agent-stuff

Reusable agent skills packaged for Claude Code, Codex, and Pi.

The repository is a monorepo: each directory under plugins/ is an independently installable plugin containing one or more related skills. The agent-workflow plugin provides a coordinator-led repository workflow for ADRs, implementation prompts, direct agent handoffs, and Linear-first work that runs on the user's Mac.

Install

Review the source before installing. Agent skills can instruct an agent to read files, run commands, and modify repositories.

Claude Code

claude plugin marketplace add robzienert/agent-stuff
claude plugin install agent-workflow@agent-stuff

Inside Claude Code, the equivalent commands are:

/plugin marketplace add robzienert/agent-stuff
/plugin install agent-workflow@agent-stuff

Codex

codex plugin marketplace add robzienert/agent-stuff
codex plugin add agent-workflow@agent-stuff

Pi

pi install git:github.com/robzienert/agent-stuff

Use -l to record the Pi package in the current project's .pi/settings.json instead of your user settings:

pi install -l git:github.com/robzienert/agent-stuff

Included plugin

agent-workflow

Skill	Purpose
bootstrap	Initialize the ADR, prompt-index, and coordinator workflow in a repository.
coordinator	Load the project's coordinator identity and orient on current design state.
new-adr	Record an architecture decision in MADR format.
draft-prompt	Convert an approved design or Linear issue into an implementation prompt.
handoff	Gate and delegate a ready prompt through Linear or a direct subagent.
linear-local-runner	Run Linear issues through local Codex or Claude with progress comments.

Linear local runner

The linear-local-runner skill makes Linear the operational interface while coding stays on your Mac. It polls a Linear project, creates an isolated Git worktree, and starts the authenticated Codex CLI or Claude Code. It does not use Linear Coding Sessions or Linear AI credits.

Linear is the only work queue. The plugin does not create a second repository inbox. Git stores ADRs, implementation prompts, code, verification evidence, and merge provenance.

Requirements

• macOS with launchd and Keychain.
• A personal Linear API key.
• git, jq, curl, and Python 3.
• An authenticated codex or claude CLI.
• Repository-specific verification commands, such as unit tests and boot smokes.

Configure a repository

Invoke the linear-local-runner skill, or run its configurator from this checkout:

python3 plugins/agent-workflow/skills/linear-local-runner/scripts/configure.py \
  --repo /path/to/project \
  --team-id "$LINEAR_TEAM_ID" \
  --project-id "$LINEAR_PROJECT_ID" \
  --verify-command 'scripts/tools/test.sh' \
  --verify-command 'scripts/tools/smokes.sh'

Set the two Linear ID variables first. Replace the verification commands with the target repository's actual checks.

The configurator copies the runner and installer into the target repository. It writes the project-specific settings to .agents/linear-runner/config.json. Review that file before you load the launch agent.

Store the Linear API key under the keychain_service value from that configuration:

security add-generic-password -U -a "$USER" \
  -s "$(jq -r '.keychain_service' .agents/linear-runner/config.json)" -w

Install the launch agent and confirm the next action without changing Linear:

scripts/tools/install_linear_local_runner.sh
scripts/tools/linear_local_runner.sh --dry-run

Linear issue contract

A runnable issue must be in the configured project and Todo state. It must have prompt and agent-ready labels. Its description must contain the complete implementation specification, not a filepath that requires another application. Add agent:codex or agent:claude to select the provider. Codex is the default. Add model:mechanical, model:moderate, or model:hard to override the configured default model tier.

The launch agent polls once per minute while the Mac is awake. The runner allows one active issue, moves claimed work to In Progress, and posts progress comments every 30 minutes. It records the provider, model, CLI session ID, worktree, and phase. It runs the configured verification commands, then asks a fresh reviewer to inspect the diff. It moves work to In Review only after a pass.

A failed run stays In Progress. Correct the cause and move the same issue back to Todo. The runner resumes the existing Codex or Claude session and worktree. It never creates a second task record for recovery. The runner never merges or moves an issue to Done.

If the repository has an origin remote and authenticated gh, the runner can push the branch and open a draft pull request. Otherwise, the clean branch remains local. The direct handoff skill remains available for prompts that do not use Linear.

Repository layout

.agents/plugins/marketplace.json       Codex marketplace catalog
.claude-plugin/marketplace.json        Claude Code marketplace catalog
package.json                           Pi package metadata
plugins/agent-workflow/
  .codex-plugin/plugin.json            Codex plugin manifest
  .claude-plugin/plugin.json           Claude Code plugin manifest
  skills/                              Shared Agent Skills-format sources
    linear-local-runner/               Local Linear runner setup and assets

The skill sources are shared across all three agents. Platform-specific files contain only catalog and presentation metadata.

Validate

claude plugin validate .
python3 /path/to/plugin-creator/scripts/validate_plugin.py plugins/agent-workflow
pi -e .

License

MIT