start

Your first task

What does one task look like end to end?

Before you start, install Stagecoach and set up your project.

1. Specify

In your agent, invoke aw-specify (or agent-workflow:specify in the native Claude plugin):

aw-specify Add CSV export to the reports endpoint.

It reads the project workflow, repository instructions, and the nearest working examples, then asks only what it can't find out. For this task it might ask whether the export must match the on-screen column order, and whether the existing JSON endpoint stays unchanged.

For the full lane it writes the plan folder:

ai-plans/csv-export/
├── prompt.md            # the original request and decisions
├── requirements.md      # outcomes and acceptance criteria
├── implementation.md    # approach, existing patterns, scope
├── deferred.md          # what is deliberately left out, and why
└── pipeline.json        # the contract the runner enforces

The heart of pipeline.json is small. One writable repository with its paths, one check with its command and expected test identity, and one outcome mapped to it:

{
  "schema_version": 1,
  "task": "csv-export",
  "plan_files": ["prompt.md", "requirements.md", "implementation.md", "deferred.md"],
  "repositories": [
    {"id": "app", "access": "write", "branch": "feature/csv-export",
     "paths": ["src/**", "tests/**"], "test_paths": ["tests/**"]},
    {"id": "checks", "access": "read"}
  ],
  "stages": ["implement-tests", "implement", "review"],
  "checks": [
    {
      "id": "csv-export-behavior",
      "repository": "app",
      "command": {"argv": ["python3", "-m", "unittest", "-v", "tests.test_reports"], "cwd": ".", "timeout_seconds": 300},
      "parser": "unittest",
      "identities": ["test_exports_csv"],
      "phases": ["red", "green"]
    }
  ],
  "outcomes": [{"id": "exports-report-as-csv", "checks": ["csv-export-behavior"]}]
}

checks is a read-only companion repository: it can supply context and checks, and it never gets a branch. The full contract is in the plan reference.

2. Confirm the plan

Approve when another engineer could implement it without inventing behavior:

• no product or implementation choice is left open;
• every outcome maps to a named check with a command, parser, and test identity;
• writable paths, the read-only repository, and deferred work are explicit;
• risky areas are named;
• stages, routes, and delivery actions are what you expect.

3. Manual lane

For a change this small you could stop here and implement in the same session:

aw-implement

implement follows the plan's path boundaries, writes the test and the code, runs csv-export-behavior, fixes what fails, and inspects the final diff. It reports what it observed and what remains. It doesn't commit or push.

4. Full lane

Validate the plan and check readiness:

agent-workflow validate-plan ai-plans/csv-export/pipeline.json

Example output

{
  "task": "csv-export",
  "valid": true
}

agent-workflow preflight ai-plans/csv-export/pipeline.json

Example output, trimmed

{
  "binding_sha256": "d0ea11d3…",
  "boundary": "Provider controls plus post-stage path/evidence guards; trusted mode is not an OS sandbox.",
  "permission_mode": "trusted",
  "routes": [
    {
      "model": "your-model",
      "model_validation": "passed to provider; account availability is verified on first request",
      "permission_mode": "trusted",
      "provider": "codex"
    }
  ],
  "side_effects": "Stage selection authorizes only the named delivery actions.",
  "stages": ["implement-tests", "implement", "review"],
  "task": "csv-export"
}

run --dry-run prints the same report and stops. When you're ready, start the run in the background:

agent-workflow run ai-plans/csv-export/pipeline.json --detach

Example output

{
  "log": "/path/to/project/.agent-workflow/local/evidence/csv-export/runner.log",
  "next": "agent-workflow --project /path/to/project status csv-export",
  "pid": 48213,
  "task": "csv-export"
}

Follow it:

agent-workflow watch csv-export --seconds 300

Example output

{"status": "running", "stage": "implement-tests", "completed": [], "pending_input": null, "error": null}
{"status": "running", "stage": "implement", "completed": ["implement-tests"], "pending_input": null, "error": null}
{"status": "running", "stage": "review", "completed": ["implement-tests", "implement"], "pending_input": null, "error": null}
{"status": "complete", "stage": "review", "completed": ["implement-tests", "implement", "review"], "pending_input": null, "error": null}

In the implement-tests stage the runner accepted the red state only because the expected test failed on an assertion:

Example output from the red check

FAIL: test_exports_csv (tests.test_reports.Csv.test_exports_csv)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "tests/test_reports.py", line 6, in test_exports_csv
    self.assertEqual(export([["a", 1]]), "a,1\r\n")
AssertionError: None != 'a,1\r\n'

A ModuleNotFoundError or a syntax error in the same place would have been rejected as a setup failure.

5. Read the evidence

Everything the runner observed stays in the project's evidence directory, by default .agent-workflow/local/evidence/csv-export/:

.agent-workflow/local/evidence/csv-export/
├── state.json           # the run journal: status, stages, attempts, delivery
├── binding.json         # the plan, profile, skills, and rules this run is bound to
├── attempts/0000.json   # each agent attempt: route, session, usage, output
├── evidence/            # each verified stage: checks, rules, fingerprint
└── runner.log           # output of a detached runner

A green record contains the command the runner ran, its parsed result, the rule findings, the changed paths, and the fingerprint of the files:

Example: evidence/implement-0001.json, trimmed

{
  "changes": {"app": ["src/reports.py"]},
  "checks": [
    {
      "command": "python3 -m unittest -v tests.test_reports",
      "exit_code": 0,
      "id": "csv-export-behavior",
      "identities": ["test_exports_csv"],
      "output": "test_exports_csv (tests.test_reports.Csv.test_exports_csv) ... ok\n\nRan 1 test in 0.000s\n\nOK",
      "parser": "unittest",
      "passed": true,
      "phase": "green",
      "problems": [],
      "repository": "app"
    }
  ],
  "fingerprint": "c0045111…",
  "passed": true,
  "rules": []
}

Each attempt also records usage and cost_usd exactly as the provider reported them. A provider that reports no cost leaves cost_usd as null, never zero.

Next steps

• Add commit-and-push and draft-pr to the stages, with an approved commit message, title, and body file, to let the runner deliver. See review and delivery.
• Learn what to do when a run pauses: when a task stops.
• Read how the evidence is judged: evidence and checks.

Checkpoints IN-1 VER-1 VER-2 VER-5

Edit this page on GitHub