aman/README.md

# aman
> Local amanuensis

Python X11 STT daemon that records audio, runs Whisper, applies local AI cleanup, and injects text.

## Target User

The canonical Aman user is a desktop professional who wants dictation and
rewriting features without learning Python tooling.

- End-user path: native OS package install.
- Developer path: Python/uv workflows.

Persona details and distribution policy are documented in
[`docs/persona-and-distribution.md`](docs/persona-and-distribution.md).

## Install (Recommended)

End users do not need `uv`.

### Debian/Ubuntu (`.deb`)

Download a release artifact and install it:

```bash
sudo apt install ./aman_<version>_<arch>.deb
```

Then enable the user service:

```bash
systemctl --user daemon-reload
systemctl --user enable --now aman
```

### Arch Linux

Use the generated packaging inputs (`PKGBUILD` + source tarball) in `dist/arch/`
or your own packaging pipeline.

## Distribution Matrix

| Channel | Audience | Status |
| --- | --- | --- |
| Debian package (`.deb`) | End users on Ubuntu/Debian | Canonical |
| Arch `PKGBUILD` + source tarball | Arch maintainers/power users | Supported |
| Python wheel/sdist | Developers/integrators | Supported |

## Runtime Dependencies

- X11
- PortAudio runtime (`libportaudio2` or distro equivalent)
- GTK3 and AppIndicator runtime (`gtk3`, `libayatana-appindicator3`)
- Python GTK and X11 bindings (`python3-gi`/`python-gobject`, `python-xlib`)

<details>
<summary>Ubuntu/Debian</summary>

```bash
sudo apt install -y libportaudio2 python3-gi python3-xlib gir1.2-gtk-3.0 libayatana-appindicator3-1
```

</details>

<details>
<summary>Arch Linux</summary>

```bash
sudo pacman -S --needed portaudio gtk3 libayatana-appindicator python-gobject python-xlib
```

</details>

<details>
<summary>Fedora</summary>

```bash
sudo dnf install -y portaudio gtk3 libayatana-appindicator-gtk3 python3-gobject python3-xlib
```

</details>

<details>
<summary>openSUSE</summary>

```bash
sudo zypper install -y portaudio gtk3 libayatana-appindicator3-1 python3-gobject python3-python-xlib
```

</details>

## Quickstart

```bash
aman run
```

On first launch, Aman opens a graphical settings window automatically.
It includes sections for:

- microphone input
- hotkey
- output backend
- writing profile
- output safety policy
- runtime strategy (managed vs custom Whisper path)
- help/about actions

## Config

Create `~/.config/aman/config.json` (or let `aman` create it automatically on first start if missing):

```json
{
  "config_version": 1,
  "daemon": { "hotkey": "Cmd+m" },
  "recording": { "input": "0" },
  "stt": {
    "provider": "local_whisper",
    "model": "base",
    "device": "cpu",
    "language": "auto"
  },
  "models": {
    "allow_custom_models": false,
    "whisper_model_path": ""
  },
  "injection": {
    "backend": "clipboard",
    "remove_transcription_from_clipboard": false
  },
  "safety": {
    "enabled": true,
    "strict": false
  },
  "ux": {
    "profile": "default",
    "show_notifications": true
  },
  "advanced": {
    "strict_startup": true
  },
  "vocabulary": {
    "replacements": [
      { "from": "Martha", "to": "Marta" },
      { "from": "docker", "to": "Docker" }
    ],
    "terms": ["Systemd", "Kubernetes"]
  }
}
```

`config_version` is required and currently must be `1`. Legacy unversioned
configs are migrated automatically on load.

Recording input can be a device index (preferred) or a substring of the device
name.
If `recording.input` is explicitly set and cannot be resolved, startup fails
instead of falling back to a default device.

Config validation is strict: unknown fields are rejected with a startup error.
Validation errors include the exact field and an example fix snippet.

