Merge pull request #2170 from dgageot/fix-thinking

Remove /think command

Author: dgageot

agent-schema.json

@@ -554,20 +554,12 @@
           "description": "Whether to track usage"
         },
         "thinking_budget": {
-          "description": "Controls reasoning effort/budget. Use 'none' or 0 to disable thinking. OpenAI: string levels ('minimal','low','medium','high'). Anthropic: integer token budget (1024-32768), 'adaptive' (lets the model decide), or effort levels ('low','medium','high','max') which use adaptive thinking with the given effort. Amazon Bedrock (Claude): integer token budget or effort levels ('low','medium','high') mapped to token budgets. Google Gemini 2.5: integer token budget (-1 for dynamic, 0 to disable, 24576 max). Google Gemini 3: string levels ('minimal' Flash only,'low','medium','high'). Thinking is only enabled when explicitly configured.",
+          "description": "Controls reasoning effort/budget. Use 'none' or 0 to disable thinking. OpenAI: string levels ('minimal','low','medium','high','xhigh'). Anthropic: integer token budget (1024-32768), 'adaptive' (adaptive thinking with high effort by default), 'adaptive/<effort>' where effort is low/medium/high/max (adaptive thinking with specified effort), or effort levels ('low','medium','high','max') which use adaptive thinking with the given effort. Amazon Bedrock (Claude): integer token budget or effort levels ('low','medium','high') mapped to token budgets. Google Gemini 2.5: integer token budget (-1 for dynamic, 0 to disable, 24576 max). Google Gemini 3: string levels ('minimal' Flash only,'low','medium','high'). Thinking is only enabled when explicitly configured.",
           "oneOf": [
             {
               "type": "string",
-              "enum": [
-                "none",
-                "minimal",
-                "low",
-                "medium",
-                "high",
-                "max",
-                "adaptive"
-              ],
-              "description": "Reasoning effort level. 'adaptive'/'max' are Anthropic-specific. Use 'none' to disable thinking."
+              "pattern": "^(none|minimal|low|medium|high|xhigh|max|adaptive(/low|/medium|/high|/max)?)$",
+              "description": "Reasoning effort level. 'adaptive' uses adaptive thinking with high effort (default). 'adaptive/<effort>' specifies the effort level (low/medium/high/max). Use 'none' to disable thinking."
             },
             {
               "type": "integer",
@@ -583,8 +575,13 @@
             "low",
             "medium",
             "high",
+            "xhigh",
             "max",
             "adaptive",
+            "adaptive/low",
+            "adaptive/medium",
+            "adaptive/high",
+            "adaptive/max",
             -1,
             1024,
             8192,

cmd/root/run.go

@@ -404,10 +404,10 @@ func (f *runExecFlags) createLocalRuntimeAndSession(ctx context.Context, loadRes
 		slog.Debug("Loaded existing session", "session_id", resolvedID, "session_ref", f.sessionID, "agent", f.agentName)
 	} else {
 		wd, _ := os.Getwd()
-		sess = session.New(f.buildSessionOpts(agent.MaxIterations(), agent.ThinkingConfigured(), wd)...)
+		sess = session.New(f.buildSessionOpts(agent.MaxIterations(), wd)...)
 		// Session is stored lazily on first UpdateSession call (when content is added)
 		// This avoids creating empty sessions in the database
-		slog.Debug("Using local runtime", "agent", f.agentName, "thinking", agent.ThinkingConfigured())
+		slog.Debug("Using local runtime", "agent", f.agentName)
 	}
 
 	return localRt, sess, nil
@@ -494,12 +494,11 @@ func (f *runExecFlags) buildAppOpts(args []string) ([]app.Opt, error) {
 // buildSessionOpts returns the canonical set of session options derived from
 // CLI flags and agent configuration. Both the initial session and spawned
 // sessions use this method so their options never drift apart.
-func (f *runExecFlags) buildSessionOpts(maxIterations int, thinking bool, workingDir string) []session.Opt {
+func (f *runExecFlags) buildSessionOpts(maxIterations int, workingDir string) []session.Opt {
 	return []session.Opt{
 		session.WithMaxIterations(maxIterations),
 		session.WithToolsApproved(f.autoApprove),
 		session.WithHideToolResults(f.hideToolResults),
-		session.WithThinking(thinking),
 		session.WithWorkingDir(workingDir),
 	}
 }
@@ -544,7 +543,7 @@ func (f *runExecFlags) createSessionSpawner(agentSource config.Source, sessStore
 		}
 
 		// Create a new session
-		newSess := session.New(f.buildSessionOpts(agent.MaxIterations(), agent.ThinkingConfigured(), workingDir)...)
+		newSess := session.New(f.buildSessionOpts(agent.MaxIterations(), workingDir)...)
 
