Repository navigation
Allow AI agents to submit feedback without confirmation - #7795
Conversation
|
Commit: Regression test reportSelected 1 changed tests (limit=20, 0 not selected) 1 tests: regression=1 unreleased=0 coverage=0 failing=0
TestAccept/cmd/feedback: branch ✅ | main (06e25a1) ❌ | latest (v1.27.0) ❌main (06e25a1): latest (v1.27.0): |
✅ Integration test reportCommit: 2be1b3f
Top 4 slowest tests (at least 2 minutes):
|
Check the harness request log after declined submissions and document interactive legacy-flag behavior.
There was a problem hiding this comment.
Automated review at 2be1b3f. Each finding was verified against the code.
No blocking issues found.
This PR removes the mandatory confirmation prompt for AI-agent-submitted feedback, lets agents submit directly, deprecates the --yes/-y flag, and updates help text with content guidance. The code changes are internally consistent: the deprecated flag still works, explicit --yes=false from agents is correctly rejected before authentication, acceptance and unit tests cover the new paths, and the removed helpers (checkApproval, confirm, escapeInvisible, IsErrorTTY) have no remaining callers. No concrete defects survived verification.
anton-107
left a comment
There was a problem hiding this comment.
A few questions inline about who decides that agent feedback gets sent.
| is set, because this command explicitly submits feedback. | ||
|
|
||
| AI agents can submit feedback directly without asking the user for approval. | ||
| Use --can-be-contacted only if the user explicitly agrees to be contacted. |
There was a problem hiding this comment.
Question: --can-be-contacted opts the authenticated user into being contacted by Databricks. Before, the prompt showed the user "Databricks may contact you: yes". Now only this help line stands between the agent and that choice:
trace AGENT=test-agent $CLI feedback --can-be-contacted "probe" < /dev/nullsends "can_be_contacted": true without the user seeing it. Should the CLI refuse --can-be-contacted from an agent unless the user has confirmed it?
There was a problem hiding this comment.
I'd keep --can-be-contacted available to agents in unattended runs. The standard Fast Feedback form starts with the contact box checked. The CLI leaves follow-up off by default, and the help tells agents to pass the flag only if the user has explicitly agreed to be contacted. I'm comfortable relying on that guidance here.
This comment was generated with GitHub MCP.
| such as --can-be-contacted, and pass --yes only after the user approves it.` | ||
| collected. Do not include customer content, personally identifiable information | ||
| (PII), credentials or other sensitive data. Follow-up contact is opt-in with | ||
| --can-be-contacted. Feedback is sent even when DATABRICKS_CLI_DISABLE_TELEMETRY |
There was a problem hiding this comment.
Question: sending despite DATABRICKS_CLI_DISABLE_TELEMETRY used to rest on "because you asked to send it". When an agent sends without a prompt, the user didn't ask, yet the behaviour stays and only the reason changes (here and in libs/telemetry/logger.go).
trace DATABRICKS_CLI_DISABLE_TELEMETRY=1 AGENT=test-agent $CLI feedback "probe" < /dev/nullposts the text to /telemetry-ext. For a user who opted out, what is left to stop an agent from filing feedback? Should agent submissions the user hasn't approved respect the opt-out?
There was a problem hiding this comment.
DATABRICKS_CLI_DISABLE_TELEMETRY turns off automatic CLI usage reporting. Calling databricks feedback is an explicit request to send the supplied message to Databricks. Fast Feedback uses /telemetry-ext for delivery, and the command already sends when usage telemetry is disabled. I'd keep that behavior when an agent invokes the command too.
This comment was generated with GitHub MCP.
Changes
Allow AI agents to submit
databricks feedbackdirectly, including when no terminal is attached. Keep--yes/-yaccepted as deprecated compatibility flags for existing callers. An explicit false value from an agent declines the submission before authentication, including when a terminal is attached.CLI help explicitly excludes customer content, personally identifiable information (PII), credentials and other sensitive data. It tells agents to use the existing
--can-be-contactedflag only with the user's explicit agreement and to notify the user with the reference ID after a successful submission. Update the telemetry documentation and remove the obsolete confirmation helpers.Why
Mandatory confirmation interrupts agent workflows and blocks noninteractive feedback. Direct submissions let agents report CLI problems while the receipt and agent instructions keep users informed.
Tests
cmd/feedback,libs/cmdioandlibs/telemetrypass. Feedback tests cover human and agent submissions, with and without a terminal, and check the receipt and agent attribution.This pull request and its description were written by Isaac.