Core Workflows

Projects Track: Standing Shifts & Multi-Agent Workflows

Organize GitHub triage, smoke tests, scheduled releases, and operations into structured projects powered by zero-token guards and multi-agent teams.

Writing code is only half of running a product. The other half is ongoing operations—watching GitHub issues and CI failures, smoke-testing builds, cutting releases, researching competitors, and answering users. The Projects track in Munk AI turns these recurring jobs into structured, multi-agent workflows.

How Projects and Tasks Are Structured

Inside each product, you can organize recurring work into one or more Projects. Every project is a plain-text directory under projects/<id>/:

products/open-compare/projects/github/
├── README.md               # Shared project context and operating notes
├── guards/                 # Zero-token preflight scripts (.mjs)
│   └── github-sentinel.mjs
├── tasks/                  # Standing task definitions (.md)
│   ├── github-issue-pr-actions.md
│   ├── l1-package-smoke.md
│   └── release-batch.md
└── .opencode/              # Project-local agents and skills (optional)

Projects and shifts

Triggers: Schedules, Chains, and Webhooks

Each task (tasks/<id>.md) declares when a Shift (one execution of a task) should start via YAML frontmatter:

  1. Cron Schedules (schedule): Run on a standard 5-field cron schedule (e.g., cut a daily release at 07:00, or check GitHub every 10 minutes).
  2. Task Chaining (after): Automatically kick off a downstream task once an upstream shift finishes cleanly (for example, trigger l3-cdp-ui right after l1-package-smoke passes).
  3. Incoming Webhooks (webhook): Trigger a shift immediately when an external HTTP request hits your local runtime's generated webhook URL.
  4. Manual Shifts: Omit trigger to run the task on demand whenever you click Run Shift.

Zero-Token Guards: Wake Agents Only When Needed

Polling GitHub or an external API every 5 minutes with a large language model would burn through tokens with nothing to show for it.

That is why high-frequency tasks pair their schedule with a Guard script (guards/*.mjs):

  • When the schedule fires, Munk AI runs the lightweight Node.js script first (typically finishing in under a second with zero LLM token cost).
  • Nothing new (exit 10): The shift quietly skips without waking an agent.
  • Action needed (exit 0): The guard outputs a brief summary of what changed, and Munk AI wakes the agent team to handle the shift.

Multi-Agent Teams & Human Assist

  • Team Delegation (team): A task can coordinate specialized agents (such as triage analysts, QA testers, or copywriters) that share the project's files, custom skills (SKILL.md), and read-only codebase access (reads.code).
  • Human Assist for Authenticated Platforms: When a shift needs to interact with a logged-in tool or web dashboard (such as gh CLI, store consoles, or social accounts), you never paste passwords into prompts. Instead, you log in once in your local terminal or dedicated browser profile, and agents reuse that authenticated session safely.