erosolar-cli 1.7.81 → 1.7.83

package/agents/erosolar-code.rules.json

@@ -43,6 +43,11 @@
       "summary": "Narrate intent before running tools/commands and summarize output afterward.",
       "severity": "required"
     },
+    {
+      "id": "guardrail.continuous_execution",
+      "summary": "CRITICAL: Continue using tools until the task is complete. Do NOT stop after reading files—immediately proceed to make the requested edits. Do NOT describe what you WILL do—just DO it by calling the appropriate tools. After each tool call, continue with the next step until the task is fully accomplished.",
+      "severity": "critical"
+    },
     {
       "id": "guardrail.manual_loop_supervision",
       "summary": "Stop if you lack evidence—surface blockers instead of guessing or editing blindly.",

package/agents/general.rules.json

@@ -42,6 +42,11 @@
       "detail": "Keep tool usage auditable by narrating the intent before running them and summarizing their output afterward.",
       "severity": "required"
     },
+    {
+      "id": "guardrail.continuous_execution",
+      "summary": "CRITICAL: Continue using tools until the task is complete. Do NOT stop after reading files—immediately proceed to make the requested edits. Do NOT describe what you WILL do—just DO it by calling the appropriate tools. After each tool call, continue with the next step until the task is fully accomplished.",
+      "severity": "critical"
+    },
     {
       "id": "guardrail.manual_loop_supervision",
       "summary": "Humans supervise the loop manually—escalate when you lack evidence or stall.",

package/dist/bin/erosolar.js

@@ -4,8 +4,6 @@
  *
  * @license MIT
  * @author Bo Shang
- *
- * Thank you to Anthropic for allowing me to use Claude Code to build erosolar-cli.
  */
 import { launchShell } from '../shell/shellApp.js';
 import { runHeadlessApp } from '../headless/headlessApp.js';

package/dist/bin/erosolar.js.map

@@ -1 +1 @@
-{"version":3,"file":"erosolar.js","sourceRoot":"","sources":["../../src/bin/erosolar.ts"],"names":[],"mappings":";AACA;;;;;;;GAOG;AACH,OAAO,EAAE,WAAW,EAAE,MAAM,sBAAsB,CAAC;AACnD,OAAO,EAAE,cAAc,EAAE,MAAM,4BAA4B,CAAC;AAE5D,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;AAEnC,IAAI,IAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,EAAE,CAAC;IAC5B,cAAc,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,EAAE;QACvC,OAAO,CAAC,KAAK,CAAC,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;QAC9D,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAClB,CAAC,CAAC,CAAC;AACL,CAAC;KAAM,CAAC;IACN,WAAW,CAAC,eAAe,EAAE,EAAE,sBAAsB,EAAE,IAAI,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,EAAE;QAC7E,OAAO,CAAC,KAAK,CAAC,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;QAC9D,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAClB,CAAC,CAAC,CAAC;AACL,CAAC"}
+{"version":3,"file":"erosolar.js","sourceRoot":"","sources":["../../src/bin/erosolar.ts"],"names":[],"mappings":";AACA;;;;;GAKG;AACH,OAAO,EAAE,WAAW,EAAE,MAAM,sBAAsB,CAAC;AACnD,OAAO,EAAE,cAAc,EAAE,MAAM,4BAA4B,CAAC;AAE5D,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;AAEnC,IAAI,IAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,EAAE,CAAC;IAC5B,cAAc,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,EAAE;QACvC,OAAO,CAAC,KAAK,CAAC,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;QAC9D,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAClB,CAAC,CAAC,CAAC;AACL,CAAC;KAAM,CAAC;IACN,WAAW,CAAC,eAAe,EAAE,EAAE,sBAAsB,EAAE,IAAI,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,EAAE;QAC7E,OAAO,CAAC,KAAK,CAAC,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;QAC9D,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAClB,CAAC,CAAC,CAAC;AACL,CAAC"}

