cancel_material
ChatGPTStop a render job the learner no longer wants. Pass material_id, job_id and the job’s current job_revision from get_material as expected_revision. A queued job is cancelled at once; a running one returns cancel_requested and may still finish first, so report the state that comes back. A finished material is never deleted by this call.
change_program
ChatGPTChange the learner’s program, only when they ask: add wishes, drop or pause a goal, more or less of something, other days or times, “I can do this now”, start over. Pass what they asked for in said, word for word, and new skills in add_wishes, also word for word; Minerva’s server re-composes the program from their words, their kept lines and results (about a minute, status rebuilding; read get_learning_context view program to wait). Days already done stay as they were. Or pass only status paused or active to pause or resume without changing anything. target is {id, expected_revision} from get_learning_context; a stale revision returns REVISION_CONFLICT and changes nothing.
correct_practice
ChatGPTFix a recorded practice event the learner says is wrong, only on their request. correction.kind: replace_observation, replace_duration, withdraw_event (it stops counting), detach_media (unlink kept media from the event; the media stays until deleted) or rescore (re-run a result event’s checks, optionally against a newer lesson_ref after you fixed an answer key). Pass target_event_id, its expected_revision and a short reason; a stale revision returns REVISION_CONFLICT and changes nothing. The correction is kept beside the original, so the history stays readable. To erase data, use delete_learning_data.
delete_learning_data
ChatGPTDelete part of the learner’s saved learning, in two steps. First call action preview with an exact scope: kind program, sessions, materials, events, evidence or media (kept photos, video and audio with every preview), plus the include_sessions and include_owned_materials flags that kind requires. A preview deletes nothing; it returns every object that would go and a confirmation_url. Show the list and the link: the signed-in learner confirms on that page, and that confirmation performs the deletion. The delete action accepts only a token issued by that page; never invent one or ask for one. Covers saved learning only, not the account or ChatGPT tasks.
get_learning_context
ChatGPTRead the signed-in learner’s saved learning. Call it first in every conversation. view: today (the program in short, today’s step with its drawn practice session_id when ready, the daily task, progress with recent_notes and latest_result), program (the whole program: their wishes word for word, goals and stages, what the goals share, the days ahead, next ideas; while the server composes it, this view waits about 25 seconds), sessions, progress, media (photos, video and audio the learner kept, as previews and numbers; media_id reads one), recipes (lesson skeletons for your own lessons), library (rendered materials) or export (all saved data; narrow it with ID or date filters). recent_notes: up to five latest notes, newest first, exactly as stored; quote only text_source learner_verbatim, or a session_end note without text_source. Reads only.
get_material
ChatGPTRead one of the learner’s materials and the state of its render job (job_state). Use it to follow a job from prepare_material until the material is ready, then use its revision in a lesson or a media pointer. Starts, retries and charges nothing; the rendered material comes back only once it exists.
open_practice
ChatGPTOpen the learner’s practice, drawn by Minerva’s server for their step: today’s step by default, another day’s with step_id, or a one-off little skill they ask for right now with ask (their words, word for word; it is not added to the program). The server draws the lesson — pictures, the thing to do, what to look for — when it is not ready yet, which can take up to a minute; then it opens in the player with a text version of the whole lesson. If drawing is still running after about 25 seconds this returns status preparing: call it again. Steps are usually drawn ahead, so the daily task can simply open them. An optional request_id makes a retry of an ask return the first lesson.
open_session
ChatGPTOpen a saved practice in the interactive player, with a readable text version of the whole lesson for any screen. Use it to reopen a practice that is already drawn — today’s step when get_learning_context view today gives its session_id (the daily task does exactly this), or any earlier one. resume: last (saved place), start or furthest; block_id jumps to one block. Returns the blocks exactly as saved, this session’s progress, recent_notes from every practice of the same goal, and latest_results: the newest result per block with its arithmetic facts. Judging technique, pictures and free text is yours. Pointers to the learner’s attempts resolve here; an unresolved one shows its alt. Opening starts no playback, microphone or saving; only the learner’s taps do. Reads only.
