A feature run unattended

The installer only scaffolds; it installs nothing into your tools. It prints a line for every file, created AGENTS.md, created BACKLOG.md and so on, and ends with:

    scaffold: 18 created, 1 merged, 0 kept
  ✓ Project scaffolded

The merged file is .gitignore. The commit then records the project files.

This route leaves docs/context/CONVENTIONS.md, GOALS.md and STATUS.md as templates; filling them in is the job of ab-project-start in a session. The run coped by picking its own check and logging the choice: Lint/verification gate with CONVENTIONS.md unfilled -> npm test (node --test); follow the code idiom. If your project has no obvious test command, run /ab-project-start in a session first, as the first two tutorials do, so the run has your real commands.

gh creates the repository, adds it as origin and pushes main:

https://github.com/Ninety2UA/ab-tutorial-scratch-20261008
To https://github.com/Ninety2UA/ab-tutorial-scratch-20261008.git
 * [new branch]      HEAD -> main
branch 'main' set up to track 'origin/main'.

If you make a throwaway repository to try this, deleting it later needs the delete_repo scope, which gh does not have by default. In the recorded run gh repo delete failed with HTTP 403: Must have admin rights to Repository. Run gh auth refresh -h github.com -s delete_repo first, then gh repo delete <owner>/<name> --yes.

git starts the branch at the commit with the project files. The runner later records that commit as the base of the run.

The runner checks everything it needs and starts nothing. After a first line naming the tool, the branch and the project folder, it printed:

  ▸ Dry run: preflight only, nothing starts.
  ▸ Posture: claude -p --permission-mode auto (auto-approve safe tools; a dangerous rm is denied)
  ▸ Helpers per wave: 20 · iteration timeout: 3600s · max iterations: 10
  ▸ claude: logged in
  ▸ claude: -p skips the workspace trust dialog; nothing to trust
  ▸ gh: authenticated
  ▸ Working tree clean
  ▸ Would record base c0cea3c, branch feat/sales-by-customer, push URL https://github.com/Ninety2UA/ab-tutorial-scratch-20261008.git, pull requests in github.com/Ninety2UA/ab-tutorial-scratch-20261008
  ▸ Would run per iteration: claude with the skill prompt for: Add a sales-by-customer report to the reports page
  ▸ Skill reference on this host: /ab-ship-pipeline
  ✓ Dry run complete

Most of the 18 seconds went to the claude and gh sign-in checks.

After the same preflight checks, the runner recorded the base, probed the posture and started the first iteration:

  ▸ Recorded base c0cea3c on feat/sales-by-customer, push URL https://github.com/Ninety2UA/ab-tutorial-scratch-20261008.git, pull requests in github.com/Ninety2UA/ab-tutorial-scratch-20261008
  ▸ Probe: the posture can write .git; the skill commits

  → iteration 1: starting claude (18:45:27)

Then nothing appeared for 39 minutes. It has not hung: the next step shows how to watch it. When the iteration finished, the runner printed the session’s last message and took over publishing:

    last message: The `/reports` page now has a "Sales by customer" table under the region table, and the work is committed on `feat/sales-by-customer`. I didn't push or open a P
  → iteration 1: status done, stage ship

  ▸ Publishing feat/sales-by-customer
  ▸ Scanning the 7 commit(s) the push would publish and the PR body for secrets
  ✓ No secrets found
  ✓ Pushed feat/sales-by-customer to https://github.com/Ninety2UA/ab-tutorial-scratch-20261008.git
  ✓ Opened pull request: https://github.com/Ninety2UA/ab-tutorial-scratch-20261008/pull/1

  Run complete: published after 1 iteration(s).
    iterations: 1 · transient retries: 0 · timeouts: 0 · failed iterations: 0 · denials logged: 0 · 39m31s

Its last line names the folder of its logs, ~/.local/state/agent-blueprint/<repo hash>/logs/. The runner’s exit code was 0, which means published.

ab-ship-pipeline keeps its state in .agent-blueprint/run/state.json. The first version it wrote began:

