Web app needs user-facing reasoning effort control (85 reasoning cap, adaptive router)

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

The web app's adaptive thinking router allocates reasoning depth based on perceived question complexity. Short questions that actually require real thinking and cross-referencing often get shallow treatment—even when project instructions say to assume all questions are complex.

This produces two related failure patterns:

1. Claude seems to reconstruct facts mid-response instead of verifying against project files.
2. Claude will contradict or ignore conclusions established a few messages earlier (in same chat). It behaves as if each turn is independent rather than building on what it already worked through.

Both patterns seem to share a root cause: the router sees a short question, allocates shallow reasoning, and Claude "shoots from the hip" instead of doing the work (e.g., reviewing earlier chat context, checking project files). Structured, longer prompts (file updates, project reviews) in the same chat session don't trigger this—it seems to come from short questions that actually require precision.

On topics with real stakes—health, compliance, financial analysis—this is a reliability problem, not a UX inconvenience. It means things with high stakes need to be closely verified and double checked.

Proposed Solution

Claude Code exposes /effort with levels up to max. Cowork has ultrathink for per-turn high effort. The API supports configurable thinking budgets. The web app does not have any such user override.

Expose a per-message or per-session reasoning effort control in the web app. It doesn't need to be prominent—an advanced setting or per-message toggle would be enough. Users who need it would opt in at the cost of more tokens per turn.

Alternative Solutions

The current workaround is switching to CLI (which has effort control) or double-checking high-stakes work on a second surface (e.g., another AI model).

Priority

High - Significant impact on productivity

Feature Category

Performance and speed

Use Case Example

_No response_

Additional Context

_No response_