A Claude Code plugin that gives an agent a palette of visual-element skills — each generates a self-contained, interactive HTML artifact (charts, diagrams, tables, slide decks, code views, wireframes, agent reports). Every element is editable, commentable, stylizable (light + dark), and exportable, and any element can be composed inside any other via nested HTML + SVG.

Features

• Interactive HTML output — every diagram, table, card and chart point is clickable; the click returns a typed selection to the agent and the browser closes itself.
• v2 modal-comment threads on agent reports — hover any paragraph, list item, table row or code block; type a comment in a right-aligned modal; agent replies inline (polling, atomic save, resume on reopen).
• v3 per-element decision toggles (approve / reject — both off = skip) on every finding — two pill-shaped switches with mutex, in a warm slate/teal/rust palette; the toggle is the outcome signal Claude reads first, the thread stays free-text for clarifications.
• 10 slash commands: amvcp-generate-web-diagram, amvcp-generate-visual-plan, amvcp-generate-slides, amvcp-diff-review, amvcp-plan-review, amvcp-project-recap, amvcp-fact-check, amvcp-interactive-report, amvcp-respond-to-comment, amvcp-share-page.
• Embedded renderers — Mermaid, Graphviz (viz.js), TikZJax, KaTeX (with the default math-macro set), a vendored regex visualiser with per-mount undo/redo, shift-click multi-select, and ⌘⇧Z redo, plus the vendored Pierre high-fidelity diff viewer (Shiki highlight, split/unified, merge-conflict resolver, huge-file virtualizer).
• Multi-click text selection on prose blocks — depth 1 (letter) → 2 (word) → 3 (block) → up to depth 7.
• Self-contained pages — single HTML file, vanilla JS runtime, WASM/CDN deps lazy-loaded only when the page actually uses them. No build step.
• Multi-harness — Claude Code, Pi, Codex CLI, OpenCode, and Cursor configs included.
• Dev-browser test suite covering every fixed bug, edit-panel surface, and the v3 decision-toggle flow (incl. mutex).

Installation

/plugin marketplace add Emasoft/ai-maestro-plugins
/plugin install ai-maestro-visual-communicator-plugin@ai-maestro-plugins

Usage

After installation, invoke any of the 10 slash commands from within a Claude Code session — each one launches the corresponding HTML page in a Chromium app window and returns the user's selection / comments back to the agent.

/amvcp-generate-web-diagram   # interactive web diagram
/amvcp-generate-visual-plan   # plan-of-work review page
/amvcp-generate-slides        # slide deck
/amvcp-diff-review            # code diff review with comments
/amvcp-plan-review            # markdown plan review with toggles
/amvcp-project-recap          # snapshot of project state
/amvcp-fact-check             # claim-by-claim fact-check page
/amvcp-interactive-report     # convert any .md report into an
                              # interactive HTML page with v2 modal
                              # comments + v3 decision toggles
/amvcp-respond-to-comment     # continue an open comment thread
/amvcp-share-page             # deploy a finished page to a public Vercel URL

Every page is a single self-contained HTML file. On Submit, the page closes itself and the agent receives the multi-select payload (picked items + comment threads + decision toggles).

Memory (recall before you render)

The plugin uses the AI-Maestro global janitor-hosted wiki memory — durable project facts live as symptom-indexed markdown notes across three scopes: LOCAL (~/.claude/projects/<slug>/memory/), PROJECT (in-repo at .claude/project/memory/, git-tracked + shared), and USER (cross-project). The protocol lives in ~/.claude/rules/markdown-memory-recall.md; the day-to-day legs are the global /janitor-memory-recall, /janitor-memory-write, and /janitor-memory-update skills (search runs on memgrep, degrading to plain grep when the binary is absent).

VISUAL-COMMUNICATOR workflow wiring:

• Before generating any page, recall house-style and confirmed-preference notes (themes, density, palettes) so prior decisions shape the output without re-asking.
• After a confirmed preference or a solved gotcha, write one symptom-indexed note (plus its MEMORY.md index line) for future sessions.

memgrep is optional: install it once from the ai-maestro-janitor repo (cargo install --path <…>/ai-maestro-janitor/scripts/memgrep); until then recall falls back to grep — it degrades, never breaks.

AI Maestro side panel (optional)

If — and only if — the plugin happens to be running inside the AI Maestro harness, a generated artifact can be pushed straight into the dashboard side panel instead of being handed over as a file path:

scripts/amvcp-panel-push.py <artifact.html> [--agent <uuid-or-name>]

This is an optional path, not a dependency. amvcp is a universal plugin: ai-maestro depends on amvcp, never the reverse. The script finds aimaestro-panel.sh on PATH and nowhere else, and when it is absent — the normal case — it prints panel: unavailable and exits 0, so amvcp behaves exactly as it does standalone. The target agent comes from --agent, $AMVCP_PANEL_AGENT, or $AIMAESTRO_AGENT_ID; with none of them it prints panel: no-target and exits 0.

Once the CLI is present and a target is known, a delivery problem is a real failure and exits non-zero. That includes delivered: 0: the panel is a live surface rather than a queue, so a zero count means no dashboard had the channel open and the artifact was dropped, not stored — reporting it as delivered would leave you believing something is on screen when nothing received it. Credentials are never handled here; aimaestro-panel.sh reads AID_AUTH (agent callers) or AIMAESTRO_SUDO_TOKEN (user callers) from the environment itself.

Platform requirements

Supported OSes: macOS, Linux, and Windows. Python 3.12+ required.

Every user-facing entry point is pure Python: the amvcp-share-page command (scripts/share.py), the test runner (tests/run-all-tests.py), the page renderer (scripts/render-interactive-report.py), and the selection server (scripts/amvcp-select.py). Windows users no longer need WSL2 — the plugin runs natively on every OS that has Python 3.12+.

Dependency

amvcp depends on the ai-maestro-plugin base plugin for the PRRD/TRDD/kanban tasks pillar scripts (get-prrd.py, prrd-edit.py, findprrd.py, findtrdd.py, kanban.py, bootstrap_design.py, prrd_lib.py). These scripts are not vendored here — they are resolved at runtime via scripts/prrd-trdd/resolve_pillar_scripts.sh, a dependency-free POSIX shim that prints the absolute path of the installed base plugin's pillar directory (no hard-coded cache path):

DIR="$(sh scripts/prrd-trdd/resolve_pillar_scripts.sh)" || exit 1
python3 "$DIR/kanban.py"          # render the TRDD kanban board
python3 "$DIR/get-prrd.py" --list # list this project's PRRD rules

Install the ai-maestro-plugin base (or set AI_MAESTRO_PRRD_SCRIPTS_DIR) so the resolver can find the scripts. The project's rules live in design/requirements/PRRD.md and its task backlog in the design/ zone tree (proposals/, tasks/, archived/, refused/).

License

MIT. Forked from nicobailon/visual-explainer v0.8.0; see CHANGELOG.md.