lonny-agent 0.1.2 → 0.1.9

package/src/agent/prompt-builder.ts

@@ -26,7 +26,19 @@ export function buildSystemPrompt(config: Config): string {
 - \`search\`: Search the web using Tavily (query: string, search_depth?: string, include_answer?: boolean, max_results?: number, topic?: string, days?: number)
 `
     }
-    const planOnly = mode === 'plan'
+    if (mode === 'plan') {
+      return `Available tools (read-only investigation + write_plan):
+- \`read\`: Read file contents (paths: string[])
+- \`glob\`: Find files by glob pattern (pattern: string)
+- \`grep\`: Search file content by regex (pattern: string, include?: string, path?: string)
+- \`ls\`: List directory (path?: string)
+- \`bash\`: Execute read-only shell commands for investigation (NEVER modify files)
+- \`find\`: Find files by name pattern (pattern: string, path?: string, maxResults?: number)
+- \`git\`: Run read-only git commands (command: string)
+- \`search\`: Search the web using Tavily (query: string, search_depth?: string, include_answer?: boolean, max_results?: number, topic?: string, days?: number)
+- \`write_plan\`: Save plan/todo markdown into .lonny/ folder
+`
+    }
     return `Available tools:
 - \`read\`: Read file contents (paths: string[])
 - \`glob\`: Find files by glob pattern (pattern: string)
@@ -34,7 +46,7 @@ export function buildSystemPrompt(config: Config): string {
 - \`ls\`: List directory (path?: string)
 - \`bash\`: Execute a shell command
 - \`edit\`: Replace text in files — call with {"edits": [{"file_path", "old_string", "new_string"}]} (array required)
-${planOnly ? '- `write_plan`: Save plan/todo markdown into .lonny/ folder\n' : ''}- \`install_skill\`: Install an npm package as a skill — fetches package info from npm, runs npm install, and creates a .lonny/skills/ file with usage instructions for the AI
+- \`install_skill\`: Install an npm package as a skill — fetches package info from npm, runs npm install, and creates a .lonny/skills/ file with usage instructions for the AI
 - \`find\`: Find files by name pattern (pattern: string, path?: string, maxResults?: number)
 - \`git\`: Run read-only git commands (command: string)
 - \`search\`: Search the web using Tavily (query: string, search_depth?: string, include_answer?: boolean, max_results?: number, topic?: string, days?: number)
@@ -59,10 +71,12 @@ ${getToolListForMode(config.mode)}`
 
 RULES (plan-specific):
 1. Read first: Use read/grep/glob tools to gather all context you need before planning.
-2. You CANNOT edit source files — you have no code edit tools. Only read and analyze.
-3. Use \`bash\` for read-only commands only.
-4. You MUST persist the final plan AND todo list to a file in \`.lonny/\` using \`write_plan\`. The \`write_plan\` content MUST include both ## Plan and ## Todo List sections.
-5. You MUST also include the todo list in your text response to the user (not just in the file).
+2. You NEVER edit source files. You ONLY use read-only tools (read/glob/grep/ls/find/git/search) and bash (read-only commands only). You do NOT have access to edit, install_skill, or exec.
+3. Use \`bash\` for investigation only — NEVER to modify files, install packages, or run write operations.
+4. Your ONLY output is a plan file saved via \`write_plan\`. You CANNOT modify the codebase directly.
+5. You MUST persist the final plan AND todo list to a file in \`.lonny/\` using \`write_plan\`. The \`write_plan\` content MUST include both ## Plan and ## Todo List sections.
+6. You MUST also include the todo list in your text response to the user (not just in the file).
+7. If the user asks you to modify files, run write commands, or install packages — refuse and explain they need to switch to code mode (\`/mode code\`).
 
