bethesda-strings-editor: Instructions file for Claude Code

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Running the Application

This project uses the conda environment bethesda-strings-editor (Python 3.10) for everything — running the app, the test suite, and any install. Do not use the system interpreter or pip outside the env; it has a different and incomplete set of packages.

conda activate bethesda-strings-editor
python main.py

# non-interactively (one-shot commands):
conda run -n bethesda-strings-editor python -m pytest tests/

Dependencies are in requirements.txt:

pip install -r requirements.txt
# Core:     PySide6>=6.11, requests>=2.33, cryptography>=48.0, anthropic>=0.104
# Optional: keyring, curl-cffi, py7zr, rarfile, hunspell/spylls — each has a
#           runtime fallback, so the app starts without them

Logging goes to both stdout and translator.log in the project root.

Ollama Models

Three Modelfiles are tracked. None of them names a real GGUF — every FROM is a /path/to/… placeholder, because the fine-tunes are not published and a machine-specific absolute path is worse than useless to anyone else (and rots here too: both previously-committed paths are gone from this machine). Point FROM at a GGUF you have, then ollama create <name> -f <file>.

Parameter precedence is the thing to get right when editing them, and it is per-model, not global. The app always overrides system, temperature, num_ctx and num_predict; for the sampling knobs it sends top_p/top_k/min_p/repeat_penalty/repeat_last_n only when the model's own profile in OllamaWorker.MODEL_CONFIGS defines that key (ollama_worker.py payload["options"] loop). No profile defines min_p or repeat_last_n, so those always come from the Modelfile; top_k comes from the Modelfile for gemma4-* (that profile omits it) but is overridden for translategemma3-st. num_predict is adaptive per string — max(profile_num_predict, min(settings.ollama_num_predict, max(512, input_len × 4))) — so a profile's num_predict acts as the floor, not the cap. Profile lookup is exact name → base-name prefix → _DEFAULT_MODEL_CONFIG, which is why a model called gemma4-opus48-st picks up the gemma4 profile (and its think=false).