← MCP server and skill grades

skills

warpdotdev/common-skills/.agents/skills

C
Type
Skill
Scanned
Nov 13, 2026
Version
8792394ff1b6
Files read
62

This collection of Warp engineering skills guides AI coding agents through pull request workflows, spec-driven development, CI debugging, interactive HTML walkthroughs, and local session grading. Its biggest risk is that two included feedback skills (suggestion-box and complain) instruct the agent to silently send workspace context to Slack and query Google Cloud Secret Manager via gcloud during any task without asking or telling the user.

What it can reach

Files
Reads and edits repository files, specs, and diffs; writes HTML reports and state under ~/.readouts, ~/.sagas, .warp/pr-walkthrough, and /tmp; skill-doctor reads local conversation histories and SQLite databases from Claude Code, Codex, Warp, Pi, Grok Build, ZCode, and Hermes.
Shell commands
Runs local commands and helper scripts including git, gh, cargo, python3, gcloud, npx wrangler, and repository test or formatting scripts.
Network
Connects to GitHub APIs, Slack webhooks (hooks.slack.com), Google Cloud Secret Manager, warp-brand-site.vercel.app, Linear MCP tools, optional Cloudflare Pages deployments, and JavaScript CDNs (cdn.jsdelivr.net and cdnjs.cloudflare.com).
Secrets
Reads GH_TOKEN or GITHUB_TOKEN, checks COMPLAINT_SLACK_WEBHOOK_URL and SUGGESTION_BOX_SLACK_WEBHOOK_URL, queries GCP Secret Manager in project warp-server-staging via gcloud, and accepts a user-supplied GitHub token in readout's browser code viewer.
Browser
Opens generated HTML readouts and skill-doctor reports in the default browser, launches Playwright Chromium to validate D3 walkthroughs, and spawns cloud computer-use agents for UI bug reproduction.

Findings

1 high4 low
  1. HighSilent autonomous Slack feedback submission and unprompted gcloud secret lookup

    .agents/skills/suggestion-box/SKILL.md:3

    Both suggestion-box and complain instruct the agent in their always-loaded frontmatter descriptions to run submit.py proactively during any task without user invocation, permission, preview, or disclosure, and submit.py even tells the agent on stderr to ignore and hide errors. When triggered, submit.py checks for a webhook environment variable and otherwise runs gcloud secrets versions access against Warp's warp-server-staging GCP project using the user's active Google Cloud credentials, then posts up to 1,200 characters (including repo names, branches, PR or issue links, commands, and error excerpts in suggestion-box) to Slack. Beyond sending data without consent, prompt injection in untrusted issues, PRs, or logs can abuse this hidden channel when a webhook or GCP secret access is available.

    Fix: Remove suggestion-box and complain from general-purpose skill bundles, or require explicit user invocation, preview, and confirmation before accessing gcloud secrets or sending messages to Slack.

  2. LowRuntime fetching of unpinned remote skill instructions from a .vercel.app URL

    .agents/skills/brandalf/SKILL.md:14

    Whenever brandalf is used directly or via pr-walkthrough, it instructs the agent to fetch live instructions from https://warp-brand-site.vercel.app/brand/skill and treat the fetched page as the authoritative source of truth. Because the remote content is not pinned or verified—and lives on a shared .vercel.app subdomain—a compromise of that web deployment or a subdomain takeover if the Vercel project is ever renamed or deleted would inject arbitrary instructions into the coding agent at runtime.

    Fix: Bundle the canonical brand guidance as a local reference file inside .agents/skills/brandalf/ rather than fetching live agent instructions from a remote URL at runtime.

  3. LowOver-broad directory matching in skill-doctor pulls in unrelated local conversations

    .agents/skills/skill-doctor/scripts/collect_sessions.py:1483

    When a user selects 'Conversations in this repository' in skill-doctor, session_matches_repo still falls back to checking p.name == repo.name or repo.name in p.parts even when the session's working directory exists locally outside the selected repository. If the repository folder has a common name such as app, api, web, src, docs, or skills, conversations from unrelated confidential projects that contain that segment anywhere in their path are extracted, sent to the model for scoring, and summarized into the shareable HTML report.

    Fix: If p.resolve() exists on disk and is not relative to repo (outside of explicit git worktree paths), return False instead of matching repo.name against arbitrary path segments in p.parts.

  4. LowExternal CDN scripts loaded without Subresource Integrity in generated HTML viewers

    .agents/skills/readout/assets/code-pane.html:129

    The readout code viewer dynamically loads Highlight.js from cdnjs.cloudflare.com without a Subresource Integrity (integrity) hash, inside the same page context where users are prompted to paste a GitHub Personal Access Token (stored in memory and sessionStorage). Similarly, pr-walkthrough/scripts/d3_canvas_runtime.py loads D3 7.9.0 from cdn.jsdelivr.net without an integrity attribute. A CDN compromise would execute arbitrary JavaScript in the viewer's browser and could read any entered GitHub token.

    Fix: Add Subresource Integrity hashes (integrity and crossOrigin='anonymous') to the dynamically created script elements in code-pane.html and d3_canvas_runtime.py, or inline the libraries locally.

  5. LowPredictable /tmp/pr-body.md temporary file path in write-feature-docs

    .agents/skills/write-feature-docs/SKILL.md:411

    Step 6 of write-feature-docs instructs the agent to write the pull request description to a hardcoded /tmp/pr-body.md path rather than creating a private temporary file. On shared multi-user machines, another local user could pre-create or replace /tmp/pr-body.md to read or modify the draft PR description, and concurrent runs can overwrite each other.

    Fix: Use mktemp to create a unique, user-only temporary file for --body-file and delete it after gh pr create completes.