Subtext

Most debugging tools for language models show you what came in and what came out. Subtext shows you what happened in between. The tool attaches to a locally-running language model and streams its J-space representations — internal states the model uses for multi-step reasoning — to a browser-based dashboard in real time. You can watch the verbal workspace shift as the model reads the prompt and before it commits to output tokens. Sessions can be recorded via `record_session.py`, exported, and replayed in the browser, so you can step through a conversation after the fact and inspect state at each turn.

The differentiating feature is the J-space instrumentation itself. Grounded in Anthropic research identifying a small set of internal representations that behave like a global workspace — contents that the model can verbally report, deliberately modulate, and causally use for reasoning — Subtext makes those representations observable rather than inferred. That is a different class of visibility than attention weights or token probabilities, and the repo includes a reference implementation and a paper link for teams who want to validate or extend the methodology.

Subtext fits tightly into one workflow: a researcher or developer with a local model who wants ground-truth visibility into reasoning steps, not just output quality. It breaks equally fast in two directions: any model running behind a remote API is outside its reach, and there are no hooks to feed observations into an external evaluation framework, a database, or a CI pipeline. Teams doing one-off debugging sessions will find it self-contained. Teams who need to track internal states across hundreds of runs and correlate them with downstream metrics will hit the export-and-do-it-yourself wall immediately.

The project ships with `server.py`, `test_client.py`, `test_multiturn.py`, and `verify_accuracy.py` alongside a `requirements.txt`, with Windows batch start support via `start.bat`. The interface is a single `index.html` served locally. There is no packaged binary and no install wizard — you clone, install dependencies, and run the server against a compatible local model.