Arcgram v2.1.2: Auditable Diagram Harness for AI Agent Reasoning

[jovesun-lab]

Arcgram makes AI agent reasoning visible and auditable as an interactive graph diagram, turning walls of prose into typed nodes and explicit edges you inspect and correct node by node. Use it as a visual agent harness: propose → draw → self-audit → fix, human in the loop. It outputs one standalone HTML file (no build step, no npm, no CDN, no runtime), is framework-agnostic and agent-neutral, and ships authoring plus Checkpoint (diagram structure), Reconcile (diagram vs code), and Validate self-check skills.

Why It Matters

An agent's plan reads fine in prose even when a step is broken. As a diagram, a missing step dangles in space and a circular dependency is a loop you catch at a glance. One skill file; runs on Claude, GPT, Gemini, DeepSeek, Kimi and Zhipu GLM, in tools like Cursor, Cline and Aider. The HTML output supports pan, zoom, hover and filter, and opens in any browser.

What's New in v2.1.2

• The authoring docs now match the v2 engine. There is no canvas size to set: when a diagram opens, the view fits the whole drawing into the window, and 0 or the fit button fits it again. The layout tips, the schema reference and the skill no longer ask for W / H, an auto-extending canvas or a canvas-edge margin.
• Spacing rules of thumb corrected: nodes are 48 px tall at their smallest, a column's y-step is about 100 px, and columns start about 300 px apart.
• Draw a diagram right in the chat: USAGE.md shows how an agent with an inline visual tool loads the engine from this repository through jsDelivr and fills in the diagram data, about 2 KB per diagram.
• Drop-in update from 2.1.1: the engine, the gates and the skills' code are unchanged. The plugin version moves to 2.1.2 so installed copies pick up the new docs.

Since v2.0

• Reconcile is back in the claude.ai plugin marketplace, and the plugin description fits the marketplace's 500-character limit (v2.1.1).
• Audit extension: a per-map SHOW-COMPLETED stamp shows resolved marks by default on maps that opt in; hidden mode prints an honest "N closed hidden" count in the header.
• Unknown BUGMARK stamps now warn by name instead of being silently ignored.
• Headless SVG export: buildExportSVG runs without a browser, wire-complete output.
• Gate bundle refreshed: 56 public gates.
• Cleaner arrival-label placement: the label at an edge's end backs off from the target node and its type tag, so labels no longer overlap the node they point to.
• Cleaner edge routing.
• Opt-in usage telemetry, OFF by default: only if you enable it, the engine sends one anonymous "diagram opened" ping. No personal data, none of your diagram's content, no cookies or fingerprinting; honours Do Not Track; does nothing offline.

Install

Install as a plugin (Claude Code / Cowork):

/plugin marketplace add jovesun-lab/arcgram
/plugin install arcgram

Already installed? Run /plugin marketplace update arcgram to get 2.1.2.

Just the engine:

curl -O https://raw.githubusercontent.com/jovesun-lab/arcgram/main/template-v2.html

Try It First (No Install)

• A game-system flow: https://jovesun-lab.github.io/arcgram/examples/example.html
• A decision flowchart: https://jovesun-lab.github.io/arcgram/examples/example-thinkflow.html
• How to use Arcgram: https://jovesun-lab.github.io/arcgram/examples/usage-workflow.html
• The self-audit (it flags its own weak nodes): https://jovesun-lab.github.io/arcgram/examples/example-audit.html

Links

• Site: https://arcgram.io
• Blog: https://arcgram.io/blog/
• Full changelog: v2.1.1...v2.1.2

License

Apache 2.0. Diagrams carry a small "Made with Arcgram" mark (keeping it is free; that is how attribution works here); a mark-free build is available under a separate commercial license. See LICENSE, NOTICE and WATERMARK-AND-COMMERCIAL-TERMS.md. "Arcgram" and the logo are trademarks of Rae Sun.

---

This discussion was created from the release Arcgram v2.1.2: Auditable Diagram Harness for AI Agent Reasoning.