Agent Workbench
Agent skills come from other people's repositories. This one pins every single one to a commit it has actually read.
The problem
An agent skill is a folder of instructions in somebody else's repository, installed by name. Install it by name twice a month apart and you get two different things, because the name is a moving pointer — and nothing tells you it moved. A skill that quietly changed its instructions is worse than one that broke: the bootstrap succeeds, the environment is different, and the first symptom is an agent behaving oddly on a project three days later.
The solution
Every external source is pinned in `manifest/sources.yaml` to the exact commit SHA that was reviewed, next to its repository URL and package id. Installs go through `npx skills add <owner>/<repo>@<sha>`, printed rather than executed silently, so what lands is auditable before it lands. A git fallback copies the same pinned folders into an isolated directory for reading, and nothing in either path writes to an upstream repository.
Outcomes
- Skill sources are pinned by commit SHA, so a bootstrap run today and one in six months install byte-identical instructions.
- `manifest/sources.yaml` records repository, reviewed ref, and package id for each source — provenance in one file rather than in a shell script's history.
- Install profiles (core, frontend, full) print the exact pinned commands, so the operator sees what will run before it runs.
+2 more in the write-up
Inside the write-up
- The thing a version number was supposed to solve
- The manifest is the whole idea
- Printing the command instead of running it
- Two paths, for two different questions
- What it deliberately does not do
- What it is actually for