Codex/topic discovery lora

.agents/skills/course-learning-coach/SKILL.md

@@ -0,0 +1,116 @@
+---
+name: course-learning-coach
+description: Guides daedalus course-learning for courses, tutorials, official learning paths, online curricula, chapter-based docs, and lesson labs. Use when the user learns a course such as Hugging Face Course, follows an official tutorial/notebook, asks what to study next in a curriculum, or says they copied course code but do not understand the underlying mechanism.
+---
+
+# Course Learning Coach
+
+Use this skill when the user is learning a course, tutorial, official learning path, online curriculum, or chapter-based technical material through daedalus.
+
+## Core Rule
+
+Course learning turns course material into an observable learning path: learning need, syllabus route, concept mechanism, lesson lab, user observation, practice transfer, review, and archive.
+
+It keeps daedalus's filesystem-first state, WIP discipline, human-owned notes, first principles, and critical lens. Its coaching center is course mastery: the user should be able to explain, reproduce, modify, and transfer the course concepts.
+
+The course-learning path is:
+
+```text
+learning need -> syllabus route -> concept mechanism -> smallest lesson lab
+-> user observation -> practice transfer -> capstone/review -> archive
+```
+
+## Startup
+
+1. Resolve active context from `workspaces/.daedalus/current.toml`, `workspaces/current-project`, or `workspaces/current-topic`.
+2. Confirm the project state uses `task.kind = "course-learning"` and `project.source_kind = "course"`.
+3. If the metadata does not match course-learning, pause course-learning routing and repair the concrete files in place with user approval.
+4. Read project `shared/syllabus-map.md`, `shared/course-progress.md`, `shared/concept-map.md`, active topic `.daedalus/state.md`, `.daedalus/outcome-map.md`, and `.daedalus/todo.md`.
+5. Route to exactly one current course-learning stage.
+
+## Stage Routing
+
+### 01 Need Aligner
+
+Load:
+
+- `system/prompts/course/01-need-aligner.md`
+
+Use when the learning goal, course role, output proof, or non-goals are unclear.
+
+### 02 Syllabus Mapper
+
+Load:
+
+- `system/prompts/course/02-syllabus-mapper.md`
+
+Use when choosing chapters, mapping a course outline, or deciding what to skip/delay.
+
+### 03 Concept Roadmap
+
+Load:
+
+- `system/prompts/course/03-concept-roadmap.md`
+
+Use when chapter titles need to become mechanism questions and checkpoints.
+
+### 04 Lesson Lab
+
+Load:
+
+- `system/prompts/course/04-lesson-lab.md`
+
+Use when the user follows course code, runs notebooks, inspects datasets/tokenizers/models/loss, or asks why the code is written that way.
+
+### 05 Mechanism Deep Dive
+
+Load:
+
+- `system/prompts/course/05-mechanism-deep-dive.md`
+
+Use only when lesson output, official docs, and small experiments are not enough to explain a hidden mechanism.
+
+### 06 Practice Transfer
+
+Load:
+
+- `system/prompts/course/06-practice-transfer.md`
+
+Use when transferring a course concept to the user's real task.
+
+### 07 Capstone Lab
+
+Load:
+
+- `system/prompts/course/07-capstone-lab.md`
+
+Use when building the final course-driven mini project.
+
+### 08 Review Loop
+
+Load:
+
+- `system/prompts/course/08-review-loop.md`
+
+Use when checking whether the user can explain, reproduce, modify, or transfer concepts without the tutorial.
+
+### 09 Closeout Archive
+
+Load:
+
+- `system/prompts/course/09-closeout-archive.md`
+
+Use when closing a course topic and mining verified knowledge.
+
+## Coaching Rules
+
+- Advance at most one lesson, mechanism, transfer checkpoint, or review checkpoint per turn.
+- Ask 1-3 high-value questions when the user's intent or understanding is still fuzzy.
+- Write Agent-generated action plans to `guides/`, not `notes/`.
+- Write `notes/` only after user answers, runs code, observes output, or explains a mechanism.
+- When a course-learning answer explains a reusable mechanism, API control panel, parameter trade-off, speed/cost lever, or "how to derive the code" frame, update the active focused `guides/` artifact in the same turn unless the user asks to only discuss; stage `README.md` files should only receive index/recovery updates.
+- Before writing guides or notes, classify the artifact shape. Stage `README.md` files are indexes/timelines for restoring the learning scene. A reusable plan, concept, mechanism explanation, comparison, debugging pattern, user recap, or Agent failure mode must become a dedicated guide/note file and be linked from the stage README plus shared maps when relevant.
+- For lesson labs, always separate prediction, minimum code, observed output, explanation, and next transfer.
+- Treat warnings, shapes, losses, artifacts, and generated files as learning evidence, not terminal noise.
+- Keep historical directories after manual reorganization unless the user approves renaming them.
+- After changing lifecycle state, evidence, risks, or next action, synchronize `todo.md`, `outcome-map.md`, and relevant shared maps.

