-
Notifications
You must be signed in to change notification settings - Fork 0
FAQ
Short answers. Each one links to the page that works it out properly.
An open-source planning skill and terminal review interface for AI coding agents. It turns an agent's plan into a versioned artifact you can read, annotate, revise, compare, approve and execute.
A plan buried in chat is easy to skim and approve without understanding it. PlanX separates planning from execution, gives the plan a review interface, and makes approval refer to an exact version. The goal is not more planning text — it is helping you decide what should be built before an agent starts building it.
Plan modes help an agent think before it codes, but the plan stays a transient message inside one conversation. PlanX adds the artifact around it: versions, word-level diffs, line comments, direct edits, folding, human approval and cross-agent hand-offs. It also expects you to leave plan mode — the accept/reject gate is where PlanX starts, not where it ends. See Planning.
Yes, MIT-licensed. The CLI is published as @thisisnsh/planx on npm, and the
source, contribution guide and security policy are in the
repository.
Yes. Start a new Codex session and use $planx <task>. See
The skill.
Yes. Start a new Claude Code session and use /planx <task>. See
The skill.
That is how each agent spells a skill invocation. PlanX uses the right one when it composes a command for you.
Yes. Any command that accepts a trailing prompt can be a hand-off target, set
with planx defaults, provided the receiving agent has the PlanX skill
installed. See Custom agents.
Yes — that is the point of storing the plan as a file. Claude Code can research and plan, Codex can build, and the version number is the whole hand-off. See Custom agents.
Start a new agent session. A session already running does not pick up a skill written underneath it. See Troubleshooting.
planx for the picker, planx <plan-id> for a plan, planx <plan-id> v<n> for
an exact version. See Reviewing.
Press ? in a review. The hint bar along the bottom shows only what works on
the row you are on. Every key is also tabulated in Reviewing.
v to start a selection, ↑ ↓ to extend, f to write the comment, enter
to save. See Feedback and edits.
n. Use line feedback for a passage and the note for something that applies
everywhere. See Feedback and edits.
e on the line, or on a selection to walk it line by line. An edit says this
is the wording, I have decided — the agent reproduces it rather than
interpreting it. See Feedback and edits.
Empty it: f, clear the text, enter.
space folds the section, note or unchanged diff run under the cursor, h
folds every note at once, j jumps between comments, and g / G reach the
ends. See Reviewing.
##, ### and #### are exactly what the review can fold, and the heading is
also the label a comment comes back under. A plan divided finely enough folds
where you need it and comes back with feedback you can place. See
Planning.
That the version was reviewed and there is nothing to change. PlanX does not create an identical revision for it — that version is ready to build. See Feedback and edits.
Open the newer one and PlanX shows the diff first; d toggles it, ← and →
move between versions. Changed words are highlighted and unchanged runs
collapse. See Versions and diffs.
planx diff <plan-id> v1 v3 --print, with --plain for a raw unified diff or
--stat for the summary line. See Versions and diffs.
planx list, or planx and → on the plan. planx <plan-id> v<n> opens one
directly. See Versions and diffs.
Because those commands arrive as a line copied out of a review of one particular
version, and the plan may have gained another since. Defaulting would silently
retarget the newest, which is how an agent ends up building a version nobody
read. latest still works — it just has to be said.
s, then pick a revise hand-off. PlanX submits the line comments, the edits and
the note together, and can resume the session that wrote the plan so the agent
revises with its repository research still in context. See
Hand-offs.
s, then an execute hand-off. It starts from the stored version rather than a
summary remembered from chat, and it records the session that ran it. See
Hand-offs.
No. Handing a plan on with its feedback open says build it, with these — the execute payload carries every comment, and the agent works them into the code.
planx, → for the versions, ctrl+r on the green one. PlanX starts that
session back up with the same flags, in the plan's own directory, and sends it
no prompt — the first thing it hears is whatever you type. See
Resuming a build.
No session was recorded, the agent is not one PlanX can launch, or the plan row
has two builds behind it — press → and pick the version. See
Resuming a build.
Separate planning from execution. Capture the plan, exit the agent, review it in the terminal, and request revisions until the diff matches your decisions. Execute only the exact version you approved.
It makes scope changes visible before code is written. When an agent revises the approach, the next version is a diff rather than a replacement in chat, and you can comment on the changed lines or rewrite them before execution.
~/.planx, one directory per plan. --dir or PLANX_DIR points at a different
store. See Installation.
No. index.json is a cache the list and the picker read instead of opening
every plan, so a file changed behind it leaves the two disagreeing. Every change
goes through the CLI, and planx doctor is what repairs the index. See
Troubleshooting.
ctrl+d in the picker, which confirms first. There is no trash behind it. See
Versions and diffs.
Start here
- Installation
- The skill
- Why review an agent's plan
- Plan mode vs. a reviewable plan
- Planning with Claude Code
- Planning with Codex
Using PlanX
Reference
Ask