prepare_material
ChatGPTRender one learning material on this server from a bounded description, only when the practice needs it or the learner asks. input.kind: audio (WAV from ABC notation, a pulse of clicks or drum voices, or MIDI notes), image (one diagram, grid, plot, fretboard or keyboard primitive as SVG and PNG), sheet (a Field Journal page of up to eight cards or steps with Markdown and LaTeX, optional pictures including the learner’s kept previews), video (captioned MP4 from pictures, optional own WAV) or authored_asset (original_four_notes or one_small_practice, revision 1). retention: library, or session_only with session_id and expires_at within 30 days. Returns a job; follow it with get_material. No speech, copyrighted songs, realistic video, charges or outside services.
record_practice
ChatGPTAppend one practice event to a saved session, only after the learner chose to keep it. event.kind: attempt (their line in observation.text, optional duration_seconds, block_id and kept media_ids), result (what a block’s interaction returned: taps, an answer, measurements, cards, moves, notes, a pitch reading; omit block_id for a number said in chat), playback or session_end. A result returns facts: arithmetic and exact-token checks only (taps and numbers within tolerance, choice ids, chess moves, MIDI notes); text and formulas come back unchecked for you to judge. Set text_source to learner_verbatim only for their exact words. Fix an event with correct_practice. An optional client_event_id dedupes a retry. Nothing awards mastery. A session_end with disposition finished on a program step marks that day done, and the next practice gets drawn ahead.
revise_session
ChatGPTChange a saved lesson, only when the learner asks. action revise saves your edited session as a new revision; archive retires the lesson. Pass target {id, expected_revision}; a stale revision returns REVISION_CONFLICT and changes nothing, so read again before retrying. Practice already recorded stays bound to the revision it was made on and is never rewritten; open the new revision for new attempts. Nothing is deleted here; to remove data, use delete_learning_data.
save_media
ChatGPTKeep photos, video or audio the learner chose to share for learning. Call it when they attach a file in chat or pick one in the player and want it used in practice. role is your call: goal (they want it like this), source (teach me from this: a recipe, a textbook page, a score, a pattern) or attempt (how theirs came out). Each file is read on this server and the original is deleted right after; only small previews and numbers stay (size, video frames, duration, pitch, tempo, pauses). Never infer identity, emotion, age or health from faces or voices. Media never goes to any model or outside service. Repeating a file_id returns the same item. A learner without a camera or microphone can describe the attempt in words instead, kept with record_practice.
save_session
ChatGPTKeep a lesson you composed in chat as a practice the learner can reopen. Supplying step_id replaces that program step’s current practice assignment; the previous lesson and its recorded progress are preserved and remain available in saved sessions. Without step_id, program assignments are unchanged. Optional goal links the new lesson to a goal. A lesson contains 1–24 practice blocks: up to six show primitives, at most one interaction in do, and look_for in your words; every non-text show primitive needs alt. Usually use open_practice, which draws the day’s practice on the server. Call this only when the learner asks to keep your lesson, including any step reassignment. Optional request_id dedupes an identical retry. The microphone opens only after the learner taps; no audio is saved.
set_daily_task
ChatGPTRecord the learner’s daily ChatGPT task for their program. Minerva cannot create, run or cancel ChatGPT tasks and sends no reminders: you create the task with ChatGPT’s own scheduler when the learner says yes (use daily_task_suggestion from start_program), then record what happened. action: created (the scheduler made it; pass its time, days and title, and its ID if it returned one), changed, removed, or unavailable (no scheduler here; the learner comes back by asking for today’s step).
start_program
ChatGPTStart the learner’s program from the little skills they want to learn. Pass every wish in their own words, exactly as they said it, one per item — never reworded, merged or translated — and what they said about their time, place, things and level in said, also word for word. Call it once they have named what they want to learn and when in their day they can practise; ask nothing more. Minerva’s server then composes the program — goals, stages, what the goals share, one small step per practice day and ideas for later — in about a minute; this returns at once with status building. While it builds, offer the daily task (daily_task_suggestion), then read get_learning_context view program, which waits for it. A learner has one program: when one exists this fails with PROGRAM_EXISTS; use change_program. An optional request_id makes a retry return the first result.