pi-prose-watch

An AI-tone check for pi that runs only while the agent writes prose.

Scanning every edit would judge code, configuration, and logs for no reason. Here the agent arms the files it drafts, and only those are checked:

1. When the agent starts drafting or revising an article, document, chapter, or note, it calls prose_watch with action on and the file's path. The tool's prompt guideline tells it to.
2. After every successful write or edit of an armed file, the text that call added (not the whole file; front matter, code fences, and tags removed) goes to Jev, TypeSafe's System One classifier.
3. Text is judged in chunks of up to 6,000 characters, at most four per call; a longer write says how much was judged. When Jev judges a chunk likely AI-written, a warning is appended to that tool's result, naming where the passage starts and the pattern that shows most: staged contrasts, filler transitions, slogans, figurative verbs, inflation, translated phrasing, or framing. The agent rereads those sentences before going on.
4. prose_watch with action off disarms the given files, or every file when no path is given. The armed set is saved in the session, so /reload, resume, and fork keep it, and a new session starts with nothing armed.

Paths are matched the way pi's own tools resolve them (a leading @, ~, ./, symlinks such as /tmp on macOS), so a file armed as post.md is still checked when written as ./post.md.

pi install npm:pi-prose-watch

It needs TYPESAFE_API_KEY in the environment. TYPESAFE_DEFAULT_MODEL and TYPESAFE_BASE_URL are optional; the base URL takes pi's form, ending at /v1 (default https://api.typesafe.ai/v1). Arming without the key fails, and a failed Jev call is reported in the tool result as "not judged", so a check that did not run never looks like a pass.

What the verdict means

Jev returns a probability and a pattern, not the sentences, so the warning is a reminder: finding and rewriting the sentences stays with the agent's own writing rules.

The threshold is 0.7. It was calibrated against jev-1.13.0 on Traditional Chinese: 17 paragraphs of one author's published essays scored 0.25–0.55, and four paragraphs of stock AI prose scored 0.80–0.92. One AI paragraph written in a translated register scored 0.56 and is not caught; the check is tuned to stay quiet on human text rather than to catch everything. Edits shorter than 40 characters of prose are not sent, and a write made by another tool (a nested call) is skipped, since its result never reaches the model.

License

MIT