Context Circuit
Get started

Your first Context Circuit workflow

Take one bounded product change through the complete human-controlled lifecycle.

Template 2.5.0CLI 2.5.0

Start with knowledge

Describe the outcome in project language:

Let customers download an invoice as a PDF. Preserve the current authorization
rules and audit trail.

The agent retrieves the smallest relevant set of accepted project knowledge. It reads the request against that context, writes an intent with success criteria and numbered open questions, and stops. The CLI allocates the intent ID and maintains the record shape; it does not decide what the outcome means.

Read the outcome, approve, then ask for plans

Answer questions by number, then approve only when the written outcome is correct:

1. Include paid and overdue invoices.
2. Use the invoice number as the filename.
3. Record every download in the existing audit trail.
The intent now matches what I want. I approve it.

The agent passes your exact approval words to the CLI, which stamps approved_at and returns planning_recommended. Approval records your decision and nothing more: the agent says that plans are the next step, recommends creating them, and stops. No code is read and no plan ID is reserved yet.

Planning starts when you ask for it:

Create the plans.

You can also say both at once — "I approve it. Create the plans." — once you have read the intent. The agent then investigates real code, decides which repositories are involved, and writes linked plans without asking for approval again. The CLI allocates plan IDs and links records; it does not inspect code or author the plan.

Commit the records you changed

Approval and completion leave the records they touched uncommitted, and the CLI reports them as commit_recommended. The agent asks you one question: commit and push those files, or commit only. Your answer authorizes that commit, and a push of the workspace's base branch to its upstream, that once. After "commit only", the agent says other clones cannot see the records yet. The CLI never stages, commits, or pushes.

Ask to execute after plans exist

Once the plans are visible, make a new request:

Execute all plans for this intent. Prepare the required worktrees, implement the
approved outcome, run the repository checks, and report anything unverified.

The agent asks the CLI for dependency order and worktree preparation. It then implements the outcome, runs the actual repository checks, and commits implementation work on the prepared plan branches. The CLI does not execute code, run tests, commit, or claim that work succeeded.

Deliver, then close the circuit

Delivery is another explicit request:

Push the completed plan branches and open pull requests against this machine's
recorded base branches.

After the work lands:

The invoice PDF changes have landed. Mark the plans complete with these exact
words and reconcile any durable knowledge that changed.

The agent uses ordinary Git and provider tools for delivery. For completion, it gives your exact note to the CLI, then judges whether the accepted project knowledge must change. The CLI records completion but cannot make that knowledge judgment. The agent then asks the commit question for the completed records, and, while you have finished records not archived, offers to move them under _archive/. Archiving is optional, keeps every ID, and a no changes nothing.

Diagram viewer
100%

Drag to move around. Use the controls to zoom.

Separate decisions

Delivery does not mark a plan complete. Completion does not remove worktrees or branches. Cleanup requires its own request.

Use the workflow guides for deeper explanations of each conversation stage, or the prompt cookbook for more supported cases.

On this page