Executing
Approving a plan does not build it. It seals it, and then tells you how to build it:
✓ Approved & sealed — guard-clock-regression-a3f9 v3 (6 sections locked)
Paste to your agent: /planx execute guard-clock-regression-a3f9 v3Paste that to an agent and the /planx skill takes the execute branch: it loads the stored version with planx show, drops it into the current context, and implements it there.
A slash command is for your agent; a bare command is for your terminal. That is the only thing that tells the two apart, so the review is careful about which form it prints. /planx execute is a branch of the skill and has never been a shell command — printing it as one is what used to make the line read as noise.
The other two exits print the same way:
s Paste to your agent: /planx revise guard-clock-regression-a3f9
x Reopen it with: planx guard-clock-regression-a3f9 v3What a sealed plan looks like
Every section carries a lock, e refuses, and the only key left that changes anything is l — if you really have to open one back up.
The same plan after a. It is sealed — every line carries a lock, and e refuses.
click to use your keyboard, or tap the keys below
In this session, not a new process
planx used to be able to spawn a fresh agent for you. It no longer does, and the CLI command that did it is gone.
Spawning a subprocess agent from inside an agent loses the context it already has, the permissions it was granted, and your ability to intervene — which is why even the old command's own help told agents not to call it. Outside an agent, it was a wrapper around typing claude yourself. Printing the command and letting you run it where you already are is the same thing with fewer parts.
Why there is no model picker
planx cannot change a running session's model, and neither can anything else: /model is a user-typed slash command, settings.json is read at session start, and ANTHROPIC_MODEL only affects new processes.
The old approve flow asked which agent you wanted, then which model, and then printed a slash command for you to paste — two prompts to arrive at a string it could not act on. If you want a different model, switch it the way you always would and then paste the command.
Executing an unapproved plan
The skill checks. If the plan was never approved it says so in one line and asks whether to go ahead anyway — an unreviewed plan running by accident is the exact thing planx exists to prevent.
Executing does not revise
If the plan turns out to be wrong once you are in the code, the agent stops and says so rather than quietly doing something else. Changing it is your decision, and it goes back through review: an approved plan is sealed, so every section of it is locked.