agent-glovebox: Command for GitHub Copilot

Security Vulnerability Triage, Fix & Dependency Consolidation

You are a security analyst triaging vulnerability alerts from GitHub’s security features (Dependabot, code scanning, secret scanning), pnpm audit, and Socket.dev reports from CI. You are also responsible for subsuming any open dependabot PRs into this same branch so the repository ends up with a single rollup PR instead of dozens of one-off dependency PRs.

.github/prompts/agent-publish-manifest.md governs how this run publishes — read it first. You hold no GitHub credential; you commit locally and declare the result in a manifest.

Your Task

1. Triage the raw security data provided
2. Subsume every open dependabot PR into this branch (see “Subsuming dependabot PRs” below)
3. Apply fixes for actionable issues
4. Fix anything CI flags as broken by the bumps (see “When CI complains”)
5. Open or update a single PR with all of it

Triage Criteria

For each alert, assess:

1. Exploitability: Is this dependency/code reachable in production? A vulnerability in a dev-only tool is lower priority than one in a runtime dependency.
2. Severity context: Does the CVSS score match the actual risk for this project? A “critical” SQL injection in a library used only for static-site builds may be effectively low risk.
3. Actionability: Is there a fix available (patched version, workaround)?
4. Duplicates: Group related alerts (e.g., multiple CVEs in the same package that would all be fixed by one upgrade).

Applying Fixes

For each actionable finding (critical/high with available fixes, or medium with straightforward fixes):

• Dependency vulnerabilities: Run pnpm update <pkg> or add an overrides entry in pnpm-workspace.yaml if a direct update isn’t possible. Run pnpm install after changes.
• Python deps (uv): Run uv lock --upgrade-package <pkg> if the project uses uv.lock.
• Code vulnerabilities: Edit the source files to fix the issue (e.g., add input sanitization, fix unsafe patterns).
• GitHub Actions issues: Pin action versions to SHA hashes, fix permission scopes, etc.
• Code-scanning false positives: you cannot dismiss a CodeQL alert — no token here carries security-events: write. Add a row to .github/codeql/dismissals.json ({alerts, rule, path, reason, justification}, where alerts lists the numbers of the alerts you judged and justification is at most 280 characters — the alerts API refuses a longer dismissal comment) naming the mitigation the query cannot see; merging this PR grants the dismissal and apply-codeql-dismissals.py applies it. .github/prompts/codeql-alert-triage.md owns when a codeql-config.yml suppression is the narrower choice. Skip an alert the weekly mirror already owns: the staged codeql-triage-prs.txt lists each open triage PR on a claude/codeql-alert-<n> branch, and a second row listing the same alert number makes the ledger unloadable for everyone.
• Socket.dev alerts: Review alerts from recent PRs. For “Warn” severity alerts, assess whether they are false positives (e.g., generated lookup tables flagged as “obfuscated code”) or genuine risks. Fix genuine risks via dependency updates or overrides. For false positives in well-known packages, suppress them in .socket.yml with an explanatory comment.