skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 22f0b463e7a5
- Files read
- 80
This skill library guides AI coding agents through structured workflows for brainstorming, planning, git worktrees, test-driven development, subagent delegation, code review, and session debugging. Its biggest risk is that the workspace setup skill instructs the agent to automatically run dependency installers (including unisolated pip installs) and test suites even if worktree creation is declined, which can run untrusted code when working on an unfamiliar repository.
What it can reach
- Files
- Reads and writes project files, git worktrees, .superpowers workspace directories, temporary directories under /tmp, and coding-agent session transcripts on disk during diagnostics.
- Shell commands
- Executes git commands, package managers (npm, pip, poetry, cargo, go), test suites, Graphviz (dot), and bundled Bash and Node.js helper scripts.
- Network
- Starts a key-protected local HTTP and WebSocket server for visual brainstorming, fetches a branding and telemetry image from primeradiant.com, and searches or files GitHub issues and pull requests via gh or curl.
- Browser
- Opens the user's default web browser to display interactive HTML mockups when the visual brainstorming companion is accepted.
- System
- Checks and signals local process IDs via ps, /proc, and kill to monitor and stop the background brainstorming server.
Findings
-
MediumWorkspace setup runs install scripts and unisolated pip install even if declined
skills/using-git-worktrees/SKILL.md:114
When setting up a workspace before feature work or plan execution, the skill asks only whether to create a git worktree—and instructs the agent that even if the user declines, it should proceed to Step 2 and Step 3, which automatically run npm install, cargo build, pip install -r requirements.txt, poetry install, and the project's test suite. In a repository cloned from an untrusted source, these commands can execute arbitrary code via package install hooks, build scripts, or tests, and bare pip install modifies the user's active or global Python environment outside any isolated virtualenv.
Fix: Require explicit user confirmation before installing packages or running tests in unfamiliar repositories, stop if workspace setup is declined, and run pip inside a virtual environment.
-
LowUndocumented telemetry request to primeradiant.com in visual companion
skills/brainstorming/scripts/server.cjs:106
When the user opens the local visual brainstorming companion in their browser, every framed page and waiting screen loads an external logo image from primeradiant.com with the installed Superpowers version in the query string. Although no-referrer prevents leaking the session key, this sends the user's IP address, browser User-Agent, and version to an external server by default without being mentioned in the skill documentation, where the opt-out environment variables (such as SUPERPOWERS_DISABLE_TELEMETRY) are also undocumented.
Fix: Serve a local SVG or PNG asset from skills/brainstorming/scripts/ instead of fetching from primeradiant.com, or document the external request and SUPERPOWERS_DISABLE_TELEMETRY opt-out in visual-companion.md.
-
LowPredictable /tmp session directory vulnerable to local pre-creation
skills/brainstorming/scripts/start-server.sh:130
When start-server.sh is run without --project-dir, it constructs a predictable /tmp/brainstorm-<pid>-<timestamp> path and creates subdirectories with mkdir -p (and server.cjs falls back to /tmp/brainstorm if run directly). On a multi-user machine, another local user can pre-create matching directories under /tmp so they own the parent session folder and can swap its content or state subdirectories to inject HTML into the browser view or spoof click events read by the agent.
Fix: Create ephemeral session directories atomically with mktemp -d "${TMPDIR:-/tmp}/brainstorm.XXXXXXXXXX" in start-server.sh and fs.mkdtempSync in server.cjs.
-
LowUnprompted GitHub issue search sends session symptom terms before redaction
skills/diagnosing-superpowers/SKILL.md:48
During local session diagnosis, Step 5 instructs the agent to automatically search GitHub issues on obra/superpowers via gh or curl whenever Superpowers involvement is rated possible or likely, even if the user only asked why a session went wrong and did not request a bug report. Because redaction and scrubbing happen later in Step 6, searching for session symptoms in Step 5 can send unredacted error text or terms to GitHub's API without asking and pulls public issue titles into the agent's context.
Fix: Run the GitHub issue search only after the user asks to check or file an issue, and scrub search terms before sending them to GitHub.
-
LowSkill creation checklist mandates git push without explicit user approval
skills/writing-skills/SKILL.md:385
The writing-skills workflow marks its deployment checklist as mandatory for every created or edited skill and instructs the agent to create a todo item to commit the skill and push to the user's fork if configured, without requiring user confirmation before pushing. Unlike finishing-a-development-branch, which waits for explicit permission before pushing, an agent following this checklist on a repository with a configured remote could push private or work-in-progress skills externally without asking.
Fix: Change the deployment checklist item to ask the user for confirmation before committing or pushing any skill changes to a remote repository.