package/dist/contracts/agent-schemas.json

@@ -61,6 +61,14 @@
   ],
 
   "models": [
+    {
+      "id": "gpt-5.1-codex-mini",
+      "label": "gpt-5.1-codex-mini",
+      "provider": "openai",
+      "description": "Lightweight GPT-5.1 Codex for fast iteration and quick tasks.",
+      "reasoningEffort": "low",
+      "capabilities": ["chat", "reasoning", "tools", "streaming"]
+    },
     {
       "id": "gpt-5.1-codex",
       "label": "gpt-5.1-codex",
@@ -111,8 +119,8 @@
       "capabilities": ["chat", "reasoning", "tools", "streaming"]
     },
     {
-      "id": "claude-opus-4.1",
-      "label": "opus-4.1",
+      "id": "claude-opus-4-20250514",
+      "label": "opus-4",
       "provider": "anthropic",
       "description": "Anthropic Opus 4.1 for the richest Claude reasoning runs.",
       "temperature": 0.7,
@@ -120,7 +128,7 @@
       "capabilities": ["chat", "reasoning", "tools", "streaming"]
     },
     {
-      "id": "claude-haiku-4.5",
+      "id": "claude-haiku-4-5-20251001",
       "label": "haiku-4.5",
       "provider": "anthropic",
       "description": "Anthropic Haiku 4.5 focused on latency-sensitive workflows.",
@@ -143,12 +151,19 @@
       "capabilities": ["chat", "tools", "streaming"]
     },
     {
-      "id": "grok-4.1-fast-reasoning",
-      "label": "grok-4.1-fast-reasoning",
+      "id": "grok-4-1-fast-reasoning",
+      "label": "grok-4-1-fast-reasoning",
       "provider": "xai",
       "description": "Grok-4.1 fast reasoning mode with improved performance and enhanced thinking capabilities.",
       "capabilities": ["chat", "reasoning", "tools", "streaming"]
     },
+    {
+      "id": "grok-4-1-fast-non-reasoning",
+      "label": "grok-4-1-fast-non-reasoning",
+      "provider": "xai",
+      "description": "Grok-4.1 fast non-reasoning mode for lower latency without thinking.",
+      "capabilities": ["chat", "tools", "streaming"]
+    },
     {
       "id": "grok-4",
       "label": "grok-4",
@@ -191,13 +206,6 @@
       "description": "Google Gemini 2.5 Flash for lower-latency edits and iterative coding.",
       "capabilities": ["chat", "tools", "streaming", "multimodal"]
     },
-    {
-      "id": "gemini-3.0-pro-preview",
-      "label": "gemini-3.0-pro-preview",
-      "provider": "google",
-      "description": "Google Gemini 3.0 Pro Preview for next-generation multimodal reasoning and advanced capabilities.",
-      "capabilities": ["chat", "reasoning", "tools", "streaming", "multimodal"]
-    },
     {
       "id": "llama3.1:8b",
       "label": "llama3.1:8b",

package/dist/contracts/unified-schema.json

@@ -22,7 +22,7 @@
         "maxDelayMs": 40000,
         "backoffMultiplier": 2
       },
-      "models": ["claude-opus-4-5-20251101", "claude-sonnet-4-5-20250929", "claude-haiku-4-5"],
+      "models": ["claude-opus-4-5-20251101", "claude-sonnet-4-5-20250929", "claude-haiku-4-5-20251001"],
       "defaultModel": "claude-sonnet-4-5-20250929",
       "status": "production"
     },

