Ship: AI-Powered Software Development Harness
An agentic development harness for Claude Code, Codex & Cursor: agent-routed workflows from raw requirement to green PR.
Ship helps agents choose and run the right amount of software delivery process: one standalone phase, a grouped quality/build bundle, or the full raw-input-to-green-PR flow.
How It Works
Ship is a harness, not a copilot. It doesn't help AI write code — it constrains AI to produce reliable results through mechanically enforced quality gates.
The problem Ship solves: AI coding agents are capable but unreliable. They skip tests, hallucinate about code they haven't read, review their own work and call it good, and declare victory without evidence. Ship makes these failure modes structurally impossible.
- Use Ship chooses the right route.
/ship:use-ship decides whether the task needs one skill, a phase bundle, or the full /ship:auto workflow.
- Production artifacts stay organized. When a task needs durable docs, agents use the repo's existing convention or create a focused
docs/ship/<task-id>/ folder for requirements, design, engineering, quality, delivery, and archive notes.
- Atomic skills stay standalone. Focused skills like
/ship:dev, /ship:e2e, /ship:review, /ship:qa, /ship:refactor, and /ship:handoff work directly without a full workflow.
- Input, state, and outputs are separate. Raw requirements live under
input/. The orchestrator keeps only minimal run state. Markdown artifacts and repository code are the deliverables.
- Every phase is isolated. The reviewer has never seen the implementation context. The QA evaluator can only see the spec, the diff, and the running application. Fresh context per phase means no accumulated bias.
- Plans are adversarially tested. An independent peer challenger produces code-grounded objections with file paths and snippets. The planner must respond with evidence, not hand-waving. Two rounds before you see anything.
- Evidence is hierarchical. L1 (screenshot, curl response, console log) is the only acceptable proof. L2 (HTTP 200, "tests passed") is insufficient. L3 ("should work based on the code") is an automatic FAIL.
- State lives on disk, not in memory. The current phase is tracked in local state. On resume, the orchestrator reads disk and picks up where it left off. A stop-gate hook blocks session exit while the workflow is active.
- The finish line is checks green, not PR created. After opening the PR, Ship enters a fix loop — read CI failures, dispatch fixes, address review comments, resolve merge conflicts — up to 3 rounds before escalating.
- Test-driven implementation. Stories follow a RED-GREEN-REFACTOR cycle with per-story code review before merge.
Installation
Claude Code
/plugin marketplace add heliohq/ship
/plugin install ship@heliohq
Codex
/plugins
Search for Ship, then install it. In Codex App, open Plugins in the sidebar and install Ship from there. See .codex/INSTALL.md for details.
Cursor
/add-plugin ship
Or search for Ship in the Cursor plugin marketplace.
Verify Installation
Open a fresh session and confirm the /ship:* skills are available — for example, run /ship:use-ship plan out a user authentication system.
Updating
/plugin update ship
Skills
Run /ship:use-ship when you want the agent to choose the right Ship route. Run /ship:auto when you explicitly want the full staged workflow. Or run individual phases when you only need one; atomic skills do not require an active auto run.
