- Home
- Skills
- Agents & Orchestration
- Ship Loop: Issue to PR, 3 Adversarial Reviews, Merge
More screenshots
Works with the AI Agents you already use
Ship Loop: Issue to PR, 3 Adversarial Reviews, Merge
Claude ships one GitHub issue end to end and merges only when three read-only reviewers approve the current commit.
$10
Ship Loop: Issue to PR, 3 Adversarial Reviews, Merge
Example session with this skill installed
/ship #42
- Read your context and instructions
- Compiled the ship loop
Summary
Shipped
#42 (CSV export of an account's transactions) in
PR #57 after 2 review rounds. The reviewers blocked a cross-organisation data leak and the missing test that would have hidden it.
| Round | review-code | review-tests | review-docs | Gate |
|---|---|---|---|---|
| 1 @9f2c1e4 | BLOCK (1) | BLOCK (1) · suite exit 0 | APPROVE | BLOCK code tests |
| 2 @c41a7b0 | APPROVE · #1 resolved | APPROVE · #1 resolved · suite exit 0 | APPROVE | MERGE |
What the reviewers caught
Code Review — VERDICT: BLOCK · src/export/csv.ts:31: the query filters by the accountId from the path but never checks that the account belongs to the caller's organisation. Any logged-in user can export any account by guessing ids.
Test Review — VERDICT: BLOCK · Criterion 4, "an account in another organisation returns 404", has no test. Re-introducing the bug above keeps this suite green.
The fix
const account = await accounts.findForOrg(req.params.id, req.user.orgId);
if (!account) return notFound(res); // same 404 as a nonexistent id
Gate output, read back from GitHub
code: APPROVE r2 @c41a7b0
tests: APPROVE r2 @c41a7b0
docs: APPROVE r2 @c41a7b0
MERGE
Next steps
- Merged with
--squash, issue #42 closed, branch deleted. - Next unblocked issue: #43, scheduled exports. Run
/ship #43.
Connects securely to your tools. The creator never sees your data.
About this skill
Most agent coding loops end with the agent deciding its own work is fine. Ship Loop takes that decision away from it./ship #42 makes Claude branch, implement the issue, run your tests and open a PR. Then three reviewer subagents that cannot edit code take over: one for code, one for tests, one for docs. Each posts a verdict comment on the PR ending in a machine-readable marker. A small script reads those markers back from GitHub and prints MERGE, WAIT or BLOCK. Only verdicts for the PR's current head commit count, so an approval from before the last push can never sneak a change through.What the reviewers catch•Code: scope creep first, then missing authorization checks, swallowed errors, injection, races, broken transactions, and whatever invariants your CLAUDE.md lists.•Tests: runs the suite itself, then asks "if this code were broken, would these tests fail?" It flags happy-path-only tests, tautologies, over-mocking and acceptance criteria with no test.•Docs: undocumented public API and, above all, documents that now say something false.Built to keep the cost bounded•One review packet per round (PR, issues, spec sections, diff, CI) instead of three agents each re-fetching everything.•Round 2+ reviews only the delta, and every earlier finding is accounted for as resolved or still open.•Turn budgets scale with diff size, stalled reviewers are detected and resumed, and a round cap turns leftover findings into issues instead of looping forever.Any language: set TEST_CMD once. Includes an installer, a config template, 3 agents, 5 POSIX shell scripts and 3 annotated example sessions. Requires Claude Code and the gh CLI.
None so far, but there might very welll be limitations due to different workflows.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 27 days ago
Requires
Permissions
Allowed Hosts
File Scopes
Shell: runs git, gh, your configured test/lint commands and the bundled POSIX scripts (gate, packet, test, check, ci-status). Read/Write: implements the issue in your repository, so it edits source files; it also writes its own scratch files under .claude/ (gitignored by install.sh). Network/GitHub: creates a branch, opens a PR, posts review comments, reads them back, can dispatch your CI workflow, and squash-merges ONLY when the gate script prints MERGE. The three reviewer agents are read-only: they cannot edit files, only comment. It never force-pushes, never commits to the base branch and never deletes a review comment.