package/dist/core/agent.d.ts

@@ -1,6 +1,7 @@
 import type { ToolRuntime } from './toolRuntime.js';
 import { type ConversationMessage, type LLMProvider, type ProviderUsage } from './types.js';
 import { ContextManager } from './contextManager.js';
+import { type PasteSummary } from './multilinePasteHandler.js';
 export interface AgentCallbacks {
     onAssistantMessage?(content: string, metadata: AssistantMessageMetadata): void;
     onStreamChunk?(chunk: string): void;
@@ -11,12 +12,30 @@ export interface AgentCallbacks {
     onContextRecovery?(attempt: number, maxAttempts: number, message: string): void;
     /** Called when agent continues after context recovery - useful for updating UI */
     onContinueAfterRecovery?(): void;
+    /** Called when auto-continuing because model expressed intent but didn't act */
+    onAutoContinue?(attempt: number, maxAttempts: number, message: string): void;
+    /** Called when multi-line paste is detected - displays summary instead of full content */
+    onMultilinePaste?(summary: string, metadata: PasteSummary): void;
+    /** Called when verification should be triggered for a final response */
+    onVerificationNeeded?(response: string, context: VerificationCallbackContext): void;
+}
+export interface VerificationCallbackContext {
+    /** Working directory for verification */
+    workingDirectory: string;
+    /** Recent conversation history for context */
+    conversationHistory: string[];
+    /** Provider ID */
+    provider: string;
+    /** Model ID */
+    model: string;
 }
 export interface AssistantMessageMetadata {
     isFinal: boolean;
     elapsedMs?: number;
     usage?: ProviderUsage | null;
     contextStats?: Record<string, unknown> | null;
+    /** True if content was already displayed via streaming chunks */
+    wasStreamed?: boolean;
 }
 interface AgentOptions {
     provider: LLMProvider;
@@ -24,6 +43,12 @@ interface AgentOptions {
     systemPrompt: string;
     callbacks?: AgentCallbacks;
     contextManager?: ContextManager;
+    /** Provider ID for verification context */
+    providerId?: string;
+    /** Model ID for verification context */
+    modelId?: string;
+    /** Working directory for verification */
+    workingDirectory?: string;
 }
 export declare class AgentRuntime {
     private readonly messages;
@@ -33,6 +58,9 @@ export declare class AgentRuntime {
     private readonly contextManager;
     private activeRun;
     private readonly baseSystemPrompt;
+    private readonly providerId;
+    private readonly modelId;
+    private readonly workingDirectory;
     constructor(options: AgentOptions);
     send(text: string, useStreaming?: boolean): Promise<string>;
     private processConversation;
@@ -48,6 +76,11 @@ export declare class AgentRuntime {
     private resolveToolCalls;
     private get providerTools();
     private emitAssistantMessage;
+    /**
+     * Trigger verification for a final response if callback is registered
+     * and response contains verifiable claims (implementation, build success, etc.)
+     */
+    private triggerVerificationIfNeeded;
     getHistory(): ConversationMessage[];
     loadHistory(history: ConversationMessage[]): void;
     clearHistory(): void;
@@ -72,9 +105,9 @@ export declare class AgentRuntime {
      *
      * This is called when an API call fails due to context length exceeding limits.
      * It performs increasingly aggressive pruning on each attempt:
-     * - Attempt 1: Remove 30% of oldest messages
-     * - Attempt 2: Remove 50% of oldest messages
-     * - Attempt 3: Remove 70% of oldest messages (keep only recent)
+     * - Attempt 1: Remove 30% of oldest messages + truncate tool outputs to 5k
+     * - Attempt 2: Remove 50% of oldest messages + truncate tool outputs to 2k
+     * - Attempt 3: Remove 70% of oldest messages + truncate tool outputs to 500 chars
      *
      * @returns true if recovery was successful (context was reduced)
      */

