sampler/runner: enable gpt-oss structured outputs

harmony/harmonyparser.go

@@ -52,6 +52,7 @@ type HarmonyParser struct {
 	MessageStartTag  string
 	MessageEndTag    string
 	HeaderEndTag     string
+	ConstrainAllowed bool
 	acc              strings.Builder
 	lifetimeAcc      strings.Builder
 }
@@ -328,6 +329,7 @@ func (h *HarmonyMessageHandler) AddContent(content string, toolParser *HarmonyTo
 				}
 			case "final":
 				h.state = harmonyMessageState_Normal
+				h.HarmonyParser.ConstrainAllowed = true
 			}
 		case HarmonyEventContentEmitted:
 			logutil.Trace("harmony event content", "content", event.Content, "state", h.state)

runner/ollamarunner/runner.go

@@ -814,7 +814,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 		req.Options.TopP,
 		req.Options.MinP,
 		req.Options.Seed,
-		grammar,
+		nil,
 	)
 
 	seq, err := s.NewSequence(req.Prompt, req.Images, NewSequenceParams{
@@ -865,6 +865,12 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	// TODO(parthsareen): generalize grammar enablement on the fly for all thinking models
+	if harmonyMessageHandler == nil {
+		seq.sampler.SetGrammar(grammar)
+	}
+
+	grammarSet := false
 	for {
 		select {
 		case <-r.Context().Done():
@@ -877,6 +883,10 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 					var toolContent string
 					content, thinking, toolContent = harmonyMessageHandler.AddContent(content, harmonyToolParser)
 					harmonyToolParser.Add(toolContent)
+					if harmonyMessageHandler.HarmonyParser.ConstrainAllowed && !grammarSet {
+						seq.sampler.SetGrammar(grammar)
+						grammarSet = true
+					}
 				}
 
 				if err := json.NewEncoder(w).Encode(&llm.CompletionResponse{

sample/samplers.go

@@ -25,6 +25,10 @@ type Sampler struct {
 	grammar     *GrammarSampler
 }
 
+func (s *Sampler) SetGrammar(grammar *GrammarSampler) {
+	s.grammar = grammar
+}
+
 func (s *Sampler) Sample(logits []float32) (int32, error) {
 	if len(logits) == 0 {
 		return -1, errors.New("sample: no logits provided to sample")