Claude Code Multimodel Workflow — Git skill for Claude Code

claude-code-multimodel-workflow

**v2 — server-enforced merge gates for semi-autonomous repos.** (v1 of this repo packaged the full Claude Code workflow as a plugin; it survives at the `plugin-v1-archive` tag. v2 narrows to the piece that proved most worth sharing: the merge-gate infrastructure.)

Templates for putting the **merge decision on the server** so that humans and coding agents can share a repo without per-PR human approval — and without trusting any agent's self-report. Proven on two private instantiations (a multi-branch product monorepo and a single-account homelab monorepo) before extraction; the knobs that differed between them are exactly the parameters these templates expose.

The problem

Path-filtered CI plus advisory bot reviews is theater the moment an agent (or a tired human) can merge: any write-permission integration can post a passing status, a suite that never fired looks the same as a suite that passed, and local git hooks don't exist on a fresh clone. If agents open PRs, the merge gate has to be something no PR author — human or agent — can forge or dodge.

The shape

One **always-run required check** (`merge-gate`) attests that every CI suite the PR's diff *requires* has succeeded on the head SHA, and a **branch ruleset** makes that check — bound to the GitHub Actions app — the only thing standing between green CI and merge. Everything else (bot reviews, coverage receipts) is deliberately advisory.

PR opened ──► suites fire (path-filtered, as usual)
         └──► merge-gate fires (NO path filter — always materializes)
                 │  reads the diff, decides which suites are required,
                 │  polls until each has a successful run for this SHA
                 ▼
          ruleset requires merge-gate (app-bound) ──► merge allowed

The attestation contract

A ruleset closes the forgery class only with **all** of the following; anything less is theater:

1. Required checks **bound to the GitHub Actions app