[BUG] Cloud Projects sessions run at low reasoning effort (15) while session_context.effort_level is “high”

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Summary
In a claude.ai Projects thread session (cloud), session_context.effort_level is "high", but the reasoning-effort value the model receives on its turns is low (15). Switching effort seems to update the stored setting without changing what the model is served.

Environment

• Surface: Claude app, Projects (cloud thread session, environment_kind anthropic_cloud)
• Model: claude-opus-5-5[1m] (last_served_model: claude-opus-5-5)
• Date: 2026-10-11, ~19:48 UTC
• Session: session_01LvH3hWKAAGYDNyMFJjMKsG (parent/coordinator: session_013WeuKgBGtceKdKpgnEAnU3)

Impact
We require every session to run at medium effort or higher because low effort causes thrashing. There's no reliable way to confirm or enforce that while the stored setting and the served value disagree, and no in-session indicator shows the served value.

What Should Happen?

Expected
The served reasoning effort matches "high".

Actual

• get_session returns session_context.effort_level: "high".
• The reasoning-effort value in the model's context on that turn is 15.
• The coordinator session showed the same pattern: after switching itself to medium or high, its turns still behaved like low effort.

Caveat
The 15 is the value the model sees in its own context. As #33542 notes, a model's report of its own effort may not be reliable, so please check the session IDs above against server-side request parameters.

Error Messages/Logs

Steps to Reproduce

Steps to reproduce

1. Set the project's default thread effort to high.
2. Have the coordinator start a thread session with effort explicitly set to high.
3. In the thread, call get_session (claude-code-remote) with no session_id, and compare it with the reasoning-effort value the model receives on that turn.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

v2.1.296

Platform

Anthropic API

Operating System

Other Linux

Terminal/Shell

Non-interactive/CI environment

Additional Information

Related issues

• #67518: cloud sessions take effort only from session_context.effort_level. Here that field is "high" and apparently isn't honored.
• #93735: effective effort varies per desktop session with no settings change.
• #72249: effort is lowered mid-session without consent.
• #33542: effortLevel "max" served as 85. By comparison, 15 is far below what "high" should give.