Merge pull request #1411 from krissetto/no-think-from-yaml

Allow disabling thinking via yaml with thinking_budget: 0 or none

Author: dgageot

cagent-schema.json

@@ -459,26 +459,29 @@
           "description": "Whether to track usage"
         },
         "thinking_budget": {
-          "description": "Controls reasoning effort/budget. OpenAI: string levels ('minimal','low','medium','high'), default 'medium'. Anthropic: integer token budget (1024-32768), default 8192. Amazon Bedrock (Claude): same as Anthropic. Google Gemini 2.5: integer token budget (-1 for dynamic, 0 to disable, 24576 max), default -1. Google Gemini 3: string levels ('minimal' Flash only,'low','medium','high'), default 'high' for Pro, 'medium' for Flash.",
+          "description": "Controls reasoning effort/budget. Use 'none' or 0 to disable thinking. OpenAI: string levels ('minimal','low','medium','high'), default 'medium'. Anthropic: integer token budget (1024-32768), default 8192. Amazon Bedrock (Claude): same as Anthropic. Google Gemini 2.5: integer token budget (-1 for dynamic, 0 to disable, 24576 max), default -1. Google Gemini 3: string levels ('minimal' Flash only,'low','medium','high'), default 'high' for Pro, 'medium' for Flash.",
           "oneOf": [
             {
               "type": "string",
               "enum": [
+                "none",
                 "minimal",
                 "low",
                 "medium",
                 "high"
               ],
-              "description": "Reasoning effort level (OpenAI, Gemini 3)"
+              "description": "Reasoning effort level (OpenAI, Gemini 3). Use 'none' to disable thinking."
             },
             {
               "type": "integer",
               "minimum": -1,
               "maximum": 32768,
-              "description": "Token budget for extended thinking (Anthropic, Bedrock Claude, Gemini 2.5)"
+              "description": "Token budget for extended thinking (Anthropic, Bedrock Claude, Gemini 2.5). Use 0 to disable thinking."
             }
           ],
           "examples": [
+            "none",
+            0,
             "minimal",
             "low",
             "medium",

cmd/root/run.go

@@ -310,10 +310,17 @@ func (f *runExecFlags) createLocalRuntimeAndSession(ctx context.Context, loadRes
 
 		slog.Debug("Loaded existing session", "session_id", f.sessionID, "agent", f.agentName)
 	} else {
+		thinking := true
+		if tb := agent.Model().BaseConfig().ModelConfig.ThinkingBudget; tb != nil {
+			if tb.Effort == "none" || (tb.Tokens == 0 && tb.Effort == "") {
+				thinking = false
+			}
+		}
 		sess = session.New(
 			session.WithMaxIterations(agent.MaxIterations()),
 			session.WithToolsApproved(f.autoApprove),
 			session.WithHideToolResults(f.hideToolResults),
+			session.WithThinking(thinking),
 		)
 		// Session is stored lazily on first UpdateSession call (when content is added)
 		// This avoids creating empty sessions in the database

docs/USAGE.md

@@ -284,12 +284,32 @@ models:
 
 Determine how much the model should think by setting the `thinking_budget`
 
+- **Disable thinking**: Use `none` (string) or `0` (integer) to explicitly disable thinking for any provider
 - **OpenAI**: use effort levels — `minimal`, `low`, `medium`, `high`. Default: `medium`
 - **Anthropic**: set an integer token budget. Range is 1024–32768; must be strictly less than `max_tokens`. Default: `8192` with `interleaved_thinking: true`
 - **Google (Gemini 2.5)**: set an integer token budget. `0` -> disable thinking, `-1` -> dynamic thinking (model decides). Default: `-1` (dynamic)
 - **Google (Gemini 3)**: use effort levels — `minimal` (Flash only), `low`, `medium`, `high`. Default: `high` for Pro, `medium` for Flash
 - **Amazon Bedrock (Claude models)**: set an integer token budget, same as Anthropic. Default: `8192` with `interleaved_thinking: true`
 