 OUTPUT FORMAT (you MUST include both in write_plan AND in your response text):
 
@@ -113,8 +127,13 @@ RULES (code-specific):
 ${isWindows ? '  - Use PowerShell. Do NOT use Unix commands like `cat`, `ls`, `grep`, `which`, `chmod`, `mv`, `cp`, `rm`, `touch`, `mkdir`, `uname`, etc.' : ''}
 ${isWindows ? '  - Use `type` instead of `cat`, `dir` instead of `ls`, `where` instead of `which`' : ''}`
 
+  // Plan mode uses its own standalone tool list inside modeInstructions — skip sharedRules
+  const rulesSection =
+    config.mode === 'plan'
+      ? ''
+      : `
+${sharedRules}`
   return `${modeInstructions}
 
-${envSection}
-${sharedRules}${skillsSection}`
+${envSection}${rulesSection}${skillsSection}`
 }

package/src/agent/providers/anthropic.ts

@@ -12,7 +12,11 @@ export class AnthropicProvider implements LLMProvider {
     this.model = model || 'claude-sonnet-4-20250514'
   }
 
-  async *chat(messages: LLMMessage[], tools: ToolDefinition[]): AsyncGenerator<LLMChunk> {
+  async *chat(
+    messages: LLMMessage[],
+    tools: ToolDefinition[],
+    signal?: AbortSignal,
+  ): AsyncGenerator<LLMChunk> {
     const anthropicTools: Anthropic.Tool[] = tools.map(t => {
       const properties: Record<string, unknown> = {}
       for (const [key, param] of Object.entries(t.parameters)) {
@@ -71,13 +75,16 @@ export class AnthropicProvider implements LLMProvider {
       return { role: 'user' as const, content: '' }
     })
 
-    const stream = this.client.messages.stream({
-      model: this.model,
-      system: systemMsg?.content || '',
-      messages: anthropicMessages,
-      tools: anthropicTools.length > 0 ? anthropicTools : undefined,
-      max_tokens: 8192,
-    })
+    const stream = this.client.messages.stream(
+      {
+        model: this.model,
+        system: systemMsg?.content || '',
+        messages: anthropicMessages,
+        tools: anthropicTools.length > 0 ? anthropicTools : undefined,
+        max_tokens: 8192,
+      },
+      signal ? { signal } : undefined,
+    )
 
     let currentToolUse: {
       id: string

package/src/agent/providers/google.ts

@@ -48,7 +48,11 @@ export class GoogleProvider implements LLMProvider {
     this.baseUrl = baseURL || 'https://generativelanguage.googleapis.com/v1beta'
   }
 
-  async *chat(messages: LLMMessage[], tools: ToolDefinition[]): AsyncGenerator<LLMChunk> {
+  async *chat(
+    messages: LLMMessage[],
+    tools: ToolDefinition[],
+    signal?: AbortSignal,
+  ): AsyncGenerator<LLMChunk> {
     const systemInstruction = messages.find(m => m.role === 'system')?.content || ''
     const nonSystemMessages = messages.filter(m => m.role !== 'system')
 
@@ -135,6 +139,7 @@ export class GoogleProvider implements LLMProvider {
         'Content-Type': 'application/json',
         'x-goog-api-key': this.apiKey,
       },
+      signal,
       body: JSON.stringify(body),
     })
 

package/src/agent/providers/ollama.ts

@@ -40,7 +40,11 @@ export class OllamaProvider implements LLMProvider {
     this.model = model || 'llama3.2'
   }
 
-  async *chat(messages: LLMMessage[], tools: ToolDefinition[]): AsyncGenerator<LLMChunk> {
+  async *chat(
+    messages: LLMMessage[],
+    tools: ToolDefinition[],
+    signal?: AbortSignal,
+  ): AsyncGenerator<LLMChunk> {
     // Build Ollama-format messages
     const ollamaMessages: OllamaMessage[] = messages.map(m => {
       if (m.role === 'system') {
@@ -108,6 +112,7 @@ export class OllamaProvider implements LLMProvider {
     const response = await fetch(`${this.baseUrl}/api/chat`, {
       method: 'POST',
       headers: { 'Content-Type': 'application/json' },
+      signal,
       body: JSON.stringify(body),
     })
 

package/src/agent/providers/openai.ts

@@ -1,4 +1,5 @@
 import OpenAI from 'openai'
+import type { RequestOptions } from 'openai/core.js'
 import type {
   ChatCompletionChunk,
   ChatCompletionMessageParam,
@@ -47,7 +48,11 @@ export class OpenAIProvider implements LLMProvider {
     this.strictTools = strictTools ?? (baseURL ? /beta/i.test(baseURL) : false)
   }
 
