dsh-ai-teacher

Readme

# dsh-ai-teacher

AI老师 (ai-teacher) agent preset for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness). Installing this plugin syncs the `ai-teacher` preset into `~/.dsh/.agent-presets/ai-teacher`, making **AI老师** selectable in the new-session preset picker.

## What AI老师 is

A full-capability teaching agent that **masters PDF documents and explains them as objective truth**:

- **Objective-truth presenter** — reads a PDF (via the host's built-in `read_pdf`), digests the content, and explains it as "this is how it is", never as "the book says…"
- **Every sentence is grounded** — statements are backed by the source, jointly-derived steps, or accepted definitions; when unsure, it checks first instead of making things up
- **One step at a time** — one formula / one equality / one logical step per message, with a confirmation question, then stops
- **Socratic guidance** — guess-before-prove, reverse engineering, meta-cognitive follow-ups
- **Motivation, invisible** — keeps learning momentum (task chunking, immediate feedback, growth visibility) without ever showing game-like UI

It also has the full `standard` preset capability (shell, filesystem, search, skills, plan mode, goals, subagents, workflows).

## Install

```sh
dsh plugin --profile web add dsh-ai-teacher
```

Or from a local checkout:

```sh
dsh plugin --profile web add link:D:/VSCode_/Project/dsh/dsh-ai-teacher
```

The preset appears in the new-session preset picker as **AI老师**.

## Prerequisites

- None beyond dsh itself. PDF reading uses the host's built-in `read_pdf` tool.

## How it works

On host startup the plugin syncs two trees into the harness home:

- `presets/ai-teacher` → `~/.dsh/.agent-presets/ai-teacher` — the preset composition
- `skills/*` → `~/.dsh/skills/` — the teaching methods as dsh skills

The preset's `agent.cordis.yml` is an AGENT-PLANE composition adapted from the builtin `standard` preset. Its persona is deliberately lean (~900 chars): identity + always-on ground rules. Specific teaching methods live in skills that the agent loads on demand:

| Skill | Loaded when |
|-------|-------------|
| `teach-new-concept` | teaching a new concept — guess-before-prove, definition-motive-justification |
| `derive-step-by-step` | doing a derivation/proof — one step per message, motive + justification each step |
| `answer-confusion` | student is confused or asks — locate the stuck point, SOCRATES guidance |
| `frankel-study` | studying Frankel's *The Geometry of Physics* — book-specific notation & structure |

The persona lists these skills with their trigger conditions, so loading is reliable while keeping the always-on prompt small.

## License

MIT