Skill publishing pipeline
A guided publishing tool and a CI check that let anyone on a team ship agent skills without breaking installs for everyone.
- Role
- Lead engineer
- Year
- 2026
- Stack
- Python, Agent skills, CI
- Code
- Private
Overview
The skill publishing pipeline lets people who are not engineers add agent skills to a library the whole team installs from. A publishing tool runs on the contributor's machine, and a CI check keeps the team's skill registry in step with what is merged. I built the tool and the first version of the check.
Problem
Publishing a skill meant creating a branch, editing two config files, bumping a version, adding a registry row, and filling in a pull request template. One bad config line broke installs for everyone. Registry rows were kept by hand, so they drifted or showed up twice.
Approach
I split the work by what each part can be trusted to do. The tool prepares a change, and a reviewer still merges it. The CI check records what changed, so the registry no longer depends on memory.
Planning and applying are separate steps. The planner only reads, and when it cannot match a skill to a category it asks instead of guessing. Its first version kept decisions in shell scripts that were hard to test, so after review I rewrote them in typed Python.
How it works
The tool is a plugin for a coding assistant, made of a playbook skill, a local command line program, and a hook that blocks commits and pushes to the main branch. It runs with the contributor's own access.
- 1The tool checks the setup and validates the skill files.
- 2It plans every file change and shows the whole plan.
- 3The contributor approves, and the tool adds the registry row.
- 4It applies the plan only if the submission, the main branch, and the plan are unchanged, then opens a pull request.
- 5The CI check works out what changed for each skill and updates its row.
The tool treats every submission as untrusted, rejecting links outside the folder, special files, and oversized archives. Config changes go to a temporary file that must parse before it replaces the original.
The CI check never deletes. A removed skill is marked deprecated, and its old status is saved so a closed pull request can restore it. A teammate later extended it to cover a new skill category.
Testing
Both parts have unit tests for their pure logic, and the CI check runs its tests before every sync. Before launch, I ran the check in dry run mode on real pull requests and against a copy of the registry, which surfaced several drifted fields.
Outcome
A contributor can now publish a skill through one guided flow that never touches the main branch, and every merged change leaves exactly one registry row. The first hook blocked merges in every repository on the machine, so it is now scoped to one. Next is cleanup when the command line program is missing, which can leave a stray branch and an orphan row.