Exam Help Agent Skill

exam-help — quick-reference card

Purpose

Render one read-only card covering the validated workflow, three modes, four time tiers, lightweight|full, chat|visual, workspace truth/views, six quiz types, provenance, and subskill routing.

Activation

Use only when the student asks how the suite works, which modes/types exist, or what workspace files mean.

Inputs

No files, arguments, or state reads. The caller supplies the selected language.

Workflow

• Emit the matching static language-pack card: persisted zh, en, or bilingual, otherwise an explicit ad-hoc request.
• Read no wiki, bank, plan, progress, or workspace file; run no script/subskill.
• Stop without tutoring, quizzing, ingesting, grading, or initialization.

The card must say:

• processing_mode=lightweight is the default/recommended startup choice. It

inventories names, then visually processes only the current-phase PDF pages or

definitely single-frame PNG/JPEG/BMP (maximum eight primary pages and one active

batch). Overview contact sheets group at most four pages at roughly 768 px/tile;

page/prompt/dependency detail and target-answer-only solution calls use

source-qualified locations and canonical PNG evidence under .lightweight/assets/.

Bind exact external answer pages with register-answer-dependency; dependency

pages are locator/detail only. Figure prompt/answer crops stay distinct. An

unfinished batch can close only through receipt-backed abandon --reason; taught

progress cannot be abandoned, while replace-taught --reason preserves it as

superseded history and plans an exact-slice successor. It keeps the page-batch /

progress state machines and creates no Study Guide/PDF. processing_mode=full is

explicit opt-in and opens the complete ingestion/review route. Input-token savings

never shorten the teaching explanation.

• artifact_mode is independent from processing intensity and never inferred from

subscription. Missing/legacy/unknown means chat. In full mode, chat stops

without PDF while standing visual additionally requires typed Guide, render,

receipt hashes, every-page QA, and artifact_ready=ready. A one-shot artifact

leaves stored state unchanged. In lightweight, a saved visual preference is

dormant and effective output remains chat; it never builds a Study Guide.

• answer_explanation_mode is also independent. Its stored-schema fallback is

ordinary, and every full-Guide item still gets a detailed beginner-first

explanation. At full-v2 Guide entry, a verified host-native child with a fresh

independent context plus exact single-item input/tool restrictions makes

isolated the default unless the learner opted out; tell the learner once about

extra host quota/time, with no second API key or external upload. Missing or

incomplete capability stays ordinary and is stated honestly. A separately

billed external Provider is explicit-request-only and retains no-upload planning,

current pricing/privacy disclosure, and exact-plan upload consent. A model name,

subscription, key, full, or visual alone proves neither route.

• ingest_course.py is the full-mode build entry; exit 10 is process success with

blocked readiness, and teaching remains forbidden. .ingest/ is full-mode

build/review truth, .lightweight/session.json is on-demand page-batch truth,

study_state.json is progress truth, and Markdown is a generated view.

• Lightweight completion uses current-phase taught-batch/notebook events and skips

typed Guide/full-build evidence; superseded predecessors/events remain history but

leave the current denominator. verified needs revision-bound checkpoints from an

unchanged bank that pre-existed lightweight initialization; startup captures only

its immutable stat baseline. Routine health checks use metadata/physical identity,

not stream hashes; exact hashes occur at transitions/completion or explicit

status --verify-live. Older taught history is unchecked_historical until that

phase is resumed.

• MinerU, Docling, and LangGraph are named-request-only and remote/cloud-only. Never

download, install, probe, import, or execute them locally.

• Provenance is exact: 🟢 来自资料 / 🟡 AI补充，可能与你老师讲的不完全一致 / ⚠️ AI生成答案，非老师/教材提供.

Output Contract

Output exactly one help card and mutate nothing. Student prose is English by default, Simplified Chinese for a Chinese opening, or explicit bilingual composition.

Language packs

• 中文 → ../../locales/zh/skills/exam-help.md
• English → ../../locales/en/skills/exam-help.md
• 双语 → compose both blockwise, zh then > EN:, under docs/language-policy.md

Display aliases are normalized to zh, en, or bilingual.

Boundaries

This card is read-only. Route actual review to exam-cram.