ship-kit

A portable agent-skills bundle for the whole dev loop: turn a reported bug into a fix, or a finished change into a green, mergeable pull request — and keep the PR honest until a reviewer approves it.

ship-kit's skills are Agent-Skills-compatible (SKILL.md, per the open Agent Skills standard) with a few Claude Code affordances such as argument-hint, which other harnesses simply ignore. So the same kit installs on any compatible agent — Claude Code, OpenAI Codex, xAI Grok, Nous Hermes, OpenClaw, VS Code / Copilot, and more. It bundles everything it needs, so it does not rely on skills being pre-installed by the host.

What's inside

Skill	What it does
fix	The inbound headliner: triage → branch → implement → ship. Diagnoses a reported bug, cuts a fix/… branch, implements the fix at the root cause, then hands off to ship for the whole way to a green PR. Pass --yolo to run hands-off.
ship	The outbound headliner: simplify → verify → code-review → smart-commit → branch hygiene → PR → babysit → reflect. Stops at approved + CI green + mergeable; only merges with --merge.
triage	Diagnoses a bug with parallel sub-agents across different angles (reproduce, recent changes, backward root-cause trace, blast-radius), adversarially verifies the leading cause, and proposes the fix — read-only, never edits code. fix is the applier.
babysit	Drives one GitHub PR to a fresh APPROVED + all CI green + mergeable state. Each pass verifies every review finding against the code, fixes the valid ones, pushes back with evidence on the invalid, and replies on every thread. Never merges — a human does that.
smart-commit	Clusters uncommitted changes into logical groups and commits each with a conventional-commit message. Syncs CHANGELOG and drift-checks README/docs against the change.
reflect	Captures session lessons (gotchas, patterns, integration quirks) into the knowledge base. Self-gating, never writes without your approval. Runs as ship's final step.
simplify	Quality-only cleanup pass (reuse / simplification / efficiency) over a change. Bundled so ship works without a pre-installed simplifier.
verify	Runs the project's tests / app and observes the change actually working. Bundled.
code-review	Reviews a diff for correctness bugs and cleanups, then applies fixes (--fix) or posts inline PR comments (--comment). Bundled.
check	The fast local loop: detects the repo's own toolchain, then deslops, type-checks, lints, builds and tests — self-healing, fixing what it finds until the run converges. --commit hands off to smart-commit when green.
merge-and-monitor	Merges a PR, then follows the deployment it triggers to completion and verifies health. Runs automatically as ship --merge's Stage 9 end state.
wrap-up	Session-closeout audit: what's still unfinished in this conversation, as a verdict — SAFE TO CLOSE or a ranked table. Writes nothing.
yolo	Autonomous mode: never asks, resolves ambiguity by evidence + research, fans out sub-agents. Two ways in — FULL ("yolo", "just do it") and DECISION ("your call", "do what you think is best"). Also the --yolo flag on ship / fix.
ask-me-questions	The deliberate inverse of yolo: multi-round structured questioning before acting on a genuinely ambiguous request.
answer-first	Output mode for scanning, not reading: verdict on line 1, evidence as file:line, no preamble or recap, bounded list lengths, explicit "what I did NOT do".

Self-contained by design

ship chains simplify, verify, and code-review. On Claude Code those exist as built-ins — but on Codex, Grok, Hermes, or OpenClaw they don't. So ship-kit ships its own portable copies of all three. ship prefers a richer native/built-in skill of the same name when the host provides one (e.g. Claude Code's code-review with its cloud ultra mode) and falls back to the bundled copy everywhere else. Result: the pipeline runs the same on every harness.

Install

Pick the line for your agent. See docs/INSTALL.md for the full matrix, project-scoped installs, and troubleshooting.

Claude Code (plugin marketplace):

/plugin marketplace add nsollazzo/ship-kit
/plugin install ship-kit

xAI Grok Build (official CLI) — install the plugin (recommended; works in every project):

grok plugin install nsollazzo/ship-kit#.agents --trust
grok inspect   # confirm ship, fix, babysit show source: plugin: ship-kit

From a local clone: grok plugin install ./.agents --trust. Skills-only alternative: bin/install.sh grok (→ ~/.grok/skills/). See docs/INSTALL.md for project-scoped installs, marketplace browse, and dev symlinks. Grok does not use Claude's /plugin commands — use the grok plugin CLI or /marketplace in the TUI.

Codex, OpenClaw, VS Code, or any other Agent-Skills agent — clone and run the installer:

git clone https://github.com/nsollazzo/ship-kit && cd ship-kit
bin/install.sh agents      # → ~/.agents/skills  (the widest-reach open-standard path)
# or: bin/install.sh all   # installs into every harness it detects under $HOME

Per-harness targets: agents (Codex / Grok community / OpenClaw / VS Code), claude, grok (Grok Build skills copy), hermes, openclaw, copilot. Use --project to drop the skills into the current repo's .agents/skills/, or --dest <dir> for an arbitrary location.

Nous Hermes — install natively as a Tap (no clone needed):

hermes skills tap add nsollazzo/ship-kit

Ask your agent to install it

Most of these agents have shell access, so you can just ask the agent to install ship-kit for you. Paste one of these:

• Grok Build:

  Run grok plugin install nsollazzo/ship-kit#.agents --trust, then grok inspect and confirm ship, fix, and babysit are listed with source plugin: ship-kit.

• Any other shell-capable agent (Codex, OpenClaw, …):

  Install the ship-kit skills for me: clone https://github.com/nsollazzo/ship-kit and run bin/install.sh with the target that matches this agent (one of: agents, claude, hermes, openclaw, copilot). Then reload skills and confirm ship and babysit are listed.

• Nous Hermes:

  Run hermes skills tap add nsollazzo/ship-kit, then list skills and confirm ship and babysit loaded.

• Claude Code: type /plugin marketplace add nsollazzo/ship-kit, then /plugin install ship-kit.

Using it

Once installed, just say what you want — the skills auto-trigger by description on every agent:

"fix this bug: …" · "triage this failure" · "check my code" · "ship it" · "babysit PR 412" · "smart commit my changes" · "merge and watch the deploy" · "are we done? safe to close?" · "be terse" · "your call"

They're also slash commands where the host exposes them: /ship-kit:ship on Claude Code, $ship on Codex, /ship on Grok / Hermes / OpenClaw. ship takes flags — --fast, --merge, --no-monitor, --no-verify, --yolo; fix takes --yolo (plus the ship flags it passes through); check takes a phase name plus --commit / --fix-only. --yolo runs hands-off (no questions); it stops at green unless you also pass --merge.

Full usage guide — the pipeline stage by stage, every flag, babysit, and worked examples: docs/USAGE.md.

Requirements

• The gh CLI, authenticated (gh auth status), for babysit and ship's PR steps.
• A git repository with a GitHub remote.

Notes

• babysit never merges and never mutates external state to force a gate green. It drives to approval; a human merges.
• Project conventions override the recipe. ship defers to a repo's own instruction file (CLAUDE.md, AGENTS.md, …) and its verify/test recipe — e.g. a project that forbids a tracked CHANGELOG.md, or runs tests in CI rather than locally.
• One source of truth. The canonical skills live in .agents/skills/. The Claude Code plugin reads them directly (the marketplace points its plugin source at ./.agents). The only generated tree is skills/ — a flat mirror for Hermes Taps, produced by bin/build-adapters.sh (run it after editing a skill; --check verifies it's in sync).

License

MIT