Mindtrain

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 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 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.