PR #15664: openai: honor reasoning_effort in /v1/responses endpoint

• Repository: ollama/ollama
• Author: balgaly
• State: open | merged: False
• Link: https://github.com/ollama/ollama/pull/15664

Description (problem / solution / changelog)

Fixes #15635.

Problem

/v1/chat/completions correctly disables thinking when reasoning_effort: "none" is passed, but /v1/responses ignores reasoning.effort entirely. Reproduction from the issue:

# Chat Completions — works (0.5s, no reasoning)
curl -s http://localhost:11434/v1/chat/completions \
  -d '{"model":"gemma4:e2b","messages":[{"role":"user","content":"Reply with ONLY: [1,3]"}],"reasoning_effort":"none","stream":false}'

# Responses — thinking NOT disabled (4.7s, full reasoning)
curl -s http://localhost:11434/v1/responses \
  -d '{"model":"gemma4:e2b","input":"Reply with ONLY: [1,3]","reasoning":{"effort":"none"},"stream":false}'

Root cause

openai/responses.go:FromResponsesRequest deserializes r.Reasoning.Effort (it is already used later for echoing reasoning config back to the client) but never converts it into the api.ChatRequest.Think field. The Chat Completions path at openai/openai.go:625-644 already performs this mapping.

Fix

Mirror the existing Chat Completions logic in FromResponsesRequest:

• effort == "none" → Think{Value: false} (suppresses thinking)
• effort == "high" | "medium" | "low" → Think{Value: <effort>}
• any other value → the same validation error the Chat Completions path returns

Validation, error message, and accepted effort values are identical across the two endpoints, so behavior stays consistent.

Tests

TestFromResponsesRequest_ReasoningEffort covers:

• "none" disables thinking (Think.Value == false)
• "low", "medium", "high" set the string value
• missing reasoning leaves Think unset
• invalid effort returns an error

All existing ./openai/... and ./middleware/... tests still pass.

Notes

• Only reasoning.effort is handled here — OpenAI's Responses API does not expose a flat reasoning_effort (that is a Chat Completions shape). Clients that sent {"reasoning_effort": "none"} on /v1/responses before this PR were effectively sending an unknown field; after the PR they should move to the canonical {"reasoning": {"effort": "none"}} form (which also did not work before).

Changed files

• openai/responses.go (modified, +14/-0)
• openai/responses_test.go (modified, +73/-0)