fix: auto-continue circuit breaker + node timeout

README.md

@@ -301,9 +301,11 @@ All configuration is controlled via environment variables in the `docker run` co
 
 **Optional — Agent:**
 
-| Variable                | Default | Description                        |
-| ----------------------- | ------- | ---------------------------------- |
-| `AGENT_RECURSION_LIMIT` | `1000`    | Max graph execution steps per call |
+| Variable                        | Default  | Description                                    |
+| ------------------------------- | -------- | ---------------------------------------------- |
+| `AGENT_RECURSION_LIMIT`         | `1000`   | Max graph execution steps per call             |
+| `AGENT_AUTO_CONTINUE_LIMIT`     | `1000`   | Max consecutive auto-continue attempts before circuit breaker triggers |
+| `AGENT_NODE_TIMEOUT`            | `600000` | Superstep timeout in milliseconds (default 10 minutes) |
 
 **Optional — Persistence:**
 
@@ -489,7 +491,7 @@ Graceful shutdown flushes all buffered log entries to disk before process exit.
 |               | `openai.base_url`                    | `https://api.openai.com/v1`              | API endpoint URL                              |
 |               | `openai.model`                       | `gpt-4o`                                 | Model name                                    |
 |               | `openai.credentials.apiKey`          | _(empty)_                                | API key for authentication                    |
-|               | `openai.temperature`                 | `0.7`                                    | Sampling temperature (0–2)                    |
+|               | `openai.temperature`                 | `0.4`                                    | Sampling temperature (0–2)                    |
 |               | `openai.maxTokens`                   | `4096`                                   | Max output tokens                             |
 |               | `openai.rateLimit.requestsPerMinute` | `60`                                     | Rate limit for API calls                      |
 | `sandbox`     | `paths`                              | `["memory/", "skills/", "src/", "/tmp"]` | Allowed filesystem paths                      |
@@ -526,6 +528,8 @@ Graceful shutdown flushes all buffered log entries to disk before process exit.
 | `tui`         | `name`                               | `madz`                                   | TUI identifier in banner                      |
 |               | `cursorChar`                         | `█`                                      | Cursor character                              |
 | `agent`       | `recursionLimit`                     | `1000`                                   | Max graph execution steps per agent call      |
+|               | `autoContinueLimit`                  | `1000`                                   | Max consecutive auto-continue attempts before circuit breaker triggers |
+|               | `nodeTimeout`                        | `600000`                                 | Superstep timeout in milliseconds (default 10 minutes) |
 | `persistence` | `mode`                               | `memory`                                 | Storage backend (`memory`, `sqlite`)          |
 |               | `sqlite_path`                        | `memory/checkpoints.db`                  | SQLite checkpointer file path                 |
 

config.yaml

@@ -6,7 +6,7 @@ providers:
     encoding: cl100k_base
     credentials:
       apiKey:
-    temperature: 0.7
+    temperature: 0.4
     maxTokens: 4096
     rateLimit:
       requestsPerMinute: 120
@@ -70,6 +70,8 @@ tui:
   cursorChar: "█"
 agent:
   recursionLimit: 1000
+  autoContinueLimit: 1000
+  nodeTimeout: 600000
 persistence:
   mode: memory
   sqlite_path: memory/checkpoints.db

index.js

@@ -192,6 +192,7 @@ const agent = createReactAgent(
 	tools,
 	checkpointer,
 	config.agent?.recursionLimit ?? undefined,
+	config.agent?.nodeTimeout ?? 600000,
 );
 
 const sessionConfig = { configurable: { thread_id: sessionState.getThreadId() } };

