Agent Bootstrap.

Problem.

Using several AI coding clients across machines turned configuration into a collection of drifting copies. Their formats and discovery paths differ, and copying one complete configuration onto another machine risks replacing settings or granting capabilities that do not belong there.

Implementation.

1. 01

   Established a canonical configuration area in the existing infrastructure repository: shared agreements, personal skills, an MCP catalog, and separate workstation and sandbox profiles.

2. 02

   Built a Python installer that generates client-specific adapters and merges only managed settings. It preserves unrelated model, trust, and approval choices, backs up changed files, and refuses conflicting skill directories before writing.

3. 03

   Locked fourteen upstream skills to exact source revisions and file hashes. Reconstructed those dependencies from their public histories instead of treating an installed cache as an unversioned source of truth.

4. 04

   Added GitHub Actions CI for manifests, generated adapters, and installer tests. Exercised all three clients in an isolated temporary home, keeping validation independent of production services.

Result.

Published the shared setup in a private repository with a passing CI run. The installer and configuration checks passed in an isolated environment, and every locked dependency was reconstructed with matching hashes. Migration of existing real-machine configurations remains an explicit, conflict-aware step.

What I learned.

Reproducibility means controlling what changes as well as what gets installed. Client adapters should be small, dependencies should be identifiable, and a dry run should expose conflicts before writing. Configuration validation can run in CI without granting the runner access to production.