+**Disabling thinking:**
+
+```yaml
+models:
+  # Using string "none"
+  gpt-no-thinking:
+    provider: openai
+    model: gpt-5
+    thinking_budget: none # or 0
+
+  # Using integer 0
+  claude-no-thinking:
+    provider: anthropic
+    model: claude-sonnet-4-5
+    thinking_budget: 0 # or none
+```
+
+Note: When thinking is disabled via config, it can still be enabled during a session using the `/think` command, which will restore the provider's default thinking configuration.
+
 Examples (OpenAI):
 
 ```yaml

pkg/model/provider/override_test.go

@@ -341,3 +341,120 @@ func TestApplyOverrides_DoesNotModifyOriginal(t *testing.T) {
 	// Result should have changes
 	assert.Nil(t, result.ThinkingBudget, "Result ThinkingBudget should be nil")
 }
+
+// TestApplyOverrides_EnableFromDisabled tests that explicitly enabling thinking when
+// the config has it disabled (Tokens=0 or Effort="none") restores provider defaults.
+// This is the key behavior that makes /think work when YAML starts with thinking_budget: 0/none.
+func TestApplyOverrides_EnableFromDisabled(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name                 string
+		config               *latest.ModelConfig
+		expectThinkingBudget *latest.ThinkingBudget
+	}{
+		{
+			name: "Anthropic: enable from Tokens=0 restores default 8192",
+			config: &latest.ModelConfig{
+				Provider:       "anthropic",
+				Model:          "claude-sonnet-4-0",
+				ThinkingBudget: &latest.ThinkingBudget{Tokens: 0},
+			},
+			expectThinkingBudget: &latest.ThinkingBudget{Tokens: 8192},
+		},
+		{
+			name: "Anthropic: enable from Effort=none restores default 8192",
+			config: &latest.ModelConfig{
+				Provider:       "anthropic",
+				Model:          "claude-sonnet-4-0",
+				ThinkingBudget: &latest.ThinkingBudget{Effort: "none"},
+			},
+			expectThinkingBudget: &latest.ThinkingBudget{Tokens: 8192},
+		},
+		{
+			name: "OpenAI: enable from Tokens=0 restores default medium",
+			config: &latest.ModelConfig{
+				Provider:       "openai",
+				Model:          "gpt-4o",
+				ThinkingBudget: &latest.ThinkingBudget{Tokens: 0},
+			},
+			expectThinkingBudget: &latest.ThinkingBudget{Effort: "medium"},
+		},
+		{
+			name: "OpenAI: enable from Effort=none restores default medium",
+			config: &latest.ModelConfig{
+				Provider:       "openai",
+				Model:          "gpt-4o",
+				ThinkingBudget: &latest.ThinkingBudget{Effort: "none"},
+			},
+			expectThinkingBudget: &latest.ThinkingBudget{Effort: "medium"},
+		},
+		{
+			name: "Gemini 2.5: enable from Tokens=0 restores default -1 (dynamic)",
+			config: &latest.ModelConfig{
+				Provider:       "google",
+				Model:          "gemini-2.5-flash",
+				ThinkingBudget: &latest.ThinkingBudget{Tokens: 0},
+			},
+			expectThinkingBudget: &latest.ThinkingBudget{Tokens: -1},
+		},
+		{
+			name: "Bedrock Claude: enable from Tokens=0 restores default 8192",
+			config: &latest.ModelConfig{
+				Provider:       "amazon-bedrock",
+				Model:          "anthropic.claude-3-sonnet",
+				ThinkingBudget: &latest.ThinkingBudget{Tokens: 0},
+			},
+			expectThinkingBudget: &latest.ThinkingBudget{Tokens: 8192},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			// Step 1: Apply provider defaults (simulating createDirectProvider flow)
+			result := applyProviderDefaults(tt.config, nil)
+
+			// Verify thinking is still disabled after defaults (because explicit 0/none is preserved)
+			require.NotNil(t, result.ThinkingBudget, "ThinkingBudget should be set")
+			assert.True(t, result.ThinkingBudget.Tokens == 0 || result.ThinkingBudget.Effort == "none",
+				"ThinkingBudget should still be disabled after defaults")
+
+			// Step 2: Apply override with thinking explicitly enabled (simulates /think toggle)
+			mo := options.ModelOptions{}
+			options.WithThinking(true)(&mo)
+			result = applyOverrides(result, &mo)
+
+			// Verify defaults were restored
+			require.NotNil(t, result.ThinkingBudget, "ThinkingBudget should be set after enable override")
+			assert.Equal(t, tt.expectThinkingBudget.Tokens, result.ThinkingBudget.Tokens, "Tokens should match default")
+			assert.Equal(t, tt.expectThinkingBudget.Effort, result.ThinkingBudget.Effort, "Effort should match default")
+		})
+	}
+}
+
+// TestIsThinkingBudgetDisabled tests the helper function.
