---
name: mindtrain
description: Configure a private MindTrain instance, query its knowledge catalog, create user-approved training domains and knowledge points from AI dialogue or local reference libraries, and run persistent conversational training through Trainer MCP. Use when the user configures MindTrain, asks what domains or topics exist, searches the catalog, wants to create or extend a training domain, selects a local document directory, confirms a domain draft, starts or continues training, generates a question, answers or rejects a question, asks follow-ups, revises a saved question, ends a session, or inspects learning progress.
---

# MindTrain

Use the bundled MindTrain bridge as the only application data interface. Do not read or write repository question, candidate, session, attempt, or mastery files.

## Configure first use

1. Call `get_mindtrain_configuration` before the first training action in a task.
2. When `configured` is false, explain that MindTrain needs the private Trainer MCP URL and the deployment's single-user Bootstrap Token.
3. Ask for the full HTTPS MCP URL. Ask for the Token only when the user is willing to provide it in the current private conversation; otherwise direct them to run `python3 scripts/mindtrain_mcp_bridge.py --configure` from the installed plugin directory.
4. Call `configure_mindtrain_instance` with the URL and Token. Never repeat, display, summarize, or persist the Token anywhere except through that configuration tool.
5. Inspect the returned `compatibility`. Continue only when its status is `compatible` or `compatible_version_difference`. For a compatible version difference, briefly recommend synchronizing Plugin and server versions without blocking the requested work. For `incompatible_or_unavailable` or a version/contract error, stop remote operations, show the sanitized upgrade guidance, and do not retry training tools until the user updates the indicated component and opens a new task.

The bridge saves configuration outside the repository with user-only file permissions. Never put the private URL or Token into Skill files, Git configuration, examples, or commits.

## Query or build the knowledge catalog

Read [knowledge-catalog.md](references/knowledge-catalog.md) when the user asks what can be learned, searches for a knowledge point, or wants to create or extend a training domain. Query the current catalog before proposing a domain so stable IDs do not collide and existing domains are extended deliberately.

For a dialogue-created domain, clarify the learning goal, audience, scope, emphasis, and difficulty only as needed. Generate one domain target with any number of root topics. Call `preview_training_domain` with `originType: ai_dialogue`, show the complete tree, diff, conflicts, and the explicit warning when no references are bound. Do not invent Source records for AI output.

## Build from local references

When the user names a local directory as a reference library, read [reference-library.md](references/reference-library.md) and follow its configure, sync, organize, preview, and confirmation workflow. The user chooses what to learn; Codex organizes the selected material into a proposed training domain, topic hierarchy, and relations. Local files, extracted text, absolute paths, and indexes must remain on the Codex host. Core receives only source metadata, hashes, the approved training structure, and generated questions.

Never call `confirm_training_domain` until the user has seen the complete training-domain preview and explicitly confirmed saving it once. Reuse the exact `proposalHash` returned by preview. A conflict requires a revised proposal and a new preview; never overwrite an existing topic implicitly. When speaking to the user, say “保存并启用训练领域”, not “import/apply a knowledge catalog.”

## Run training

1. Call `list_knowledge_domains` before starting. Resolve a user-named domain and pass its exact ID. When exactly one domain exists, use it automatically. When none exist, guide the user through domain creation. When multiple exist and the user did not choose one, ask them to choose; never select the first domain silently.
2. Call `create_training_session` with the resolved `domainId`; Core always uses the application-configured `questionCount`. Use provider ID `weighted`, displayed to users as `加权调度`. One session trains exactly one domain.
3. Call `get_next_assignment`.
4. When it returns `assignment`, show only the stem and A-D options. Use exactly: `请回复选项字母，可用逗号分隔。`
5. Treat a clear option selection as a formal answer and call `submit_choice_answer` once.
6. Treat concept questions, uncertainty, challenges, and hint requests as interactions. Call `record_interaction`, answer conversationally, and keep the assignment pending.
7. After successful grading, render: result, correct answer, conclusion, option analysis, mechanism, pitfalls, version notes, related topics, and sources.
8. Offer exactly: `下一题`, `深入追问`, `结束总结`.
9. Call `finish_training_session` when the target is complete or the user ends early.

