agent-plugins

Agent skills for making a coding agent verify its own work and deliver it.

Written to the Agent Skills standard (SKILL.md), so the same skills run in Claude Code, OpenAI Codex, Kimi, pi, and the 30+ other tools that read the format — and with GLM, Kimi K2, or any other model, since the skills live in the harness, not the model. There is one copy of each skill; nothing is forked per agent.

Plugins

Plugin	Owns	Does
feedback-loops	Code quality ("is it correct?")	Encode your verification process as skills so the agent self-verifies. Skills: setup-feedback-loop (once per project — detects the stack, writes docs/verification.md), green-loop (every change — runs the recorded checks until green). Based on Anthropic's Feedback loops post.
ship	Delivery ("is it delivered?")	Take a green change out the door. Skill: ship (/ship) — verify → feature branch → rebase → /code-review → push → PR → watch. Plus an opt-in, fail-open Stop-hook gate that refuses "done" while delivery is incomplete.

The split — one source of truth for checks

The two plugins are deliberately separate, with no duplicated check commands:

• feedback-loops owns code quality. The verification checks are defined once in the project's docs/verification.md and run by green-loop (run → observe → fix until green).
• ship owns delivery only. /ship takes an already-green change through feature branch → rebase onto the latest base → /code-review → push → PR → watch. It delegates the actual checks back to green-loop; it never re-implements lint/test/build commands.
• The ship-gate Stop hook checks delivery state only — tree clean → not on a protected branch → pushed → PR green → not conflicted → review requested — and runs no lint/tests. It fails open (missing config / jq / python3 / gh → allow) and is opt-in via .claude/ship.config.json.

green-loop  →  (green)  →  /ship  →  branch → rebase → review → push → PR → watch
   ▲                                    │
   └──── re-run Layer 1 if the ─────────┘
         rebase moved the base

The rebase → re-verify rule: when the rebase pulls in new base commits (or you resolve a conflict), the old green is stale, so /ship re-runs Layer 1 via green-loop before continuing.

Install

Claude Code (plugin marketplace)

Gets the skills and the optional ship-gate Stop hook:

claude plugin marketplace add ddalgrande/agent-plugins
claude plugin install feedback-loops@ddalgrande-plugins
claude plugin install ship@ddalgrande-plugins

Or from a local clone: claude plugin marketplace add /path/to/agent-plugins.

Codex, Kimi, pi, and other SKILL.md agents

git clone https://github.com/ddalgrande/agent-plugins
cd agent-plugins
./install.sh

That symlinks every skill into ~/.agents/skills/ — the shared discovery directory Codex, Kimi, and pi all read — so one install covers them, and git pull updates them in place. Then restart the agent:

Agent	Invoke	Reads
Claude Code	/green-loop	plugin, or ~/.claude/skills/
Codex	$green-loop, or picked up automatically	~/.agents/skills/
Kimi	/skill:green-loop	~/.agents/skills/
pi	/skill:green-loop, or picked up automatically	~/.agents/skills/

Other options:

./install.sh --list                     # what this repo provides
./install.sh --dry-run                  # show the plan, change nothing
./install.sh --copy                     # copy instead of symlink
./install.sh --project ~/code/my-app    # into my-app/.agents/skills, committable
./install.sh claude                     # ~/.claude/skills (no hooks)

The installer never overwrites a directory it didn't create — if something else already owns a skill name, it says so and skips it.

GLM, and other models

GLM is a model, not an agent — it runs inside a harness (Claude Code via Z.AI's Anthropic-compatible API, Cline, Roo Code, Kilo Code, TRAE). Install for the harness you use and the skills work with whatever model is behind it. Nothing here is model-specific.

What's portable, and what isn't

Piece	Claude Code	Codex / Kimi / pi / others
The three skills	✅	✅
docs/verification.md contract	✅	✅ — plain Markdown
ship-gate Stop hook	✅	❌ — hooks have no cross-agent equivalent

Only the hook is Claude-specific. Everywhere else the skills are advisory rather than enforced, which is how they behave on Claude Code too unless you opt into the gate.

Migrating from ship@ship-tools? Uninstall it first (claude plugin uninstall ship@ship-tools) — otherwise both plugins register a Stop hook and the gate runs twice. The config schema also changed: the old gates/paths keys are removed; the gate now reads only gate.enabled and gate.max_blocks from .claude/ship.config.json.

See each plugin's README for full docs: feedback-loops · ship.

Repository layout

.claude-plugin/marketplace.json       # marketplace manifest (lists both plugins)
AGENTS.md                             # instructions for agents working on this repo
install.sh                            # portable installer for non-Claude agents
plugins/feedback-loops/               # code-quality plugin
├── .claude-plugin/plugin.json
├── README.md
└── skills/
    ├── setup-feedback-loop/SKILL.md
    └── green-loop/
        ├── SKILL.md
        └── references/e2e-recipes.md
plugins/ship/                         # delivery plugin
├── .claude-plugin/plugin.json
├── README.md
├── skills/ship/SKILL.md
└── hooks/
    ├── hooks.json                    # registers the Stop-hook gate
    └── ship-gate.sh                  # delivery-state gate (fail-open, opt-in)

License

MIT