Self-Healing CI/CD Pipeline

Tests fail in CI → architect automatically fixes them → creates a PR with the fix.

The problem

The most frustrating CI/CD pattern: a developer pushes, tests fail due to an edge case, and the pipeline stalls waiting for human attention. At 3am nobody is watching. At 9am there are 5 blocked PRs and the team starts the day resolving broken tests instead of building features.

In 2026 this has a name: the “Pipeline Doctor” or “Interceptor pattern”. A failure is not a stop signal — it is a trigger for a repair agent. GitHub, GitLab, and the major platforms are converging toward this model.

Where architect fits in

Architect positions itself as the Repair Agent between the test failure and the creation of a PR with the fix. Its Ralph Loop is exactly the primitive this pattern needs: fix→test→verify in a loop until it passes or the budget runs out.

Diagram

flowchart TD
    A["👨‍💻 Developer Push"] --> B["GitHub Actions / GitLab CI"]
    B --> C{"🧪 Tests"}
    C -->|"✅ Pass"| D["Deploy / Merge"]
    C -->|"❌ Fail"| E["architect loop\n--check 'pytest'\n--budget 0.50"]

    E --> F{"Ralph Loop"}
    F --> G["LLM analyzes\nerror logs"]
    G --> H["Applies fix\n(guardrails active)"]
    H --> I{"pytest passes?"}
    I -->|"❌ No"| G
    I -->|"✅ Yes"| J["Generates JSON report"]

    J --> K["git commit + push\nnew branch"]
    K --> L["Creates automatic PR\nwith attached report"]
    L --> M["👨‍💻 Human Code Review"]
    M --> D

    F -->|"Budget exhausted\nor max_iterations"| N["❌ Escalation\nNotify team"]

    style E fill:#2563eb,color:#fff,stroke:#1d4ed8
    style F fill:#2563eb,color:#fff,stroke:#1d4ed8
    style H fill:#7c3aed,color:#fff,stroke:#6d28d9
    style J fill:#059669,color:#fff,stroke:#047857
    style N fill:#dc2626,color:#fff,stroke:#b91c1c

Implementation

GitHub Actions workflow

# .github/workflows/self-healing.yml
name: Self-Healing Tests
on: [push, pull_request]

jobs:
  test:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: '3.12'

      - name: Run tests
        id: tests
        run: pytest tests/ -q
        continue-on-error: true

      - name: Auto-fix with architect
        if: steps.tests.outcome == 'failure'
        run: |
          pip install architect-ai-cli
          architect loop "Fix the failing tests. \
            Read the pytest output to understand what fails \
            and apply the minimal fix needed." \
            --check "pytest tests/ -q" \
            --config .architect.yaml \
            --confirm-mode yolo \
            --budget 0.50 \
            --max-iterations 5 \
            --report-file fix-report.json \
            --exit-code-on-partial 1
        env:
          OPENAI_API_KEY: ${{ secrets.LLM_KEY }}

      - name: Create PR with fix
        if: steps.tests.outcome == 'failure' && success()
        uses: peter-evans/create-pull-request@v6
        with:
          title: "[architect] Auto-fix: tests corrected"
          body-path: fix-report.json
          branch: architect/auto-fix-${{ github.sha }}
          commit-message: "fix: auto-remediation via architect Ralph Loop"

Architect configuration

# .architect.yaml
llm:
  model: openai/gpt-4.1
  api_base: https://api.openai.com/v1
  api_key_env: OPENAI_API_KEY

guardrails:
  protected_files:
    - ".env"
    - "*.pem"
    - "*.key"
    - "docker-compose.yml"
    - "Dockerfile"
    - ".github/**"
  max_files_modified: 10
  code_rules:
    - pattern: 'eval\('
      severity: block
    - pattern: 'exec\('
      severity: block

costs:
  budget_usd: 0.50

Architect features used

Feature	Role in this architecture
Ralph Loop	Core: fix→test→verify cycle until pytest passes
Guardrails	Protects CI/CD files, secrets, and Dockerfiles
Budget	Hard limit to prevent runaway costs at 3am
Reports	JSON attached to the PR as evidence of what changed
Exit codes	--exit-code-on-partial so CI knows if it was successful
.architect.md	Project conventions respected in the fix

Escalation flow

If architect cannot fix the tests (budget exhausted or max_iterations reached), the workflow must escalate:

      - name: Notify failure
        if: steps.tests.outcome == 'failure' && failure()
        uses: slackapi/slack-github-action@v1
        with:
          payload: |
            {
              "text": "⚠️ Auto-fix failed in ${{ github.repository }}. Tests still broken after 5 attempts ($0.50 budget). Requires manual intervention.",
              "blocks": [...]
            }

Differential value

Without architect, implementing this pattern requires:

• Custom retry loop
• Pytest error parsing
• Secure code execution with sandboxing
• Cost tracking per iteration
• Iteration limits with fallback
• Report generation for the PR

Architect packages it all into a single command with guardrails included. The GitHub Actions workflow goes from ~100 lines of custom script to ~15 lines.