claude-code

Delegate to this agent during Claude Code setup (local `/rdl-team:cc-setup`) when you need to discover what plugins are available across the RDL marketplace and the team's extra marketplaces, and decide which ones THIS repo should enable. It locates the team's tracked marketplace list, enumerates each marketplace's plugin catalog (live), inspects the repo's languages and tooling, and returns a ranked suggestion set: the always-useful baseline (pr-review-toolkit, gh@rdl-agent-extensions, worktrunk, the applicable LSP) plus language/stack-matched picks, each with id, marketplace, and a one-line reason. It only researches and recommends — it does not install or edit settings.

Delegate to this agent to engineer, improve, and validate prompts using a dual-persona Prompt Builder / Prompt Tester methodology. Analyzes sources, applies imperative-language best practices, and runs mandatory validation cycles before finalizing any prompt.

Read-only skill-quality auditor. Applies the non-inferable-value rubric (Biggs + SkillsBench, version pins, verify-canonical guard) to a target SKILL.md or session diff and emits severity-rated findings with file:line anchors and keep/cut/pin recommendations. Never modifies files.

Add a new subagent to this agent-extensions catalog and take it all the way to a CI-green, commit-ready change. Use when the user wants to create, add, import, adapt, or normalize a catalog agent — a subagent authored under agents/<name>/agent.md — including adapting an upstream github/awesome-copilot .agent.md into repo conventions, or writing one from scratch. Walks the full registry → plugin-tree → manifest → changelog → validation pipeline: writes agents/<name>/agent.md, registers it in a registry bundle, runs sync-plugins, generate_manifests, changie, and validate-plugins. A session-scoped Stop hook it installs (and tears down) keeps the turn alive until the agent is fully wired, so the session cannot end with a half-wired agent. Covers both import-and-normalize (the common path) and from-scratch authoring.

Claude Code agent teams — coordinate multiple independent Claude Code sessions working in parallel with shared task lists and inter-agent messaging. Use when the user asks to create a team, spawn teammates, coordinate parallel work, or describes work that benefits from parallel agents. Also triggers on TeamCreate, SendMessage, teammate mode, team lead, agent coordination, parallel sessions, or 'can we parallelize this with agents?'. Helps with planning parallel execution strategies and structuring team-based work.

Create, save, and reuse Claude Code dynamic workflows — JavaScript orchestration scripts that Claude writes and a runtime executes to fan out to many subagents at scale. Use when the user wants to build a workflow, automate a repeatable multi-agent process, codify an orchestration, or save a run for later. Also triggers on `ultracode`, `/workflows`, `.claude/workflows/`, `/deep-research`, "orchestrate subagents", "codebase audit", "large migration", "cross-checked research", or "make this a reusable command". The skill first decides whether a workflow is even the right tool (vs a subagent, skill, or agent team), then drives the author-and-save loop and the `.claude/workflows/` distribution rules.

Create, manage, and debug Claude Code hooks — user-defined commands that run at specific points in Claude Code's lifecycle for deterministic control. Use when the user asks about hooks, guardrails, pre/post tool execution, safety rules, command blocking, context injection, auto-formatting, permission auto-approval, or completion checklists for Claude Code. Also trigger on mentions of PreToolUse, PostToolUse, PermissionRequest, SessionStart, Stop, UserPromptSubmit, settings.json hooks, prompt-based or agent-based hooks, or when the user worries that hook output looks like prompt injection. Covers all five hook types (command, http, mcp_tool, prompt, agent), the full event lifecycle, the JSON output contract (permissionDecision / decision / hookSpecificOutput), exit codes, and safe context-injection patterns.

Design reproducible multi-step flows in Claude Code — decide between a single ordered skill, multiple step commands, hook-gated state, and when to escalate a step to subagents, an agent team, or a dynamic workflow. Use when the user wants a repeatable procedure (step 1 → 2 → 3), a "pipeline", enforced step ordering, a gated process where a later step must check an earlier one, or asks how to package a multi-step process as a plugin. Also triggers on "make this reproducible", "marker file", "state machine", "step 2 needs step 1", "fan out", or "skill vs command vs workflow". Picks the lightest structure that fits and shows the marker-file + hook gating pattern.

1 hook across 1 event