Plan review
Use this for several decisions that can change a proposed plan. For one question, ask in the conversation. The page is a review surface; the final implementation plan must still follow the planning skill.
This skill is based on Fluidplan by morganhub (MIT licence). Its workflow of decision cards that a person answers in rounds, ending in a DECISIONS.md file, is the model for this one. The page styles and fonts in assets/fluidplan/ are copied from it; see the README.md and LICENSE there. The CLI, server and review page were written for this repository and do less than Fluidplan's own app.
Run the bundled CLI from the project root with node ~/.agents/skills/plan-review/scripts/plan-review.mjs. Node 20 or later is required. If the shared install path is unavailable, locate this skill's scripts/plan-review.mjs beside this file.
- Inspect the request, any source plan, and relevant project files. Identify decisions the person can genuinely make. Write plain language about the consequence of each option.
- Run
new <id> --title "..."to create.plan-review/<id>/plan.json. Replace its example decision with pages and decisions in the data format. Use stable IDs. Include enough context, reasons, and option consequences for a person to answer. Runcheck <id>and fix errors. - Run
serve <id>in the background. Then runnode <skill folder>/scripts/tailnet-proxy.mjs <printed URL>in the background. On a machine running Tailscale it prints ahttp://100.x.x.x:<port>/p/...address that the person's other devices on their tailnet can open; give them that address as well as the local one. Elsewhere it prints a note and exits, and the local URL is the only one. The proxy opens the page to the tailnet, so if your permissions ask for approval, ask the person rather than looking for another route. The browser saves answers on the local machine. Keep the agent turn open when possible. If the browser cannot reach the URL, runpage <id>and give the person the standalone HTML file. Ask them to copy or download its answers and paste or attach them. Record a downloaded packet withimport <id> <answers.json>, then runsubmit <id>. - Once the person selects Send answers, run
digest <id>. If the agent cannot be woken automatically, the person can say “sent” in the conversation. Treat rewrites and questions as requests for revision. Preserve accepted decisions and all answers that did not need revision. - Update
plan.jsonfor every decision listed underrevise, adding"revision": { "round": <next round>, "note": "What changed" }. Runcheck <id>, thennext <id>. The page reloads to show the revised cards. Repeat untildigesthas no revisions or unanswered decisions. - Run
finish <id>to writeDECISIONS.md. Loadplanningand write one implementation-ready Markdown plan based on the settled decisions. Keep that plan in the location and format required byplanning; do not treat the review JSON as the final plan. Report both paths and the plan overview addresses thatplanningasks for.
The CLI owns state.json; the browser or import owns answers.json; the agent owns plan.json. Never silently treat an unanswered card as agreement. Do not use next --force or edit state files by hand. The server binds to 127.0.0.1, and its tokenized URL is local to the current review. The Tailscale proxy listens only on the machine's Tailscale address and passes on only requests made to that address by the review's own page; tailscale serve is not used because it needs root.