Verification Gates

Overview

Verification gates are mandatory quality checkpoints that must pass before any implementation work is considered complete. These gates prevent broken code from being committed, ensure type safety, maintain code style consistency, and validate that all tests pass.

Owner Agent: Alphonse (Verifier Agent)

Key Principles

1. Non-Negotiable Quality: Every code change must pass all applicable gates
2. Automated Enforcement: Gates are triggered automatically via hooks
3. Clear Feedback: Failures provide actionable error messages
4. Project-Aware: Commands adapt to detected project type and tooling
5. Fail-Fast, Fix-First: When any gate fails, work stops until resolved

Verification Hierarchy

Loid (Executor)          Alphonse (Verifier)
     |                         |
Quick sanity tests       Full verification gate
during implementation    before completion
     |                         |
     +----------> Handoff >----+

Gate Types

Pre-Commit Gates

Pre-Complete Gates

Security Gates

Triggered by changes to sensitive files or security-critical code:

Verification Commands

Quick Reference by Language

For complete command reference, see references/verification-commands.md.

Verification Process

Alphonse Verification Steps

1. Identify project type (Node.js, Python, Go, Rust, Java)
2. Run test suite with appropriate framework
3. Execute type checking if configured
4. Run linters if present
5. Attempt production build
6. Report structured results

Standard Output Format

## Verification Results

### Tests
- Status: [PASS | FAIL]
- Output: [Summary of test execution]

### Type Check
- Status: [PASS | FAIL | SKIPPED (<why>) | COVERED (Lawliet)]
- Errors: [List of type errors if any]

### Lint
- Status: [PASS | FAIL | SKIPPED (<why>) | COVERED (Lawliet)]
- Warnings: [Count and summary]

### Build
- Status: [PASS | FAIL | SKIPPED (<why>)]
- Issues: [Build errors if any]

### Overall: [VERIFIED | FAILED | ENVIRONMENT_BLOCKED]

COVERED (Lawliet): in the parallel /orchestrate Phase 4+5 flow, Lawliet runs type check and lint concurrently, so Alphonse reports them as covered instead of re-running them.

Failure Handling

Failure Severity

These are gate-failure labels for Alphonse's verification. They are separate from the ERROR/WARNING/INFO review-finding scale. The machine-read verdict is Alphonse's ### Overall: VERIFIED | FAILED | ENVIRONMENT_BLOCKED.

Escalation Path

Alphonse's ### Overall verdict decides the route:

• VERIFIED: proceed.
• FAILED: Loid fixes within the capped review-fix rounds, then re-verify.
• ENVIRONMENT_BLOCKED: warn and proceed with the caveat (not routed to Loid).

At the round cap, open failures go to the Phase 6 report, and the orchestrator pauses only when one is unsafe to ship.

For detailed failure protocols, see references/failure-handling.md.

Best Practices

DO

• Run verification after every significant change
• Fix failures immediately
• Trust Alphonse's verdict
• Add tests for new features
• Run full suite before PR

DON'T

• Skip verification to save time
• Suppress type errors with any
• Disable lint rules without justification
• Commit with failing tests
• Use --no-verify flag
• Mark tasks complete without verification

Gate Selection Guide

Override Rules

Verification gates can ONLY be skipped when:

1. User explicitly requests skip - Must be documented
2. No code changes were made - Documentation-only tasks
3. Task is purely exploratory - No production impact

Override Documentation Required:

## Verification Override

Reason: [Explicit reason for skipping]
Requested by: [User or escalation source]
Risk Assessment: [Low/Medium/High]
Compensating Controls: [What will catch issues later]

Multi-reviewer disagreement (Codex co-review)

Disagreement rule: See the canonical truth table and Divergence Cap in references/codex-co-review.md. The summary: any cited Lawliet ERROR/WARNING blocks (even under APPROVED); Codex flips the verdict only with a file:line-cited ERROR/WARNING — INFO is advisory; Codex runs only under standard/thorough with codex.available: true.

Additional Resources

Reference Files

• references/verification-commands.md - Complete command reference
• references/project-detection.md - Project type detection details
• references/failure-handling.md - Failure handling protocols
• references/codex-co-review.md - /orchestrate Phase 4 Codex co-review, truth table, Divergence Cap
• references/intent-ledger.md - /orchestrate Phase 6 Intent Ledger template

Examples

• examples/verification-scenarios.md - Worked verification examples

Related Files

Related Skills

• task-classification: Determines when verification is required
• agent-behavior-constraints: Model routing and tool access