+func TestIsThinkingBudgetDisabled(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name     string
+		budget   *latest.ThinkingBudget
+		expected bool
+	}{
+		{"nil budget", nil, false},
+		{"Tokens=0", &latest.ThinkingBudget{Tokens: 0}, true},
+		{"Effort=none", &latest.ThinkingBudget{Effort: "none"}, true},
+		{"Tokens=8192", &latest.ThinkingBudget{Tokens: 8192}, false},
+		{"Effort=medium", &latest.ThinkingBudget{Effort: "medium"}, false},
+		{"Tokens=-1 (dynamic)", &latest.ThinkingBudget{Tokens: -1}, false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+			assert.Equal(t, tt.expected, isThinkingBudgetDisabled(tt.budget))
+		})
+	}
+}

pkg/model/provider/provider.go

@@ -351,8 +351,13 @@ func applyOverrides(cfg *latest.ModelConfig, opts *options.ModelOptions) *latest
 	// Create a copy to avoid modifying the original
 	enhancedCfg := *cfg
 
+	t := opts.Thinking()
+	if t == nil {
+		return &enhancedCfg
+	}
+
 	// If thinking is explicitly disabled (e.g., via /think command), clear thinking configuration
-	if t := opts.Thinking(); t != nil && !*t {
+	if !*t {
 		enhancedCfg.ThinkingBudget = nil
 		if enhancedCfg.ProviderOpts != nil {
 			delete(enhancedCfg.ProviderOpts, "interleaved_thinking")
@@ -361,11 +366,40 @@ func applyOverrides(cfg *latest.ModelConfig, opts *options.ModelOptions) *latest
 			"provider", cfg.Provider,
 			"model", cfg.Model,
 		)
+		return &enhancedCfg
+	}
+
+	// If thinking is explicitly enabled (e.g., via /think command) but the config has it
+	// disabled (Tokens == 0 or Effort == "none"), clear ThinkingBudget and re-apply defaults
+	// so the provider gets its standard thinking configuration.
+	if enhancedCfg.ThinkingBudget != nil && isThinkingBudgetDisabled(enhancedCfg.ThinkingBudget) {
+		enhancedCfg.ThinkingBudget = nil
+		applyModelDefaults(&enhancedCfg)
+		slog.Debug("Override: thinking enabled - restored default thinking configuration",
+			"provider", cfg.Provider,
+			"model", cfg.Model,
+		)
 	}
 
 	return &enhancedCfg
 }
 
+// isThinkingBudgetDisabled returns true if the thinking budget is explicitly disabled.
+// NOT disabled when:
+// - Tokens > 0 or Tokens == -1 (explicit token budget)
+// - Effort is set to something other than "none" (e.g., "medium", "high")
+func isThinkingBudgetDisabled(tb *latest.ThinkingBudget) bool {
+	if tb == nil {
+		return false
+	}
+	if tb.Effort == "none" {
+		return true
+	}
+	// Tokens == 0 with no Effort means explicitly disabled (thinking_budget: 0)
+	// Tokens == 0 with Effort set (e.g., "medium") means Effort-based config, not disabled
+	return tb.Tokens == 0 && tb.Effort == ""
+}
+
 // applyModelDefaults applies provider-specific default values for model configuration.
 // These defaults are applied only if the user hasn't explicitly set the values.
 //

pkg/runtime/runtime.go

@@ -500,6 +500,22 @@ func getAgentModelID(a *agent.Agent) string {
 	return ""
 }
 
+// isModelThinkingDisabled checks if the model's thinking configuration is explicitly disabled
+// (thinking_budget: 0 or thinking_budget: none).
+func isModelThinkingDisabled(model provider.Provider) bool {
+	if model == nil {
+		return false
+	}
+	tb := model.BaseConfig().ModelConfig.ThinkingBudget
+	if tb == nil {
+		return false
+	}
+	if tb.Effort == "none" {
+		return true
+	}
+	return tb.Tokens == 0 && tb.Effort == ""
+}
+
 // agentDetailsFromTeam converts team agent info to AgentDetails for events
 func (r *LocalRuntime) agentDetailsFromTeam() []AgentDetails {
 	agentsInfo := r.team.AgentsInfo()
@@ -754,6 +770,12 @@ func (r *LocalRuntime) RunStream(ctx context.Context, sess *session.Session) <-c
 			if !sess.Thinking {
 				model = provider.CloneWithOptions(ctx, model, options.WithThinking(false))
 				slog.Debug("Cloned provider with thinking disabled", "agent", a.Name(), "model", model.ID())
+			} else if isModelThinkingDisabled(model) {
+				// If thinking is enabled for this session but the model config has it disabled
+				// (e.g., thinking_budget: 0 or thinking_budget: none), clone with explicit enable
+				// so that applyOverrides restores provider defaults.
+				model = provider.CloneWithOptions(ctx, model, options.WithThinking(true))
+				slog.Debug("Cloned provider with thinking enabled (restoring defaults)", "agent", a.Name(), "model", model.ID())
 			}
 
 			modelID := model.ID()