●Blog

Notes from the lockfile

On agent configuration, skill distribution, and the supply chain underneath them: what is happening in the ecosystem and what it means in practice. Subscribe via RSS.

A CLI's skill can now follow the version of the CLI you installed

Skillfold's @installed ref pins a skill shipped inside an npm package to the version your project has installed, and fails the check when they drift.

Read →

Claude Code now labels what a subagent tells you as subagent output

Claude Code 2.1.277 marks subagent results and cleans invisible Unicode from prompts. What that closes, and what pinning skill content still doesn't.

Read →

One manifest can now install skills and rules for Cursor

Skillfold's new Cursor target installs project skills and always-on rules in Cursor's native paths, with the same lockfile and drift checks.

Read →

Claude Code can score whether a plugin helps. It doesn't ask which copy you ran.

Claude Code's new plugin eval command measures a plugin's effect with a no-plugin baseline and gates CI on it - but the score never records which revision produced it.

Read →

Claude Code will tell you which skills are dead weight. It won't tell you if they drifted.

Claude Code's new /skill-doctor flags unused skills and their context cost per session - a usage read, not a check on what's actually declared or installed.

Read →

npm built a pause button for publishing. Most CI still skips it.

npm's staged publishing adds a human approval step before a package goes live, and a fresh trusted-publishing compromise shows what that step would catch.

Read →

Five vendors agreed on a plugin folder. Distribution is still your problem.

Agent Plugins 1.0 gives skills and MCP servers one folder shape across six clients, then leaves versioning, pinning, and installation to someone else.

Read →

NVIDIA is signing its agent skills. Here is what a signature buys you

NVIDIA now ships CUDA-X agent skills with a scan, a signature, and a machine-readable card. It is a real trust stack, and it stops well short of "safe."

Read →

A valid provenance attestation shipped malware. Here is how.

The July 2026 AsyncAPI npm compromise passed trusted publishing and provenance checks. What those checks actually proved, and what they never claimed to.

Read →

A study mapped 1.4 million agent skills. Most declare no dependencies at all.

A July 2026 paper built the first dependency graph across 1.4 million agent skills, and found most carry no declared dependency metadata at all.

Read →

Microsoft ships stable Agent Skills for .NET and Python

Microsoft's Agent Framework made SKILL.md a stable, two-language API this month, and what that does and does not mean for anyone managing skills.

Read →

MCP's stateless rewrite ships July 28: what actually changes

The Model Context Protocol drops server sessions for a stateless core and writes down a deprecation policy for the first time. Here is what changes.

Read →

What the skill supply-chain research actually recommends

Six months of disclosures converged on the same mitigations for agent skills. Here is what pinning and content hashes buy, and what they plainly do not.

Read →

One manifest, two agents

Claude Code and Codex read the same SKILL.md format but look for it in different places, and handle rules completely differently. The targets key makes one manifest install for both.

Read →

Composed skills, or how to stop copy-pasting prompts

Composition concatenates skills into one generated SKILL.md, carries their supporting files along, and regenerates whenever an input changes. Here is what it does and where the edges are.

Read →

Your skills directory is undeclared state

A .claude/skills directory has every property we spent twenty years teaching people to avoid in dependency management. Here is the case for a manifest and a lockfile.

Read →