BYOK model limits catalog is missing gpt-5.5

組み込みの BYOK モデル制限カタログとその lookup を見つけ、近接する gpt-5.4、gpt-5.4-mini、gpt-5.4-nano のエントリを比較します。手動のトークンオーバーライドを行わずに、提供されている Copilot CLI の再現手順を再実行します。gpt-5.5 が fallback 警告なしでカタログの制限を解決できれば完了です。

Describe the bug

When using BYOK / a custom OpenAI-compatible provider with gpt-5.5, Copilot CLI treats the model as missing from the built-in limits catalog and falls back to default token limits.

This appears to be a catalog mismatch: gpt-5.5 can be listed/used by the provider, but the BYOK limits catalog does not include an entry for it. As a result, context/compaction uses a much smaller default prompt budget unless users manually set COPILOT_PROVIDER_MAX_PROMPT_TOKENS and COPILOT_PROVIDER_MAX_OUTPUT_TOKENS.

Affected version

GitHub Copilot CLI 1.0.40

Steps to reproduce the behavior

Use an OpenAI-compatible BYOK provider that exposes gpt-5.5, then run Copilot CLI with the model selected but without manual token limit overrides:

export COPILOT_PROVIDER_TYPE=openai
export COPILOT_PROVIDER_BASE_URL=<redacted>/v1
export COPILOT_PROVIDER_API_KEY=<redacted>
export COPILOT_PROVIDER_MODEL_ID=gpt-5.5
export COPILOT_PROVIDER_WIRE_API=responses
unset COPILOT_PROVIDER_MAX_PROMPT_TOKENS
unset COPILOT_PROVIDER_MAX_OUTPUT_TOKENS

copilot --log-level debug -p 'Reply with exactly OK and nothing else.'

Observed debug log excerpts:

[WARNING] Model "gpt-5.5" is not in the built-in catalog. Using defaults for: prompt tokens (COPILOT_PROVIDER_MAX_PROMPT_TOKENS), output tokens (COPILOT_PROVIDER_MAX_OUTPUT_TOKENS). Run `copilot help providers` for configuration details.
[INFO] Using custom provider: type=openai, wireApi=responses
[DEBUG] Using model: gpt-5.5
[DEBUG] Listed models: [gpt-5.5,gpt-5.5]
"max_prompt_tokens": 128000,
"max_context_window_tokens": 200000,
CompactionProcessor: Utilization 15.4% (19674/128000 tokens) below threshold 80%

If manual overrides are provided, they do take effect:

export COPILOT_PROVIDER_MAX_PROMPT_TOKENS=400000
export COPILOT_PROVIDER_MAX_OUTPUT_TOKENS=128000

Then the compaction log uses the manual prompt limit:

CompactionProcessor: Utilization 4.9% (19666/400000 tokens) below threshold 80%

Expected behavior

gpt-5.5 should either:

1. be included in the BYOK built-in model limits catalog with appropriate prompt/context/output limits, or
2. have a documented way to inherit/resolve limits for this model without falling back to the generic defaults.

Actual behavior

gpt-5.5 is usable from the provider model list, but the BYOK model limits lookup misses it and falls back to defaults:

max_prompt_tokens: 128000
max_context_window_tokens: 200000

Additional context

The nearby GPT-5.x entries in the BYOK catalog appear to include models such as gpt-5.4, gpt-5.4-mini, and gpt-5.4-nano with larger limits. The issue seems specific to gpt-5.5 being absent from the built-in BYOK limits catalog.