Agent Workflows

Reusable patterns for AI-agent skills, cron jobs, and automated workflows — self-hosted, designed to run on a local appliance (an agent that owns a chat brain + scripted automations).

Everything here is anonymized and product-agnostic: repo names, client names, people, and machine paths are generalized so you can adapt a pattern to your own stack.

Layout

agent-workflows/
├── cron-issue-workup/   # Issue-triage → work-up → morning-report workflow
│   ├── README.md        # What it does, how it's wired, safety model
│   ├── prompt.md        # The cron prompt (the agent's instructions)
│   ├── config.json      # Routing / thresholds / branch map
│   ├── scripts/
│   │   ├── collect.py       # Deterministic issue collector (stdout → agent prompt)
│   │   └── prepare_clone.sh # Builds an isolated working clone
│   └── examples/
│       └── report.md        # Sample morning write-up
├── headless-agent-handoff/  # Chat agent → CLI coding agent delegation
│   ├── README.md        # When to hand off, when not to, how it's wired
│   ├── SKILL.md         # The agent-facing instruction block (load as a skill)
│   ├── bin/
│   │   └── handoff-agent    # Backend-agnostic wrapper: timeout + permission mode + logging
│   └── examples/
│       └── handoff-log.md   # Sample smoke test + real read-only handoff

Contributing a workflow

Each workflow directory is self-contained:

1. README.md — trigger, what it does, inputs/outputs, safety rails.
2. prompt.md — the exact instruction block handed to the agent on each run.
3. Supporting scripts/config — deterministic, side-effect-free where possible.
4. examples/ — a realistic sample of the produced output.

Design principles

• Non-destructive by default. Unless explicitly gated on, an automation never pushes, merges, closes, or deletes anything. Output is a report a human reviews.
• Never touch a live working tree. Work happens in an isolated clone.
• Deterministic collector. A plain script gathers state and hands the agent a JSON brief; the agent applies judgment on top, but the data is reproducible.
• Explicit, gated escalation. Dangerous capabilities (opening PRs, mutating trackers) live behind a config flag, off by default.
• Self-review on every run. The output includes a validation checklist confirming the safety rails held.