Cursor is now publishing plugins into Claude Code's marketplace, and the newest one asks the agent to formally verify the hardest thing in your repo
The cursor/plugins repo hit GitHub's daily trending board over the weekend with 15 commits in a day and 12,705 stars, and the thing worth noticing is not any single plugin. It is the file at .claude-plugin/marketplace.json, owner Cursor, contact [email protected], which exposes two of Cursor's plugins to Claude Code's plugin marketplace: origin-apps, for building on the Cursor Origin API, and pstack, whose source path is harness/pstack/claude-code. That directory was first committed on October 16 with a PR titled install pstack in Claude Code from a generated tree, and has had twelve commits since, the latest on Saturday night. Android published its skills into the Claude Code and Codex marketplaces last week. Cursor doing it is a different signal, because Cursor sells the competing harness.
The repo itself is Cursor's official plugin spec plus about seventeen first-party plugins and fifty-five third-party integrations, each a directory with a .cursor-plugin/plugin.json manifest. The first-party list reads like a catalog of this year's agent patterns: continual-learning does incremental transcript-driven memory updates to AGENTS.md using high-signal bullet points only, thermos is a thermo-nuclear branch review with parallel subagents, ralph-loop is the Ralph Wiggum iterative loop, orchestrate fans large tasks across parallel cloud agents with planners, workers, verifiers and structured handoffs, advisor consults a stronger model before major decisions, cli-for-agent is a pattern guide for CLIs that agents can run reliably. All of these are text files. The spec is the product; the plugins are examples of what a vendor thinks the spec should carry.
The PR that landed on Saturday is the one to read. Number 709 adds /create-formal-verification-skill to pstack, and the body, written by a Cursor agent, opens with: a test checks the cases that someone thought to write; races, retries, state machines, parsers and unsafe code often break on a case that nobody wrote. The skill chooses the strongest target in the repo, not the easiest one to verify, picks a verification tool and says why, verifies one property and shows that property fail on a deliberately broken variant of the code, then writes a project-local skill so the next agent can tell whether a change touches the verified code, rerun the check with one command, and tell a real bug from a modeling error. It then follows its own new skill once on a scratch change that breaks the property, confirms the check catches it, and drops the change. It does not weaken the property or the model to get a pass. A follow-up commit on Saturday night added formal verification as an option during pstack setup.
Read that next to the week's other items and the direction is obvious. Sphinx gates the human's understanding at push time. The Claude Code team's review post and the Postgres 19 delay say review is the bottleneck. vLLM says agent-written kernels outrun its reviewers and wants automatic verification. Cursor's answer, shipped as a markdown skill into Anthropic's marketplace, is to have the agent build a verifier for the one invariant that would hurt most, and leave it behind as a skill for the next agent. The harness that owns the editor is distributing its verification habits through the rival's package manager, because the skill file is now the unit of distribution and the harness is just where it runs.
Links: https://github.com/cursor/plugins , https://github.com/cursor/plugins/pull/709 , https://github.com/cursor/plugins/blob/main/.claude-plugin/marketplace.json
The repo itself is Cursor's official plugin spec plus about seventeen first-party plugins and fifty-five third-party integrations, each a directory with a .cursor-plugin/plugin.json manifest. The first-party list reads like a catalog of this year's agent patterns: continual-learning does incremental transcript-driven memory updates to AGENTS.md using high-signal bullet points only, thermos is a thermo-nuclear branch review with parallel subagents, ralph-loop is the Ralph Wiggum iterative loop, orchestrate fans large tasks across parallel cloud agents with planners, workers, verifiers and structured handoffs, advisor consults a stronger model before major decisions, cli-for-agent is a pattern guide for CLIs that agents can run reliably. All of these are text files. The spec is the product; the plugins are examples of what a vendor thinks the spec should carry.
The PR that landed on Saturday is the one to read. Number 709 adds /create-formal-verification-skill to pstack, and the body, written by a Cursor agent, opens with: a test checks the cases that someone thought to write; races, retries, state machines, parsers and unsafe code often break on a case that nobody wrote. The skill chooses the strongest target in the repo, not the easiest one to verify, picks a verification tool and says why, verifies one property and shows that property fail on a deliberately broken variant of the code, then writes a project-local skill so the next agent can tell whether a change touches the verified code, rerun the check with one command, and tell a real bug from a modeling error. It then follows its own new skill once on a scratch change that breaks the property, confirms the check catches it, and drops the change. It does not weaken the property or the model to get a pass. A follow-up commit on Saturday night added formal verification as an option during pstack setup.
Read that next to the week's other items and the direction is obvious. Sphinx gates the human's understanding at push time. The Claude Code team's review post and the Postgres 19 delay say review is the bottleneck. vLLM says agent-written kernels outrun its reviewers and wants automatic verification. Cursor's answer, shipped as a markdown skill into Anthropic's marketplace, is to have the agent build a verifier for the one invariant that would hurt most, and leave it behind as a skill for the next agent. The harness that owns the editor is distributing its verification habits through the rival's package manager, because the skill file is now the unit of distribution and the harness is just where it runs.
Links: https://github.com/cursor/plugins , https://github.com/cursor/plugins/pull/709 , https://github.com/cursor/plugins/blob/main/.claude-plugin/marketplace.json
β Previous
Anthropic measured its own supervision bottleneck: agent hours outnumber employee hours six to one, and 80 percent of those hours are the agent waiting for a human
Next β
Vals.ai gave six assistants an impossible refund and no supervision: Grok wrote to the MP 50 times, Fable leaked a family illness to five organisations, all six asked the bank for a chargeback
Comments