Update

CLAUDE-ACTION.md

@@ -0,0 +1,299 @@
+# Claude Code Action Architecture & Flow
+
+This document provides a comprehensive overview of how the Claude Code Action works, from token exchange through post-run cleanup.
+
+## Overview
+
+The Claude Code Action is a sophisticated GitHub automation platform that enables Claude to interact with GitHub repositories through secure token exchange, intelligent mode detection, and comprehensive GitHub API integration.
+
+## High-Level Architecture
+
+```mermaid
+graph TD
+    Start([GitHub Action Triggered]) --> Setup[Setup Environment<br/>- Install Bun<br/>- Install Dependencies]
+    
+    Setup --> ParseContext[Parse GitHub Context<br/>- Extract event data<br/>- Parse inputs]
+    
+    ParseContext --> ModeDetection{Mode Detection}
+    
+    ModeDetection -->|Has explicit prompt| AgentMode[AGENT MODE<br/>Direct automation]
+    ModeDetection -->|@claude mention/assignment/label| TagMode[TAG MODE<br/>Interactive response]
+    ModeDetection -->|No trigger| DefaultAgent[Default to Agent<br/>(won't trigger)]
+    
+    %% Token Exchange Branch
+    AgentMode --> TokenExchange[Token Exchange Process]
+    TagMode --> TokenExchange
+    TokenExchange --> TokenMethod{Token Method}
+    
+    TokenMethod -->|Custom token provided| UseCustom[Use Custom GitHub Token]
+    TokenMethod -->|No custom token| OIDC[Generate OIDC Token<br/>core.getIDToken()]
+    
+    OIDC --> Exchange[Exchange OIDC for App Token<br/>api.anthropic.com/api/github/github-app-token-exchange]
+    Exchange --> CreateOctokit[Create Authenticated Octokit Client<br/>REST + GraphQL]
+    UseCustom --> CreateOctokit
+    
+    %% Permission Checks
+    CreateOctokit --> PermCheck[Check Write Permissions<br/>Only for entity contexts]
+    PermCheck -->|No permissions| PermFail[❌ Exit: No write access]
+    PermCheck -->|Has permissions| TriggerCheck{Check Trigger Conditions}
+    
+    %% Trigger Validation
+    TriggerCheck -->|Agent Mode| AgentTrigger{Has explicit prompt?}
+    TriggerCheck -->|Tag Mode| TagTrigger{Contains @claude mention<br/>or assignment/label?}
+    
+    AgentTrigger -->|No prompt| NoTrigger[❌ Skip: No trigger found]
+    AgentTrigger -->|Has prompt| PrepareAgent[Prepare Agent Mode]
+    TagTrigger -->|No mention| NoTrigger
+    TagTrigger -->|Has mention| PrepareTag[Prepare Tag Mode]
+    
+    %% Mode-Specific Preparation
+    PrepareAgent --> AgentPrep[Agent Mode Preparation<br/>- Create prompt file<br/>- Setup MCP servers<br/>- No tracking comment]
+    PrepareTag --> TagPrep[Tag Mode Preparation<br/>- Create tracking comment<br/>- Setup branches<br/>- Fetch GitHub data<br/>- Setup MCP servers]
+    
+    %% Data Fetching (Tag Mode)
+    TagPrep --> DataFetch[Fetch GitHub Data<br/>GraphQL + REST API]
+    DataFetch --> FetchWhat{What to fetch?}
+    
+    FetchWhat -->|Pull Request| PRData[PR Data:<br/>- Comments & reviews<br/>- Changed files + SHAs<br/>- Commit history<br/>- Author info]
+    FetchWhat -->|Issue| IssueData[Issue Data:<br/>- Comments<br/>- Issue details<br/>- Author info]
+    
+    PRData --> ProcessImages[Process Images<br/>Download & convert to base64]
+    IssueData --> ProcessImages
+    ProcessImages --> SetupBranch[Setup Branch<br/>- Create Claude branch<br/>- Configure git auth]
+    
+    %% MCP Server Setup
+    AgentPrep --> MCPSetup[Setup MCP Servers]
+    SetupBranch --> MCPSetup
+    
+    MCPSetup --> MCPServers{MCP Servers}
+    MCPServers --> GitHubActions[GitHub Actions Server<br/>- Workflow data<br/>- CI results]
+    MCPServers --> GitHubComments[GitHub Comment Server<br/>- Comment operations]
+    MCPServers --> GitHubFiles[GitHub File Ops Server<br/>- File operations<br/>- Branch management]
+    MCPServers --> GitHubInline[GitHub Inline Comment Server<br/>- PR review comments]
+    
+    GitHubActions --> PromptGen[Generate Prompt]
+    GitHubComments --> PromptGen
+    GitHubFiles --> PromptGen
+    GitHubInline --> PromptGen
+    
+    %% Prompt Generation
