TL;DR If you launch Claude Code with
--allow-dangerously-skip-permissions, plan mode in the terminal only asks Claude not to edit: nothing stops it. Leave with/exitand come back withcommand claude --continue: same conversation, still in plan mode, no bypass, and the blocks are back.
Plan mode is two things at once: an instruction telling Claude to research and propose without touching anything, and blocks Claude Code puts on every edit and every command until you approve the plan. In an interactive terminal session with bypass permissions available, Claude Code drops the blocks and keeps only the instruction. If Claude tries to edit a file or run a command while it plans, it runs without asking you.
When bypass is available
You never have to enter bypass mode. It's enough for your session to have it in the Shift+Tab cycle, and that happens if you launch with any of these:
--allow-dangerously-skip-permissions. It's the way to keep bypass close without turning it on, and the CLI reference shows it asclaude --permission-mode plan --allow-dangerously-skip-permissions: exactly the combination where plan mode stops holding Claude back.--dangerously-skip-permissionsor--permission-mode bypassPermissions, thenShift+Tabround to plan."defaultMode": "bypassPermissions"in~/.claude/settings.json, in managed settings, or in a file you pass with--settings. In the project's.claude/settings.jsonit no longer counts.
The status bar won't tell you: it says ⏸ plan mode on either way. To find out which one you're in, press Shift+Tab all the way round the cycle. If at any point you see ⏵⏵ bypass permissions on, your session has bypass available and your plan mode blocks nothing.
It only happens in the interactive terminal. In claude -p, the Agent SDK, and the VS Code extension's chat panel, plan mode keeps its blocks even if you launch with --allow-dangerously-skip-permissions: there the flag only leaves bypass to pick later.
The note that gives it away
Since v2.1.299, Claude Code tells you when you enter plan mode. This is the text, verbatim from the v2.1.328 binary:
Session was started with bypass permissions, so plan mode won't ask before most edits or commands. Start a new session without bypass permissions for prompts in plan mode.
In the binary, the condition that shows the note is the same one that drops the blocks: plan mode, bypass available, interactive session. That's why it says "started with bypass permissions" even if you only had it in the cycle. Three details from the code:
- It shows once per session, under a second after you enter plan mode. The second time you
Shift+Tabround to plan, it's gone. - It isn't saved in the transcript: scroll back through a resumed session and it isn't there.
- It's a note in the conversation, not a dialog. It asks you nothing and doesn't stop Claude.
What still asks
With bypass available, plan mode behaves like bypass. The only things still holding Claude back are the ones bypass doesn't skip either:
- Explicit
askrules in your settings. rmorrmdiron critical paths, such as/or your home directory.denyrules, which block in every mode.
Claude still gets the instruction to plan without editing, and it almost always follows it. But if you use plan mode to force yourself to understand the change before anything is touched, "almost always" isn't the point: you want an edit to be impossible until you say yes.
How to get the blocks back
1. Resume the same conversation without the flag
Leave with /exit and, in the same folder:
command claude --continue
--continue resumes the session in the mode you left it in, and a session that ended in plan mode comes back in plan mode. What it doesn't bring back is the launch flags, so bypass is no longer available and edits wait for you to approve the plan again. command skips your alias if that's where the flag lives; with no alias, plain claude --continue does it.
2. Take it out of your everyday launch
If the flag is in your alias, move it to a separate one for when you work inside a container:
alias claude-bypass='claude --allow-dangerously-skip-permissions'
If what you have is "defaultMode": "bypassPermissions" in ~/.claude/settings.json, step 1 isn't enough: --continue reads that file again. Remove the line, or switch to another default mode.
3. Turn it off on that machine
If you never use bypass outside a container, lock it in ~/.claude/settings.json:
{
"permissions": {
"disableBypassPermissionsMode": "disable"
}
}
With this setting Claude Code rejects --dangerously-skip-permissions and bypass stops being available, so plan mode always blocks. The strictest value wins for this setting, so a repository can't open it back up. It's the same setting an admin puts in managed settings for a whole team.
4. If what you wanted was to stop approving every command
Bypass is usually there so you don't approve commands one by one. Auto mode gets you that without giving up plan mode: with auto mode available and useAutoModeDuringPlan on, as it is by default, a classifier reviews shell commands in plan mode instead of asking you, and edits stay blocked until you approve the plan.
Reference
| How you launch, in the terminal | Bypass available? | Plan mode |
|---|---|---|
claude |
No | Blocks edits and commands |
claude --allow-dangerously-skip-permissions |
Yes | Doesn't block |
claude --dangerously-skip-permissions |
Yes, and you start in it | Doesn't block |
defaultMode: "bypassPermissions" in ~/.claude/settings.json |
Yes, and you start in it | Doesn't block |
Any of these, with disableBypassPermissionsMode: "disable" |
No | Blocks |
command claude --continue after a session with the flag |
No | Blocks |
claude -p, the Agent SDK, or the VS Code chat |
Whatever the flag says | Always blocks |
All six modes, and how Shift+Tab cycles them, are in the 6 permission modes.
Official docs: Analyze before you edit with plan mode · Skip all checks with bypassPermissions mode
Requirements
- The note arrives in v2.1.299.