mp-ai-module-controller/README.md

# mp-ai-module-controller

Rust controller for starting a local `llama.cpp` server, asking it to map natural-language requests to known local actions, and dispatching exact-match Rust scripts. It supports both verbose action JSON and compact dictionary-coded outputs for hypersmall local models.

## Setup

```sh
cp .env.example .env
# edit LLAMA_MODEL_PATH in .env
cargo build
cargo run -- dictionary generate
git config core.hooksPath .githooks
```

The controller expects `llama-server` from `llama.cpp` to be installed. On this machine the planned binary is `/opt/homebrew/bin/llama-server`; if that path is not present, the controller falls back to `llama-server` on `PATH`.

## Commands

```sh
cargo run -- dictionary generate
cargo run -- serve
cargo run -- query "log hello from codex"
cargo run -- run logger --payload '{"message":"hello from direct dispatch"}'
cargo run -- run calculate --payload '{"expression":"2 + 2 * 3"}'
cargo run -- run to_slug --payload '{"text":"Hello, World!"}'
cargo run -- run write_note --payload '{"filename":"hello.txt","content":"hello"}'
scripts/precommit-check.sh
python3 scripts/generate_training_data.py --dry-run
scripts/deploy_training_data.sh
```

`serve` starts:

```sh
llama-server --model "$LLAMA_MODEL_PATH" --host "$LLAMA_HOST" --port "$LLAMA_PORT"
```

Optional environment variables:

- `LLAMA_CONTEXT_SIZE`: appended as `--ctx-size`.
- `LLAMA_EXTRA_ARGS`: shell-split and appended to `llama-server`.
- `CONTROLLER_LOG_LEVEL`: tracing log level, default `info`.

## Dictionary

Run `cargo run -- dictionary generate` to create:

- `dictionary/actions.jsonl`: canonical training dictionary, one action record per line.
- `dictionary/actions.index.json`: compact runtime lookup index.
- `dictionary/model.codebook.json`: static-plus-adaptive compact codebook.
- `dictionary/model.examples.jsonl`: hand-authored examples/templates for future synthetic-data tooling.

`dictionary/static-base.json` is the tracked static base dictionary. Generated metadata includes `dictionary_version`, `registry_checksum`, `static_base_checksum`, and `codebook_checksum`.

The runtime only dispatches exact `action_id` matches from `actions.index.json`. Aliases and examples are training hints, not executable ids. Compact model outputs must match the generated codebook version and checksum before expansion.

## LLM Response Contract

The local model must return compact JSON:

```json
{"action_id":"logger","payload":{"message":"hello"}}
```

It may also return compact codebook JSON:

```json
{"v":"1.0.0","c":"<codebook_checksum>","a":"A001","p":{"F001":"hello"}}
```

Malformed JSON, missing action codes, stale versions/checksums, aliases, and unknown action ids are rejected without running anything.

## Initial Script

The first registered script was `logger`. It writes JSONL records to `logs/controller-actions.jsonl`. Dispatch is detached: after the controller spawns the script process, it does not monitor completion.

## Action Suites

Math & Conversion:

- `calculate`: `{ "expression": "2 + 2 * 3" }`
- `convert_units`: `{ "value": 100, "from": "km", "to": "miles" }`
- `generate_uuid`: `{}`
- `random_number`: `{ "min": 1, "max": 100 }`

Text Utilities:

- `count_words`: `{ "text": "..." }`
- `to_slug`: `{ "text": "..." }`
- `hash_string`: `{ "text": "...", "algorithm": "sha256|md5|sha1" }`
- `truncate_text`: `{ "text": "...", "max_chars": 100 }`

File Operations:

- `write_note`: `{ "filename": "...", "content": "..." }`
- `append_to_file`: `{ "filename": "...", "content": "..." }`
- `read_file`: `{ "filename": "..." }`
- `delete_file`: `{ "filename": "..." }`

File operations are confined to project-local `notes/` and reject absolute paths or path traversal.

## CI And Hooks

The project includes:

- `.codex/environments/environment.toml` for Codex setup and run actions.
- `.gitea/workflows/ci.yml` for Gitea Actions.
- `.woodpecker.yml` for Woodpecker.
- `.githooks/pre-commit` for local pre-commit checks.
- `scripts/precommit-check.sh` as the shared validation entrypoint.

Enable the hook after cloning:

```sh
git config core.hooksPath .githooks
```

## SDG Bridge

`scripts/generate_training_data.py` generates instruction-format seed data from current dictionary artifacts. The default export root is `/Volumes/Zoe/custom-local-llm`.

Dry run:

```sh
python3 scripts/generate_training_data.py --dry-run
```

Local dataset generation:

```sh
python3 scripts/generate_training_data.py \
  --dictionary-dir dictionary/ \
  --output-dir /Volumes/Zoe/custom-local-llm/training-data/ \
  --manifest /Volumes/Zoe/custom-local-llm/manifest.llm.json
```

Optional Mimir/Data Designer expansion is enabled only when `--sdg-host`, `--sdg-user`, and `--sdg-key` are supplied. The bridge does not start or restart Mimir inference services.

If remote expansion fails, the timestamped local seed dataset remains in place, `SDG_ERROR.txt` is written in that dataset directory, and the manifest entry is marked `generated_local_remote_failed`.

Woodpecker deploys a new training-data export to Zoe on `main` pushes by running:

```sh
scripts/deploy_training_data.sh
```