.agents/skills/repo-learning-coach/SKILL.md

@@ -123,6 +123,6 @@ For long-running work:
 - Tie every reading step to a future output artifact.
 - Do not let chat become the only learning record. If the user answers, corrects, or validates an important point, update the relevant `notes/` file in the same turn.
 - Preserve the distinction between user understanding, Agent calibration, and verified source-code conclusions.
-- Do not accept a code-reading note that only explains call chains. It must explain production constraints, failure handling, invariants, trade-offs, and migration limits.
+- Do not accept a code-reading note that only explains call chains. It must explain production constraints, failure handling, invariants, trade-offs, and transfer limits.
 - Do not mark the task complete until demo/business transfer/knowledge archival are addressed.
 - Do not archive unverified summaries as knowledge.

.claude/skills/repo-learning-coach/SKILL.md

@@ -20,6 +20,7 @@ Non-negotiables:
 - Keep stable projects under `workspaces/projects`; complete context lives in `workspaces/.daedalus/current.toml`, while top-level symlinks show only action entries: `current-project` / `current-topic` for active learning and `closeout-topic` for pending reflection.
 - Resolve active context from `workspaces/.daedalus/current.toml` or `workspaces/current-topic` before reading project/topic state. Resolve pending closeout from `pending_closeout_topic` or `workspaces/closeout-topic`. The repo root intentionally has no `.daedalus/state.toml`.
 - Keep at most one active topic and at most one pending closeout topic. When no active topic exists, do not treat a selected project as the current learning entry.
+- Treat `workspaces/discovery` as pre-topic exploration space for unclear motivations, competing materials, and need hypotheses; do not create active lifecycle state from discovery alone.
 - Prefer workspace artifacts over chat memory.
 - Output drives input: every reading, debugging, review, or implementation step must advance a final artifact or a blocked decision.
 - Ground progress in code, tests, runtime evidence, and git diff before trusting markdown maps.
@@ -80,6 +81,7 @@ Closeout progression must follow `system/prompts/common/closeout-flow.md`: once
 Load only the prompt(s) needed for the current action:
 
 - CLI lifecycle, render, validate, review, and deterministic knowledge checks: `system/prompts/repo/repo-learning-cli-contract.md`
+- Topic selection discovery before task cards or gatekeeping: `system/prompts/common/topic-discovery.md`
 - Human-owned notes, notes review, topic closeout retrospective, and knowledge-base gate: `system/prompts/common/human-owned-notes.md`
 - Resume or "where are we": `system/prompts/common/resume.md`
 - Checkpoint, cursor sync, pre-commit sync, post-commit orientation: `system/prompts/common/checkpoint-lifecycle.md`

.codex/plans/course-learning-system.md

@@ -0,0 +1,35 @@
+---
+name: course-learning-system
+overview: 追踪 daedalus course-learning 一等学习类型建设；详细方案以 docs/plan/15-course-learning-system.md 为准。
+status: completed
+todos:
+  - id: design-course-learning-plan
+    content: 编写 course-learning 一等学习类型正式计划。
+    status: completed
+  - id: implement-course-templates
+    content: 新增 course project/topic templates 与 state/source kind。
+    status: completed
+  - id: implement-course-cli
+    content: 新增 init course-learning、validate/render 支持。
+    status: completed
+  - id: add-course-prompts-skill
+    content: 新增 system/prompts/course 与 course-learning-coach skill。
+    status: completed
+  - id: organize-hf-course-project
+    content: 手动整理 hugging-face-llm-course，使其成为 course-learning project。
+    status: completed
+isProject: false
+---
+
+# Codex Tracking Card
+
+详细计划不在此处展开。本文件只用于 Codex `status` / `todos` 追踪。
+
+## Canonical Plan
+
+- [Course Learning 一等学习类型实现方案](../../docs/plan/15-course-learning-system.md)
+
+## Tracking Rule
+
+- `docs/plan/`：正式详细计划，作为 daedalus 项目的长期设计资料。
+- `.codex/plans/`：Codex 执行追踪卡，只保留 `status`、`todos` 和正式计划链接。

