Skip to content

PlanBefore v0.4.0

Choose a tag to compare

@kevincoder91 kevincoder91 released this 21 Sep 03:41

Highlights

  • Planning documents in your repository. You can now let PlanBefore write
    Markdown planning documents (requirements, tasks, progress, reviews) into
    one folder you authorize per project:
    pbnode expose <path> --spec-root specs. Writes are limited to .md files
    inside that folder, never follow symlinks and are atomic. Remove the
    permission with --clear-spec-root; pbnode projects shows it. Projects
    without a spec folder stay read-only.
  • Tasks run by your local developer agent. An authorized MCP client can
    ask your own Codex installation to implement one task on your machine. The
    agent receives only that task, works on a dedicated branch, commits its
    work and returns a structured report; PlanBefore checks the reported commit
    against Git and never accepts a task on its own. Runs are asynchronous and
    can be cancelled without touching your working tree.
  • New MCP tools: spec.write, task.start, task.status, task.cancel,
    rf.checkpoint and rf.status. The existing analysis tools are unchanged
    and read-only, and there is still no tool for arbitrary file writes, shell
    commands or Git changes.
  • pbnode status now says whether task execution is available.

Upgrade notes

Run the install command again, then restart the Node:
pbnode stop && pbnode serve --background.

Task execution is optional. Without Codex, everything else keeps working and
the task tools report that execution is unavailable. To enable it, add an
executor: block to your config.yaml (see the README). If codex is not on
your PATH —the Codex desktop app ships it inside its own bundle— set
executor.command to the full path of the binary.

With the default sandbox (workspace-write), current Codex versions let the
agent read .git but not write it, so a task that needs a branch or a commit
stops as blocked without changing your repository. Letting the agent create
branches and commit requires executor.sandbox: danger-full-access, which
gives the agent full access to your machine, not just the project. That is
your decision; PlanBefore never changes it for you.

Local task state is stored only on your machine, next to your configuration
(workflow/state.json, readable only by you). The agent's report is returned
to the MCP client that asked for it, like any other tool result; PlanBefore
does not store it on its servers.


Install

macOS / Linux:

curl -fsSL https://planbefore.kvnlabs.xyz/install.sh | REPOFLOW_VERSION=v0.4.0 sh

Windows (PowerShell):

$env:REPOFLOW_VERSION = 'v0.4.0'; irm https://planbefore.kvnlabs.xyz/install.ps1 | iex

Both installers verify the SHA-256 of the download against checksums.txt.
Manual installation, requirements (git, rg) and first steps:
https://planbefore.kvnlabs.xyz/install

The binaries are not code-signed yet; macOS and Windows may show a warning
for files downloaded with a browser. See the installation guide.