Repository navigation
Reasoning effort 'medium' is not supported for model 'claude-haiku-4.5' #4345
Description
Describe the bug
When both copilot_cli_opus_medium_effort_default and copilot_cli_gpt_5_4_mini_for_explore feature flags are active (assigned server-side), the CLI throws the following error repeatedly during sub-agent execution:
Execution failed: Error: Reasoning effort 'medium' is not supported for model 'claude-haiku-4.5'.
The explore and task built-in agents are defined with model: claude-haiku-4.5 in definitions/explore.agent.yaml and definitions/task.agent.yaml . The copilot_cli_opus_medium_effort_default flag applies reasoning_effort: medium globally — including to these sub-agents — before the copilot_cli_gpt_5_4_mini_for_explore swap has a chance to replace the model. As a result, claude-haiku-4.5 receives a reasoning_effort value it does not support.
Affected version
1.0.77
Steps to reproduce the behavior
- Be assigned both flags: copilot_cli_opus_medium_effort_default: true and copilot_cli_gpt_5_4_mini_for_explore: true (visible in config.json under expAssignmentsCache )
- Trigger any task that uses the explore or task sub-agent
- Observe the error in output
Expected behavior
Either the model swap happens before the reasoning effort is applied, or reasoning_effort is omitted for models that do not support it (as already fixed in main: "Allow reasoning effort to be omitted so the server can select the default").
Additional context
Platform:
Windows ( win32-x64 )
Activity
indeherb commentedon Aug 5, 2026
Some copilot sessions ignore this error message and continue working, other times a session just doesn't do anything anymore.
dlemstra commentedon Aug 5, 2026
This can also happen when using auto mode and 'claude-haiku-4.5' is picked.
indeherb commentedon Aug 10, 2026
Hi @stephentoub — this is still occurring on v1.0.78 (current latest as of 2026-08-10), which is after this issue was closed on 2026-08-07.
Execution failed: Error: Reasoning effort 'medium' is not supported for model 'claude-haiku-4.5'.
We didn't receive any feedback or explanation for why this was closed — no comment was left, and the two follow-up reports before closure (confirming it was still reproducing, including in auto mode) don't appear to have been addressed. Given it's still happening on the latest version, could this be reopened, or could you clarify what fix was expected to resolve it?
stephentoub commentedon Aug 10, 2026
It was fixed in 1.0.79-7.
dlemstra commentedon Aug 13, 2026
@stephentoub It is still happening in 1.0.79: #4473. Also happened to me today.
aexklon commentedon Aug 14, 2026
Still happening
trcat commentedon Aug 21, 2026
Still happening，v1.0.80
dlemstra commentedon Aug 21, 2026
Looks like the latest
v1.0.81pre-releasehas this fixed.patniko commentedon Oct 26, 2026
Thanks everyone for the reports. Sub-agents and Auto mode now resolve reasoning effort against the model they actually run on, so they no longer send an effort level such as
mediumto a model likeclaude-haiku-4.5that doesn't support it; as noted above, this stopped reproducing as of Copilot CLI v1.0.81, and the duplicate #4473 has also been closed. If you still see "Reasoning effort ... is not supported" on the latest version, please reopen with your version and the model involved.Triaged as part of a Copilot SDK backlog review, with help from GitHub Copilot.