GoalBuddy Conveyor
Runs broad work as sequential child boards with one active child at a time.
Local Codex manager / worker control plane
Turn broad Codex work into repeatable manager recipes: one manager supervises one worker through Dispatch, evidence gates, handoffs, and an auditable SQLite trail.
npm install -g agent-conveyor
conveyor manager-recipes --show goalbuddy-conveyor --json
conveyor pair --manager-recipe goalbuddy-conveyor ...
Selects a recipe, checks evidence, records decisions, and nudges only when the gate allows it.
Routes notifications, blocks unsafe continuations, records command attempts, and keeps delivery mechanical.
Implements one slice, returns receipts, consumes inbox messages, and hands off before context cleanup.
The setup conversation can stay free-form, but before the manager starts it resolves to a named recipe or explicit custom mode. Each recipe has settings, permissions, evidence gates, and cleanup rules.
Runs broad work as sequential child boards with one active child at a time.
Improves or proves test confidence before another worker pass can begin.
Iterates visible quality using screenshots, browser evidence, and visual diffs.
Observes, negotiates criteria, and asks focused next-step questions without taking over.
Drives PR readiness, CI monitoring, green merge, handoff, and fresh-worker replay.
Agent Conveyor makes the manager commit to a mode before it gets cut loose, then keeps every loop grounded in Dispatch and receipts.
conveyor manager-config "$TASK" --recipe goalbuddy-conveyor --mode strict
conveyor cycle "$TASK"
conveyor replay "$TASK"
The manager turns free-form intent into `goalbuddy-conveyor`, `ux-polish-loop`, or `custom`.
Mode, objective, permissions, tools, epilogues, and cleanup policy are saved in SQLite.
Notifications are delivered by push or inbox, with command attempts and side effects recorded.
Completion needs receipts: tests, screenshots, CI, handoff, audit, or an explicit blocker.
Recipe behavior is not hidden in chat memory. The control plane records enough state to replay a run, report a bug, and improve the recipes when dogfooding exposes a weak gate.
Recipe name, supervision mode, permissions, tools, epilogues, and ack policy.
Manager requests, Dispatch claims, retries, blocks, and side-effect risk.
Worker and manager messages, delivery mode, consumed state, and failures.
Living criteria, deferred follow-ups, satisfied proof, and rejected scope.
Observation history and the manager context used for each decision.
Dispatch heartbeat, routing issues, stale work, and searchable issue clues.
Agent Conveyor is for local, auditable supervision: a manager that can watch useful work, ask the next good question, and prove why it continued, stopped, compacted, or finished.