package/dist/core/agent.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"agent.d.ts","sourceRoot":"","sources":["../../src/core/agent.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,kBAAkB,CAAC;AACpD,OAAO,EACL,KAAK,mBAAmB,EACxB,KAAK,WAAW,EAGhB,KAAK,aAAa,EACnB,MAAM,YAAY,CAAC;AACpB,OAAO,EAAE,cAAc,EAAE,MAAM,qBAAqB,CAAC;AAuBrD,MAAM,WAAW,cAAc;IAC7B,kBAAkB,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,QAAQ,EAAE,wBAAwB,GAAG,IAAI,CAAC;IAC/E,aAAa,CAAC,CAAC,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;IACpC,eAAe,CAAC,CAAC,YAAY,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,IAAI,CAAC;IAC7E,8DAA8D;IAC9D,kBAAkB,CAAC,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3C,8DAA8D;IAC9D,iBAAiB,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IAChF,kFAAkF;IAClF,uBAAuB,CAAC,IAAI,IAAI,CAAC;CAClC;AAED,MAAM,WAAW,wBAAwB;IACvC,OAAO,EAAE,OAAO,CAAC;IACjB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,KAAK,CAAC,EAAE,aAAa,GAAG,IAAI,CAAC;IAC7B,YAAY,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,IAAI,CAAC;CAC/C;AAED,UAAU,YAAY;IACpB,QAAQ,EAAE,WAAW,CAAC;IACtB,WAAW,EAAE,WAAW,CAAC;IACzB,YAAY,EAAE,MAAM,CAAC;IACrB,SAAS,CAAC,EAAE,cAAc,CAAC;IAC3B,cAAc,CAAC,EAAE,cAAc,CAAC;CACjC;AAED,qBAAa,YAAY;IACvB,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAA6B;IACtD,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAc;IACvC,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAc;IAC1C,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAiB;IAC3C,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAwB;IACvD,OAAO,CAAC,SAAS,CAAsC;IACvD,OAAO,CAAC,QAAQ,CAAC,gBAAgB,CAAgB;gBAErC,OAAO,EAAE,YAAY;IAa3B,IAAI,CAAC,IAAI,EAAE,MAAM,EAAE,YAAY,UAAQ,GAAG,OAAO,CAAC,MAAM,CAAC;YAqBjD,mBAAmB;YAwDnB,4BAA4B;IA2E1C;;;;;;;OAOG;YACW,gBAAgB;IAgE9B,OAAO,KAAK,aAAa,GAExB;IAED,OAAO,CAAC,oBAAoB;IAY5B,UAAU,IAAI,mBAAmB,EAAE;IAInC,WAAW,CAAC,OAAO,EAAE,mBAAmB,EAAE,GAAG,IAAI;IAajD,YAAY,IAAI,IAAI;IAOpB;;;;;;OAMG;YACW,qBAAqB;IAiCnC;;OAEG;IACH,OAAO,CAAC,eAAe;IAOvB;;OAEG;IACH,iBAAiB,IAAI,cAAc,GAAG,IAAI;IAI1C;;;;;;;;;;OAUG;YACW,0BAA0B;CAmJzC"}
+{"version":3,"file":"agent.d.ts","sourceRoot":"","sources":["../../src/core/agent.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,kBAAkB,CAAC;AACpD,OAAO,EACL,KAAK,mBAAmB,EACxB,KAAK,WAAW,EAGhB,KAAK,aAAa,EACnB,MAAM,YAAY,CAAC;AACpB,OAAO,EAAE,cAAc,EAAE,MAAM,qBAAqB,CAAC;AACrD,OAAO,EAAkC,KAAK,YAAY,EAAE,MAAM,4BAA4B,CAAC;AA2E/F,MAAM,WAAW,cAAc;IAC7B,kBAAkB,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,QAAQ,EAAE,wBAAwB,GAAG,IAAI,CAAC;IAC/E,aAAa,CAAC,CAAC,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;IACpC,eAAe,CAAC,CAAC,YAAY,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,IAAI,CAAC;IAC7E,8DAA8D;IAC9D,kBAAkB,CAAC,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3C,8DAA8D;IAC9D,iBAAiB,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IAChF,kFAAkF;IAClF,uBAAuB,CAAC,IAAI,IAAI,CAAC;IACjC,gFAAgF;IAChF,cAAc,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7E,0FAA0F;IAC1F,gBAAgB,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,QAAQ,EAAE,YAAY,GAAG,IAAI,CAAC;IACjE,wEAAwE;IACxE,oBAAoB,CAAC,CAAC,QAAQ,EAAE,MAAM,EAAE,OAAO,EAAE,2BAA2B,GAAG,IAAI,CAAC;CACrF;AAED,MAAM,WAAW,2BAA2B;IAC1C,yCAAyC;IACzC,gBAAgB,EAAE,MAAM,CAAC;IACzB,8CAA8C;IAC9C,mBAAmB,EAAE,MAAM,EAAE,CAAC;IAC9B,kBAAkB;IAClB,QAAQ,EAAE,MAAM,CAAC;IACjB,eAAe;IACf,KAAK,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,wBAAwB;IACvC,OAAO,EAAE,OAAO,CAAC;IACjB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,KAAK,CAAC,EAAE,aAAa,GAAG,IAAI,CAAC;IAC7B,YAAY,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,IAAI,CAAC;IAC9C,iEAAiE;IACjE,WAAW,CAAC,EAAE,OAAO,CAAC;CACvB;AAED,UAAU,YAAY;IACpB,QAAQ,EAAE,WAAW,CAAC;IACtB,WAAW,EAAE,WAAW,CAAC;IACzB,YAAY,EAAE,MAAM,CAAC;IACrB,SAAS,CAAC,EAAE,cAAc,CAAC;IAC3B,cAAc,CAAC,EAAE,cAAc,CAAC;IAChC,2CAA2C;IAC3C,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,wCAAwC;IACxC,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,yCAAyC;IACzC,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC3B;AAED,qBAAa,YAAY;IACvB,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAA6B;IACtD,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAc;IACvC,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAc;IAC1C,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAiB;IAC3C,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAwB;IACvD,OAAO,CAAC,SAAS,CAAsC;IACvD,OAAO,CAAC,QAAQ,CAAC,gBAAgB,CAAgB;IACjD,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAS;IACpC,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAS;IACjC,OAAO,CAAC,QAAQ,CAAC,gBAAgB,CAAS;gBAE9B,OAAO,EAAE,YAAY;IAgB3B,IAAI,CAAC,IAAI,EAAE,MAAM,EAAE,YAAY,UAAQ,GAAG,OAAO,CAAC,MAAM,CAAC;YA+BjD,mBAAmB;YAwFnB,4BAA4B;IA4G1C;;;;;;;OAOG;YACW,gBAAgB;IAgE9B,OAAO,KAAK,aAAa,GAExB;IAED,OAAO,CAAC,oBAAoB;IAY5B;;;OAGG;IACH,OAAO,CAAC,2BAA2B;IAyCnC,UAAU,IAAI,mBAAmB,EAAE;IAInC,WAAW,CAAC,OAAO,EAAE,mBAAmB,EAAE,GAAG,IAAI;IAajD,YAAY,IAAI,IAAI;IAOpB;;;;;;OAMG;YACW,qBAAqB;IAiCnC;;OAEG;IACH,OAAO,CAAC,eAAe;IAOvB;;OAEG;IACH,iBAAiB,IAAI,cAAc,GAAG,IAAI;IAI1C;;;;;;;;;;OAUG;YACW,0BAA0B;CA4MzC"}

