Add TOML config, text sanitization, custom endpoints

[Martin-Atrin]

Summary

• TOML configuration (~/.config/fnkey/config.toml) with auto-generated template on first launch, Settings menu item to open it
• Custom API endpoints — any OpenAI-compatible transcription and chat completions API (Groq, OpenAI, vLLM, local llama.cpp, etc.)
• Always-on text sanitization — LLM-powered cleanup of filler words, repeated words, grammar, and misheard terms after Whisper transcription. Uses a lightweight model (as small as 0.6B params) for sub-200ms latency
• Separate API keys — api_key for Whisper, polish_api_key for sanitizer, supporting mixed providers (e.g. Groq STT + local sanitizer)
• Custom system prompt (polish_prompt) for domain-specific term correction via replacement dictionaries
• Configurable hotkey — fn, option, control, shift, or command
• Language hint — ISO-639-1 code passed to Whisper for better accuracy
• Max tokens cap on sanitizer output to prevent hallucination runaway on small models
• Debug logging (stderr) for audio duration, WAV size, Whisper response, and sanitizer output
• CLAUDE.md — agent-facing documentation with architecture, setup guide, and code map
• App icon and code signing support

Test plan

• Fresh install: verify config.toml template is auto-created on first launch
• Groq cloud: set api_key and default endpoints, verify transcription works
• Local endpoints: run llama.cpp whisper + Qwen3-0.6B sanitizer, verify sub-200ms polish
• Mixed providers: different api_key and polish_api_key, verify both endpoints authenticate correctly
• Custom polish_prompt: verify domain-specific term corrections
• always_polish = true: verify hotkey gives polished text, hotkey+modifier gives raw
• always_polish = false: verify inverse behavior
• All hotkey options: fn, option, control, shift, command
• Language hint: set language = "de" etc., verify Whisper respects it
• Long recordings (>10s): verify full audio is captured and transcribed via debug logs

🤖 Generated with Claude Code

[evoleinik]

Code Review

Must Fix

1. API keys stored world-readable — std::fs::write() creates config.toml with default 0644 permissions. Should chmod 0600 after creation since the file holds API keys:

use std::os::unix::fs::PermissionsExt;
std::fs::set_permissions(&toml_path, std::fs::Permissions::from_mode(0o600))?;

2. Default TOML template duplicated — The exact same template string appears in both load_config() and open_settings(). Extract to a const or function to avoid them drifting apart.

3. /no_think hardcoded in default prompts — This is Qwen3-specific. Users switching to GPT-4, Claude, or Llama will have /no_think literally in their system prompt, which models may echo back or misinterpret. Remove from hardcoded defaults; users who need it can add it to their custom polish_prompt.

Should Fix

4. Fn auto-detection fallback removed — The current code has a 5-second fallback from Fn to Option for keyboards that don't emit Fn events (common with external keyboards). This PR removes it silently. That's a breaking change for existing users. Consider keeping the fallback when hotkey = "fn" (the default).

5. Debug logging may leak PII — eprintln!("[fnkey] whisper text: {}", text) logs full transcriptions to stderr → Console.app/system logs. Could contain sensitive dictated content. Consider truncating or making debug logging opt-in via a debug = true config flag.

6. Silent truncation — If the sanitizer hits the max_tokens cap, the user gets half a sentence pasted silently. Should check finish_reason == "length" in the API response and fall back to raw text when truncated.

Minor

7. Config struct verbosity — The 10 separate default_*() functions could be replaced with #[serde(default)] on the struct + a single Default impl.

8. PR scope — 10+ features, +925/-160 lines in one PR. This would be easier to review and safer to merge as 3-4 smaller PRs (config migration, custom endpoints, sanitizer enhancements, cosmetic).

What's Good

• TOML config with backward compat (legacy api_key file, env var fallback) is well designed
• Custom endpoints are a real need for local/self-hosted setups
• JSON response fallback for non-compliant Whisper servers is practical
• Settings menu item with ObjC delegate is properly implemented
• The always_polish inversion logic is correct
• README expansion is thorough

---

🤖 Generated with Claude Code

[evoleinik]

Product Feedback

Thanks for the PR — there's some genuinely useful work here. Wanted to share some thoughts on the product direction before we go further.

Love these — clear wins

• TOML config — needed it, the plain api_key file doesn't scale
• Custom endpoints — this is the biggest value-add. Opens fnkey to OpenAI, local llama.cpp, vLLM, etc. Huge for adoption
• Language hint — zero-cost accuracy win for non-English users
• Settings menu item — good discoverability, users don't need to know the config path

Concerns about scope

fnkey's identity is "hold Fn, speak, paste." One thing, done well. This PR pulls it toward a configurable platform, and I want to be careful about that.

always_polish = true as the default — Today polish is opt-in (Fn+Ctrl). Flipping to always-on changes every user's first experience: added latency, LLM rewrites that may alter meaning or drop content. I'd prefer keeping always_polish = false as default and letting users opt in.

5 hotkey options — The Fn key is the product (it's called "fnkey"). The existing Fn-with-Option-fallback covers the real use case (external keyboards that don't emit Fn). Do we need control/shift/command? Each adds testing surface and edge cases (like the polish modifier collision when hotkey=control).

Separate polish_api_key — Adds config complexity for a niche case. Can add it later when someone asks.

Debug logging always on — Logs full transcriptions to stderr/Console.app. Should be behind a debug = true flag or removed.

Suggested path forward

Would you be open to splitting this up? Something like:

1. Config migration — TOML config + backward compat + Settings menu item
2. Custom endpoints — configurable URLs + models + language hint + JSON response handling
3. Sanitizer enhancements — custom prompt, max_tokens, always_polish (as opt-in)

That way we can land the structural wins quickly and iterate on the behavioral changes.

---

🤖 Generated with Claude Code