lesson-language-phone-practice

Lesson Language Phone Practice

Create one target-language phone role-play from a completed lesson, aiming for about 30 seconds and ending by 60 seconds. Default to preview only. Do not place a live call unless learner consent, destination, exact call plan, provider readiness, and one-call approval are all explicit in the current workflow.

Read references/safety.md before preparing any live path. Read references/examples.md when shaping a preview or interpreting a result.

Required inputs

Require all of these from an authoritative lesson or trainer-owned record:

• completed lesson/activity reference;
• lesson aim, stage aim, Can-do objective, and grammar target;
• one scenario; the lesson-plan phrases themselves are the narrow communication goals and assessment targets;
• two or three exact taught phrases selected from the lesson record;
• optional learner support or caller wording;
• trainer note or observed practice need;
• learner consent covering an AI call and transcript/result handling;
• one authorized E.164 destination, its supported region, and locale;
• a configured phone provider whose credentials stay in the host secret store.

Do not invent taught phrases, use unrelated learner data, scrape a number, infer a country code, or accept a browser/client result as evidence.

Workflow

1. Freeze a preview containing the lesson reference, aims, Can-do, grammar, scenario, exact lesson-plan phrases, support, roughly 30-second target, 60-second maximum, one opportunity per phrase, and at most one contextual repair. Never ask the teacher to invent or replace assessment phrases in the phone task.
2. Show the complete generated brief and success conditions to the trainer. Permit edits to the scenario, fixed facts and learner support, but never to provenance, lesson-plan phrases, call aims or phrase-derived success evidence. Require explicit trainer approval before assignment.
3. Validate the destination against canonical E.164: a plus sign followed by a nonzero country code and 8–15 digits total. Validate region and locale against the provider's current supported list. Mask the destination in output, for example +1415•••0100.
4. Confirm the learner explicitly consented to this purpose, destination, AI caller, timing, and transcript/result retention. One lesson or earlier call does not imply consent now.
5. Produce the dry-run preview. State clearly: No call has been placed.
6. Before live dispatch, show the exact masked recipient, purpose, phrases, provider, timing, data handling, one-call rule, and disable behavior. Wait for fresh approval for exactly one call.
7. Reserve a durable operation before calling. Derive one stable idempotency key from the lesson-practice authorization, not a click, timestamp, retry counter, or random value. Persist the provider call ID immediately when returned.
8. Submit once. A timeout, restart, ambiguous response, failure, voicemail, wrong person, or no answer never authorizes a new key or redial. Recover only the same operation through documented idempotent replay or the stored call ID.
9. Poll the stored call ID to a documented terminal state. Do not treat provider task completion as proof that the language goal was demonstrated.
10. Reconcile structured output against speaker-labelled transcript turns. Only learner speech may support phrase use or appropriateness. Unknown speakers, missing transcript, unsupported phrases, out-of-window duration, or ambiguous endpoint must return insufficient evidence.
11. Return the bounded result and masked operation summary. Do not turn it into a generic score, proficiency level, mastery claim, credential, reward, readiness decision, or progression change.