feat(pr-review): load connector layer review skills from c1-engineering marketplace

[luisina-santos]

What

Complements the Claude PR review with the layer review skills from the agentic-connector-development plugin in ConductorOne/claude-marketplace.

The pieces were mostly already in place: Skill and Task are in the action's --allowedTools, and base-pr-review.md Step 5 already tells the reviewer it may use them. What was missing is the plugin those skills live in. claude-code-action@661a6fef exposes plugins / plugin_marketplaces inputs, so this wires them up.

Dispatch is by changed path, so a config-only PR does not pay for four skills:

Changed paths	Skill
pkg/config/**	review-config-layer
pkg/client/**, pkg/connector/client*.go	review-client-layer
pkg/connector/**	review-connector-layer
pkg/connector/actions.go, *_actions.go	review-actions-layer

Required before this helps: CLAUDE_MARKETPLACE_TOKEN

claude-marketplace is private, and the runner's automatic GITHUB_TOKEN is an installation token scoped to the repository running the workflow — same-org membership does not extend it. (github-workflows is public, which is why the existing Checkout workflow repo step needs no token and this asymmetry has not come up before.)

So loading the skills needs an org-level CLAUDE_MARKETPLACE_TOKEN secret: a fine-grained PAT or GitHub App token with contents: read on ConductorOne/claude-marketplace and nothing else.

Until that secret exists, this PR is a no-op — every connector review logs a warning and proceeds exactly as it does today. Nothing breaks in the meantime, which is what makes on-by-default safe here.

Decisions taken

• On by default for the connector profile. general never loads skills and needs no token.
• Degrades, never fails. No token, expired token, or a clone that does not land → warning, review continues on base prompt + mixins. Each summary now carries a **Skills:** line, so a missing secret shows up on the PR instead of silently downgrading every review.
• Connector mixin left intact. C1–C8 / R1–R13 overlap the client and connector skills substantially. Rather than strip them now, connector-skills.md tells the reviewer to post an overlapping finding once, at the higher severity. Worth revisiting once we can see real output.
• Focused three + actions layer, not deep-code-review — its 8-finder fan-out does not fit this budget. --max-turns 100 → 150 and timeout-minutes 15 → 30 to fit up to four skills.

Two implementation notes worth a look

1. Cloned, not actions/checkout. Composite action steps do not support continue-on-error, so a checkout with a bad token would fail the whole review rather than degrade it. The step clones with || warn instead.
2. .git is deleted after cloning. The clone URL embeds the token in .git/config, and the reviewer has Read/Grep over the workspace. Only the plugin files are needed, so the git metadata goes. _marketplace/ is also added to the "ignore these directories" instruction in Step 1 so it is not mistaken for PR content.

Skills run read-only

These skills were written for interactive development and some steps assume they can build, test, edit, and fetch vendor docs. connector-skills.md makes the Step 5 read-only contract win explicitly: no edits, no build/test commands, no network fetches, no posting, and a skill's own approval verdict does not clear the PR. Skill output is analysis that still goes through Step 6 validation before it becomes a comment.

Verification

• action.yml and both caller workflows parse; every composite step uses only keys composite actions accept (this is how the continue-on-error problem surfaced).
• Prompt assembly dry-run in both branches: skills-available appends connector-skills.md (540 lines) and skills-unavailable does not (494), with the Skills status: line and heredoc terminator correct in both.
• Not yet exercised end-to-end — that needs the secret plus a real connector PR. Suggest landing this, adding the secret, then watching the first review's **Skills:** line and turn count before enabling more.

🤖 Generated with Claude Code