Agent Patterns Plugin

Audit Claude subagent configs for completeness, security, and best practices. Use when reviewing agents/ for missing frontmatter, overprivileged tools, or bad model choices.

Configure Claude Code agent teams (implicit team, SendMessage, TaskUpdate). Use when running parallel agents, coordinating with messaging, or setting up a lead/teammate architecture.

Promote rules, skills, or agents from project scope to parent or user-global scope. Use when reorganizing .claude/ directories or when sibling repos hold near-duplicate rules.

Gate outward-bound text (upstream issues, docs, PR bodies) through isolated haiku fresh-reader critique before publishing. Use when an artifact must survive a reader with zero project context.

Install, configure and troubleshoot MCP servers. Use when adding/enabling servers, editing .mcp.json, fixing OAuth, or when a server runs stale code after an upstream fix.

Copy, generalize, or merge a project's .claude/{agents,commands} into user-scoped configuration. Use when assimilating another project's Claude setup or generalizing an agent.

Configure per-project plugin settings via .claude/plugin-name.local.md files. Use when building plugins with user-configurable behavior, storing agent state, or controlling hooks.

Audit a CLAUDE.local.md — verify each claim live, fix drifted environment facts, cut findings duplicating versioned docs. Use when local notes are stale, or before appending a finding.

Audit CLAUDE.md and .claude/rules for always-loaded content that should become an on-demand skill. Use when CLAUDE.md feels bloated, trimming context, or promoting a rule into a skill.

Tool results that mean something other than they look — a pattern that never compiled, a glob that cannot match a directory, one tier searched but reported absent, a rejected flag reading as "no results". Use when an empty result is about to gate an action, land in a doc, or be reported as done.

Adversarial second-pass review that tries to break code, designs, plans, or ADRs. Use when stakes are high and a normal review already ran.

Scaffold the code execution pattern for MCP-based agents. Use when agents call many MCP tools, intermediate data exceeds context, you need loops, or PII must stay out of context.

Verify orchestrator premises (file counts, build state, artefact presence) before dispatching parallel subagents. Use when a wave's briefs cite a number, path, or behaviour not yet checked.

Expose a live-adjustable control instead of guessing a magic constant for a parameter the agent cannot itself perceive. Use when tuning visual, audio, or UX output (color, size, position, timing, volume) that only the user can judge.

Sequential-wave dispatch for WO chains where output of one feeds the next, shared locks, or shared files prevent fan-out. Use when planning dependent multi-WO landings.

FastMCP authoring for Python MCP servers — tools, resources, prompts, TDD, release-please. Use when building or extending an MCP server, adding a tool/resource, or scaffolding a new server.

Multi-model design consults via PAL (kimi, glm, gemini, gpt). Use when asking other models to brainstorm a design or reconciling their split answers.

Pre-dump-then-dispatch for tools holding an exclusive lock (Ghidra, migrations, single-writer caches). Use when fanning out parallel agents needing a non-concurrent resource.

Dispatch contract for spawning parallel agents covering worktree collisions, scope overflow, and silent exits. Use when fanning out concurrent agents or authoring a lead prompt.

Write and configure custom agent definitions in Claude Code agents/ directory. Use when creating an agent .md file, defining a specialized agent, or configuring agent tools.

Data-loss invariants for a CLI that runs an agent SDK in a git worktree then force-removes it. Use when building or reviewing such a CLI, or when an agent run reports "no changes" after writing files.

Execution-grounded review: run tests first, trace each acceptance criterion to execution evidence. Use when verifying an implementation meets spec.