⚙️ FEATURE: CI/CD integration — structured exit codes, JSON output, and GitHub Actions step

[FernandoCelmer]

Problem

pycodeloop run always exits with code 0, even when the agent reports an error, hits max turns, or regressions are detected (#38). This makes it impossible to use pycodeloop reliably in CI pipelines — a failed agent run looks identical to a successful one from the shell's perspective. The "Code as Agent Harness" literature identifies CI/CD pipeline integration as a core harness deployment pattern.

Expected behavior

Exit codes

Code	Meaning
0	Agent completed successfully, verification passed
1	Agent error (exception, max turns reached)
2	Regression detected (baseline check failed post-run)
3	Budget exceeded

Machine-readable output

pycodeloop run "fix the auth bug" --ci --output json

{
  "status": "success",
  "turns": 4,
  "tokens": {"input": 12400, "output": 2100},
  "cost_usd": 0.042,
  "regression": false,
  "files_modified": ["src/auth.py", "tests/test_auth.py"]
}

GitHub Actions step

- name: Run pycodeloop agent
  uses: dotflow-io/pycodeloop-action@v1
  with:
    prompt: "Fix failing tests in src/auth.py"
    provider: claude
    model: claude-sonnet-4-6
    check_cmd: "pytest tests/ -q"
  env:
    ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}

Suggested implementation

• Add --ci flag to pycodeloop run CLI command
• Map agent outcomes to exit codes in cli/commands/run.py
• Add --output [text|json] flag; JSON mode serializes RunResult dataclass
• Add files_modified field to RunResult by reading from FileAccessLog
• Create dotflow-io/pycodeloop-action repository with composite action wrapping pip install pycodeloop && pycodeloop run

References

• "Code as Agent Harness" — arXiv 2605.18747, Section 5 (Deployment patterns)