Cognitive Fab
Claude Code · plugins
Polytrion Poly Plugins Tools Services Blog Get started
CLAUDE CODE · OPEN SOURCE

Two plugins we use every day

Review what Claude writes.
Map what it builds.

polywatch reads each change Claude makes in the background and hands Claude back only the defects that survive a second look. archlens turns a codebase, or a design document, into architecture diagrams that carry their evidence. Both are open source, and both come from the same stack as Poly.

polywatch → archlens → Poly, the language →
Plugin · code review

polywatch: a background bug hunter for the code Claude writes

Claude writes code faster than you can review it. After each turn, a reviewer from another model family lists candidate defects, Claude Opus checks the serious ones against the code, and on your next prompt you see at most two ranked findings. Confirmed defects go back to Claude to fix.

Apache-2.0Claude Code pluginhooks + background worker
  • Claude never waits. Nothing runs inside the turn except writing a small file; the review runs in the background.
  • Not a verdict, not a wall of warnings. On 50 replayed commits, a pass/fail verdict separated later-fixed from never-fixed commits no better than chance (AUC 0.48). So polywatch ignores the verdict and shows specific claims instead.
  • A ranking you train. You mark each finding real or false; each outcome moves the score for its kind, in a ledger that stays in your project.
  • Cheap, and capped. A reviewed turn typically costs $0.01 to $0.15, with limits of $0.50 per turn and $5 per day that a repository cannot raise.
  • Clear about what leaves your machine. The changed code of each turn goes to the reviewer's API; files such as .env and *.pem are never sent. The full policy is in the repository.

Install

in Claude Code
/plugin marketplace add cognitive-fab/polywatch
/plugin install polywatch@polywatch

Then

add your Anthropic key under
/plugin → polywatch → Configure
How it works, and the evidence → GitHub ↗
Plugin · architecture

archlens: write the architecture analysis, and the diagrams compile from it

You, or Claude, write one structured analysis of a system. archlens compiles it into validated, interactive diagrams and a markdown document that cannot disagree with them, and nobody types a coordinate.

MITClaude Code plugin and Node.js toolNode 22+
  • One diagram per question. Ask /archlens how does the harness mount the gate? and get a single diagram that answers it, joined to any analysis you already have.
  • Map the whole system. Ask Claude to map the architecture, and it reads the code, writes the analysis, validates it and renders one diagram per question.
  • Start from a design document. Name the document as the only source, and components it describes but the code does not contain are marked planned, not built.
  • Evidence, not decoration. Source links are checked against git, and everything a picture had to drop is reported rather than hidden.
  • The same output, every time. The analysis file is the thing worth reviewing; the diagrams and the document are generated from it, so they cannot drift apart.

Example: the Poly architecture pages are generated from an analysis of the Poly repository with archlens.

Install

the renderer, once
npx skills add tt-a1i/archify -g

in Claude Code
/plugin marketplace add cognitive-fab/arch-lens
/plugin install archlens@arch-lens

Then

map this architecture with archlens
archlens on GitHub ↗ See an example →

Part of a larger open-source stack

Both plugins come out of the work on Poly, a language and runtime for the stateful core of software written by coding agents, and the tools around it: Polygraph, Polyflow and the rest. The same stack sits underneath the digital twins we build.

Poly →   All the tools →   The technology behind the twin →