package/dist/core/agent.js

@@ -1,7 +1,54 @@
+import { isMultilinePaste, processPaste } from './multilinePasteHandler.js';
 /**
  * Maximum number of context overflow recovery attempts
  */
 const MAX_CONTEXT_RECOVERY_ATTEMPTS = 3;
+/**
+ * Maximum number of auto-continuation attempts when model expresses intent but doesn't act
+ */
+const MAX_AUTO_CONTINUE_ATTEMPTS = 3;
+/**
+ * Patterns that indicate the model intends to take action but hasn't yet
+ * These suggest the model should be prompted to continue
+ */
+const INTENT_WITHOUT_ACTION_PATTERNS = [
+    // "Let me X" patterns - model is stating what it will do
+    /\blet me\s+(create|write|implement|add|update|edit|modify|fix|build|make|refactor|read|check|look|search|find|analyze|examine|review)/i,
+    // "I'll X" / "I will X" patterns
+    /\bi['']ll\s+(create|write|implement|add|update|edit|modify|fix|build|make|refactor|read|check|look|search|find|analyze|start|begin|now)/i,
+    /\bi will\s+(create|write|implement|add|update|edit|modify|fix|build|make|refactor|read|check|look|search|find|analyze|start|begin|now)/i,
+    // "I'm going to X" patterns
+    /\bi['']m going to\s+(create|write|implement|add|update|edit|modify|fix|build|make|refactor|read|check|look|search|find|analyze)/i,
+    // "Now I'll X" / "First, I'll X" patterns
+    /\b(now|first|next)\s*(,)?\s*i['']ll\s+/i,
+    // Explicit continuation signals
+    /\bhere['']s (the|my) (plan|approach|solution|implementation)/i,
+    // Numbered steps suggesting action to come
+    /^\s*\d+\.\s+(create|write|implement|add|update|edit|modify|fix|show|read|check)/im,
+    // Bullet points suggesting planned actions
+    /^[\s•\-\*]+\s*(create|write|implement|add|update|edit|modify|fix|shows?|reads?|checks?)\s/im,
+];
+/**
+ * Check if response indicates intent to act without actually acting
+ * This detects when the model says "let me do X" but doesn't call any tools
+ */
+function shouldAutoContinue(content, hasToolCalls) {
+    // If there are tool calls, no need to auto-continue
+    if (hasToolCalls) {
+        return false;
+    }
+    // If content is very short, likely not an incomplete intent
+    if (content.length < 50) {
+        return false;
+    }
+    // Check for intent patterns
+    for (const pattern of INTENT_WITHOUT_ACTION_PATTERNS) {
+        if (pattern.test(content)) {
+            return true;
+        }
+    }
+    return false;
+}
 /**
  * Check if an error is a context overflow error
  */
@@ -24,11 +71,17 @@ export class AgentRuntime {
     contextManager;
     activeRun = null;
     baseSystemPrompt;
+    providerId;
+    modelId;
+    workingDirectory;
     constructor(options) {
         this.provider = options.provider;
         this.toolRuntime = options.toolRuntime;
         this.callbacks = options.callbacks ?? {};
         this.contextManager = options.contextManager ?? null;
+        this.providerId = options.providerId ?? 'unknown';
+        this.modelId = options.modelId ?? 'unknown';
+        this.workingDirectory = options.workingDirectory ?? process.cwd();
         const trimmedPrompt = options.systemPrompt.trim();
         this.baseSystemPrompt = trimmedPrompt || null;
         if (trimmedPrompt) {
@@ -40,7 +93,18 @@ export class AgentRuntime {
         if (!prompt) {
             return '';
         }
-        this.messages.push({ role: 'user', content: prompt });
+        // Handle multi-line paste: show summary to user, send full content to AI
+        if (isMultilinePaste(prompt)) {
+            const processed = processPaste(prompt);
+            // Notify UI about the paste summary
+            this.callbacks.onMultilinePaste?.(processed.displaySummary, processed.metadata);
+            // But send the full content to the AI
+            this.messages.push({ role: 'user', content: processed.fullContent });
+        }
+        else {
+            // Single-line or short text: send as-is
+            this.messages.push({ role: 'user', content: prompt });
+        }
         const run = { startedAt: Date.now() };
         this.activeRun = run;
         try {
@@ -57,6 +121,7 @@ export class AgentRuntime {
     }
     async processConversation() {
         let contextRecoveryAttempts = 0;
+        let autoContinueAttempts = 0;
         while (true) {
             // Prune messages if approaching context limit (BEFORE generation)
             await this.pruneMessagesIfNeeded();
@@ -80,13 +145,40 @@ export class AgentRuntime {
                     }
                     this.messages.push(assistantMessage);
                     await this.resolveToolCalls(response.toolCalls);