src/agent/react.js

@@ -19,16 +19,20 @@ const CONTEXT_TOO_LONG_MESSAGE = "The conversation is too long. Please start a n
  * @param {unknown[]} [tools=[]] - Optional array of LangChain tool definitions
  * @param {import("@langchain/langgraph").BaseCheckpointSaver | null} [checkpointer=null] - Optional LangGraph checkpointer for persistent conversation memory
  * @param {number} [recursionLimit] - Optional LangGraph recursion limit for the agent graph
+ * @param {number} [timeout] - Optional timeout in milliseconds for superstep execution (default: 600000 / 10 minutes)
  * @returns {ReturnType<typeof createReactAgentGraph>} A compiled ReAct agent
  */
 /* node:coverage ignore next */
-export function createReactAgent(model, tools = [], checkpointer = null, recursionLimit = null) {
-	return createReactAgentGraph({
+export function createReactAgent(model, tools = [], checkpointer = null, recursionLimit = null, timeout = 600000) {
+	const agent = createReactAgentGraph({
 		llm: model,
 		tools,
 		...(checkpointer && { checkpointer }),
 		...(recursionLimit !== null && { recursionLimit }),
 	});
+	// Set superstep timeout on the compiled graph to prevent nodes from hanging
+	agent.stepTimeout = timeout;
+	return agent;
 }
 
 /**

src/config/schemas.js

@@ -70,7 +70,7 @@ const _OpenaiProviderConfigSchema = z.object({
 	model: z.string().min(1),
 	encoding: z.string().optional(),
 	credentials: OpenAICredentialsSchema,
-	temperature: z.number().min(0).max(2).default(0.7),
+	temperature: z.number().min(0).max(2).default(0.4),
 	maxTokens: z.number().int().positive().default(4096),
 	rateLimit: RateLimitSchema.default({ requestsPerMinute: 60 }),
 });
@@ -187,6 +187,8 @@ export const TuiSchema = z.object({
 
 export const AgentSchema = z.object({
 	recursionLimit: z.number().int().positive().default(1000),
+	autoContinueLimit: z.number().int().positive().default(1000),
+	nodeTimeout: z.number().int().positive().default(600000),
 });
 
 // --- Persistence schemas ---
@@ -263,7 +265,7 @@ export const DEFAULT_CONFIG = {
 		redact: { paths: ["credentials.apiKey"] },
 	},
 	schedules: { maxConcurrent: 1, mode: "inprocess", syncOnInit: true, entries: [] },
-	agent: { recursionLimit: 1000 },
+	agent: { recursionLimit: 1000, autoContinueLimit: 1000, nodeTimeout: 600000 },
 	tui: { name: "madz", cursorChar: "\u2588" },
 	persistence: { mode: "memory", sqlite_path: "memory/checkpoints.db" },
 };

src/tui/app.js

@@ -44,6 +44,8 @@ export default function App({
 	const [isCompacting, setIsCompacting] = useState(false);
 	const scrollRef = useRef(null);
 	const isQuittingRef = useRef(false);
+	const autoContinueCountRef = useRef(0);
+	const isAutoContinuingRef = useRef(false);
 	const { exit } = useApp();
 	const exitRef = useRef(exit);
 	exitRef.current = exit;
@@ -335,6 +337,9 @@ export default function App({
 					let responseContent = committedContent;
 
 					// Auto-continue if the agent stalled with zero text output
+					// Circuit breaker: configurable limit (default 1000) of consecutive
+					// empty responses to prevent infinite loops when the model generates
+					// thinking but no text
 					if (!responseContent.trim() && !isQuittingRef.current) {
 						// Show tool results so the user knows work happened
 						if (lastToolCallDisplay) {
@@ -348,8 +353,28 @@ export default function App({
 							});
 						}
 
