Lab Director Plugin

A Kimi-native plugin that brings the Reynoza Brothers' pipeline-driven harness to anyone with a homelab account. Works on Kimi CLI; legacy Claude Code support included.

What it does

Drives every actionable request through a Conductor-orchestrated parallel swarm pipeline:

Conductor → Classifier → Router → Research Swarm → Plan Swarm → Review Board → Execution Swarm → Continuous Verification → Tracker(end)

Integrates with Plane (tickets), Outline (docs), Forgejo (PRs), Langfuse (telemetry), and Kubernetes (deployments).

Three modes

• pipeline-only — workflow stages without homelab dependency
• homelab-only — Plane/Outline/Forgejo helpers without pipeline ceremony
• collab — pipeline + homelab fused (default)

See docs/modes.md for details.

---

Quick Start (Kimi)

1. Install the plugin

kimi plugin install https://forgejo.aaron.reynoza.org/aaron/lab-director-plugin.git

Verify it loaded:

kimi plugin list
# → lab-director-plugin@0.4.0

2. Add hooks to ~/.kimi/config.toml

Copy-paste this into ~/.kimi/config.toml (create it if it doesn't exist). Important: extra_skill_dirs must come before any [[hooks]] arrays in TOML:

extra_skill_dirs = ["$HOME/.kimi/plugins/lab-director-plugin/skills"]

[[hooks]]
event = "SessionStart"
command = "/Users/aaronvaldez/.kimi/plugins/lab-director-plugin/scripts/session-start-hook.sh"
timeout = 5

[[hooks]]
event = "PostToolUse"
command = "/Users/aaronvaldez/.kimi/plugins/lab-director-plugin/scripts/pre-commit-checks.sh"
matcher = "Shell|WriteFile|StrReplaceFile"
timeout = 10

Why hooks? The SessionStart hook injects the pipeline context into every new session. The PostToolUse hook runs safety checks before every file write. Why extra_skill_dirs first? In TOML, keys after [[arrays]] get parsed as belonging to the last array entry. If extra_skill_dirs comes after [[hooks]], Kimi silently ignores it. Note: Kimi parses / commands before hooks fire. Use /skill:name or natural language instead.

3. Run post-install (optional but recommended)

Symlink the plugin's skills so Kimi discovers them automatically, even without extra_skill_dirs:

bash ~/.kimi/plugins/lab-director-plugin/scripts/post-install.sh

Tip: After installing the plugin, you can also run scripts/post-install.sh to symlink the skills into ~/.kimi/skills/lab-director-plugin/. This avoids the need for extra_skill_dirs.

4. Run setup

In any project directory:

/skill:setup

Or use natural language: Run the lab-director-plugin setup skill.

The setup skill:

• Verifies connectivity to your services (Plane, Outline, K8s, etc.)
• Walks you through identity, mode, and credentials
• Generates ~/.kimi/mcp.json with MCP servers
• Creates a per-project config at .lab-director-plugin/config.jsonc
• Writes a personalized onboarding doc for you

5. Verify everything works

/skill:setup-verify

Or use natural language: Run the setup-verify tool.

This checks:

• ✅ Plane API reachable
• ✅ Outline API reachable
• ✅ Langfuse API reachable (if configured)
• ✅ Kubernetes cluster reachable
• ✅ MCP config valid
• ✅ Hooks configured

6. Start working

Pipeline: Deploy the new monitoring dashboard

The pipeline takes over: creates a ticket, researches, plans, executes, verifies, and closes it out.

---

Commands

Use /skill:name or natural language for everything:

Intent	Command
Start pipeline	Pipeline: <context> or /skill:pipeline <context>
Run setup	/skill:setup
Check services	/skill:setup-verify or Run setup-verify tool.
Switch mode	Switch to mode: <name>
Conductor entry	/skill:conductor <context>
Classify work	/skill:classifier <context>
Route swarms	/skill:router <mode>

---

For Friends

Forwarding this to someone? Here's the one-liner they need:

kimi plugin install https://forgejo.aaron.reynoza.org/aaron/lab-director-plugin.git

Then they paste the hooks block above into ~/.kimi/config.toml, add extra_skill_dirs, run the setup skill, and they're good to go.

The plugin is zero-build — no npm, no cargo, no docker. Just markdown, bash, and Kimi.

---

Parallel Swarm Architecture

v0.4.0 introduces a Conductor-driven swarm model:

• Conductor — single entry agent that manages pipeline state and swarm lifecycle
• Classifier — analyzes complexity and selects mode (lightweight / standard / heavy)
• Router — dispatches to parallel swarm factories (Research, Plan, Review, Execution, Verification)
• Swarms — domain-specific subagent groups that run in parallel, reporting summaries back to the Conductor

This replaces the sequential 10-stage flow with parallel execution where safe, reducing wall-clock time for multi-file and multi-repo work while preserving stage-gate quality checks.

Read the full architecture in specs/pipeline-v0.4.0.md.

---

Legacy Claude Code Support

If you still use Claude Code, the plugin installs the same way as before:

claude marketplace add https://forgejo.aaron.reynoza.org/aaron/lab-director-plugin.git
claude plugin enable lab-director-plugin@lab-director-marketplace

The same repo serves both platforms. Skills and scripts are shared; only the packaging differs.

---

What's in the box

Component	Count	Purpose
Skills	30+	Workflow stages, homelab helpers, setup, constraints, swarm factories
Tools	4	Pipeline trace, Plane comment, pre-commit checks, setup verify
Hooks	2	SessionStart (context injection), PostToolUse (safety checks)
Specs	30+	Stage specifications, swarm architecture, verifier sub-specs
Templates	6	Project identity, spec, MCP config, orchestrator

---

License

MIT.