+                    // Reset auto-continue counter since model is actively working
+                    autoContinueAttempts = 0;
                     continue;
                 }
                 const reply = response.content?.trim() ?? '';
+                // Check if model expressed intent to act but didn't call tools
+                // This catches "Let me create..." without actual tool calls
+                if (shouldAutoContinue(reply, false) && autoContinueAttempts < MAX_AUTO_CONTINUE_ATTEMPTS) {
+                    autoContinueAttempts++;
+                    // Emit the planning content but mark as non-final
+                    if (reply) {
+                        this.emitAssistantMessage(reply, { isFinal: false, usage, contextStats });
+                    }
+                    this.messages.push({ role: 'assistant', content: reply });
+                    // Auto-prompt with increasingly direct instructions
+                    const prompts = [
+                        'Continue. Execute the actions you described using the available tools.',
+                        'You MUST use tools NOW. Call write_file or edit_file immediately. Do not explain - just call the tool.',
+                        'CRITICAL: Call a tool right now. Use write_file with file_path and content parameters. No more text - only tool calls.',
+                    ];
+                    this.messages.push({
+                        role: 'user',
+                        content: prompts[Math.min(autoContinueAttempts - 1, prompts.length - 1)],
+                    });
+                    const autoContinueMessage = `Model expressed intent but didn't use tools. Auto-prompting to continue...`;
+                    this.callbacks.onAutoContinue?.(autoContinueAttempts, MAX_AUTO_CONTINUE_ATTEMPTS, autoContinueMessage);
+                    continue;
+                }
                 if (reply) {
                     this.emitAssistantMessage(reply, { isFinal: true, usage, contextStats });
                 }
                 this.messages.push({ role: 'assistant', content: reply });
+                // Trigger verification for final responses with verifiable claims
+                this.triggerVerificationIfNeeded(reply);
                 return reply;
             }
             catch (error) {
@@ -111,6 +203,7 @@ export class AgentRuntime {
             return this.processConversation();
         }
         let contextRecoveryAttempts = 0;
+        let autoContinueAttempts = 0;
         while (true) {
             // Prune messages if approaching context limit (BEFORE generation)
             await this.pruneMessagesIfNeeded();
@@ -138,7 +231,8 @@ export class AgentRuntime {
                 if (toolCalls.length > 0) {
                     const narration = fullContent.trim();
                     if (narration) {
-                        this.emitAssistantMessage(narration, { isFinal: false, usage, contextStats });
+                        // Mark as wasStreamed since content was already output via onStreamChunk
+                        this.emitAssistantMessage(narration, { isFinal: false, usage, contextStats, wasStreamed: true });
                     }
                     const assistantMessage = {
                         role: 'assistant',
@@ -147,14 +241,42 @@ export class AgentRuntime {
                     };
                     this.messages.push(assistantMessage);
                     await this.resolveToolCalls(toolCalls);
+                    // Reset auto-continue counter since model is actively working
+                    autoContinueAttempts = 0;
                     continue;
                 }
-                // Final message
+                // Check if model expressed intent to act but didn't call tools
+                // This catches "Let me create..." without actual tool calls
                 const reply = fullContent.trim();
+                if (shouldAutoContinue(reply, false) && autoContinueAttempts < MAX_AUTO_CONTINUE_ATTEMPTS) {
+                    autoContinueAttempts++;
+                    // Emit the planning content but mark as non-final
+                    // Mark as wasStreamed since content was already output via onStreamChunk
+                    if (reply) {
+                        this.emitAssistantMessage(reply, { isFinal: false, usage, contextStats, wasStreamed: true });
+                    }
+                    this.messages.push({ role: 'assistant', content: reply });
+                    // Auto-prompt with increasingly direct instructions
