Add ucode doctor command to diagnose and fix the local setup

[max-rozen-oss-db]

Summary

Adds ucode doctor — a health check for the local ucode setup, modeled on brew doctor / flutter doctor / npm doctor. It runs a series of independent checks, prints a per-check status line, and for any problem it knows how to fix, prompts to apply the fix and reports whether it worked.

Checks

Check	On problem
uv on PATH	error — needed to install/upgrade ucode
npm on PATH	warn — needed to install agent CLIs automatically
Databricks CLI installed	offers to install
Databricks CLI version vs the databricks aitools floor	offers to upgrade
Workspace configured	warn — points at ucode configure
Databricks credentials valid	offers to run databricks auth login (the usual launch failure is an expired/invalid token surfacing as a 403 from the agent)
ANTHROPIC_AUTH_TOKEN / ANTHROPIC_API_KEY in the env	warn — it shadows the credential ucode configures for Claude Code (advisory; can't unset a parent shell's env)
Each configured coding agent's CLI	offers to install (if missing) or update (if outdated)
mlflow CLI, when tracing is enabled	offers to install the pinned version
ucode's installed build	offers a reinstall-to-latest from GitHub

Checks that don't apply are skipped: credentials/tracing only run when a workspace is configured / tracing is enabled, and the env-collision check only appears when one of those vars is set.

Behavior

• Opt-in fixes. The command is read-only until the user says yes to a specific suggestion. Each prompt defaults to no, so a declined or non-interactive (piped) run changes nothing.
• Per-fix outcome. After each accepted fix it prints ✓ <check>: fixed or ! <check>: fix did not complete. A failed fix doesn't abort the run or the remaining checks.
• Output. A ✓/!/✗/• status glyph per line, then a one-line summary (No problems detected. or N issues found; M fix(es) applied.).

Example (healthy machine, piped so prompts default to no):

ucode doctor

  ✓ uv: found on PATH
  ✓ npm: found on PATH
  ✓ Databricks CLI: v1.2.1
  ✓ Workspace: https://…databricks.com
  ✓ Databricks auth: credentials are valid
  ✓ Claude Code: installed and up to date
  • ucode: v0.1.0+… (installed from GitHub)
✔ No problems detected.

Implementation notes

New logic lives in src/ucode/doctor.py; cli.py gets a thin doctor command wrapper. To avoid reaching into privates, added read-only public helpers:

• databricks.py: databricks_cli_version() (report-only, never upgrades) and upgrade_databricks_cli().
• agents/__init__.py: tool_binary_installed(), tool_update_available(), update_tool_binary(), tracing_mlflow_ok(), ensure_tracing_mlflow_cli().

Testing

• uv run pytest tests/test_doctor.py tests/test_databricks.py tests/test_cli.py — passing.
• uv run ruff check . and uv run ruff format --check src/ tests/ — clean.
• tests/test_doctor.py covers each check's classification (ok/warn/error + whether a fix is offered), the skip-when-not-applicable paths (no workspace, tracing disabled, env var unset/blank), and the doctor flow (applies on accept, skips on decline, reports failure without raising, never prompts on a healthy check).
• One unrelated test (test_mcp ordering assertion) fails identically on main — pre-existing, not touched by this PR.