Configures Claude Code permissions for low-prompt work inside a project; changes nothing under Codex. Creates or merges allow/deny rules in .claude/settings.json and installs a PreToolUse hook gating out-of-project writes and deletes, precious untracked files, and history-changing git on protected branches. Writes only under .claude/, including an ownership record. Use after /optimus:init or standalone.
Installs into .claude/skills of the current project.
Are you the author of Permissions?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/oprogramadorreal-permissions)
---
description: Configures Claude Code permissions for low-prompt work inside a project; changes nothing under Codex. Creates or merges allow/deny rules in .claude/settings.json and installs a PreToolUse hook gating out-of-project writes and deletes, precious untracked files, and history-changing git on protected branches. Writes only under .claude/, including an ownership record. Use after /optimus:init or standalone.
disable-model-invocation: true
---
# Optimus Permissions
Under Codex, stop without changing files: this skill configures Claude Code only. Recommend configuring Codex's own sandbox and approval policy instead.
Configure permission rules and a path-restriction hook so Claude Code agents can work autonomously inside the project without constant prompts, while destructive operations outside it stay gated.
To remove this setup while keeping init's artifacts, use `/optimus:reset permissions`.
`$CLAUDE_PLUGIN_ROOT` below means this skill's base directory minus `/skills/permissions`. Write that absolute path into every Bash command, where the variable can be empty. Read `$CLAUDE_PLUGIN_ROOT/skills/init/references/managed-files.md` and apply it to the hook and every settings entry this run adds or removes. Do not write `.claude/.optimus-version`.
Security model in brief: the installed hook prompts on writes and blocks deletes outside the project (Claude's memory store and session scratchpad are exempt), asks before editing any precious unversioned file and blocks deleting them (backups and IDE scratch files stay deletable), and blocks history-modifying git operations on protected branches. Inside the project, operations not on the deny list run without prompts.
## Step 1: Detect existing configuration
1. If `.claude/settings.json` exists, read it in full; if it is not valid JSON, show the parse problem and ask the user how to proceed before changing any file.
2. Note whether `.claude/hooks/restrict-paths.sh` already exists (fresh install vs update).
3. If `.mcp.json` exists at the project root, extract the top-level MCP server names for Step 3.
## Step 2: Install the path-restriction hook
If an existing `.claude/hooks/restrict-paths.sh` differs from the template, apply managed-files.md's **Overwrite and migration** rules: replace a recorded `refresh: "template"` copy whose hash still matches without asking. Otherwise read it, show a concrete comparison that separates customizations from older-template drift, and use `AskUserQuestion` for its Merge / Keep existing / Replace choice. Keep existing also reports the remaining version/behavior difference; Replace discards only the differences explicitly shown.
When creating or replacing as selected, `cp` `$CLAUDE_PLUGIN_ROOT/skills/permissions/templates/hooks/restrict-paths.sh` to `.claude/hooks/restrict-paths.sh` (never re-type it through Read/Write), then apply only approved customizations. Keep-existing does not authorize a later verification step to overwrite that file.
## Step 3: Create or update settings.json
Base configuration: `$CLAUDE_PLUGIN_ROOT/skills/permissions/templates/settings.json`.
If `.claude/settings.json` does not exist, create it from the template. If it exists, **merge** — read, merge, write, never blind-overwrite:
1. **permissions.allow** — add template entries not already present. Never remove existing entries.
2. **permissions.deny** — add template entries not already present. If the existing settings have git deny entries (git as a command, not part of words like `github`) beyond the template's set, they may block the feature-branch workflow (commit/push) that skills like /optimus:tdd need — list them and use `AskUserQuestion`: **Replace with template set (Recommended)** — remove only the extra git deny entries and use the template's (branch protection is still enforced by the hook); non-git deny entries untouched — or **Keep all**.
3. **hooks.PreToolUse** — add the template's hook entry as a separate matcher group, appending to any existing array; skip if an entry already references `restrict-paths.sh`.
4. **Preserve everything else** — existing PostToolUse hooks, custom sections, all unrelated configuration.
In either case, if `.mcp.json` was found, add `mcp__<server-name>` entries to `permissions.allow` for each server. The result must be valid JSON.
## Step 4: Verify and report
Fix any issue before reporting:
1. For an installed/replaced hook, `.claude/hooks/restrict-paths.sh` matches the template plus only approved customizations — `diff` against it and run `bash -n` on the installed copy. Correct accidental copy errors; preserve a file the user chose to keep.
2. `.claude/settings.json` has a PreToolUse entry that resolves to the installed `restrict-paths.sh` — a merge that drops or misspells it leaves the project unprotected with no other symptom.
3. Scan for precious unversioned files. Get the rules with `sed -n '/^strip_backup_suffix() {/,/^}/p;/^is_precious_name() {/,/^}/p;/^is_recoverable_precious_name() {/,/^}/p' .claude/hooks/restrict-paths.sh` instead of reading the whole hook. Exclude `.git/`, dependencies, build output, and `.claude/.optimus-reset-backups/`. Report untracked matches as protected, distinguishing the intentionally deletable backup/IDE category. Offer specific custom patterns for sensitive files the scan misses.
4. `.claude/.optimus-managed.json` matches the hook written and the settings entries this run added.
Report: files created or updated (fresh install vs update), allow/deny counts, detected MCP servers, and the one-line security summary from the top of this skill. Point to this skill's README for the trust model, the auto-mode layering, and the not-OS-sandboxing caveat.
If `.claude/CLAUDE.md` does not exist, recommend `/optimus:init` next; otherwise `/optimus:unit-test` to establish coverage or `/optimus:tdd` to start developing. Suggest a fresh conversation for it.