Live Console

live-console

Spin up a console: a zero-dependency local web app, rendered from a declarative spec, that persists every response the user gives to a store on disk and self-terminates on Submit — so you read their answers back without copy-paste. The engine (scripts/console_server.py) owns all page behavior; you only author the spec.

Run loop

1. Choose the surface. Start from a template — list templates/ and

read each file's _about line for its purpose — or write a spec from scratch. Copy your starting point into the session scratchpad and fill in real items. Done when: a spec file exists in the scratchpad with a title and ≥1 item, every field has a type + key, and any referenced images sit beside the spec. Anything the schema can't express → raw-HTML escape hatch (see AUTHORING.md).

1. Launch (background). Run, as a background process:

python3 "${CLAUDE_PLUGIN_ROOT}/skills/live-console/scripts/console_server.py" Done when: the printed url opened in the browser and you've noted the printed results file path.

1. Hand off and wait. Tell the user the console is open and to click

Submit & done when finished. Do NOT poll. Done when: the background process exits (Submit) — that exit is your wake signal — or the user says to check now.

1. Read and report — exhaustive, visually grouped. Read the results file.

Report as a compact table or status-grouped list (✅/❌/⏭) rather than a paragraph per item. Account for every item in the spec: state each item's response, and explicitly flag any left unanswered or marked fail/reject, drilling into fails. Done when: no spec item is silently omitted — each is named as answered (with its value) or flagged unanswered/failed.

1. Clean up. If the server is still running (the user read early without

submitting), stop it. Done when: no console process running, its port free.

Field types (detail in AUTHORING.md)

toggle (pass/fail/skip, approve/reject) · comment · choice (single/multi) · rating (stars/number) · rank (drag-reorder) · item visuals (compare HTML snippets or image files → pick one + comment each; a single entry doubles as a plain illustration — a diagram beats a paragraph, see AUTHORING.md).

Visual-first: if an item's content is inherently spatial (a flow, a layout, a before/after, an architecture) — show it, don't describe it. See AUTHORING.md's "Visual-first authoring."

Facts

• Store: .results.json beside the spec — {_meta:{done,submittedAt}, responses:{itemId:{key:value,updatedAt}}}. Auto-saved on every change, so it's readable mid-session; _meta.done flips to true only on Submit.
• Notify: the server runs in the background and exits on Submit — no polling; the exit is the signal to read the store.
• Defaults: binds 127.0.0.1 only, OS-assigned free port, auto-opens the browser (macOS), theme-aware, ephemeral scratchpad store. Flags: --port N, --no-open, --results PATH.
• Schema + raw-HTML escape hatch (Console.set/get/submit): see [AUTHORING.md](AUTHORING.md).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: punkaze
• Source: punkaze/skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.