π-plans plugin
A planning extension for the Pi coding agent: reviewed Markdown plans first, and code changes only after your approval
- 238 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The extension controls file writes and the launch of delegated subagents
- After approval workers may edit the project on the model you choose
- Plan state and cache live in .git/pi-plans/ and ~/.cache/pi-plans/
Install
Manual install
pi install npm:pi-plansTo try without installing: pi -e npm:pi-plans. Needs Node.js 22.6+ with the --experimental-strip-types flag.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The extension controls file writes and the launch of delegated subagents
- After approval workers may edit the project on the model you choose
- Plan state and cache live in .git/pi-plans/ and ~/.cache/pi-plans/
This is third-party code. Review the repository files before installing.
What it does
π-plans is a planning extension for the Pi coding agent built around plan, review, execute. Before approval writing is closed: planning, questions and refinement run read-only, and `edit` and `write` are blocked outside the plan files until a handoff that never completes automatically. It ships 7 skills, 25 slash commands, 8 tools and 2 templates. Questions from one round gather into a single tabbed form, with an advantage and a drawback stated for every option. Reviewers work read-only and return severity-graded findings and up to five questions: `plan-normal` runs two reviewers, `plan-big` and `plan-huge` three, and the planner writes each one's direction. A plan for several product versions is kept in one directory. After approval you choose where it runs: this session, one delegated subagent or several, on a model and effort you pick. Workers report through the `plans_update_task` tool, and a dashboard shows waves, status and evidence. An independent reviewer verifies the `VC-###` check conditions. Context compaction is deterministic, with no model, and a code graph is built on tree-sitter.
Who it is for. Developers on the Pi agent who need a controlled process: a plan for approval, review and execution with a report on every task.
Good fit when
- A big task needs the agent held back from edits until the plan is approved
- Reviewers should find hidden coupling and weak checks before work starts
- The plan should run on a different, cheaper model than the review
- Work is split across several subagents with live per-task reporting
- A product ships in versions and needs one plan across 2 to 10 versions
Not a fit when
- You work in Claude Code or Codex: the extension is written for Pi
- The task is small and needs no plan: the review overhead will not pay off right away
- You need an answer in one turn: the process has approval pauses
Example request
Draft a plan for migrating to the new logging library, run the reviewers and show the question form before executionLimitations
It needs the Pi agent and Node.js 22.6 or newer with `--experimental-strip-types`: the extension and checks run TypeScript directly and have no runtime dependencies. The author publishes the result of one paired Terminal-Bench 2.0 run (18 of 36 with the extension against 3 of 36 for plain Pi) and calls it exploratory and unstable across seeds. Writing is blocked until the plan is approved, so urgent edits mean aborting the process. The last commit is from October 2026.
How to disable. Remove the package through the Pi package manager. State in .git/pi-plans/ and plan artifacts in the run folder stay on disk; delete them by hand.
FAQ
How do I install it?
With pi install npm:pi-plans. To try it without installing use pi -e npm:pi-plans. The package entry point index.ts registers the tools, commands, skills and the write guard.
What is blocked before approval?
The edit and write tools outside the plan files. Planning, questions and refinement run read-only until the handoff, which never completes automatically.
How many reviewers are there?
Two for plan-normal and three for plan-big and plan-huge. They are read-only reviewers that return severity-graded findings and up to five questions.
Where does the plan run?
You choose after approval: the current session, one delegated subagent or two to four. Model and effort come from the same picker as the reviewer's.
Are there benchmark results?
Yes, one paired Terminal-Bench 2.0 run on 36 tasks: 18 solved with the extension and 3 with plain Pi. The author calls it exploratory and publishes the numbers, seed instability and caveats.
Discussion
No questions or reviews yet.
Related
PluginsMore than thirty separately installed extensions for the Pi Coding Agent: goals, plan mode, LSP and subscription limits
planr keeps agent-driven development tasks in Git Markdown files and shows what is ready, blocked and waiting on you
The /improve skill: a strong model audits a repository and writes self-contained fix plans, and you hand the execution to a cheaper model
A local browser UI to annotate agent plans and review diffs, then send the feedback back to the agent
A Claude Code skill that drives a hard task through explore, plan, execute and reflect, writing every decision to disk.
Writ turns workflow rules into checks at tool-call time: Claude Code writes no code until you approve the plan and tests
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| π-plans (this card) | no | 238 | |
| pi-fabric | no | 335 | another Pi extension with a different job; π-plans specializes in reviewed plans with a write guard |
| pi-codex-goal | no | 207 | a Pi extension for Codex-style goals; π-plans builds the plan, review and per-task reporting |
README in short
The description promises plan, review, execute. It lists principles (writes closed until approval, a question form, converging reviewers, execution on another model, worker reports), what is new in 0.9.0, an exploratory benchmark result and install and safety details.