+                    const prompts = [
+                        'Continue. Execute the actions you described using the available tools.',
+                        'You MUST use tools NOW. Call write_file or edit_file immediately. Do not explain - just call the tool.',
+                        'CRITICAL: Call a tool right now. Use write_file with file_path and content parameters. No more text - only tool calls.',
+                    ];
+                    this.messages.push({
+                        role: 'user',
+                        content: prompts[Math.min(autoContinueAttempts - 1, prompts.length - 1)],
+                    });
+                    const autoContinueMessage = `Model expressed intent but didn't use tools. Auto-prompting to continue...`;
+                    this.callbacks.onAutoContinue?.(autoContinueAttempts, MAX_AUTO_CONTINUE_ATTEMPTS, autoContinueMessage);
+                    continue;
+                }
+                // Final message - mark as streamed to avoid double-display in UI
                 if (reply) {
-                    this.emitAssistantMessage(reply, { isFinal: true, usage, contextStats });
+                    this.emitAssistantMessage(reply, { isFinal: true, usage, contextStats, wasStreamed: true });
                 }
                 this.messages.push({ role: 'assistant', content: reply });
+                // Trigger verification for final responses with verifiable claims
+                this.triggerVerificationIfNeeded(reply);
                 return reply;
             }
             catch (error) {
@@ -249,6 +371,46 @@ export class AgentRuntime {
         }
         this.callbacks.onAssistantMessage?.(content, payload);
     }
+    /**
+     * Trigger verification for a final response if callback is registered
+     * and response contains verifiable claims (implementation, build success, etc.)
+     */
+    triggerVerificationIfNeeded(response) {
+        if (!this.callbacks.onVerificationNeeded) {
+            return;
+        }
+        // Only trigger verification for responses that likely contain verifiable claims
+        // These patterns indicate the model is claiming to have completed work
+        const verifiablePatterns = [
+            /\b(implemented|created|wrote|added|fixed|built|deployed|completed|refactored)\b/i,
+            /\b(tests?\s+(are\s+)?pass(ing)?|build\s+succeed)/i,
+            /\b(file|function|class|module|component)\s+(has been|is now|was)\s+(created|updated|modified)/i,
+            /✅|✓|\[done\]|\[complete\]/i,
+            /\bcommit(ted)?\b.*\b(success|done)\b/i,
+        ];
+        const hasVerifiableClaims = verifiablePatterns.some(pattern => pattern.test(response));
+        if (!hasVerifiableClaims) {
+            return;
+        }
+        // Build conversation history for context (last 5 user/assistant exchanges)
+        const conversationHistory = [];
+        const recentMessages = this.messages.slice(-10);
+        for (const msg of recentMessages) {
+            if (msg.role === 'user' || msg.role === 'assistant') {
+                const content = typeof msg.content === 'string' ? msg.content : '';
+                if (content.length > 0) {
+                    conversationHistory.push(`${msg.role}: ${content.slice(0, 500)}`);
+                }
+            }
+        }
+        // Trigger verification callback
+        this.callbacks.onVerificationNeeded(response, {
+            workingDirectory: this.workingDirectory,
+            conversationHistory,
+            provider: this.providerId,
+            model: this.modelId,
+        });
+    }
     getHistory() {
         return this.messages.map(cloneMessage);
     }
@@ -323,9 +485,9 @@ export class AgentRuntime {
      *
      * This is called when an API call fails due to context length exceeding limits.
      * It performs increasingly aggressive pruning on each attempt:
-     * - Attempt 1: Remove 30% of oldest messages
-     * - Attempt 2: Remove 50% of oldest messages
-     * - Attempt 3: Remove 70% of oldest messages (keep only recent)
+     * - Attempt 1: Remove 30% of oldest messages + truncate tool outputs to 5k
+     * - Attempt 2: Remove 50% of oldest messages + truncate tool outputs to 2k
+     * - Attempt 3: Remove 70% of oldest messages + truncate tool outputs to 500 chars
      *
      * @returns true if recovery was successful (context was reduced)
      */
@@ -333,6 +495,9 @@ export class AgentRuntime {
         // Calculate reduction percentage based on attempt
         const reductionPercentages = [0.3, 0.5, 0.7];
