Repository navigation
Restore OpenAI flagship tier: gpt-5.6-sol requires Responses API for tool use #653
Description
Found by the 2026-08-08 full-test-suite run: both gpt-medium and gpt-large (→ openai/gpt-5.6-sol) failed at iteration 1 with:
OpenAIException - Function tools with reasoning_effort are not supported for gpt-5.6-sol in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'.
OpenAI has moved tool-calling + reasoning for the 5.6 frontier family to the Responses API; litellm's standard openai/ route uses chat completions, so Sol currently cannot run rdb's agent loop at all. gpt-5.3-codex (the Codex line) still supports chat completions with tools and passed.
Interim fix (separate PR): gpt-medium/gpt-large co-point at openai/gpt-5.3-codex per the alias rules (capability at agentic coding is the admission test; every alias must resolve to a runnable model).
To restore a distinct OpenAI flagship tier, investigate:
- litellm's Responses API bridge (whether
litellm.completioncan routegpt-5.6-*via/v1/responsestransparently, and whether tool-calling round-trips through it) — this would be the model-agnostic path - explicitly NOT recommended: passing
reasoning_effort: "none"— a provider-specific shim that disables the model's headline capability - whichever path works should be validated with
e2e.ymlusing thetest=gpt-mediumfilter before repointing
Activity
gnovak commentedon Aug 8, 2026
Sizing update after research — this is small, not a loop rewrite:
litellm.completion()calls are transparently translated to/v1/responsesfor models that need it. No rdb loop changes required (https://docs.litellm.ai/docs/response_api and litellm PR #11632).toolsandreasoning_effortare set. rdb sets tools but never reasoning_effort, so the bridge doesn't engage. A small provider-gated change (passreasoning_effortforopenai/gpt-5.6*models in the three loops' completion calls) would activate it. Keep it gated — litellm mapsreasoning_effortto provider-specific thinking params for anthropic/gemini, so passing it unconditionally would change behavior on the other providers.e2e.ymldispatch withall_models=true, test=gpt-medium(~$0.15).Recommendation: watch #33221 first; if it stalls and we want Sol back sooner, do the gated one-liner.