Luiz Thiago de Souza

Problem

Routine tickets kept waiting for attention that more important work deserved. Handing them to an agent only made sense if it could run unattended without ever being able to publish, push or touch the tracker on its own.

Constraints

• It runs unattended on the owner's machine during work hours, inside a daily budget.
• There is no path to the shared remote or the tracker without a human click.
• The owner's own checkouts must never be touched.

How it works

• A heartbeat wakes a coordinator every 15 minutes; it consults a ten-row decision table and takes at most one action.
• A ticket is selected with one allowed query, inside the work window and the day's budget.
• A spec phase runs first; risky work stops at a go or drop gate.
• Execution happens in an isolated worktree, followed by a review and a counter-review from a second model.
• A rejected change gets one repair attempt; a second rejection parks it with both verdicts for the human.
• The human's ship-it on the dashboard triggers the credential scan, the text validator, and a squashed push as a draft PR.

Decisions

• The coordinator is an AI session confined to a deterministic CLI; it never edits code.
• One ship-it click authorizes commit, push and a draft PR, and the draft is itself a second gate.
• Worktrees live on a separate drive, so the owner's own checkouts are never touched.
• Phases are driven by a deterministic layer, which removed idle gaps between them.

Trade-offs

• One task at a time and hard phase timeouts trade throughput for predictability.
• A second model reviews every change, doubling review cost to catch errors before a human sees them.

Safeguards

• The coordinator is denied git, network, file edits and publishing; workers are denied push, fetch, branch, the tracker and network.
• Seven credential-scan rules on every diff, with values masked.
• A daily budget priced by task size, never two tasks at once, a 150-minute ceiling per task and timeouts per phase.
• Three human gates: go or drop, clone a repository, ship-it.

Limits

• Routine tickets only: risky work stops at the go or drop gate and waits for the human.

In numbers

• 15-minute heartbeat
• 10-row decision table
• 324 tests in 24 files
• 5 ADRs