[elainewang-db]

Summary

agentbricks --output json deploy printed more than its JSON payload on stdout, so a script reading the payload failed with JSONDecodeError: Extra data. Two sources:

• deploy runs databricks sync and databricks apps deploy without capturing their output, so the sync progress (Action: PUT: …, Uploaded …, Initial Sync Complete) and the App object apps deploy prints went to stdout.
• When it creates an App through the CLI (no tool needs request-user scopes), or resizes it, deploy relayed the App object apps create / create-update printed through Reporter.echo, which writes to stdout.

The fix sends both to stderr:

• The databricks CLI runner gets a to_stderr flag that sends an uncaptured command's output to stderr. sync, apps deploy, apps start, apps stop and apps delete use it, so deploy and deployments start/stop/delete keep stdout for their --output json payload. apps logs and apps run-local still print to stdout, since their output is what the user asked for.
• AppProvisioner relays the apps create / create-update output with Reporter.note (stderr) instead of Reporter.echo.

In normal output, all of this still shows in the terminal.

Before / after

The same project deployed with agentbricks --output json deploy ab762demo --source <project>, stdout and stderr captured to separate files, then stdout parsed with json.load.

Before (main at 966efe7): stdout is 117 lines. The App object from apps create, then the databricks sync progress, then the deploy payload:

$ agentbricks --output json deploy ab762demo --source ./ab762demo > out.json 2> err.txt
$ cat out.json
{
  "app_status": {
    "message": "App has status: App has not been deployed yet. Run your app by deploying source code",
    "state": "UNAVAILABLE"
  },
  ...
  "url": "https://agent-bricks-ab762demo-3272836215725701.aws.databricksapps.com"
}
Action: PUT: .agent/skills/agent-bricks-workflow/SKILL.md, .agents/skills/agent-bricks-workflow/SKILL.md, ...
Uploaded agent/tools
Uploaded .agent/skills/agent-bricks-workflow
Uploaded runtime
...
{
  "deployment": "agent-bricks-ab762demo",
  ...
}
$ python -c 'import json; json.load(open("out.json"))'
json.decoder.JSONDecodeError: Extra data: line 30 column 1 (char 1070)

After (this PR at 6f5555a, redeploying the same App): stdout is only the deploy payload, and it parses. The sync progress and the App object from apps deploy go to stderr:

$ agentbricks --output json deploy ab762demo --source ./ab762demo > out.json 2> err.txt
$ cat out.json
{
  "deployment": "agent-bricks-ab762demo",
  "url": "https://agent-bricks-ab762demo-3272836215725701.aws.databricksapps.com",
  "memory_store": "ab762demo-hhknkf-memory",
  "session_store": "ab762demo-hhknkf-sessions",
  "trace_grant": "granted",
  "store_grant": "granted",
  ...
}
$ python -c 'import json; json.load(open("out.json"))'   # no error
$ cat err.txt
Action: PUT: app.yaml
Uploaded app.yaml
Initial Sync Complete
{
  "create_time": "2026-10-16T21:24:02Z",
  "creator": "elaine.wang@databricks.com",
  ...

The before run was the App's first deploy (full upload); the after run was a redeploy, so its sync is shorter.

Testing

• pytest tests/unit_tests: 1878 passed. New tests: the runner sends output to fd 2 when to_stderr is set; each of the five lifecycle commands sets it; the provisioner relays the apps create and create-update output with note, not echo. Two existing tests' expected apps delete call now include to_stderr=True.
• ruff passes; ty reports the same diagnostics as main.
• E2E, from [l] [agentbricks] Add pytest E2E suites #706's suites on the integration workspace, which read deploy --output json: without this PR, every deploy test failed to parse deploy's output. With it, all 6 tests pass (test_deploy, test_memory, test_sessions, test_tracing, test_dev, test_tools_remove) in 6 min 5 s.

This pull request and its description were written by Isaac.

[jamesbxwu]

to help understand the behavioral change better, can you include a before / after screenshot in the PR description?