Skip to main content
Project To-dos are the required actions an employer assigns to hired AI trainers on one job: acknowledgements, external training, native forms and quizzes, and work submitted for employer review. The developer surface manages the same list the employer web app shows — there is no separate agent-only store. Everything below applies identically to the HTTP API, the CLI, and the MCP tools.

​
The data model

  • An item is one logical to-do on a job, organized under an optional category.
  • Every edit publishes a new immutable version. History is never rewritten.
  • An assignment is one person’s obligation for one item version, with an append-only event and evidence trail.
  • Quizzes are to-do items pinned to one immutable published form version.

​
Audiences are always explicit

Every create, assign, or reissue names its audience. There is no default: omitting audience fails with 400 instead of assigning anyone. SELECTED fails closed. The list must be non-empty and duplicate-free, and every ID must be an active contract of that exact job. An unknown, ended, or foreign ID rejects the whole request with each offender named — nothing is silently dropped, and SELECTED can never turn on future-hire auto-assignment.

​
Preflight receipts for broad assignments

ACTIVE, FUTURE, and ACTIVE_AND_FUTURE fan out beyond an explicit list, so they require a two-step confirmation:
  1. Call the no-write preflight with the exact request you intend to send. It returns the live recipients (masked names, contract IDs, and counts), future-hire semantics, projected due and blocking behavior, duplicate candidates, and a signed receipt.
  2. Send the mutation unchanged with preflightReceipt set to that token.
The receipt is signed and short-lived (about 15 minutes) and is bound to your token’s account, the job, the action, the audience, the exact request payload, and the live roster. If anything changes — the payload, the roster, the audience — the mutation fails with 409 and a PREFLIGHT_RECEIPT_* reason, and you preflight again. Roster changes that land between the receipt check and the write itself also fail closed with 409 RECIPIENTS_CHANGED. SELECTED does not need a receipt: the explicit ID list is already the exact confirmation.

​
Idempotency is bound to the whole request

Every mutation requires an idempotencyKey (8–200 characters). For item creation the key is bound to the complete request — definition, audience, selected contracts, placement, and future-hire semantics:
  • The same key with the same request replays: you get the committed result with replayed: true, and no new assignments are created. A replay never re-resolves the roster, so a person hired after the original create is not silently added.
  • The same key with a changed request fails with 409 — a reused key can never widen a fan-out.
  • A different key with an identical active request fails with 409 LIKELY_DUPLICATE unless you explicitly pass allowDuplicate: true.
Item creation and its assignment fan-out commit in one transaction: you never get an item without its assignments or a partial fan-out.
update_item publishes a new immutable version on every call and is not retry-convergent. Do not blind-retry it; verify the result with a read instead.

​
Normalized results

Mutations return an explicit, machine-readable outcome: the item and version IDs, the audience echo with future-hire semantics, a fan-out summary (matchedContracts, createdAssignments, assignmentIds), the masked recipients the request was verified against, a replayed flag, and warnings. Archiving returns the exact number of open assignments it closed, and that archiving is not reversible through this surface while history is retained.

​
Provider-backed obligations use the same control plane

Configuring a provider connection does not invite or enroll anyone. It publishes the provider-verified to-do dormant, with no audience and future-hire assignment off. Assigning that normal Project To-do is the only action that provisions the selected people with the provider. Closing one person’s provider to-do is the canonical unassignment: future sync and provider enrollment stop while the person’s event history remains available for audit. There is no parallel manual-invite path. WAIVE remains the audited administrative override, but it terminates only the OpenTrain obligation and does not stop provider sync; use CLOSE whenever the intent is to remove provider access.

​
Selective coaching is not a full retake

Selective coaching lets an employer draft threaded feedback and choose the exact quiz questions a worker must redo. The employer first creates a no-write send-back plan; confirmation applies only the signed selection and rejects review-state drift. The worker can then start or resume, edit only the requested answers, reply to visible threads, and resubmit. A full retake reopens the whole quiz and grants another attempt on the same immutable version. Batch full retakes use a separate no-write plan and checksum-bound apply. Keeping these workflows separate prevents a targeted correction from silently widening into a full reassessment.

​
Privacy

Agent responses identify people by contract ID, user ID, and a masked display name (first name plus last initial). Personal emails never appear in list responses, and the CSV export serves the worker_email column blank.

​
Permissions

Employer reads require project_todos:read; employer mutations require project_todos:write, employer access to the job, and a claimed account. Those two scopes are available on fine-grained employer keys and included in employer Full access. Worker coaching instead requires a worker-minted project_todos:respond key and exact assignment ownership; that worker-only scope is deliberately excluded from employer Full access. The account must also have the default-off project_todos_integrations agent-surface family enabled; a full-access key does not bypass that separate gate. Static coaching-capability discovery is the only ungated exception and returns no tenant data. See Scopes and capabilities.

​
Where to go next

  • API reference — endpoints, actions, and errors.
  • Quiz surface — authoring, assigning, grading, and retakes.
  • Selective coaching — targeted redo, threads, worker responses, exports, and feedback images.
  • CLI — preview-first opentrain todos commands.
  • MCP tools — the same safety contract for MCP agents.