Apiv2 cmlapi methods

[Khauneesh-AI]

Summary

• Migrated all 98 tool functions from hand-rolled requests calls to the official cmlapi Python SDK using CMLServiceApi
• Added shared setup_client() and serialize_result() helpers in http_helpers.py
• Net change: -7,320 lines, +2,245 lines across 103 files

---

Why

The previous implementation manually constructed URLs, headers, payloads, and error handling in every function. Most tool functions had roughly 80–130 lines of repeated request-handling logic.

This caused several issues:

• Silent failures, for example delete_application returning success without actually deleting the application
• Wrong URL patterns, such as /stop vs :stop and /restart vs :restart
• Missing fields in update payloads
• Inconsistent error handling across tools
• High maintenance overhead when adding or fixing API operations

Moving to the official SDK makes the implementation smaller, more consistent, and less prone to endpoint or payload drift.

---

What changed

• All tool functions now use the common SDK flow:

  client = setup_client(host, api_key)
  result = client.method_name(...)
  return serialize_result(result)

• Added shared SDK helpers in http_helpers.py:

  • setup_client(host, api_key)
  • serialize_result(result)

• create_registered_model now uses typed SDK objects:

  • CreateRegisteredModelRequest
  • Tag

  Raw dictionaries do not serialize correctly for this endpoint.

• upload_file remains requests-based because multipart PUT upload is not supported by the SDK.

• stdio_server.py fixes:

  • Fixed 4 tool wrappers that were returning raw dictionaries instead of json.dumps(...)

• Tests updated to mock cmlapi instead of requests.

• README updated with:

  • Architecture section
  • Updated prerequisites
  • Troubleshooting guidance

---

Test plan

• All unit tests pass:

• Syntax check completed on all 107 function files.

• MCP live testing against Workbench completed for tools, including:

  • Applications CRUD
  • Jobs CRUD
  • Experiments
  • Models
  • Registered models
  • Runtimes
  • Project files
  • Collaborators
  • Admin/quota endpoints

Known limitations

• list_experiment_runs does not support the sort parameter. The API returns 400, which appears to be an upstream limitation.
• create_experiment_run via API creates metadata only. Model artifacts still require MLflow running inside a CML session or job.
• Some admin endpoints, such as quotas and accelerator labels, return 403 for non-admin users. This is expected behavior.