Profile options:

- `ux.profile=default`: baseline cleanup behavior.
- `ux.profile=fast`: lower-latency AI generation settings.
- `ux.profile=polished`: same cleanup depth as default.
- `safety.enabled=true`: enables fact-preservation checks (names/numbers/IDs/URLs).
- `safety.strict=false`: fallback to safer draft when fact checks fail.
- `safety.strict=true`: reject output when fact checks fail.
- `advanced.strict_startup=true`: keep fail-fast startup validation behavior.

Transcription language:

- `stt.language=auto` (default) enables Whisper auto-detection.
- You can pin language with Whisper codes (for example `en`, `es`, `pt`, `ja`, `zh`) or common names like `English`/`Spanish`.
- If a pinned language hint is rejected by the runtime, Aman logs a warning and retries with auto-detect.

Hotkey notes:

- Use one key plus optional modifiers (for example `Cmd+m`, `Super+m`, `Ctrl+space`).
- `Super` and `Cmd` are equivalent aliases for the same modifier.

AI cleanup is always enabled and uses the locked local Qwen2.5-1.5B GGUF model
downloaded to `~/.cache/aman/models/` during daemon initialization.
Prompts are structured with semantic XML tags for both system and user messages
to improve instruction adherence and output consistency.
Cleanup runs in two local passes:
- pass 1 drafts cleaned text and labels ambiguity decisions (correction/literal/spelling/filler)
- pass 2 audits those decisions conservatively and emits final `cleaned_text`
This keeps Aman in dictation mode: it does not execute editing instructions embedded in transcript text.
Before Aman reports `ready`, local llama runs a tiny warmup completion so the
first real transcription is faster.
If warmup fails and `advanced.strict_startup=true`, startup fails fast.
With `advanced.strict_startup=false`, Aman logs a warning and continues.
Model downloads use a network timeout and SHA256 verification before activation.
Cached models are checksum-verified on startup; mismatches trigger a forced
redownload.

Provider policy:

- `Aman-managed` mode (recommended) is the canonical supported UX:
  Aman handles model lifecycle and safe defaults for you.
- `Expert mode` is opt-in and exposes a custom Whisper model path for advanced users.
- Editor model/provider configuration is intentionally not exposed in config.
- Custom Whisper paths are only active with `models.allow_custom_models=true`.

Use `-v/--verbose` to enable DEBUG logs, including recognized/processed
transcript text and llama.cpp logs (`llama::` prefix). Without `-v`, logs are
INFO level.

Vocabulary correction:

- `vocabulary.replacements` is deterministic correction (`from -> to`).
- `vocabulary.terms` is a preferred spelling list used as hinting context.
- Wildcards are intentionally rejected (`*`, `?`, `[`, `]`, `{`, `}`) to avoid ambiguous rules.
- Rules are deduplicated case-insensitively; conflicting replacements are rejected.

STT hinting:

- Vocabulary is passed to Whisper as compact `hotwords` only when that argument
  is supported by the installed `faster-whisper` runtime.
- Aman enables `word_timestamps` when supported and runs a conservative
  alignment heuristic pass (self-correction/restart detection) before the editor
  stage.

Fact guard:

- Aman runs a deterministic fact-preservation verifier after editor output.
- If facts are changed/invented and `safety.strict=false`, Aman falls back to the safer aligned draft.
- If facts are changed/invented and `safety.strict=true`, processing fails and output is not injected.

## systemd user service

```bash
make install-service
```

Service notes:

- The user unit launches `aman` from `PATH`.
- Package installs should provide the `aman` command automatically.
- Inspect failures with `systemctl --user status aman` and `journalctl --user -u aman -f`.

## Usage

- Press the hotkey once to start recording.
- Press it again to stop and run STT.
- Press `Esc` while recording to cancel without processing.
- `Esc` is only captured during active recording.
- Recording start is aborted if the cancel listener cannot be armed.
- Transcript contents are logged only when `-v/--verbose` is used.
- Tray menu includes: `Settings...`, `Help`, `About`, `Pause/Resume Aman`, `Reload Config`, `Run Diagnostics`, `Open Config Path`, and `Quit`.
- If required settings are not saved, Aman enters a `Settings Required` tray mode and does not capture audio.

