internal: Materialize skills during agent run by timflannagan · Pull Request #215 · agentregistry-dev/agentregistry

timflannagan · 2026-02-25T16:43:25Z

Description

Follow-up to #198 that added support for adding metadata to the agent.yaml. This commit adds runtime skill resolution and materialization for agent run workflows. When skills have been configured for an agent, we resolve those [ image, registryURL ] input sources to images, and then leverage docker create/cp/etc. to fetch those contents and extract directly to disk.

We leverage existing kagent-adk functionality here where it supports reading skills from an envvar so I only had to extend the docker compose template to mount that directory correctly.

Change Type

/kind feature

Changelog

NONE

timflannagan · 2026-02-25T17:00:43Z

internal/cli/agent/run.go

 // Note: The below implementation may be redundant in most cases.
-// It allows for registry-type MCP server resolution at run-time, but in doing so, it regenerates folders for servers which were already accounted for (i.e. command-type get generated during their `add-cmd` command)
+// It allows for registry-type MCP server resolution at run-time, but in doing so, it
+// regenerates folders for servers which were already accounted for (i.e. command-type get
+// generated during their `add-cmd` command).


TODO: I think this comment is potentially stale?

internal/cli/agent/run.go

internal/cli/agent/run_test.go

internal/cli/agent/skills_runtime.go

AGENTS.md

internal/cli/agent/skills_runtime.go

This adds runtime skill resolution and extraction for agent runs, including registry-backed skill references, so local and registry runs can mount usable skill content consistently. Previously, agent run only handled MCP runtime artifacts. Now we resolve skill image references, enforce strict image-versus-registry input validation at runtime, materialize skill files into the runtime config path, and cover the behavior with focused tests. Follow-up to #198.

Signed-off-by: timflannagan <[email protected]>

# Description Follow-up to agentregistry-dev#198 that added support for adding metadata to the agent.yaml. This commit adds runtime skill resolution and materialization for `agent run` workflows. When skills have been configured for an agent, we resolve those [ image, registryURL ] input sources to images, and then leverage docker create/cp/etc. to fetch those contents and extract directly to disk. We leverage existing kagent-adk functionality here where it supports reading skills from an envvar so I only had to extend the docker compose template to mount that directory correctly. # Change Type /kind feature # Changelog ```release-note NONE ``` --------- Signed-off-by: timflannagan <[email protected]>

github-actions bot added do-not-merge/description-invalid do-not-merge/kind-invalid do-not-merge/release-note-invalid labels Feb 25, 2026