bug: repair_run triggers full job run instead of repairing failed tasks

[sgarla]

Summary

When a user asks Claude to perform a task repair (re-run only the failed tasks of a previous job run), the skill triggers a full run_now instead of a repair_run. This is because repair_run is not implemented in the MCP tools.

Steps to Reproduce

1. Run a Databricks job where one or more tasks fail
2. Ask Claude Code (with ai-dev-kit) to repair the failed run
3. Observe that a new full job run is triggered instead of repairing the failed tasks

Expected Behavior

Claude should call jobs.repair_run(run_id=<failed_run_id>, ...) to re-run only the failed/skipped tasks from the original run, preserving the successful task outputs.

Actual Behavior

Claude falls back to manage_job_runs(action='run_now', job_id=...), which starts a brand new full run from scratch.

Root Cause

The manage_job_runs MCP tool only exposes these actions: run_now, get, get_output, cancel, list, wait.

A repair action is missing. The Databricks SDK supports w.jobs.repair_run() but it has not been implemented in:

• databricks-mcp-server/databricks_mcp_server/tools/jobs.py
• databricks-tools-core/databricks_tools_core/jobs/runs.py

Proposed Fix

1. Add a repair_run() function in databricks_tools_core/jobs/runs.py using w.jobs.repair_run()
2. Add a repair action to manage_job_runs in databricks_mcp_server/tools/jobs.py
3. Update SKILL.md to document the repair workflow

Impact

Customers using ai-dev-kit for job orchestration and failure recovery are inadvertently re-running entire jobs, wasting compute and time.