.codex/plans/discovery-driven-topic-selection.md

@@ -0,0 +1,37 @@
+---
+name: discovery-driven-topic-selection
+overview: 追踪 daedalus 选题发现能力改进和 LoRA Fine-tuning Feedback Loop Lab 启动计划；详细方案以 docs/plan 为准。
+status: completed
+todos:
+  - id: formalize-plan-location-rule
+    content: 确认正式详细计划放在 docs/plan，.codex/plans 只做带 status/todos 的 Codex 追踪卡并引用正式计划。
+    status: completed
+  - id: write-daedalus-improvement-plan
+    content: 编写 daedalus 澄清、挖掘、选题能力改进正式计划。
+    status: completed
+  - id: write-lora-topic-plan
+    content: 编写 LoRA Fine-tuning Feedback Loop Lab 新 topic 启动正式计划。
+    status: completed
+  - id: implement-topic-discovery
+    content: 按 docs/plan/13 实现 topic discovery prompt、artifact 和 clarify/gatekeeper 集成。
+    status: completed
+  - id: start-lora-topic
+    content: 按 docs/plan/14 使用 daedalus lifecycle 创建新 project/topic。
+    status: completed
+isProject: false
+---
+
+# Codex Tracking Card
+
+详细计划不在此处展开。本文件只用于 Codex `status` / `todos` 追踪。当前 daedalus 改进和新 topic 初始化均已完成。
+
+## Canonical Plans
+
+- [澄清、挖掘、选题能力改进方案](../../docs/plan/13-discovery-driven-topic-selection.md)
+- [LoRA Fine-tuning Feedback Loop Lab 启动计划](../../docs/plan/14-lora-finetuning-feedback-loop-topic.md)
+
+## Tracking Rule
+
+- `docs/plan/`：正式详细计划，作为 daedalus 项目的长期设计资料。
+- `.codex/plans/`：Codex 执行追踪卡，只保留 `status`、`todos` 和正式计划链接。
+- `workspaces/`：实际进入 daedalus lifecycle 后的 project/topic 状态。

.codex/plans/tui-learning-type-adapters.md

@@ -0,0 +1,35 @@
+---
+name: tui-learning-type-adapters
+overview: 追踪 daedalus TUI 对 repo-learning / course-learning 的解耦适配；详细方案以 docs/plan/16-tui-learning-type-adapters.md 为准。
+status: completed
+todos:
+  - id: write-tui-adapter-plan
+    content: 编写 TUI learning type adapter 正式计划。
+    status: completed
+  - id: implement-learning-type-adapter
+    content: 在 TUI app 层引入 learning-type adapter，并接入 repo/course。
+    status: completed
+  - id: update-tui-screens
+    content: 让 TUI screens 展示 learning type、source reference 和类型化进度文案。
+    status: completed
+  - id: add-course-tui-tests
+    content: 补充 course-learning TUI fixture 和 guide 查找测试。
+    status: completed
+  - id: write-changelog
+    content: 实现完成后写 changelog 总结。
+    status: completed
+isProject: false
+---
+
+# Codex Tracking Card
+
+详细计划不在此处展开。本文件只用于 Codex `status` / `todos` 追踪。
+
+## Canonical Plan
+
+- [TUI Learning Type Adapter 实现方案](../../docs/plan/16-tui-learning-type-adapters.md)
+
+## Tracking Rule
+
+- `docs/plan/`：正式详细计划，作为 daedalus 项目的长期设计资料。
+- `.codex/plans/`：Codex 执行追踪卡，只保留 `status`、`todos` 和正式计划链接。

.gitignore

@@ -43,5 +43,3 @@ coverage/
 tmp/
 temp/
 .cache/
-.daedalus-migration-backup/
-**/.daedalus-migration-backup/

.pre-commit-config.yaml

@@ -43,7 +43,7 @@ repos:
         pass_filenames: false
       - id: commit-msg-format
         name: commit msg format
-        description: Check commit message follows daedalus convention.
+        description: Check commit message follows topic/scope convention.
         entry: system/bin/check-commit-msg
         language: system
         stages: [commit-msg]

.vscode/extensions.json

@@ -0,0 +1,6 @@
+{
+  "recommendations": [
+    "ms-python.python",
+    "ms-python.vscode-python-envs"
+  ]
+}

AGENTS.md

@@ -0,0 +1 @@
+CLAUDE.md