Copy raw transcript before LLM step (#54)

## Summary
- copy the raw ASR transcript to the clipboard immediately when LLM cleanup is enabled
- leave the existing LLM clipboard update in place so the cleaned text still overwrites the raw transcript
- update the LLM-enabled transcribe test to cover the new clipboard behavior and log output

## Testing
- uv run pytest tests/agents/test_transcribe.py::test_transcribe_main_llm_enabled

Fix server crash in finally block when transcription fails early (#50)

* Fix server crash in finally block when transcription fails early

The server would crash with UnboundLocalError when an exception occurred
early in the transcribe_audio function before raw_transcript was initialized.
This happened because the finally block always tried to log the transcription
without checking if the variables existed.

Changes:
- Initialize raw_transcript and cleaned_transcript at the start of the function
- Add safety check in finally block to only log if there's content to log
- Wrap logging operation in try-catch to prevent crashes from logging failures
- Add comprehensive tests for various failure scenarios

This ensures the server remains stable even when errors occur during
transcription processing.

* Update README.md

* Fix additional server crash issue with LLM errors

Found and fixed a second crash issue where the server would exit when
LLM connections failed. The service was calling sys.exit(1) which
terminated the entire server process.

Changes:
- Set exit_on_error=False in process_and_update_clipboard for web API context
- Add proper handling for partial success (transcription ok, cleanup failed)
- Add test for LLM connection errors to ensure server stays running
- Return informative error messages when cleanup fails but transcription succeeds

This ensures the server remains stable even when LLM services are unavailable.

* fix test_process_and_update_clipboard

---------

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

feat: add support for llama-server via OpenAI-compatible API (#45)

* feat: add support for llama-server via OpenAI-compatible API

- Add openai_base_url configuration option to OpenAILLM config
- Update OpenAI provider to support custom base URLs (e.g., llama-server)
- Make API key optional when using custom base URLs
- Add --openai-base-url CLI option to all agents
- Update example configuration with llama-server usage instructions

This allows using llama-cpp's llama-server by setting:
  --llm-provider openai --openai-base-url http://localhost:8080/v1

* Update README.md

* test: fix OpenAILLM config instantiations in tests

Add missing openai_base_url parameter to all OpenAILLM instantiations
in test files to match the updated config schema.

---------

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

feat: add audio recovery mechanism for transcription failures (#44)

fix(transcribe): improve system prompt to prevent unwanted prefixes

The transcribe agent sometimes added unwanted prefixes like "Sure. Here's
the cleaned-up text:" or wrapped output in quotes despite instructions.

Enhanced the system prompt with:
- CRITICAL directive emphasizing output format
- Specific examples of wrong responses to avoid
- Clear example of correct response format
- Reinforced reminders in agent instructions

This addresses the root cause through better prompting rather than
post-processing fixes.

Fix test transcription_log

Fix correctly parsing subcommand

feat: Add Gemini support with LLMs (#21)

Make all CLI options consistent (#20)

Add extra-instructions to the config