BYOK gpt-5.5 --effort xhigh falls back to medium for individual plan

[Mapleeeeeeeeeee]

Summary

When using Copilot CLI with a BYOK/custom provider for gpt-5.5, --effort xhigh falls back to medium if the user is signed in with an individual Copilot plan.

With the same BYOK provider, the same gpt-5.5 model, and the same command, xhigh works as expected when COPILOT_OFFLINE=true is set.

Environment

• Copilot CLI: 1.0.40
• Platform: macOS arm64
• Provider mode: BYOK/custom provider
• Provider wire API: responses
• Model: gpt-5.5 through a custom provider
• GitHub Copilot plan: individual

Relevant environment variables:

COPILOT_MODEL=gpt-5.5
COPILOT_PROVIDER_BASE_URL=<custom OpenAI-compatible endpoint>
COPILOT_PROVIDER_API_KEY=<redacted>
COPILOT_PROVIDER_WIRE_API=responses

Reproduction steps

1. Configure a BYOK/custom provider:

export COPILOT_MODEL=gpt-5.5
export COPILOT_PROVIDER_BASE_URL=<custom OpenAI-compatible endpoint>
export COPILOT_PROVIDER_API_KEY=<redacted>
export COPILOT_PROVIDER_WIRE_API=responses

2. Run:

copilot --effort xhigh --log-level debug --prompt "Reply with exactly: OK" --silent --allow-all-tools

3. Check the debug log.

Actual behavior

Copilot CLI uses the custom provider, but the effective reasoning effort becomes medium:

Using custom provider: type=openai, wireApi=responses
"defaultReasoningEffort": "medium"
"reasoning_effort": "medium"

The debug log also shows that the GitHub plan context is individual:

github_copilotcli_copilotplan=individual

Expected behavior

When COPILOT_PROVIDER_BASE_URL is set and inference is routed to a custom provider, --effort xhigh should be preserved if the selected provider/model supports it. It should not be downgraded based on the GitHub Copilot account plan.

Expected debug output:

"defaultReasoningEffort": "xhigh"
"reasoning_effort": "xhigh"

Control test

With the same BYOK provider, the same gpt-5.5 model, and the same command, adding COPILOT_OFFLINE=true makes xhigh work:

COPILOT_OFFLINE=true copilot --effort xhigh --log-level debug --prompt "Reply with exactly: OK" --silent --allow-all-tools

Debug log:

Running in offline mode. GitHub features are unavailable.
Using custom provider: type=openai, wireApi=responses
"defaultReasoningEffort": "xhigh"

Additional check

With the same BYOK provider, --effort high works correctly:

"defaultReasoningEffort": "high"
"reasoning_effort": "high"