Never infer the correct answer before `submit_choice_answer` returns it. Invalid answer input does not consume the question.

## Revise a flawed saved question

When the user reports that a displayed question is unclear, incorrect, outdated, or poorly sourced, record the feedback with `record_interaction` and discuss the issue first. Never change the question merely because the user challenged it.

Call `get_question_revision_context` before revising so the complete current version and pending Assignment are authoritative. Call `revise_saved_question` only after the user explicitly asks to update the question bank and the intended correction is clear. The phrase `不合适，修改后再出` (or an unambiguous equivalent) is explicit authorization for one revision of the displayed active question and does not require a second confirmation. Use the question ID and version from the assignment, include its assignment ID as `sourceAssignmentId`, set `applyToPendingAssignment: true`, and submit only changed fields. Preserve option IDs and the correct option set unless authoritative sources support a scoring correction. Explain the revision and report the new version returned by Core.

On `question_version_conflict`, do not retry with a guessed version. Call `get_question_revision_context` with the pending Assignment to retrieve authoritative current content; if the Assignment no longer matches, explain that it must be refreshed rather than silently replacing it.

## Generate a missing question

When `get_next_assignment` returns `generation_required`, read [candidate-policy.md](references/candidate-policy.md), follow the returned `generationProfile` exactly, and generate one compliant question. Do not choose a different type, difficulty, or primary topic. If the profile references a local library, search it and read only the smallest relevant passages. When local evidence is insufficient, ask the user each time before using external authoritative sources. Call `create_candidate_question`, then call `get_next_assignment` again only after the candidate is accepted.

Before it is answered, the generated candidate is temporary and usable only by its owning session. A successful answer activates it for ordinary cross-session scheduling.

## Reject a generated question

When the user clearly rejects the currently displayed AI-generated question before answering, call `reject_generated_question` with its `assignmentId`. Do not submit an answer and do not record the rejected question as an interaction. After Core confirms physical deletion, call `get_next_assignment` and present a materially different replacement. Never call this tool for an imported, previously answered, review, or ordinary new question.

## Follow up deeply

For an explanatory follow-up, call `record_interaction` and answer in the current conversation.

For a deeper training question, generate another four-option question on the same topic and call `create_candidate_question` with `attemptType: follow_up` and the graded parent attempt ID. Present the returned follow-up assignment. It must increment only the follow-up count, never the main target. Do not count conversational explanations as attempts.

## Report learning progress

When the user asks for a learning report, call `get_learning_report`. Lead with `todayCompletedMainQuestions` / `dailyTarget`, `todayAccuracy`, `todayReviewCompleted`, `todayNewItemsIntroduced`, the due backlog, and `schedulerStatus`; these are the live daily snapshot. Treat `newBudget` as a per-session policy limit and never present `newItemAllowance` as a daily remaining quota. Then present `weakTopics` as 待加强 and `strongTopics` as 擅长, including their domain, topic path, mastery score, and correct/wrong counts. Report `insufficientEvidenceTopicCount` separately as 数据积累中; never label those topics weak or strong before they meet the server-defined sample threshold.

## Recover safely

- On a Plugin/server version or contract mismatch, do not bypass the check. Ask the user to upgrade or reinstall the MindTrain Plugin, upgrade the deployed services when indicated, and open a new Codex task so the updated Plugin is loaded.
- On `configuration_required`, return to the first-use configuration flow.
- On `no_training_domains`, guide the user through creating and confirming a training domain.
- On `training_domain_selection_required`, list the available domains and ask the user to choose one.
- On `training_domain_not_found`, refresh the domain list instead of guessing a replacement.
- On `session_domain_invalid`, explain that the session's domain no longer exists and start a new session only after the user chooses a valid domain.
- On `answer_unparseable`, ask for option letters again without revealing answer count.
- On `no_available_items`, explain the scheduler reason and offer to finish or inspect backlog.
- On Core or MCP unavailability, retain the current visible question in conversation and retry the tool; do not fabricate persistence success.
- On a local parser warning, report the affected file and continue with usable documents; never claim an unreadable document was indexed.
- Use a new idempotency key for a new user action and reuse it only when retrying that same action.

Read [tool-contract.md](references/tool-contract.md) when tool inputs, statuses, or retry behavior are unclear.