Agent Team

agent-team — Claude Code Plugin

Coordinated multi-agent workflow with skills, hooks, and structured output gates. Plugin version of the standalone .claude/ config at mike-diff/ai-coding-configs.

Install

/plugin marketplace add mike-diff/ai-coding-configs
/plugin install agent-team@mike-diff

After install, all commands are namespaced: /agent-team:discuss, /agent-team:dev, /agent-team:spec, etc.

Prerequisites

• Claude Code 2.1.108 or later (verify with claude --version).

• CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 set in your user-level ~/.claude/settings.json:

  { "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }
  

  Without this, multi-agent workflows silently fail. The plugin's SessionStart hook will warn you if it's missing.

What This Plugin Provides

• 15 skills — ask, dev, discuss, issue, loop-patterns, orient, primitives, review-patterns, skill, slop-check, spec, team-orchestration, testing-patterns, ticket, to-dos
• 5 specialized agents — explorer, implementer, qa, reviewer, skill-author
• 13 hooks across 11 lifecycle events (SessionStart, PreToolUse, PostToolUse, PermissionDenied, TeammateIdle, TaskCompleted, PreCompact, FileChanged, CwdChanged, TaskCreated, StopFailure)
• 1 output style — teaching
• 2 rules — coding-standards, mcp-caching (auto-injected via SessionStart hook)

What This Plugin Does NOT Provide

These shipped with the standalone .claude/ but cannot ship via the plugin schema:

• statusline.sh — plugin schema doesn't accept arbitrary statuslines (only subagentStatusLine, different shape). To get the same statusline, copy .claude/statusline.sh from the source repo into your project and configure it manually in .claude/settings.json.
• env.ENABLE_PROMPT_CACHING_1H — plugin settings.json doesn't accept env vars. Set this in your user-level ~/.claude/settings.json if desired.

Logs

Plugin hook logs are written to ~/.claude/plugins/data/agent-team-mike-diff/logs/:

• hooks.log — most events
• tasks.log — task-related events
• permission-denied.log — denied tool calls
• stop-failures.log — API errors

The directory is created automatically on first run.

Hacking on This Plugin

Clone the source repo, then point Claude Code at your local copy:

git clone https://github.com/mike-diff/ai-coding-configs.git
cd ai-coding-configs
claude --plugin-dir ./plugins/agent-team

Live edits are picked up via /reload-plugins (no restart needed).

To sync changes from the standalone .claude/ into the plugin:

bash scripts/sync-plugin.sh

To run the smoke test:

bash tests/smoke.sh

Known Limitations

• SessionStart hook on brand-new sessions (issue #10373): some Claude Code versions only fire SessionStart on /clear, /compact, or resume. If rules don't appear in your context, switch the inject-rules.sh hook to UserPromptSubmit in hooks/hooks.json (one-line change).
• Per-agent Stop hooks unavailable: plugin agents cannot declare hooks: frontmatter (security restriction). Structured output is enforced via inlined <output_gate> assertions in each agent prompt and the project-level TeammateIdle hook in hooks/hooks.json.

See Also

• Top-level repo README — both install paths
• Claude Code plugin docs