board (all agents) ├── context.md # Human-readable report ├── wave-{N}.csv # Temporary per-wave input (csv-wave only) ├── interactive/ # Interactive task artifacts │ ├── fixer-result.json # Per-task results │ └── cache-index.json # Shared exploration cache └── agents/ └── registry.json # Active interactive agent tracking
Implementation
Session Initialization
// Parse arguments const args = parseArguments($ARGUMENTS) const AUTO_YES = args.yes || args.y || false const CONCURRENCY = args.concurrency || args.c || 3 const CONTINUE_SESSION = args.continue || null const MODE = args.full ? 'full' : args.fix ? 'fix-only' : args.quick || args.q ? 'quick' : 'default' const DIMENSIONS = args.dimensions || 'sec,cor,prf,mnt' const TARGET = args._[0] || null
// Generate session ID
const sessionId = RV-${slugify(TARGET || 'review')}-${formatDate(new Date(), 'yyyy-MM-dd')}
const sessionDir = .workflow/.csv-wave/${sessionId}
// Create session structure
Bash({ command: mkdir -p "${sessionDir}/interactive" "${sessionDir}/agents" })
Write(${sessionDir}/discoveries.ndjson, '')
Write(${sessionDir}/agents/registry.json, JSON.stringify({ active: [], closed: [] }))
Phase 0: Pre-Wave Interactive
Objective: Parse arguments, validate target, detect pipeline mode
Execution:
-
Parse command-line arguments for mode flags (--full, --fix, -q)
-
Extract target path/pattern from arguments
-
Validate target exists and resolve to file list
-
Detect pipeline mode based on flags
-
Store configuration in session metadata
Success Criteria:
-
Refined requirements available for Phase 1 decomposition
-
Interactive agents closed, results stored
Phase 1: Requirement → CSV + Classification
Objective: Generate task breakdown based on pipeline mode and create master CSV
Decomposition Rules:
Mode Tasks Generated
quick SCAN-001 (quick scan only)
default SCAN-001 → REV-001
full SCAN-001 → REV-001 → FIX-001
fix-only FIX-001 (requires existing review report)
Classification Rules:
-
Scanner tasks: exec_mode=csv-wave (one-shot toolchain + LLM scan)
-
Reviewer tasks: exec_mode=csv-wave (one-shot deep analysis)
-
Fixer tasks: exec_mode=interactive (multi-round with rollback)
Wave Computation: Kahn's BFS topological sort with depth tracking (csv-wave tasks only).
User Validation: Display task breakdown with wave + exec_mode assignment (skip if AUTO_YES).
Success Criteria:
-
tasks.csv created with valid schema, wave, and exec_mode assignments
-
No circular dependencies
-
User approved (or AUTO_YES)
Phase 2: Wave Execution Engine (Extended)
Objective: Execute tasks wave-by-wave with hybrid mechanism support and cross-wave context propagation.
// Load master CSV
const masterCSV = readCSV(${sessionDir}/tasks.csv)
const maxWave = Math.max(...masterCSV.map(t => t.wave))
for (let wave = 1; wave <= maxWave; wave++) { // Execute pre-wave interactive tasks const preWaveTasks = masterCSV.filter(t => t.wave === wave && t.exec_mode === 'interactive' && t.position === 'pre-wave' ) for (const task of preWaveTasks) { const agent = spawn_agent({ message: buildInteractivePrompt(task, sessionDir) }) const result = wait({ ids: [agent], timeout_ms: 600000 }) close_agent({ id: agent }) updateTaskStatus(task.id, result) }
// Build wave CSV (csv-wave tasks only)
const waveTasks = masterCSV.filter(t => t.wave === wave && t.exec_mode === 'csv-wave')
if (waveTasks.length > 0) {
// Inject prev_context from context_from tasks
for (const task of waveTasks) {
if (task.context_from) {
const contextIds = task.context_from.split(';')
const contextFindings = masterCSV
.filter(t => contextIds.includes(t.id))
.map(t => [Task ${t.id}] ${t.findings})
.join('\n\n')
task.prev_context = contextFindings
}
}
// Write wave CSV
writeCSV(`${sessionDir}/wave-${wave}.csv`, waveTasks)
// Execute wave
spawn_agents_on_csv({
csv_path: `${sessionDir}/wave-${wave}.csv`,
instruction_path: `${sessionDir}/instructions/agent-instruction.md`,
concurrency: CONCURRENCY
