Skip to content
m4bwav edited this page Nov 21, 2026 · 2 revisions

trustwright is a Claude Code plugin that keeps your own dated record of how AI agents and their harnesses behave, product by product and version by version, and gives checklists that stop a wrong answer from being accepted. It has one skill, trustwright, and one standard-library script, scripts/trust.py. These pages describe version 0.1.1.

Vendor changelogs rarely say "the model now refuses more" or "it started claiming test runs it never made". trustwright keeps first-hand entries so you can tell.

What it does

  • Logs a behaviour observation: product, version, model, kind, the observation and evidence someone else could open.
  • Registers a false-positive stop or refusal: harness, version, model, task, stop source, category, the verbatim message and what got it through.
  • Scans a Claude Code session transcript for candidate stops, skipping what you typed yourself.
  • Shows entries per product and compares two versions: what is new, what stayed and what was not seen again.
  • Prints a verification checklist for a fact, code, numbers, a citation or a claimed action, with extra steps for high stakes.

Entries live in private JSONL files on your machine, never in the plugin.

Not for Windows host quirks (quirkwright), context size (context-health) or whether a skill is worth its tokens (evergreen-worth).

Pages

Clone this wiki locally