Wayland note:

- Running under Wayland currently exits with a message explaining that it is not supported yet.

Injection backends:

- `clipboard`: copy to clipboard and inject via Ctrl+Shift+V (GTK clipboard + XTest)
- `injection`: type the text with simulated keypresses (XTest)
- `injection.remove_transcription_from_clipboard`: when `true` and backend is `clipboard`, restores/clears the clipboard after paste so the transcript is not kept there

Editor stage:

- Canonical local llama.cpp editor model (managed by Aman).
- Runtime flow is explicit: `ASR -> Alignment Heuristics -> Editor -> Fact Guard -> Vocabulary -> Injection`.

Build and packaging (maintainers):

```bash
make build
make package
make package-deb
make package-arch
make release-check
```

`make package-deb` installs Python dependencies while creating the package.
For offline packaging, set `AMAN_WHEELHOUSE_DIR` to a directory containing the
required wheels.

Benchmarking (STT bypass, always dry):

```bash
aman bench --text "draft a short email to Marta confirming lunch" --repeat 10 --warmup 2
aman bench --text-file ./bench-input.txt --repeat 20 --json
```

`bench` does not capture audio and never injects text to desktop apps. It runs
the processing path from input transcript text through alignment/editor/fact-guard/vocabulary cleanup and
prints timing summaries.

Model evaluation lab (dataset + matrix sweep):

```bash
aman build-heuristic-dataset --input benchmarks/heuristics_dataset.raw.jsonl --output benchmarks/heuristics_dataset.jsonl
aman eval-models --dataset benchmarks/cleanup_dataset.jsonl --matrix benchmarks/model_matrix.small_first.json --heuristic-dataset benchmarks/heuristics_dataset.jsonl --heuristic-weight 0.25 --output benchmarks/results/latest.json
aman sync-default-model --report benchmarks/results/latest.json --artifacts benchmarks/model_artifacts.json --constants src/constants.py
```

`eval-models` runs a structured model/parameter sweep over a JSONL dataset and
outputs latency + quality metrics (including hybrid score, pass-1/pass-2 latency breakdown,
and correction safety metrics for `I mean` and spelling-disambiguation cases).
When `--heuristic-dataset` is provided, the report also includes alignment-heuristic
quality metrics (exact match, token-F1, rule precision/recall, per-tag breakdown).
`sync-default-model` promotes the report winner to the managed default model constants
using the artifact registry and can be run in `--check` mode for CI/release gates.

Control:

```bash
make run
make run config.example.json
make doctor
make self-check
make eval-models
make sync-default-model
make check-default-model
make check
```

Developer setup (optional, `uv` workflow):

```bash
uv sync --extra x11
uv run aman run --config ~/.config/aman/config.json
```

Developer setup (optional, `pip` workflow):

```bash
make install-local
aman run --config ~/.config/aman/config.json
```

CLI (internal/support fallback):

```bash
aman run --config ~/.config/aman/config.json
aman doctor --config ~/.config/aman/config.json --json
aman self-check --config ~/.config/aman/config.json --json
aman bench --text "example transcript" --repeat 5 --warmup 1
aman build-heuristic-dataset --input benchmarks/heuristics_dataset.raw.jsonl --output benchmarks/heuristics_dataset.jsonl --json
aman eval-models --dataset benchmarks/cleanup_dataset.jsonl --matrix benchmarks/model_matrix.small_first.json --heuristic-dataset benchmarks/heuristics_dataset.jsonl --heuristic-weight 0.25 --json
aman sync-default-model --check --report benchmarks/results/latest.json --artifacts benchmarks/model_artifacts.json --constants src/constants.py
aman version
aman init --config ~/.config/aman/config.json --force
```