[FEATURE] Implement "Auto" Effort With Low/Med/High/Max

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Effort levels need more guidance and transparency. Suggest implementing "Auto" effort level along side the current Low/Medium/High/Max effort settings. Auto would be more effort aware of the scope involved and potentially adjust to match.

I find it difficult to judge what the effort level should be before a prompt and what Claude actually has loaded in current cache or required reasoning. My workflow might be simple narrow focus website design, where low and medium efforts work great. But even within that narrow work it can be easy to get stuck in a situation where fix-this-break-that loops occur and higher reasoning/scope may have been needed or things are just missed. Max effort rarely encounters this, but is a greater and unnecessary demand on Anthropic's compute. The choices are run Medium which works fine until a frustrating issue occurs requiring manual inspection of the result or run Max full time and avoid the issue.

Proposed Solution

I suggest implementing an "Auto" effort setting that could adjust effort level based on perceived prompt, scope of work or perceived reasoning required. Without going into detail on what I'm not sure of, I'm sure there's plenty here that can be leveraged to achieve great quality results per the compute. The effort could simply adjust up/down a notch or two based on previous prompts, it could detect scope required before the task and adjust, or it could score user frustration and adjust as many examples. It could even adjust per prompt if done right.

Ideally the Auto effort would reach a happy balance where the quality is still high, but the compute and usage isn't taxed unnecessarily. I think having the ability to still select Max as needed should be available.

Unfortunately, there appears to be a noticeable hit to recalculate the cache when effort levels are changed, so this would need to be addressed first. Maybe Auto wouldn't suffer the recalculation penalty.

Alternative Solutions

Run Medium and run into frustrating, lower quality results.
Run Max and burn tokens unnecessarily.

Use another AI model to double-check work and present solution has worked instead of changing effort level.

Priority

Medium - Would be very helpful

Feature Category

Performance and speed

Use Case Example

1. Need a change to website table.
2. Change impacts other table alignments.
3. Ask to fix downstream alignment problems.
4. Fix to downstream alignments breaks original change.
5. Round' we go.

1. Documentation out of date.
2. Ask to update with latest changes.
3. Doesn't realize full scope of documentation even if referenced.
4. Only updates some files, not others.
5. Missed if not manually inspected.
6. Forgets rules and adds bloat.
7. Documentation drifts.

These aren't the only scenarios, it's one of many simple layouts where Max effort typically gets right.

Additional Context

_No response_