Plan Guard

plan-guard

Status: Abandoned experiment. This plugin was an attempt to warn users when entering plan mode with a non-Opus model. Due to Claude Code's plan mode being largely handled through internal UI flows (not hookable tool calls), the interception is unreliable — most plan mode entries and finalizations bypass the plugin entirely.

A simpler alternative: configure the status line to display the current model. This makes it easy to glance at the model before planning, without needing a hook. Example status line config in ~/.claude/settings.json:

"statusLine": {
  "type": "command",
  "command": "echo $input | jq -r '.model.display_name // .model.id'"
}

Asks for confirmation when entering plan mode with a non-Opus model. Planning works best with Opus — this plugin helps prevent accidentally generating plans with less capable models.

How it works

The plugin uses a combination of hooks to detect and warn about non-Opus plan mode usage:

Hook	Event	What it does
save-model.sh	SessionStart	Captures the current model and saves it to a temp file
check-plan-model.sh	PreToolUse on EnterPlanMode	Asks for confirmation if the model isn't Opus
check-plan-model.sh	PreToolUse on ExitPlanMode	Warns that the plan was generated by a non-Opus model
tag-plan-file.sh	PostToolUse on Write	Prepends a warning header to plan files in ~/.claude/plans/
cleanup.sh	SessionEnd	Removes the temp file

Limitations

The plugin provides best-effort protection, not guaranteed interception. Claude Code's plan mode can be entered and finalized through multiple paths, not all of which are hookable:

• Agent calls EnterPlanMode — hooks fire reliably. This is the main case (agent autonomously decides to plan).
• Agent explicitly calls Write + ExitPlanMode — both hooks fire. Plan files get the warning header, and the agent is told to warn the user.
• Built-in plan approval UI ("Ready to code?") — hooks are bypassed. When Claude Code uses its internal plan finalization flow, neither ExitPlanMode nor Write go through hookable tool calls.
• /plan command — enters plan mode immediately without triggering hooks. The subsequent prompt submission is not intercepted.
• shift+tab toggle — a UI-level toggle that doesn't fire any hooks.

In practice, the plugin is most effective when the agent initiates plan mode on its own (which is the case most likely to catch you off guard). User-initiated plan mode (/plan, shift+tab) is harder to intercept because it bypasses the tool call layer.

Installation

claude --plugin-dir ~/dev/cc-plugins/plan-guard

Dependencies

• jq — JSON parsing in hook scripts
• grep -P — PCRE support for model name extraction