Skill Installation¶

Skills from the opendatahub-io/skills-registry are pre-installed into every runner and sandbox image at build time. Each harness (Claude Code, OpenCode) uses a different installation mechanism that matches the tool's native plugin system.

Claude Code: plugin seed¶

Claude Code images use the official plugin seed mechanism. At build time, native CLI commands populate a seed directory that Claude Code reads at startup without re-cloning anything.

Build-time flow¶

The Containerfile runs two steps as the non-root user:

RUN git config --global url."https://github.com/".insteadOf "git@github.com:" \
    && export CLAUDE_CODE_PLUGIN_CACHE_DIR=/home/agent-ci/.claude-seed \
    && export DISABLE_AUTOUPDATER=1 \
    && claude plugin marketplace add opendatahub-io/skills-registry \
    && agentic-ci install-plugins \
    && git config --global --unset-all url."https://github.com/".insteadOf

claude plugin marketplace add — registers the skills-registry marketplace and clones it into the seed directory.
agentic-ci install-plugins — reads the marketplace.json, calls claude plugin install <name>@<marketplace> for each plugin, and generates the plugin-skills manifest.

The git config override forces HTTPS cloning since the container has no SSH keys. It is removed at the end of the same RUN layer so the final image has clean git config.

Runtime¶

Two environment variables control plugin loading:

Variable	Purpose
`CLAUDE_CODE_PLUGIN_SEED_DIR`	Points to the pre-built seed directory. Claude Code reads it on startup (read-only, no auto-updates).
`CLAUDE_CODE_SYNC_PLUGIN_INSTALL`	Set to `1`. Required for skills to load synchronously in non-interactive (`-p`) mode.

Seed directory structure¶

~/.claude-seed/
  known_marketplaces.json
  installed_plugins.json
  marketplaces/
    opendatahub-skills/           # full git clone of skills-registry
      .claude-plugin/
        marketplace.json
  cache/
    opendatahub-skills/
      odh-ai-helpers/0.1.0/       # full git clone of plugin repo
        helpers/skills/
        helpers/agents/
      code-review-skills/0.1.0/
        skills/gitlab-code-review/
          SKILL.md
          scripts/review.py
      ...

OpenCode: file-based skill discovery¶

OpenCode has no native marketplace or seed mechanism. It discovers skills by scanning directories for SKILL.md files. The build copies skill directories from each plugin repo into OpenCode's global skills path.

Build-time flow¶

RUN git clone --depth 1 --quiet https://github.com/opendatahub-io/skills-registry.git /tmp/skills-registry \
    && agentic-ci install-plugins --harness opencode \
       --marketplace-json /tmp/skills-registry/.claude-plugin/marketplace.json \
    && rm -rf /tmp/skills-registry

agentic-ci install-plugins --harness opencode reads the marketplace.json, then for each plugin:

Clones the plugin's source repo.
Copies skill directories to ~/.config/opencode/skills/. Uses explicit skills paths from the marketplace entry when present (e.g. "skills": ["./helpers/skills"] for odh-ai-helpers), otherwise falls back to standard paths (.claude/skills/, .opencode/skills/, skills/).
Records the plugin→skill mapping in the plugin-skills manifest.

Skills directory structure¶

~/.config/opencode/skills/
  git-shallow-clone/
    SKILL.md
  code-review/
    SKILL.md
  gitlab-code-review/
    SKILL.md
    scripts/review.py
  ...

All skills are flat in one directory — no plugin namespacing. OpenCode discovers them by scanning for SKILL.md files.

Plugin-skills manifest¶

Both install scripts generate a manifest at /usr/local/share/agentic-ci/plugin-skills.manifest.json that maps plugin names to the skills they contain:

{
  "odh-ai-helpers": ["git-shallow-clone", "code-review", "cve-scan", ...],
  "code-review-skills": ["gitlab-code-review"],
  "rfe-creator": ["rfe.create", "rfe.review", ...]
}

For Claude Code, the manifest is informational (debugging, auditing). For OpenCode, the manifest is functional — enable-plugins uses it to apply per-skill filtering at runtime (see below).

Filtering plugins at runtime¶

The AGENT_ENABLED_PLUGINS environment variable controls which plugins are active. If unset, all plugins are enabled. If set to a comma-separated list of plugin names, only those plugins are active.

# Enable only odh-ai-helpers — all other plugins are disabled
AGENT_ENABLED_PLUGINS=odh-ai-helpers agentic-ci run "do something" ...

# Enable two plugins
AGENT_ENABLED_PLUGINS=odh-ai-helpers,code-review-skills agentic-ci run ...

How filtering works¶

The enable-plugins script runs at container startup (called by entrypoint.sh for Podman, or by the env script for OpenShell). It reads AGENT_ENABLED_PLUGINS and AGENT_TOOL to decide which mechanism to use:

Claude Code (AGENT_TOOL=claude): Sets enabledPlugins entries to false in ~/.claude/settings.json for unwanted plugins. Claude Code reads this at startup and skips disabled plugins.

OpenCode (AGENT_TOOL=opencode): Reads the plugin-skills manifest to find which skills belong to unwanted plugins, then writes "skill-name": "deny" entries to opencode.json's permission.skill section. OpenCode hides denied skills from the agent.

Error handling¶

If a requested plugin name doesn't match any installed plugin, the script exits with an error and reports which names were unmatched:

Matched: odh-ai-helpers
ERROR: unknown plugin(s) in AGENT_ENABLED_PLUGINS: nonexistent-plugin

CLI subcommands¶

Plugin installation and filtering are agentic-ci subcommands (implemented in src/agentic_ci/plugins.py):

Command	When	Purpose
`agentic-ci install-plugins`	Build time	Install plugins for Claude Code (reads seed dir) or OpenCode (`--harness opencode --marketplace-json <path>`)
`agentic-ci enable-plugins`	Runtime	Filter active plugins based on `AGENT_ENABLED_PLUGINS`

The container entrypoint (images/runner/shared/entrypoint.sh) calls agentic-ci enable-plugins at startup. For OpenShell sandboxes, the env script calls it before the agent runs.

Adding skills without rebuilding¶

Mount a directory containing SKILL.md files and point Claude Code at it:

podman run --rm \
  -v ./my-skills:/opt/extra-skills:ro \
  -e CLAUDE_CODE_PLUGIN_SEED_DIR=/opt/extra-skills \
  quay.io/aipcc/agentic-ci/claude-runner:latest

For OpenCode, mount into the skills directory directly:

podman run --rm \
  -v ./my-skills:/home/agent-ci/.config/opencode/skills:ro \
  quay.io/aipcc/agentic-ci/opencode-runner:latest