})
// Merge results back to master
const waveResults = readCSV(`${sessionDir}/wave-${wave}.csv`)
for (const result of waveResults) {
const masterTask = masterCSV.find(t => t.id === result.id)
Object.assign(masterTask, result)
}
writeCSV(`${sessionDir}/tasks.csv`, masterCSV)
// Cleanup wave CSV
Bash({ command: `rm "${sessionDir}/wave-${wave}.csv"` })
}
// Execute post-wave interactive tasks const postWaveTasks = masterCSV.filter(t => t.wave === wave && t.exec_mode === 'interactive' && t.position === 'post-wave' ) for (const task of postWaveTasks) { const agent = spawn_agent({ message: buildInteractivePrompt(task, sessionDir) }) const result = wait({ ids: [agent], timeout_ms: 600000 }) close_agent({ id: agent }) updateTaskStatus(task.id, result) }
// Check for failures and skip dependents const failedTasks = masterCSV.filter(t => t.wave === wave && t.status === 'failed') if (failedTasks.length > 0) { skipDependents(masterCSV, failedTasks) } }
Success Criteria:
-
All waves executed in order
-
Both csv-wave and interactive tasks handled per wave
-
Each wave's results merged into master CSV before next wave starts
-
Dependent tasks skipped when predecessor failed
-
discoveries.ndjson accumulated across all waves and mechanisms
-
Interactive agent lifecycle tracked in registry.json
Phase 3: Post-Wave Interactive
Objective: Generate final review report and fix summary
Execution:
-
Aggregate all findings from scan and review tasks
-
Generate comprehensive review report with metrics
-
If fixer ran, generate fix summary with success/failure rates
-
Write final reports to session directory
Success Criteria:
-
Post-wave interactive processing complete
-
Interactive agents closed, results stored
Phase 4: Results Aggregation
Objective: Generate final results and human-readable report.
// Export results.csv
const masterCSV = readCSV(${sessionDir}/tasks.csv)
writeCSV(${sessionDir}/results.csv, masterCSV)
// Generate context.md
const contextMd = generateContextReport(masterCSV, sessionDir)
Write(${sessionDir}/context.md, contextMd)
// Cleanup interactive agents
const registry = JSON.parse(Read(${sessionDir}/agents/registry.json))
for (const agent of registry.active) {
close_agent({ id: agent.id })
}
Write(${sessionDir}/agents/registry.json, JSON.stringify({ active: [], closed: registry.closed }))
// Display summary
const summary = {
total: masterCSV.length,
completed: masterCSV.filter(t => t.status === 'completed').length,
failed: masterCSV.filter(t => t.status === 'failed').length,
skipped: masterCSV.filter(t => t.status === 'skipped').length
}
console.log(Pipeline complete: ${summary.completed}/${summary.total} tasks completed)
Success Criteria:
-
results.csv exported (all tasks, both modes)
-
context.md generated
-
All interactive agents closed (registry.json cleanup)
-
Summary displayed to user
Shared Discovery Board Protocol
Discovery Types:
Type Dedup Key Data Schema Description
finding
file+line+dimension
{dimension, file, line, severity, title}
Code issue discovered by scanner
root_cause
finding_id
{finding_id, description, related_findings[]}
Root cause analysis from reviewer
fix_applied
file+line
{file, line, fix_strategy, status}
Fix application result from fixer
pattern
pattern_name
{pattern, files[], occurrences}
Code pattern identified across files
Discovery NDJSON Format:
{"ts":"2026-03-08T14:30:22Z","worker":"1","type":"finding","data":{"dimension":"sec","file":"src/auth.ts","line":42,"severity":"high","title":"SQL injection vulnerability"}} {"ts":"2026-03-08T14:35:10Z","worker":"2","type":"root_cause","data":{"finding_id":"SEC-001","description":"Unsanitized user input in query","related_findings":["SEC-002"]}} {"ts":"2026-03-08T14:40:05Z","worker":"3","type":"fix_applied","data":{"file":"src/auth.ts","line":42,"fix_strategy":"minimal","status":"fixed"}}
Both csv-wave and interactive agents read/write the same discoveries.ndjson file.
