Skip to main content

​
Overview

After an AI agent implements a ticket, Draft runs a configurable verification pipeline to ensure the changes are correct before human review.

​
Configuration

Define verification commands in draft.yaml: 
verify_config:
  commands:
    - "pytest tests/"
    - "npm run lint"
    - "npm run typecheck"

​
How It Works

  1. A ticket moves to VERIFYING after successful execution
  2. Each command runs sequentially in the ticket’s isolated worktree
  3. Commands stop on first failure — remaining commands are skipped
  4. Evidence is captured for every command (stdout, stderr, exit code)

​
Outcomes

ResultTicket StateNext Step
All commands passNEEDS_HUMANHuman reviews the changes
Any command failsBLOCKEDAgent can retry or human intervenes

​
Evidence

Every verification command produces an Evidence record containing:
  • stdout — Standard output captured to a file
  • stderr — Standard error captured to a file
  • exit code — Command return code
  • command — The exact command that was run

​
Viewing Evidence

In the UI, expand a ticket’s detail panel to see evidence for each command:
  • Green checkmark = passed
  • Red X = failed (with exit code)

​
Evidence API

# List all evidence for a ticket
curl http://localhost:8000/tickets/{id}/evidence

# Get stdout for a specific evidence record
curl http://localhost:8000/evidence/{id}/stdout

# Get stderr for a specific evidence record
curl http://localhost:8000/evidence/{id}/stderr

​
Best Practices

Verification runs on every iteration. Prefer targeted test suites over full test runs.
AI-generated code can have type errors. Add tsc --noEmit or equivalent.
Catch formatting and style issues automatically with ruff check, eslint, etc.
Put fast-failing commands first (lint before tests) to get quicker feedback.