-  async *chat(messages: LLMMessage[], tools: ToolDefinition[]): AsyncGenerator<LLMChunk> {
+  async *chat(
+    messages: LLMMessage[],
+    tools: ToolDefinition[],
+    signal?: AbortSignal,
+  ): AsyncGenerator<LLMChunk> {
     const openAIFormattedTools: ChatCompletionTool[] = tools.map(t => {
       const properties: Record<string, unknown> = {}
       for (const [key, param] of Object.entries(t.parameters)) {
@@ -117,18 +122,22 @@ export class OpenAIProvider implements LLMProvider {
     const stream: Stream<ChatCompletionChunk> = await (
       this.client.chat.completions.create as (
         params: ExtendedCreateParams,
+        options?: RequestOptions,
       ) => Promise<Stream<ChatCompletionChunk>>
-    )({
-      model: this.model,
-      messages: openAIMessages,
-      tools: openAIFormattedTools.length > 0 ? openAIFormattedTools : undefined,
-      stream: true,
-      stream_options: { include_usage: true },
-      ...(this.thinking
-        ? { thinking: { type: 'enabled' }, reasoning_effort: this.reasoningEffort || 'high' }
-        : {}),
-      ...(this.enableCache ? { enable_cache: true } : {}),
-    })
+    )(
+      {
+        model: this.model,
+        messages: openAIMessages,
+        tools: openAIFormattedTools.length > 0 ? openAIFormattedTools : undefined,
+        stream: true,
+        stream_options: { include_usage: true },
+        ...(this.thinking
+          ? { thinking: { type: 'enabled' }, reasoning_effort: this.reasoningEffort || 'high' }
+          : {}),
+        ...(this.enableCache ? { enable_cache: true } : {}),
+      },
+      signal ? { signal } : undefined,
+    )
 
     let currentToolCall: {
       id: string

package/src/agent/session.ts

@@ -155,7 +155,7 @@ function printToolResult(tc: ToolCall, result: ToolResult, output?: SessionOutpu
     writeOut(`  ${GY}│${RS}  ${GR}✔${RS} edit\n`, output)
     if (result.output) {
       for (const l of result.output.split('\n')) {
-        if (l.trim()) writeOut(`  ${GY}│${RS}    ${GY}${l.trim()}${RS}\n`, output)
+        if (l.trim()) writeOut(`  ${GY}│${RS}  ${l.trim()}\n`, output)
       }
     }
   } else if (tc.name === 'write_plan') {
@@ -259,6 +259,7 @@ export class Session {
   turnApiCalls: number = 0
   totalApiCalls: number = 0
   private stopped: boolean = false
+  private abortController: AbortController | null = null
 
   constructor(config: Config, output?: SessionOutput) {
     this.config = config
@@ -368,6 +369,8 @@ export class Session {
   /** Stop the current conversation gracefully */
   stop(): void {
     this.stopped = true
+    // Abort any in-flight LLM stream to stop token consumption immediately
+    this.abortController?.abort()
   }
 
   /** Check if the session was stopped */
@@ -396,6 +399,9 @@ export class Session {
 
     // Reset stopped flag for new conversation
     this.resetStopped()
+    // Create a new AbortController for this chat invocation
+    // (a new controller is needed each time because abort() is one-shot)
+    this.abortController = new AbortController()
 
     // Declare toolCalls outside the loop so we can reference it in stop check
     let toolCalls: ToolCall[] = []
@@ -419,7 +425,11 @@ export class Session {
       bus.emit(EventChannels.TURN_START, { prompt: userPrompt, iteration: iterations })
       bus.emit(EventChannels.LLM_STREAM_START, { iteration: iterations })
 
-      const stream = this.provider.chat(this.messages, this.registry.getDefinitions())
+      const stream = this.provider.chat(
+        this.messages,
+        this.registry.getDefinitions(),
+        this.abortController.signal,
+      )
 
       for await (const chunk of stream) {
         if (chunk.reasoning_content) {