[Jordan Harband (ljharb)]

Adds the eight skills the "Socket Skills for Coding Agents" spec lists as gaps. Each one is a single skills/<name>/SKILL.md, checked against socket@1.14.1 source (the current npm latest; GitHub's "latest release" still points at v1.1.146), Socket Firewall, socket-mcp, and docs.socket.dev.

Skill	What it does
socket-firewall	Runs installs through sfw, explains a block and what to do next, and covers Enterprise wrapper and registry modes, CI, and Docker.
socket-pr-alerts	Triages Socket for GitHub checks and comments, checks reachability, then fixes, replaces, or accepts an alert for one PR or for good.
socket-reachability	Runs full application reachability, reads .socket.facts.json, and fixes reachable vulnerabilities first.
socket-incident-response	Confirms a malware report in the threat feed, finds usage and dependency paths, checks for install-time execution, contains it, rotates secrets, and reports.
socket-policy	Edits socket.yml and changes org alert actions, dry-running where the API supports it and asking before any real write.
socket-sbom	Generates CycloneDX locally with cdxgen and exports CycloneDX or SPDX for an existing scan.
socket-vet-agent-tools	Vets an MCP server, skill, or plugin before install using existing Socket checks.
knowledge-update	Dated corrections for outdated model knowledge about Socket products, each with a source.

Also: Claude marketplace entries (all depend on socket-connect except knowledge-update), the skill discovery test's expected list, the README's installable-names line, and the regenerated agents/README.md and Codex manifest.

Points for review

• socket-vet-agent-tools: the spec marks this as needing product confirmation. No Socket product scans an MCP server, skill, or plugin as a unit, so the skill combines existing checks (package scores and alerts, published-file reads, the skills.sh Socket audit, a --no-auto-manifest --tmp scan, install through sfw) and says so.
• Plan gates: the spec says the plan feature gates are unconfirmed, so no skill states which plan unlocks a feature. When the CLI or API reports a feature is unavailable, the skills say so plainly and link to https://socket.dev/pricing.
• socket-policy uses an org API token from SOCKET_CLI_API_TOKEN for the Policies REST API, because the CLI has no Policies command. The skill says this differs from the browser sign-in in socket-connect, dry-runs the endpoints that support it, never prints the token, and asks before the legacy security-policy write, which has no dry run.
• .socket.facts.json has no published schema. socket-reachability reads only the fields used in socket-cli's own tests and says the shape may change.

Checks

• pnpm test: 57 files, 1,376 tests pass. The baseline on main is 1,264; the increase is the per-skill structural tests.
• pnpm run check passes: format, lint, type, paths, release-kit, shipped content, generate, and scan.
• Each skill went through a separate verification pass that tried to refute every command, flag, and claim against source, and the confirmed issues were fixed. One example: an outdated socket fix flag that the CLI would silently ignore and then apply fixes.

---

Note

Low Risk
Changes are new agent documentation and marketplace metadata; no runtime application or security-critical code paths are modified.

Overview
Adds eight new agent skills (socket-firewall, socket-pr-alerts, socket-reachability, socket-incident-response, socket-policy, socket-sbom, socket-vet-agent-tools, and knowledge-update), each as a new skills/<name>/SKILL.md with step-by-step workflows aligned to Socket CLI 1.14.x, Firewall, MCP, and current docs.

Wiring and discovery are updated so the skills are installable everywhere the repo ships: new entries in .claude-plugin/marketplace.json (most depend on socket-connect), the Codex bundle in skills/.codex-plugin/plugin.json, regenerated agents/README.md and README skill tables/install name list, and an expanded EXPECTED_TOP_LEVEL list in the skill-discovery structural test.

^{Reviewed by Cursor Bugbot for commit df424d9. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{Bugbot Autofix is ON, but it could not run because the branch was deleted or merged before autofix could start.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit df424d9. Configure here.}

[cursor]

Firewall test leaves temp working directory

Medium Severity

The “confirm it blocks” steps create a temp directory and cd into it, then never restore the original working directory. Later steps still change project dependencies and rerun installs, so those commands can run against the empty temp project instead of the user’s repo.

 

^{Reviewed by Cursor Bugbot for commit df424d9. Configure here.}