Cross-Mechanism Context Bridging
Interactive Result → CSV Task
When a pre-wave interactive task produces results needed by csv-wave tasks:
// 1. Interactive result stored in file
const resultFile = ${sessionDir}/interactive/${taskId}-result.json
// 2. Wave engine reads when building prev_context for csv-wave tasks // If a csv-wave task has context_from referencing an interactive task: // Read the interactive result file and include in prev_context
CSV Result → Interactive Task
When a post-wave interactive task needs CSV wave results:
// Option A: Include in spawn message
const csvFindings = readMasterCSV().filter(t => t.wave === currentWave && t.exec_mode === 'csv-wave')
const context = csvFindings.map(t => ## Task ${t.id}: ${t.title}\n${t.findings}).join('\n\n')
spawn_agent({
message: ...\n### Wave ${currentWave} Results\n${context}\n...
})
// Option B: Inject via send_input (if agent already running)
send_input({
id: activeAgent,
message: ## Wave ${currentWave} Results\n${context}\n\nProceed with analysis.
})
Error Handling
Error Resolution
Circular dependency Detect in wave computation, abort with error message
CSV agent timeout Mark as failed in results, continue with wave
CSV agent failed Mark as failed, skip dependent tasks in later waves
Interactive agent timeout Urge convergence via send_input, then close if still timed out
Interactive agent failed Mark as failed, skip dependents
Pre-wave interactive failed Skip dependent csv-wave tasks in same wave
All agents in wave failed Log error, offer retry or abort
CSV parse error Validate CSV format before execution, show line number
discoveries.ndjson corrupt Ignore malformed lines, continue with valid entries
Lifecycle leak Cleanup all active agents via registry.json at end
Continue mode: no session found List available sessions, prompt user to select
Target path invalid AskUserQuestion for corrected path
Scanner finds 0 findings Report clean, skip review + fix stages
Core Rules
-
Start Immediately: First action is session initialization, then Phase 0/1
-
Wave Order is Sacred: Never execute wave N before wave N-1 completes and results are merged
-
CSV is Source of Truth: Master tasks.csv holds all state (both csv-wave and interactive)
-
CSV First: Default to csv-wave for tasks; only use interactive when interaction pattern requires it
-
Context Propagation: prev_context built from master CSV, not from memory
-
Discovery Board is Append-Only: Never clear, modify, or recreate discoveries.ndjson — both mechanisms share it
-
Skip on Failure: If a dependency failed, skip the dependent task (regardless of mechanism)
-
Lifecycle Balance: Every spawn_agent MUST have a matching close_agent (tracked in registry.json)
-
Cleanup Temp Files: Remove wave-{N}.csv after results are merged
-
DO NOT STOP: Continuous execution until all waves complete or all remaining tasks are skipped
Coordinator Role Constraints (Main Agent)
CRITICAL: The coordinator (main agent executing this skill) is responsible for orchestration only, NOT implementation.
Coordinator Does NOT Execute Code: The main agent MUST NOT write, modify, or implement any code directly. All implementation work is delegated to spawned team agents. The coordinator only:
-
Spawns agents with task assignments
-
Waits for agent callbacks
-
Merges results and coordinates workflow
-
Manages workflow transitions between phases
Patient Waiting is Mandatory: Agent execution takes significant time (typically 10-30 minutes per phase, sometimes longer). The coordinator MUST:
-
Wait patiently for wait() calls to complete
-
NOT skip workflow steps due to perceived delays
-
NOT assume agents have failed just because they're taking time
-
Trust the timeout mechanisms defined in the skill
Use send_input for Clarification: When agents need guidance or appear stuck, the coordinator MUST:
-
Use send_input() to ask questions or provide clarification
-
NOT skip the agent or move to next phase prematurely
-
Give agents opportunity to respond before escalating
-
Example: send_input({ id: agent_id, message: "Please provide status update or clarify blockers" })
No Workflow Shortcuts: The coordinator MUST NOT:
-
Skip phases or stages defined in the workflow
-
Bypass required approval or review steps
-
Execute dependent tasks before prerequisites complete
-
Assume task completion without explicit agent callback
-
Make up or fabricate agent results
Respect Long-Running Processes: This is a complex multi-agent workflow that requires patience:
-
Total execution time may range from 30-90 minutes or longer
-
Each phase may take 10-30 minutes depending on complexity
-
The coordinator must remain active and attentive throughout the entire process
-
Do not terminate or skip steps due to time concerns