content.generate

Binding

Job

Generate or refactor one educational content unit in one linear pass: a teacher brief that constrains everything after it, one written edition per declared language, images made to a stated claim, code and executable checks that actually run, and an independent review that receives the artifacts without the producer’s rationale.

The brief comes first and constrains what follows

The teacher brief is written before anything else, from curriculum and source evidence, in one fresh execution that inherits no turns. It publishes the learner inputs, the observable outcomes, the claims a visual may encode, the examples, and the explicit add, change, and remove dispositions, and it is then fingerprinted so nothing downstream may extend it. Everything after it is measured against it: an edition may only claim coverage of an outcome the brief published, an image may only encode a claim the brief published, and every declared edition must cover the whole published outcome set before the unit can ship. An edition that covers part of the brief is a partial lesson in one language and a complete one in another, which is exactly the silent defect this check exists to catch. A change or remove disposition names what it acts on, because a disposition with no target is a wish.

Generated code must actually run

Every implementation track is built with the exact command commands names for it, and the exit code is read and recorded in response/data/e2e.json. A non-zero exit code cannot ship, because the lesson would tell a learner that code works on nobody’s evidence. Step 7 runs the declared command per track inside a bounded run, read and repair loop whose ceiling is maxE2eIterations; the loop lives inside the step, so E2E_FAILED is raised once, when the iterations are spent, rather than on the first red run. The loop may repair the implementation or the harness. It may never move the contract it is measured by, so the contract fingerprint is taken before the first run and compared after the last one: a test deleted, skipped, loosened, or special cased to make the run green is the failure that comparison exists to catch, and it is CONTRACT_WEAKENED. A track carries its build command and its exit code, a run carries its command, its exit code, and its named assertions, and neither carries a sentence claiming success.

The image is made to a stated intent

The visual is derived from the brief’s claims, the prompt that states that intent is persisted beside the result, and the result is inspected for legibility, hierarchy, and claim fidelity. Step 5 runs only when stageModes turns the image on; with the image off the step produces nothing at all, no image and no prompt, and the receipt records STAGE_DISABLED, because a decision that leaves no record reads later as an omission. A claim the brief never published cannot appear, and a visual that fails its own claim-fidelity inspection cannot ship. An image is not decoration added at the end; it is one of the claims, drawn.

The review is a nested exchange, and revision is a fallback with a ceiling

After the executable check, the branch pauses: it emits response/response.json with status waiting and awaiting { exchange: review, kind: content-review }. The orchestrator writes review/request/request.json with the produced artifacts and the claims they make as inputs, never the producer’s rationale, and runs a fresh agent on this operator’s own profile with no inherited turns; that agent is never the execution that wrote the brief or an edition, and it writes only review/response/. It scores correctness, pedagogy, interview value and language, and, where those stages ran, visual fidelity, code quality and executable proof. Approval requires every applicable score at or above 85 and no open error finding. A returned revision is not a stop: it is the fallback REVIEW_REVISION_REQUIRED, and the branch repairs exactly the artifacts the review’s findings name, by owning stage, then reopens the exchange for the next round. The ceiling is maxReviewRounds; when those rounds are spent and the review still returns a revision, REVIEW_ROUNDS_EXHAUSTED terminates, because a unit that has been rewritten to a reviewer’s satisfaction after unbounded rounds was rewritten by the reviewer. A unit that passes its own author’s review has not been reviewed, which is why shared executions, inherited turns, and producer rationale are three separate refusals.

Boundary

Context is read-only. The operator writes only response/ of its own branch: the brief, the receipt, the machine record of the build and the executable check, and the editions, image, prompt and tracks under response/artifacts/; the review agent writes only review/response/. It runs only the declared build and executable-check commands. It never writes an article before the brief is frozen or beyond what the brief published, claims a learning outcome, example, or visual claim the brief does not publish, reports a build or executable check that was not run or whose exit code was not read, changes the executable contract during a repair loop, lets the producer of an artifact perform, steer, or brief its own review, or approves a unit while any applicable score sits below the published minimum. It does not edit the curriculum, publish the unit to the served object, or claim any readiness beyond the checks it actually ran.

Context

Inputs

Requirements

Steps

Under the defaults the unit is one Vietnamese edition with no image, no track and no executable check, and its quality rests entirely on step 8. A resume begins again at step 1, reuses only unchanged fingerprinted observations, and consumes the exact delta; a revised brief arrives as a new fingerprint, because the same fingerprint cannot yield a different answer. The receipt claims no publication, no learner outcome in production, and no acceptance beyond the checks actually run.

Outputs

Stops

Next

Source: operators/content-generate/operator.md.