Work orders and planning
Track a goal across dependent work items and multiple workflows.
A work order groups the cards needed to achieve a goal. Its rollup includes progress, active runs, blocked work and tokens across workflows.
Plan the work
Describe the goal and constraints. A planner can propose items and dependencies. Review the proposal before turning it into cards: check the scope, acceptance criteria, destination workflows and ordering.
A work order can also arrive from a linked OpenBook space as an epic: the planner writes its stories, and they wait for a person's approval before any card exists. See epics from OpenBook. The concierge can open a work order for you as well, with its plan lines as proposals.
Templates make recurring work orders reusable. Dependencies determine which items are waiting on others. The critical-path and risk views help explain why a goal may be delayed, and the report can be printed for review.
Planning history
Open a work order and expand Planning history in its plan. The latest 50 invocations show the planner, plan or replan, instruction version, status, attempted model calls and recorded tokens. A replan links to the preceding invocation. Failed or partial attempts remain visible with their reason; proposals saved before a provider failure remain in the plan.
The planner uses the instruction release configured in Agent studio. It selects a version before the first model call and retains it throughout that invocation. A later replan selects the release current when it starts. Without a release, the history labels the invocation Live instructions. Historical work orders from before this feature have no invented history.
Each invocation also has Model input history. It lists the latest 50 prepared inputs, including fallback attempts: turn, model, input bytes, shortened earlier items, approved-memory inclusion and omission, request hash, and the person or factory-supervisor replan that authorized it. These are fitted logical requests prepared for Bookbag SSO. They do not prove provider delivery or replay historical prompt bodies.
Approved memory refreshes before each turn. CodeBook checks current access, goal, pause state, tool permissions and applicable supervisor authority before a call and again before using its answer. If those change while an answer is pending, planning stops; earlier proposals remain inert and autonomous replanning does not create cards from them. Ask for a new plan after resolving the change.
Completion records that planning ran, not that its proposal was correct. Review the proposed scope and acceptance criteria before creating cards. Planner execution counts stay separate from station quality results.
Control execution
Pause or resume a work order when priorities change. Cancellation includes cleanup of the tracked work, but cannot undo external effects already performed. Inspect active external runs and repository changes before assuming a cancellation reversed them.
The planner proposes work. A workflow supervisor handles bounded exceptions during execution, a specialist can take one class of failure, and a factory supervisor sits above every workflow with a closed list of options, including replanning the unmet rest of a partially met work order. See factory supervision. A person chooses the permitted actions and reviews exceptions beyond those limits.
Use operations for active work and events to connect external outcome signals.
Run a supervised overnight order
- In Work orders, create a cloud order with one goal, the repository or service, what is out of scope, and how a reviewer can tell it worked. For a migration, name the existing fixtures and the outputs they must preserve. Enter Spend limit (USD, optional), Deadline or both before creating the order. The date field uses your local time zone. When proposing the order through chat, ask to see the exact stop time and zone before approving.
- Review the planner’s proposed milestones and story dependencies. Look for one observable behavior and one acceptance check per pull request. Confirm where people approve changes, who verifies the merged result and which prerequisites must finish before parallel work starts. Approve the items you want materialized as cards.
- Open the order and choose Open a manager for a dedicated conversation. Review its separate token ceiling before starting it. Give the manager the order reference, your acceptance criteria and which decisions require a person. The manager follows the order’s stories and can report what needs you. An overnight deadline is a stop time. It does not promise an uninterrupted run of that duration.
- Return to the same order reference, such as
WO-7, in the morning. Read Executive report, inspect the proof and PR commit for each completed milestone, and check any blocked or unverified items. In a connected chat, ask “Show WO-7 in Payments. Which checks passed on which commit, what did it cost, and which decisions need me?” Reading the report does not start another worker.
Read the limits
| On the order | What it tells you |
|---|---|
| Spend limit at configured rates | USD ceiling shared by the planner, manager, workers and eligible hosted proofs. The manager’s token ceiling can stop its own session earlier. External CI, deployment charges and billing under your provider contract remain separate. |
| Reserved | Funds set aside before an operation starts. Concurrent workers share the balance. A timeout or an incomplete provider usage report can leave the reservation outstanding. |
| Recorded and available | Known spend and the amount other work can still reserve. An outstanding reservation reduces what is available. |
| Last operation held | The reason a step could not run, for example an unpriced model, an unfundable reservation, or an executor unable to enforce the deadline or spend cap. |
| The deadline has passed | New work is held. Inspect completed work and any provider request already sent. If more work is needed, create a follow-up order with a newly approved limit. |
A limited order uses text matching for knowledge retrieval. Hosted proofs with a spend cap require a prepared Modal environment and a configured sandbox rate. Choose a cloud executor that can enforce the approved limits if a local or external runner is held. Overnight command reference
Watch production after a merge
Open the order’s Production alerts → Watch production alerts. Enter optional text from the incident title and choose Watch as owner. The match ignores case. A blank filter includes all qualifying alerts tied to work shipped by this order. Enable the watch before the incident begins.
A qualifying production incident must meet the factory’s severity threshold and start within 24 hours after the deployment that first made the order’s merged PR live. CodeBook pauses follow-up work, records the affected pull request, commit, alert and deployment, and addresses the owner on the card. Inspect the provider alert, deployed SHA and customer impact. If you decide to roll back, carry out the rollback in your deployment system.
Under Rollback decision, the watch owner or order creator chooses Rolled back or Kept deployment for each distinct deployment on hold. Optionally enter up to 1,000 characters in Decision note (optional), then choose Record and resume. Other viewers can read the recorded choice and who resumed the work. Older acknowledgments can show Decision not recorded. Repeated delivery of the same incident leaves the acknowledgment in place. Connect the monitoring source and verify the deployment

Review the change
The GitHub draft shows the checks known at its opening commit and the decision requested from the reviewer. After another push, open GitHub Checks for the current head. A reviewer can end a submitted GitHub review with Review minutes: 8. The dashboard shows the average reported time and how many merged PRs contributed. Reviewing a pull request · Productivity metrics
For a recurring production bug, add an incident check to the repository’s CI.