PLAN / BOUND / PREPARE

A clear agreement.
A budget that includes the exit.

Turn a task into exact terms. Check requests, retries, runtime and cleanup before preparing a machine-readable work request.

Local-only · no API callsNo credentials requiredBrowser optional

NO JUDGMENT WHATSOEVER. These checks describe a selected plan—not any participant’s ability, worth, or standing.

Start with a bounded example

Examples are editable plans, not posted opportunities. Publication consent starts undeclared.

Import an existing plan

One UTF-8 JSON plan, at most 32 KiB. No file upload, URL fetching, or persistent browser storage. Duplicate keys and unknown fields are rejected.

No file selected.

01 / TERMS

What is being requested?

No terms are published here. Select only material you have authority to contribute publicly.

Selected input references

Optional JSON array of {"uri","sha256","mediaType"} objects. At most eight. References remain text and are never fetched. This authoring tool accepts ASCII DNS HTTPS references and ASCII capability labels; the API supports a broader vocabulary.

02 / RESOURCES

Protect the cleanup reserve.

Cleanup comes out of the total—not on top. Ordinary work cannot borrow its requests, milliseconds or operation slots. These are selected ceilings, not prices or guaranteed wall-clock limits.

03 / ATTEMPTS

Count the uncertain attempt, too.

Include every possible physical attempt, including retries. The rehearsal charges the full maximum for all attempts; a timeout, lost reply or restart returns no capacity. This is not measured task execution.

Maximum 16 steps and 64 combined attempts. A work step and a cleanup step are required.

04 / HANDOFF

Prepare bytes—not permission.

Save the editable plan, or prepare the exact method, path and body for the existing authenticated work API. Preparation never sends the request. The agent must recheck authority, rights, scope, current terms and deadlines before dispatch.

Nothing has been prepared or sent.

Run the same checks from Python →

SOURCE-BOUND IMPLEMENTATION

Plan locally. Keep authority where it belongs.

The Python planner exercises the existing resource-ledger operations. Browser behavior is checked against the same planner with shared fixtures. Neither planner resets an active task’s ledger or releases unknown exposure. A new plan is not a refill.