Agent Workflow Bundle

A reusable, project-agnostic version of the multi-agent dev workflow: orchestrator skills (/bugfix, /feature, /enhancement, /troubleshoot) that fetch a GitHub issue, create an isolated git worktree, and drive a chain of sub-agents (explore → plan → implement → test → review) that each get a fresh context window.

See docs/agent-workflow.md for the full design and docs/agent-workflow.puml for the diagram.

What's in here

agent-workflow.config.sh        # ← per-project parameters (the ONE file you must edit)
install.sh                      # symlinks .agents/skills/* into .claude/skills/*
dev.sh                          # optional: parameterized sandbox-container launcher
.devcontainer/                  # optional: minimal generic image + egress firewall
  Dockerfile                    #   extend with your toolchain
  init-firewall.sh              #   default-deny egress + FIREWALL_EXTRA_DOMAINS allowlist
  README.md                     #   what to change per project
docs/
  agent-workflow.md             # design doc (project-agnostic)
  agent-workflow.puml           # diagram
.agents/skills/
  bugfix/ feature/ enhancement/ troubleshoot/   # orchestrators (generic)
  explore/ plan/ review/                        # generic pipeline steps
  implement-be/ implement-fe/                   # TDD steps (workflow generic, stack detail
                                                #   delegated to a project reference skill)
  test-integration/ test-e2e/ test-visual/      # test steps (test-visual is opt-in)
  add-skill/                                    # meta: how to add a skill
  _project-templates/                           # copy these, rename, fill in per project:
    backend-reference/  frontend-reference/  testing-reference/

How to reuse in another project

1. Copy what you want into the target repo:

   cp -r agent-workflow-bundle/.agents              <target>/.agents
   cp    agent-workflow-bundle/agent-workflow.config.sh  <target>/
   cp    agent-workflow-bundle/install.sh           <target>/
   cp -r agent-workflow-bundle/docs/agent-workflow.* <target>/docs/
   # optional, only if you want the sandboxed container runner:
   cp    agent-workflow-bundle/dev.sh               <target>/
   cp -r agent-workflow-bundle/.devcontainer        <target>/.devcontainer   # then extend the Dockerfile

2. Edit agent-workflow.config.sh — set your dir names, build/test commands, and the role → skill name mapping (see ADAPTING.md).
3. Turn the _project-templates/* skills into real reference skills for your stack (e.g. backend-reference → firebase-functions, frontend-reference → angular-app), or point the config at reference skills you already have.
4. Delete any skills you don't use (test-visual if you have no design specs, test-integration if you have no integration suite, etc.).
5. Run ./install.sh to create the .claude/skills/* symlinks.
6. In Claude Code: /bugfix 123.

Why .agents/ and .claude/?

Claude Code only discovers skills under .claude/skills/ (plus ~/.claude/skills/ and installed plugins). It does not scan .agents/. This bundle keeps the source-of-truth skills in .agents/skills/ (the "open agent standard" layout, so they're not Claude-Code-specific) and install.sh creates .claude/skills/<name> symlinks pointing back at them. Commit both the .agents/ tree and the symlinks.