Add TOML config, text sanitization, custom endpoints

.gitignore

@@ -1,3 +1,5 @@
 /target/
 FnKey.app/
 .DS_Store
+.venv/
+__pycache__/

CLAUDE.md

@@ -0,0 +1,113 @@
+# FnKey — Agent Instructions
+
+## What is this
+
+FnKey is a macOS menu bar app (Rust) that records audio while a hotkey is held, transcribes it via Whisper, sanitizes the text with a small LLM, and pastes the result at the cursor. Single file: `src/main.rs`.
+
+## Architecture
+
+```
+Hotkey held → mic records → release → WAV encoded → Whisper STT → LLM sanitizer → clipboard → ⌘V paste
+```
+
+- **Transcription**: Any OpenAI-compatible `/v1/audio/transcriptions` endpoint
+- **Sanitization**: Any OpenAI-compatible `/v1/chat/completions` endpoint
+- **Config**: `~/.config/fnkey/config.toml` (TOML, deserialized into `Config` struct)
+- **No runtime dependencies** — single static binary, no frameworks beyond macOS system libs
+
+## Text Sanitization Pipeline
+
+The sanitizer is the key differentiator. Raw Whisper output is messy — filler words, repeated words, misheard technical terms. A small LLM (0.6B–3B params) cleans it up in real time (<200ms).
+
+### Setup for contributors
+
+After cloning this repo, set up local inference for development:
+
+1. **Whisper STT** — download and serve a Whisper model:
+   ```bash
+   # Option A: llama.cpp (any platform with Vulkan/CUDA/Metal)
+   whisper-server -m ggml-large-v3-turbo.bin --port 8100
+
+   # Option B: faster-whisper-server (CUDA)
+   pip install faster-whisper-server
+   faster-whisper-server --model large-v3-turbo --port 8100
+
+   # Option C: MLX (macOS Apple Silicon)
+   pip install mlx-whisper
+   # (use via mlx-whisper Python API, or wrap in a server)
+   ```
+
+2. **Text sanitizer** — download and serve a small LLM:
+   ```bash
+   # Option A: llama.cpp
+   llama-server -m Qwen3-0.6B-Q6_K.gguf --port 8108 -ngl 999 --jinja -c 4096
+
+   # Option B: MLX
+   pip install mlx-lm
+   mlx_lm.server --model mlx-community/Qwen3-0.6B-4bit --port 8108
+   ```
+
+3. **Configure** `~/.config/fnkey/config.toml`:
+   ```toml
+   api_key = "not-needed"
+   transcription_url = "http://localhost:8100/v1/audio/transcriptions"
+   whisper_model = "whisper-large-v3-turbo"
+   polish_url = "http://localhost:8108/v1/chat/completions"
+   polish_model = "qwen3-0.6b"
+   always_polish = true
+   ```
+
+### How the system prompt works
+
+The `polish_prompt` config field overrides the built-in system prompt. For 0.6B models, the prompt must be:
+- **Short** — the model has limited instruction-following at this scale
+- **Explicit** — use a replacement dictionary for domain terms the STT engine misrecognizes
+- **End with `/no_think`** — disables Qwen3 reasoning mode, keeping latency under 200ms
+
+Example for a codebase that uses llama.cpp, Claude, and Tailscale:
+
+```toml
+polish_prompt = """Fix dictation. Remove filler words. Fix grammar and punctuation.
+Replace misheard terms:
+- clod dot MD/cloud dot MD → CLAUDE.md
+- agents dot MD → agents.md
+- lama dot CPP → llama.cpp
+- quan three/qan three → Qwen3
+- tail scale → Tailscale
+Output ONLY the corrected text. /no_think"""
+```
+
+**When helping users customize their prompt**: identify the technical terms they use frequently, figure out how Whisper is likely to misrecognize them, and add those as `misheard → correct` pairs. This is the single most impactful thing you can do to improve dictation accuracy.
+
+## Build & test
+
+```bash
+cargo build --release          # Build
+./build-app.sh                 # Create .app bundle
+cargo check                    # Type check only
+```
+
+After rebuilding, macOS invalidates permissions — user must re-add the app in System Settings (Input Monitoring, Accessibility).
+
+## Key code locations
+
+- `Config` struct and `load_config()` — config parsing and defaults
+- `polish_text()` — LLM sanitization call
+- `transcribe_and_paste()` — main pipeline: encode WAV → call Whisper → sanitize → paste
+- `run_event_tap()` — hotkey detection via CGEventTap
+- `enhance_audio()` — audio preprocessing (DC removal, high-pass filter, normalization)
+
+## Config fields
+
+| Field | Default | Purpose |
+|-------|---------|---------|
+| `api_key` | `""` | Bearer token for Whisper transcription (also used for sanitizer if `polish_api_key` is empty) |
+| `polish_api_key` | `""` | Separate bearer token for sanitizer endpoint (empty = falls back to `api_key`) |
+| `transcription_url` | Groq | Whisper endpoint |
+| `polish_url` | Groq | Chat completions endpoint for sanitizer |
+| `whisper_model` | `whisper-large-v3` | Model name sent to STT endpoint |
+| `polish_model` | `llama-3.3-70b-versatile` | Model name sent to sanitizer endpoint |
+| `hotkey` | `fn` | Trigger key (fn/option/control/shift/command) |
+| `language` | `""` (auto) | ISO-639-1 hint for Whisper |
+| `always_polish` | `true` | Sanitize every dictation by default |
+| `polish_prompt` | `""` (built-in) | Custom system prompt for sanitizer |

Cargo.toml

@@ -28,3 +28,6 @@ hound = "3.5"
 # JSON parsing for LLM API
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
+
+# Config file parsing
+toml = "0.8"

Info.plist

@@ -14,6 +14,8 @@
     <string>1.0.0</string>
     <key>CFBundleShortVersionString</key>
     <string>1.0.0</string>
+    <key>CFBundleIconFile</key>
+    <string>AppIcon</string>
     <key>CFBundlePackageType</key>
     <string>APPL</string>
     <key>LSMinimumSystemVersion</key>