Verification System Integration Guide

Overview

The verification system provides real-time validation of agent work and task execution. It integrates with swarm commands, non-interactive mode, and the training system.

How It Actually Works

1. Pre-Task Verification

Before any task executes, the system checks:

Git repository state (clean/dirty)
Dependencies installed
Environment ready

2. Post-Task Verification

After task completion, it verifies:

Code compiles (TypeScript/JavaScript)
Tests pass
Linting succeeds
Documentation exists (for architect agents)

3. Automatic Rollback

If verification fails and rollback is enabled:

Resets to last git commit
Preserves verification history
Feeds failure data to training system

Integration with Commands

Swarm Command with Verification

bash

# Enable verification for swarm execution
export VERIFICATION_MODE=strict
export VERIFICATION_ROLLBACK=true

# Run swarm with verification hooks
claude-flow swarm "Build REST API" --verify

# The swarm will:
# 1. Run pre-task verification
# 2. Execute the objective
# 3. Run post-task verification
# 4. Rollback if verification fails

Non-Interactive Mode with Verification

bash

# Non-interactive with verification
claude-flow swarm "Build feature" \
  -p \
  --output-format stream-json \
  --verify \
  --threshold 0.95

# Verification results appear in JSON output

Manual Verification

bash

# Pre-task check
node src/cli/simple-commands/verification-hooks.js pre task-123 coder

# Post-task check
node src/cli/simple-commands/verification-hooks.js post task-123 coder

# Feed to training
node src/cli/simple-commands/verification-hooks.js train task-123 coder

# Check status
node src/cli/simple-commands/verification-hooks.js status

Integration with Training System

Verification results automatically feed into the training system:

bash

# View training data
cat .claude-flow/training/verification-data.jsonl

# Sample output:
{"taskId":"task-123","agentType":"coder","preScore":1,"postScore":0.75,"success":false,"timestamp":"2025-01-12T15:00:00Z"}

Verification Scores

Agent Type	Checks Performed	Threshold
coder	typecheck, tests, lint	0.85
researcher	output completeness	0.85
tester	coverage threshold	0.85
architect	documentation exists	0.85

Environment Variables

bash

# Set verification mode (strict/moderate/development)
export VERIFICATION_MODE=strict

# Enable automatic rollback on failure
export VERIFICATION_ROLLBACK=true

# Custom threshold (0.0-1.0)
export VERIFICATION_THRESHOLD=0.95

Real Example: Verifying a Coder Agent

bash

# 1. Initialize verification
./claude-flow verify init strict

# 2. Run a coder task with verification
./claude-flow swarm "Implement user authentication" --verify

# Output:
🔍 Pre-task verification: swarm-123 (coder)
  ✅ git-status: 1.00
  ✅ npm-deps: 1.00
✅ Pre-task verification passed (1.00)

[... task execution ...]

🔍 Post-task verification: swarm-123 (coder)
  ✅ typecheck: 1.00
  ❌ tests: 0.40
  ✅ lint: 0.80
❌ Post-task verification failed (0.73 < 0.95)
🔄 Attempting rollback...
✅ Rollback completed

Verification Memory

Verification results are stored in .swarm/verification-memory.json:

json

{
  "history": [
    {
      "taskId": "swarm-123",
      "phase": "post",
      "passed": false,
      "score": 0.73,
      "checks": [
        {"name": "typecheck", "passed": true, "score": 1.0},
        {"name": "tests", "passed": false, "score": 0.4},
        {"name": "lint", "passed": true, "score": 0.8}
      ],
      "agentType": "coder",
      "timestamp": "2025-01-12T15:30:00Z"
    }
  ],
  "tasks": {
    "swarm-123": {
      "pre": {"passed": true, "score": 1.0},
      "post": {"passed": false, "score": 0.73}
    }
  }
}

Current Limitations

Not Yet Fully Integrated: The verification hooks exist but aren't automatically called by all commands yet
Basic Checks: Currently runs simple npm scripts, not deep code analysis
Manual Setup: Requires environment variables to be set manually
Limited Agent Types: Only 4 agent types have specific verification logic

Future Improvements

Deep Integration: Automatic verification for all agent operations
Smart Rollback: Selective rollback of only failed changes
Learning System: Use verification history to improve agent selection
Custom Checks: Allow project-specific verification rules
Real-time Monitoring: Dashboard showing verification metrics

Using Verification Today

While not fully integrated, you can use verification today by:

Running verification hooks manually before/after tasks
Setting environment variables for automatic checks
Using the --verify flag with swarm commands (when implemented)
Checking verification history with truth command

The verification system provides the foundation for ensuring quality in AI-generated code, even if it's not yet fully automated.