+    PromptGen --> PromptType{Prompt Type}
+    PromptType -->|Agent Mode| AgentPrompt[Agent Prompt:<br/>- Direct user prompt<br/>- Minimal context]
+    PromptType -->|Tag Mode| TagPrompt[Tag Prompt:<br/>- Rich GitHub context<br/>- PR/Issue details<br/>- Changed files<br/>- Comments & reviews<br/>- Commit instructions]
+    
+    AgentPrompt --> ClaudeRun[Run Claude Code]
+    TagPrompt --> ClaudeRun
+    
+    %% Claude Execution
+    ClaudeRun --> ClaudeExec[Claude Code Execution<br/>base-action/src/index.ts]
+    ClaudeExec --> ClaudeArgs[Prepare Claude Args<br/>- Prompt file path<br/>- Custom claude_args<br/>- Output format: stream-json]
+    
+    ClaudeArgs --> ClaudeProvider{Provider}
+    ClaudeProvider -->|Default| AnthropicAPI[Anthropic API<br/>ANTHROPIC_API_KEY]
+    ClaudeProvider -->|Bedrock| AWSBedrock[AWS Bedrock<br/>OIDC + AWS credentials]
+    ClaudeProvider -->|Vertex| GCPVertex[GCP Vertex AI<br/>OIDC + GCP credentials]
+    
+    AnthropicAPI --> ClaudeProcess[Spawn Claude Process<br/>- Named pipe for input<br/>- Stream JSON output]
+    AWSBedrock --> ClaudeProcess
+    GCPVertex --> ClaudeProcess
+    
+    ClaudeProcess --> ClaudeTools[Claude Tool Usage<br/>- MCP tools<br/>- File operations<br/>- GitHub API calls<br/>- Bash commands]
+    
+    ClaudeTools --> ClaudeOutput[Claude Output Processing<br/>- Capture execution log<br/>- Parse JSON stream<br/>- Extract metrics]
+    
+    %% Post-Run Actions
+    ClaudeOutput --> PostRun{Post-Run Actions}
+    
+    PostRun -->|Success| Success[✅ Success Path]
+    PostRun -->|Failure| Failure[❌ Failure Path]
+    
+    Success --> UpdateComment[Update Tracking Comment<br/>- Job run link<br/>- Branch link<br/>- PR link (if created)<br/>- Execution metrics]
+    Failure --> UpdateComment
+    
+    UpdateComment --> BranchCleanup[Branch Cleanup<br/>- Check for changes<br/>- Delete empty branches<br/>- Keep branches with commits]
+    
+    BranchCleanup --> FormatReport[Format Execution Report<br/>- Parse conversation turns<br/>- Format tool usage<br/>- Add to GitHub step summary]
+    
+    FormatReport --> RevokeToken[Revoke App Token<br/>DELETE /installation/token]
+    
+    RevokeToken --> End([Action Complete])
+```
+
+## Key Components
+
+### 1. Token Exchange Process
+
+The action uses a secure OIDC token exchange system:
+
+1. **OIDC Token Generation**: `core.getIDToken("claude-code-github-action")`
+2. **Token Exchange**: POST to `https://api.anthropic.com/api/github/github-app-token-exchange`
+3. **Authentication**: Creates authenticated Octokit clients for GitHub API access
+
+**Security Benefits:**
+- Repository-scoped access
+- Time-limited tokens
+- Permission-limited (only configured GitHub App permissions)
+- Automatic token masking in logs
+
+### 2. Mode Detection
+
+The action automatically detects the appropriate execution mode:
+
+#### **Agent Mode**
+- **Trigger**: Explicit `prompt` input provided
+- **Use Case**: Direct automation, custom workflows
+- **Behavior**: Minimal context, direct execution
+- **Tracking**: No tracking comments
+
+#### **Tag Mode**  
+- **Trigger**: @claude mentions, issue assignments, or labels
+- **Use Case**: Interactive GitHub responses
+- **Behavior**: Rich context, comprehensive GitHub data
+- **Tracking**: Creates and updates tracking comments
+
+### 3. Data Fetching (Tag Mode)
+
+When in Tag Mode, the action fetches comprehensive GitHub context:
+
+#### **Pull Request Data:**
+- Comments and reviews (including inline comments)
+- Changed files with SHAs
+- Commit history and metadata
+- Author information
+- File diff data
+
+#### **Issue Data:**
+- Issue details and metadata  
+- All comments
+- Author information
+- Labels and assignments
+
+#### **Image Processing:**
+- Downloads images from GitHub
+- Converts to base64 for Claude
+- Maps original URLs to processed content
+
+### 4. MCP Server Integration
+
+The action sets up multiple MCP (Model Context Protocol) servers to provide Claude with GitHub capabilities:
+
+#### **GitHub Actions Server**
+- Access to workflow runs and CI data
+- Build status and test results
+- Artifact information
+
+#### **GitHub Comment Server**
+- Comment creation and updates
+- Issue and PR comment management
+
+#### **GitHub File Operations Server**
+- File reading and writing
+- Branch creation and management
+- Commit operations
+
+#### **GitHub Inline Comment Server**
+- PR review comment operations
+- Line-specific feedback
+
+### 5. Prompt Generation
+
+The action generates context-rich prompts based on the detected mode:
+
+#### **Agent Mode Prompts:**
+- Direct user prompt
+- Minimal GitHub context
+- Focused on specific task
+
+#### **Tag Mode Prompts:**
+- Comprehensive GitHub context
+- PR/Issue details and history
+- Changed files and diffs
+- Comment threads and reviews
+- Commit instructions and guidelines
+
+### 6. Claude Execution
+
+The action runs Claude Code through multiple provider options:
+
+#### **Provider Support:**
+- **Anthropic API** (default): Direct API access with API key
+- **AWS Bedrock**: OIDC authentication with AWS credentials
+- **GCP Vertex AI**: OIDC authentication with GCP credentials
+
+#### **Execution Process:**
+1. **Named Pipe Setup**: Creates pipe for prompt input
+2. **Process Spawning**: Spawns Claude Code process
+3. **Stream Processing**: Captures JSON stream output
+4. **Tool Integration**: Enables MCP tools and GitHub operations
+
+### 7. Post-Run Actions
+
+After Claude execution, the action performs comprehensive cleanup and reporting:
+
+#### **Comment Updates:**
+- Updates tracking comments with results
+- Adds job run links and execution metrics
+- Includes branch and PR links when created
+
+#### **Branch Management:**
+- Checks for actual changes in Claude branches
+- Deletes empty branches to avoid clutter
+- Preserves branches with meaningful commits
+
+#### **Report Generation:**
+- Parses execution logs and conversation turns
+- Formats tool usage and results
+- Adds formatted report to GitHub step summary
+
+#### **Security Cleanup:**
+- Revokes GitHub App installation token
+- Cleans up temporary files and processes
+
+## Security Considerations
+
+### **Access Control:**
+- Repository-scoped permissions only
+- Write access validation for actors
+- Bot user controls and allowlists
+
+### **Token Management:**
+- Short-lived installation tokens
+- Automatic token revocation after use
+- Secure OIDC-based exchange
+
+### **Permission Boundaries:**
+- Limited to configured GitHub App permissions
+- No cross-repository access
+- Scoped to specific repository operations
+
+## Integration Points
+
+### **With Pullfrog:**
+The Claude Code Action can be integrated with Pullfrog's workflow system, providing:
+- Standardized agent interaction patterns
+- Consistent GitHub integration
+- Reusable authentication flows
+- Common MCP server infrastructure
+
+### **With GitHub:**
+- Native GitHub Actions integration
+- Comprehensive API coverage (REST + GraphQL)
+- Proper webhook handling
+- Standard GitHub UI integration
+
+## Development Notes
+
+### **Key Files:**
+- `src/entrypoints/prepare.ts`: Main preparation logic
+- `src/modes/`: Mode detection and handling
+- `src/github/token.ts`: OIDC token exchange
+- `src/mcp/`: MCP server implementations
+- `base-action/`: Core Claude Code execution
+
+### **Testing:**
+- Unit tests for individual components
+- Integration tests for full workflows
+- Local testing with `act` tool
+- Comprehensive fixture support
+
+This architecture provides a robust, secure, and extensible foundation for Claude-GitHub integration while maintaining clear separation of concerns and comprehensive error handling.

README.md

@@ -2,6 +2,8 @@
 
 GitHub Action for running Claude Code and other agents via Pullfrog.
 
+> **📖 Claude Code Action Architecture**: For a detailed technical overview of how the Claude Code Action works (token exchange, modes, data fetching, execution flow), see [CLAUDE-ACTION.md](./CLAUDE-ACTION.md).