{
  "status": "running",
  "stage": "plan",
  "iteration": 1,
  "host": "claude",
  "driver": "runner",

During the run its stage moved from plan through execute, review and verify to ship, and the commits showed up in git log as they landed. Inside the one headless session, by local time:

• 18:46, requirements. With nobody to ask, it locked the requirements as decisions in a new docs/context/DECISIONS.md: a second table on /reports, the same ?month= filter, a footer total, an h2 for the new table.
• 18:47, plan. Three research helpers ran in parallel, then ab-writing-plans saved docs/plans/2026-10-08-sales-by-customer-report.md with three tasks. A plan-checker helper found no blocking issue, 2 warnings and 3 suggestions, and all five were applied.
• 18:53, deepen. ab-deepen-plan ran five more research helpers in parallel and a second plan check. The plan was committed at 18:58.
• 18:58, execute. ab-orchestrate ran the tasks as plain helpers in their own worktrees: task 1 in the first wave, tasks 2 and 3 in parallel in the second. A helper checked the result of each wave; 19 tests passed after the second.
• 19:04, review. ab-iterative-refinement ran ab-review-swarm: six reviewers in parallel, then a validator and a synthesizer. It converged in the first round with no P1 finding. ab-resolve-in-parallel then added three tests for gaps the review had found.
• 19:21, knowledge capture. Skipped: the skill records lessons only when the work solved a non-trivial problem, and nothing went into docs/solutions/.
• 19:21, ship. It closed the plan and recorded one open security decision. ab-pr-workflow had a helper audit the plan against the diff and wrote the pull request body to .agent-blueprint/run/pr-body.md. The state was set to done at 19:24, and the runner took over.

At the end, git log --oneline showed the run’s seven commits on top of the project files:

c8b6b71 docs(plans): close the sales-by-customer plan and record the exposure decision
4fe0517 test(reports): pin the region footer and cover & escaping in customer names
df76088 test(reports): cover float drift in salesByCustomer totals
55ba36c docs(readme): mention the sales-by-customer table
645f0fa feat(reports): show sales by customer on the reports page
3a98ec2 feat(reports): add salesByCustomer report function
1d2a1de docs(plans): plan the sales-by-customer report

The pull request, “Add a sales-by-customer report to the reports page”, goes from feat/sales-by-customer into main: 7 commits, 9 files, 485 lines added and 11 removed. The body that ab-pr-workflow wrote has the sections What, Why, How, Testing, Review, Plan audit and Checklist, and one more headed “Needs an owner decision: who may read /reports”. From the body:

• Testing: “npm test (node --test): 22 passing, 0 failing on the branch head. That is 9 existing tests plus 13 new”.
• Review: “1 of at most 3 iterations and converged in fast mode (P1 = 0)”, and “12 raw findings. A validator rejected 5, and 7 were validated: 2 P2 and 1 P3, plus 1 present-tier decision. 3 more fell below the confidence gate.”
• The open decision: “/reports has no authentication, and the server listens on every interface”. With the customer table, the page lists every customer’s name, order count and revenue to anyone who can reach the port. The fixes are in categories the skill must ask about, “so this unattended run changed no code for it”. The question is also in docs/context/DECISIONS.md and BACKLOG.md.
• Plan audit: every task DONE, and a list of the work the plan did not name, such as the three tests the review added and five BACKLOG.md entries for problems that were already there.

The first command prints nothing. After a successful publish the runner deletes .agent-blueprint/run/state.json, pr-body.md and the provenance records; its code says “Only the runner deletes run files, and only here. Logs stay.” So after a published run there is no state.json to open. What stays:

• the runner’s logs, probe.log and iteration-1.log. The iteration log holds the headless session’s result: in this run 121 turns, 26 helper sessions and no permission denials, and a list-price cost of 18.65 US dollars. The recorded account runs on a subscription, so that figure is what the API would have charged.
• the team ledger, .agent-blueprint/team/20261008-1858-sales-by-customer-report/ledger.md.
• the review run, .agent-blueprint/review-runs/review-20261008-190509/, with one JSON file per reviewer and synthesis.md.