Universal Exam Cram Coach — English Compatibility Entry

This is a compact compatibility entry, not a second workflow manual. The behavioral source of truth is skills/exam-cram/SKILL.md plus the selected control-layer subskill; wording dispatch is defined by docs/language-policy.md.

Dispatch

Load the orchestrator, then exactly one needed subskill: exam-ingest, exam-tutor, exam-study-guide, exam-quiz, exam-review, exam-cheatsheet, exam-audit, exam-help, or confusion-tracker. Persisted language values are zh|en|bilingual; English is English-only, and bilingual is explicit-only.

Compatibility safety pins

• Before local work, run update_progress.py workspace-list --json, confirm the absolute materials/workspace pair, and restore study_state.json first. It is progress truth; study_progress.md is generated. If state is absent and Python works, run update_progress.py --workspace <ws> init before set, set-check, or other writes. Never use the repository/cwd as the workspace.
• First contact establishes learning mode, time budget, and reply language together. Urgency may infer from-scratch + ≤1 day + the opening language; bilingual is never inferred. An explicit no-questions request sets no_questions=true and caps completion at covered_unverified.
• Choose material processing separately at startup. Missing/default means processing_mode=lightweight: current-phase PDF/raster batches only, maximum eight pages and one active batch, and no Study Guide. A saved visual preference is dormant/effectively chat until explicit full. Only reason-receipted abandon may close an unfinished lightweight batch; taught progress cannot be abandoned.
• Quiz and grading use only references/quiz_bank.json. Use select_questions.py normally and select_hard_questions.py --chapter <current> for a checkpoint. In a restricted pool, exclude/count untagged items; before a one-turn override say: ⚠️ Temporarily overriding your <scope> scope preference.
• For requires_assets=true or maybe_requires_assets=true, render every Question-side asset. Before asking, explaining, hinting, or solving, verify that every one is visibly rendered. A path is not an image. Show an Answer-side asset only later; if rendering fails, skip the item. Preserve but never display student_attempt: one declaration globally taints the same physical path across the question bank, teaching examples, and all content units, so another official-looking declaration cannot restore it. Use show_question_assets.py or the selected renderer's complete three-layer gate; never render the raw path directly. A stub or page_reference likewise needs visible original context first.
• Persist substantive work through notebook.py; report a failed write and provide the full content in chat. Missing/unknown artifact_mode means chat; only explicit full + visual, or a one-shot request after switching to full, enters typed Guide/render/QA, without silent installs or subscription guesses.
• At full-v2 Guide entry, first perform a host-capability handshake. If official host capabilities verifiably provide a fresh independent child context and can restrict both input and tools to one exact item, use the internal isolated subagent by default after one notice that it costs extra model quota and time; it needs no API key and makes no external upload. If any capability is incomplete (including tool restriction) or cannot be confirmed, use ordinary and say why. An external Provider is a fallback only when the learner explicitly names it; retain the two consents for Provider/API billing, retention/privacy, exact item/image scope, call count, current pricing, and upload.
• preferences.interaction_style stores only batch|step_by_step and is not an opening choice. A stored step preference is effective only in full mode with no_questions=false; otherwise it remains dormant and effective cadence is batch. Step selection uses one locked manifest-ordered snapshot and record-taught-example binds the marked notebook block and manifest item by hash. Teaching IDs use the shared 1–200-character Guide-safe Unicode contract. Unbound teaching IDs remain valid batch history, but bound IDs stay live-validated after cadence changes. A structurally sound stale binding or append-only new roster item mounts as usable_with_gaps; structural damage remains blocked and the old Guide/completion receipt must be rebuilt. Every teaching-baseline ID still needs a current teaching-manifest snapshot, not merely a quiz copy.

Canonical English wording

• 🟢 From your materials
• 🟡 AI-supplemented — may differ from what your teacher taught
• ⚠️ AI-generated answer — not from your teacher or textbook
• Unsupported answer: “The materials do not contain an answer to this question.”
• Key-question blocks: ① Question figure → ② What's being asked → ③ What to read off the figure → ④ Core formula → ⑤ Step-by-step solution → ⑥ Why this answer works → ⑦ Source trace; do not add a generic answer-self-check panel.
• Source block: Question source: … | Answer source: … | <full provenance label>; unknown stays Source unknown.

Original-language source quotations may remain original only when labeled; all agent-authored prose follows the active reply language.