Plan Mode
ArmorClaude integrates with Claude Code's built-in plan mode for user-approved intent plans
Plan Mode Integration
ArmorClaude works with Claude Code's built-in plan mode, where Claude writes a plan for your approval before executing.
How to Use
- Press Shift+Tab twice in Claude Code. You will see "plan mode on" at the bottom
- Type your prompt (e.g. "Add a summary file for /etc")
- Claude researches (read-only), then writes a markdown plan
- ArmorClaude instructs Claude to include a fenced JSON intent block in the plan
- You approve the plan
- Claude calls
register_intent_planwith the approved plan - ArmorClaude checks policy and, when intent enforcement is active, whether the tool is in the approved plan
What the Plan Looks Like
A plan file can include a JSON block such as:
{
"goal": "Create etc-notes.md with a one-line summary of /etc",
"steps": [
{
"action": "Write",
"description": "Create the summary file",
"arguments": {
"file_path": "/tmp/demo/etc-notes.md",
"content": "/etc contains system-wide configuration files.\n"
}
},
{
"action": "Read",
"description": "Verify the file was written",
"arguments": { "file_path": "/tmp/demo/etc-notes.md" }
}
]
}Two Plan Paths, Same Enforcement
| Path | When | How plan is captured |
|---|---|---|
| MCP tool (default) | Regular prompts | Claude calls register_intent_plan to record its intended actions |
| Plan mode | Shift+Tab x2, then approve | ArmorClaude reads the JSON block from the plan file |
Both paths produce a normalized plan. Read tools still pass policy checks without needing a plan. Policies with default allow and only permit statements, such as all-allow, record intent without blocking drift. Empty default-deny policies do not qualify. Input parameters are advisory in the current configuration.
In practice, Claude often uses both paths: it writes the JSON block in the plan file AND calls register_intent_plan after approval. The MCP path wins (consumed first), and the plan file path is a backup.