         const reductionPercent = reductionPercentages[attempt - 1] ?? 0.7;
+        // Increasingly aggressive tool output truncation limits
+        const toolOutputLimits = [5000, 2000, 500];
+        const toolOutputLimit = toolOutputLimits[attempt - 1] ?? 500;
         // Notify UI about recovery attempt
         const message = `Context overflow detected. Auto-squishing context (attempt ${attempt}/${MAX_CONTEXT_RECOVERY_ATTEMPTS}, removing ${Math.round(reductionPercent * 100)}% of history)...`;
         this.callbacks.onContextRecovery?.(attempt, MAX_CONTEXT_RECOVERY_ATTEMPTS, message);
@@ -433,6 +598,39 @@ export class AgentRuntime {
         // Flatten valid turns back to messages
         const keepMessages = validTurns.flat();
         const actualRemoveCount = conversationMessages.length - keepMessages.length;
+        // Aggressively truncate tool outputs in remaining messages
+        let truncatedCount = 0;
+        for (const msg of keepMessages) {
+            if (msg.role === 'tool' && msg.content) {
+                const content = typeof msg.content === 'string' ? msg.content : JSON.stringify(msg.content);
+                if (content.length > toolOutputLimit) {
+                    // Truncate with smart ending
+                    const truncated = content.slice(0, toolOutputLimit);
+                    const lastNewline = truncated.lastIndexOf('\n');
+                    const cutPoint = lastNewline > toolOutputLimit * 0.7 ? lastNewline : toolOutputLimit;
+                    msg.content = truncated.slice(0, cutPoint) + `\n\n[... truncated ${content.length - cutPoint} chars for context recovery ...]`;
+                    truncatedCount++;
+                }
+            }
+            // Also truncate very long assistant messages
+            if (msg.role === 'assistant' && msg.content && msg.content.length > toolOutputLimit * 2) {
+                const content = msg.content;
+                const limit = toolOutputLimit * 2;
+                const truncated = content.slice(0, limit);
+                const lastNewline = truncated.lastIndexOf('\n');
+                const cutPoint = lastNewline > limit * 0.8 ? lastNewline : limit;
+                msg.content = truncated.slice(0, cutPoint) + `\n\n[... truncated for context recovery ...]`;
+                truncatedCount++;
+            }
+        }
+        // Also truncate system messages if they're huge (except first system prompt)
+        for (let i = 1; i < systemMessages.length; i++) {
+            const sys = systemMessages[i];
+            if (sys && sys.content && sys.content.length > toolOutputLimit) {
+                sys.content = sys.content.slice(0, toolOutputLimit) + '\n[... truncated ...]';
+                truncatedCount++;
+            }
+        }
         // Rebuild message array
         this.messages.length = 0;
         // Add system messages
@@ -442,7 +640,7 @@ export class AgentRuntime {
         // Add summary notice
         this.messages.push({
             role: 'system',
-            content: `[Auto Context Recovery] Removed ${actualRemoveCount} earlier messages to stay within token limits. Conversation context has been compressed.`,
+            content: `[Auto Context Recovery] Removed ${actualRemoveCount} messages and truncated ${truncatedCount} large outputs to stay within token limits.`,
         });
         // Add remaining conversation (maintaining tool call/result pairing)
         for (const msg of keepMessages) {
@@ -456,7 +654,24 @@ export class AgentRuntime {
             attempt,
             removedPercent: reductionPercent * 100,
             turnsRemoved: targetTurnsToRemove + startIndex,
+            truncatedOutputs: truncatedCount,
+            toolOutputLimit,
         });
+        // Check if we're still over limit after all reductions
+        const newStats = this.contextManager?.getStats(this.messages);
+        if (newStats && newStats.percentage > 100) {
+            // Still over limit - do one more aggressive pass
+            // Truncate ALL tool outputs to absolute minimum
+            const minLimit = 200;
+            for (const msg of this.messages) {
+                if (msg.role === 'tool' && msg.content) {
+                    const content = typeof msg.content === 'string' ? msg.content : JSON.stringify(msg.content);
+                    if (content.length > minLimit) {
+                        msg.content = content.slice(0, minLimit) + '\n[... severely truncated ...]';
+                    }
+                }
+            }
+        }
         return true;
     }
 }