[michaelp-db]

What did you change, and why?

Raised the databricks --version probe timeout from 10s to 30s (new constant _DATABRICKS_CLI_VERSION_TIMEOUT_SECONDS). Both _read_databricks_cli_version and the re-probe that ensure_databricks_cli_version runs for its error message now use it.

On a loaded arca machine, databricks --version through the dbexec wrapper can take longer than 10s. On arca, /usr/local/bin/databricks is a wrapper (exec /usr/local/bin/dbexec repo run databricks "$@"), not the raw binary, so every call pays dbexec overhead and contends with other dbexec tools that start at the same time.

Failure seen on arca on 2026-10-27, when isaac was launched from OpenUI:

ERROR Could not parse Databricks CLI version from `databricks --version` output: 'Databricks CLI v1.22.1'
Claude Code exited with code 1

The version string is valid: the parser regex handles it and the minimum is 1.0.0. What happened is that _read_databricks_cli_version returns None on TimeoutExpired, just as it does for unparseable output. ensure_databricks_cli_version then saw no parsed versions and ran --version again only to fill in the message. That second run succeeded, so the message shows a good version and blames parsing. The first probe had timed out.

Measured on arca (32 cores, load avg ~14 at the time):

• A single call takes ~0.03s. _discover_databricks_clis(use_cache=False) took 0.07–1.1s when idle.
• 16 concurrent databricks --version calls: min 0.10s, max 7.24s.
• 24 concurrent calls, bimodal: 13 at 0.27–0.41s, 7 at ~1.3s, 4 at 7.49–7.76s.

So moderate concurrency alone reaches ~7.7s, before any other machine load. When several isaac/OpenUI sessions start at once on a loaded box, the probe goes past 10s. The normal case is unaffected because the timeout only matters when the CLI is slow.

Possible follow-up (not in this PR): when the probe times out, the error should say "timed out" instead of "could not parse".

How do you know it works?

A new unit test checks that _read_databricks_cli_version passes the new timeout. tests/test_databricks.py and tests/test_desktop_auth.py pass (509 tests), and just lint is clean.

This pull request and its description were written by Isaac.

[michaelp-db]

Closing in favor of #1252 / #1326, which handle the slow first probe with a retry. Root cause (dbexec blocking on Arca port allocation) is being fixed in dbexec per ES-2318563. Arca data point: 24 concurrent databricks --version calls through dbexec took up to 7.8s.