[DOCS] Auto mode docs don't explain binding user constraints

Documentation Type

Missing documentation (feature not documented)

Documentation Location

https://code.claude.com/docs/en/permission-modes

Section/Topic

The Eliminate prompts with auto mode section, especially How actions are evaluated and the Allowed by default guidance.

Current Documentation

The auto mode page currently says:

The classifier receives user messages and tool calls as input, with Claude's own text and tool results stripped out.

It also says the classifier checks whether an action is:

an overeager escalation beyond what you asked for ... or a sudden departure from your stated intent

And the page lists this as allowed by default:

Pushing to the branch you started on or one Claude created

No documentation currently states that an explicit user instruction like "don't push", "wait for approval before merging", or "wait for X before Y" overrides an action that would otherwise be allowed by default.

What's Wrong or Missing?

The docs imply that auto mode reads the user's messages and reasons about intent, but they never make the behavioral contract explicit: direct user prohibitions stated in the conversation are supposed to be binding, even when the action would otherwise fall under the default allow list.

This leaves three gaps:

1. Users are not told that explicit negative constraints such as "don't push" or "wait until I say so" override default allowed actions.
2. The docs do not distinguish these user-imposed boundaries from the classifier's built-in block list.
3. There are no examples showing how users should phrase such constraints when working in auto mode.

The v2.1.90 release note confirms this is intended product behavior, not an edge case:

"Fixed auto mode not respecting explicit user boundaries (\"don't push\", \"wait for X before Y\") even when the action would otherwise be allowed"

Suggested Improvement

Add a short note in the auto mode page explaining that explicit user boundaries stated in the conversation override default allowed actions.

For example:

Explicit user boundaries override default allow behavior: If the user says "don't push", "wait for approval before merging", or "don't deploy until I say so", auto mode should treat that instruction as binding even if the action would otherwise be allowed by default.

Also add one or two concrete examples near the Allowed by default list showing that users can narrow Claude's authority with follow-up instructions in the conversation.

Impact

Medium - Makes feature difficult to understand

Additional Context

Affected Pages:
| Page | Context |
|------|---------|
| https://code.claude.com/docs/en/permission-modes | Auto mode classifier behavior is documented, but the page does not explicitly say that direct user prohibitions override default allowed actions |

Total scope: 1 page affected

Source: Changelog v2.1.90

"Fixed auto mode not respecting explicit user boundaries (\"don't push\", \"wait for X before Y\") even when the action would otherwise be allowed"