+						if (autoContinueCountRef.current >= (config?.agent?.autoContinueLimit ?? 1000)) {
+							// Circuit breaker: model is stuck in thinking-only loop
+							setStatusMessage("Model appears stuck — starting fresh.");
+							setMessages((prev) => {
+								const cloned = [...prev];
+								const last = cloned[cloned.length - 1];
+								if (last.role === "assistant" && last.streaming) {
+									last.streaming = false;
+								}
+								return cloned;
+							});
+							autoContinueCountRef.current = 0;
+							addMessage({
+								role: "system",
+								content: `I've tried to continue ${config?.agent?.autoContinueLimit ?? 1000} times with no text output. The model may be stuck in a reasoning loop. Please try a new conversation or rephrase your request.`,
+							});
+							return;
+						}
+
 						// Send a quiet continuation signal to the agent
 						setStatusMessage("Continuing...");
+						isAutoContinuingRef.current = true;
 						try {
 							await dispatchProvider(
 								"Please continue.",
@@ -358,15 +383,17 @@ export default function App({
 									if (isQuittingRef.current) return;
 									try {
 										if (event.type === "text") {
-											responseContent += event.text;
+											committedContent = (committedContent || "") + event.text;
 											setMessages((prev) => {
 												const cloned = [...prev];
 												const last = cloned[cloned.length - 1];
 												if (last.role === "assistant" && last.streaming) {
-													last.content = responseContent + "\u2588";
+													last.content = committedContent + "\u2588";
 												}
 												return cloned;
 											});
+											// Reset flag — text arrived, not stuck anymore
+											isAutoContinuingRef.current = false;
 										} else if (event.type === "tool_start") {
 											setMessages((prev) => {
 												const cloned = [...prev];
@@ -416,6 +443,9 @@ export default function App({
 							setStatusMessage("Done");
 						} catch (contErr) {
 							setStatusMessage(`Error continuing: ${contErr.message}`);
+						} finally {
+							isAutoContinuingRef.current = false;
+							autoContinueCountRef.current++;
 						}
 					}
 
@@ -630,6 +660,9 @@ export default function App({
 			let responseContent = committedContent;
 
 			// Auto-continue if the agent stalled with zero text output
+			// Circuit breaker: configurable limit (default 1000) of consecutive
+			// empty responses to prevent infinite loops when the model generates
+			// thinking but no text
 			if (!responseContent.trim() && !isQuittingRef.current) {
 				// Show tool results so the user knows work happened
 				if (lastToolCallDisplay) {
@@ -643,8 +676,28 @@ export default function App({
 					});
 				}
 
+				if (autoContinueCountRef.current >= (config?.agent?.autoContinueLimit ?? 1000)) {
+					// Circuit breaker: model is stuck in thinking-only loop
+					setStatusMessage("Model appears stuck — starting fresh.");
+					setMessages((prev) => {
+						const cloned = [...prev];
+						const last = cloned[cloned.length - 1];
+						if (last.role === "assistant" && last.streaming) {
+							last.streaming = false;
+						}
+						return cloned;
+					});
+					autoContinueCountRef.current = 0;
+					addMessage({
+						role: "system",
+						content: `I've tried to continue ${config?.agent?.autoContinueLimit ?? 1000} times with no text output. The model may be stuck in a reasoning loop. Please try a new conversation or rephrase your request.`,
+					});
+					return;
+				}
+
 				// Send a quiet continuation signal to the agent
 				setStatusMessage("Continuing...");
+				isAutoContinuingRef.current = true;
 				try {
 					await dispatchProvider(
 						"Please continue.",
@@ -653,15 +706,17 @@ export default function App({
 							if (isQuittingRef.current) return;
 							try {
 								if (event.type === "text") {
-									responseContent += event.text;
+									committedContent = (committedContent || "") + event.text;
 									setMessages((prev) => {
 										const cloned = [...prev];
 										const last = cloned[cloned.length - 1];
 										if (last.role === "assistant" && last.streaming) {
-											last.content = responseContent + "\u2588";
+											last.content = committedContent + "\u2588";
 										}
 										return cloned;
 									});
+									// Reset flag — text arrived, not stuck anymore
+									isAutoContinuingRef.current = false;
 								} else if (event.type === "tool_start") {
 									setMessages((prev) => {
 										const cloned = [...prev];
@@ -715,6 +770,9 @@ export default function App({
 					setStatusMessage("Received response");
 				} catch (contErr) {
 					setStatusMessage(`Error continuing: ${contErr.message}`);
+				} finally {
+					isAutoContinuingRef.current = false;
+					autoContinueCountRef.current++;
 				}
 			}