 		// Create cleanup function
 		cleanup := func() {

docs/configuration/models/index.md

@@ -110,13 +110,6 @@ Works for all providers:
 thinking_budget: none # or 0
 ```
 
-<div class="callout callout-info">
-<div class="callout-title">ℹ️ Runtime Toggle
-</div>
-  <p>Even when thinking is disabled in config, you can enable it during a session using the <code>/think</code> command in the TUI.</p>
-
-</div>
-
 ## Interleaved Thinking
 
 For Anthropic and Bedrock Claude models, interleaved thinking allows tool calls during model reasoning. This is enabled by default:

docs/features/api-server/index.md

@@ -49,7 +49,6 @@ All endpoints are under the `/api` prefix.
 | `PATCH`  | `/api/sessions/:id/permissions`     | Update session permissions                          |
 | `POST`   | `/api/sessions/:id/resume`          | Resume a paused session (after tool confirmation)   |
 | `POST`   | `/api/sessions/:id/tools/toggle`    | Toggle auto-approve (YOLO) mode                     |
-| `POST`   | `/api/sessions/:id/thinking/toggle` | Toggle thinking/reasoning mode                      |
 | `POST`   | `/api/sessions/:id/elicitation`     | Respond to an MCP tool elicitation request          |
 
 ### Agent Execution

docs/features/tui/index.md

@@ -41,7 +41,6 @@ Type `/` during a session to see available commands, or press <kbd>Ctrl</kbd>+<k
 | `/sessions` | Browse and load past sessions                  |
 | `/model`    | Change the model for the current agent         |
 | `/theme`    | Change the color theme                         |
-| `/think`    | Toggle thinking/reasoning mode                 |
 | `/yolo`     | Toggle automatic tool call approval            |
 | `/title`    | Set or regenerate session title                |
 | `/attach`   | Attach a file to your message                  |

e2e/debug_title_test.go

@@ -19,13 +19,13 @@ func TestDebug_Title(t *testing.T) {
 		{"OpenAI_gpt52pro", "openai/gpt-5.2-pro", "Assistant Capabilities Overview\n"},
 		{"OpenAI_gpt52codex", "openai/gpt-5.2-codex", "AI Assistant Capabilities\n"},
 
-		// Anthopic
+		// Anthropic
 		{"Anthropic", "anthropic/claude-haiku-4-5", "AI Assistant Capabilities Overview\n"},
 		{"Anthropic_Sonnet45", "anthropic/claude-sonnet-4-5", "What can you do?\n"},
 		{"Anthropic_Opus46", "anthropic/claude-opus-4-6", "AI Assistant Capabilities Overview\n"},
 
 		// Google
-		{"Google_Gemini25FlashLite", "google/gemini-2.5-flash-lite", "AI Capabilities Overview\n"},
+		{"Google_Gemini25FlashLite", "google/gemini-2.5-flash-lite", "AI capabilities overview\n"},
 		{"Google_Gemini3ProPreview", "google/gemini-3-pro-preview", "AI Capabilities Inquiry\n"},
 	}
 

e2e/exec_test.go

@@ -22,7 +22,7 @@ func TestExec_OpenAI_V3Config(t *testing.T) {
 }
 
 // TestExec_OpenAI_WithThinkingBudget tests that when thinking_budget is explicitly configured
-// in the YAML, thinking is enabled by default (without needing /think command).
+// in the YAML, thinking is enabled by default.
 func TestExec_OpenAI_WithThinkingBudget(t *testing.T) {
 	out := runCLI(t, "run", "--exec", "testdata/basic_with_thinking.yaml", "What's 2+2?")