Autonomy and /goal

Teams don't jump straight from pair-programming with an agent to a "dark factory". Buildline lets you raise autonomy one level at a time, as your gates and acceptance scenarios earn trust. Set autonomy in .buildline/buildline.json.

At every level, a human merges. Approvals are recorded in feature.json, and agents are instructed never to record one on someone's behalf.

#Autopilot with /goal

Claude Code's /goal keeps a session working across turns until a condition holds. After each turn a small evaluator model checks the condition against the conversation. It can't run commands itself, so the condition must be provable from printed output.

npx buildline run <id> opens Claude Code with a condition like:

/goal Feature 012 (Saved carts) reaches status "ship": following .buildline/stations/autopilot.md,
every task in .buildline/features/012-saved-carts/feature.json has "passes": true, the last
`node .buildline/bin/gates.mjs` run printed "ALL GATES PASSED", and the buildline-verifier subagent
reported every acceptance scenario as PASS. Stop early if blocked as defined in the playbook,
or after 60 turns.

The autopilot playbook loops build → verify → review, printing a status line after each step. It stops and writes the blocker to progress.md when:

• a task fails after two genuinely different approaches,
• the spec looks wrong or contradictory,
• it needs secrets, paid services, production or destructive actions,
• metrics.reworks reaches 5.

#Running unattended safely

• Use auto-accept or auto permission mode, ideally in a container, VM or git worktree.
• Accept the workspace trust dialog once interactively. Until you do, Claude Code ignores the project's permissions.allow entries.
• Headless: npx buildline run 012 --headless (claude -p … --permission-mode acceptEdits).
• Review the branch like any other PR; the ## Review section is written for you.

Edit this page on GitHub