Claude Code Auto-Fix Meets API Governance

Claude Code can now auto-fix CI failures. It reads the error, patches the code, and pushes a fix — all without you touching the keyboard. But what happens when the CI check IS the governance layer?

You get a closed loop: Delimit flags the breaking change, Claude Code reads the structured error, fixes the code, and the PR stays green. No human in the middle. No Slack thread asking "is this safe to merge?"

The loop

1. Developer opens a PR that modifies an API endpoint — renames a field, removes a parameter, changes a response type.
2. Delimit Action runs in CI. It diffs the OpenAPI spec against the base branch. It finds a breaking change: response_field_removed on GET /users. CI fails.
3. Claude Code Action triggers on the failure. It reads Delimit's structured output — change type, severity, affected path, migration guidance.
4. Claude fixes the code. It restores backward compatibility (adds the old field back as an alias, or marks it deprecated instead of removing it). Pushes the fix.
5. Delimit runs again. No breaking changes. CI passes. PR is green.

Real example: PR #11

We tested this on our own repo. PR #11 introduced a breaking API change. Delimit caught it, posted a structured comment with the exact change type and migration path, and Claude Code resolved it — all in CI, no manual intervention.

Setup

Two actions in your workflow. That's it.

name: API Governance + Auto-Fix
on: pull_request

jobs:
  check:
    runs-on: ubuntu-latest
    permissions:
      pull-requests: write
      contents: write
    steps:
      - uses: actions/checkout@v4

      # Step 1: Delimit catches breaking changes
      - uses: delimit-ai/delimit-action@v1
        with:
          spec: api/openapi.yaml

      # Step 2: Claude fixes what Delimit flags
      - uses: anthropics/claude-code-action@beta
        if: failure()
        with:
          claude_api_key: ${{ secrets.CLAUDE_API_KEY }}

Why this matters

Auto-fix without governance is dangerous — the AI might "fix" a failing test by deleting the test. Auto-fix with governance is powerful — the AI knows exactly what broke, why it broke, and what backward-compatible fix to apply.

Delimit gives Claude Code the structured signal it needs. Not a vague error message. A typed change classification: response_field_removed, parameter_type_changed, endpoint_removed. 27 change types, each with severity and migration guidance.

Install Delimit. Let your AI fix what your governance catches. Keep building.