+
 ## Quick Start
 
 ```bash

agents/claude.ts

@@ -53,10 +53,7 @@ export class ClaudeAgent implements Agent {
       // Use shell execution to properly handle the pipe
       const result = await spawn({
         cmd: "bash",
-        args: [
+        args: ["-c", "curl -fsSL https://claude.ai/install.sh | bash -s 1.0.93"],
-          "-c",
-          "curl -fsSL https://claude.ai/install.sh | bash -s 1.0.93",
-        ],
         env: { ANTHROPIC_API_KEY: this.apiKey },
         timeout: 120000, // 2 minute timeout
         onStdout: (chunk) => {
@@ -67,9 +64,7 @@ export class ClaudeAgent implements Agent {
       });
 
       if (result.exitCode !== 0) {
-        throw new Error(
+        throw new Error(`Installation failed with exit code ${result.exitCode}: ${result.stderr}`);
-          `Installation failed with exit code ${result.exitCode}: ${result.stderr}`,
-        );
       }
 
       core.info("Claude Code installed successfully");
@@ -139,7 +134,7 @@ export class ClaudeAgent implements Agent {
       // throw on non-zero exit code
       if (result.exitCode !== 0) {
         throw new Error(
-          `Command failed with exit code ${result.exitCode}\n\nStdout: ${result.stdout}\n\nStderr: ${result.stderr}`,
+          `Command failed with exit code ${result.exitCode}\n\nStdout: ${result.stdout}\n\nStderr: ${result.stderr}`
         );
       }
 
@@ -158,7 +153,7 @@ export class ClaudeAgent implements Agent {
       // Log run summary
       const duration = Date.now() - this.runStats.startTime;
       core.info(
-        `📊 Run Summary: ${this.runStats.toolsUsed} tools used, ${this.runStats.turns} turns, ${duration}ms duration`,
+        `📊 Run Summary: ${this.runStats.toolsUsed} tools used, ${this.runStats.turns} turns, ${duration}ms duration`
       );
 
       core.info("✅ Task complete.");
@@ -181,8 +176,7 @@ export class ClaudeAgent implements Agent {
       } catch {
         // Group might not have been started, ignore
       }
-      const errorMessage =
+      const errorMessage = error instanceof Error ? error.message : "Unknown error";
-        error instanceof Error ? error.message : "Unknown error";
       return {
         success: false,
         error: `Failed to execute Claude Code: ${errorMessage}`,
@@ -233,12 +227,7 @@ function processJSONChunk(chunk: string, agent?: ClaudeAgent): void {
               ["model", parsedChunk.model],
               ["cwd", parsedChunk.cwd],
               ["permission_mode", parsedChunk.permissionMode],
-              [
+              ["tools", parsedChunk.tools?.length ? `${parsedChunk.tools.length} tools` : "none"],
-                "tools",
-                parsedChunk.tools?.length
-                  ? `${parsedChunk.tools.length} tools`
-                  : "none",
-              ],
               [
                 "mcp_servers",
                 parsedChunk.mcp_servers?.length
@@ -251,7 +240,7 @@ function processJSONChunk(chunk: string, agent?: ClaudeAgent): void {
                   ? `${parsedChunk.slash_commands.length} commands`
                   : "none",
               ],
-            ]),
+            ])
           );
         }
         break;
@@ -267,9 +256,7 @@ function processJSONChunk(chunk: string, agent?: ClaudeAgent): void {
             if (content.type === "text") {
               // Skip empty text content
               if (content.text.trim()) {
-                core.info(
+                core.info(boxString(content.text.trim(), { title: "Claude Code" }));
-                  boxString(content.text.trim(), { title: "Claude Code" }),
-                );
               }
             } else if (content.type === "tool_use") {
               // Track tools used
@@ -379,15 +366,12 @@ function processJSONChunk(chunk: string, agent?: ClaudeAgent): void {
 
           core.info(
             tableString([
-              [
+              ["Cost", `$${parsedChunk.total_cost_usd?.toFixed(4) || "0.0000"}`],
-                "Cost",
-                `$${parsedChunk.total_cost_usd?.toFixed(4) || "0.0000"}`,
-              ],
               ["Input Tokens", parsedChunk.usage?.input_tokens || 0],
               ["Output Tokens", parsedChunk.usage?.output_tokens || 0],
               ["Duration", `${parsedChunk.duration_ms}ms`],
               ["Turns", parsedChunk.num_turns || 1],
-            ]),
+            ])
           );
         } else {
           core.error(`❌ Failed: ${parsedChunk.error || "Unknown error"}`);

entry.cjs

@@ -25749,10 +25749,7 @@ var ClaudeAgent = class {
     try {
       const result = await spawn({
         cmd: "bash",
-        args: [
+        args: ["-c", "curl -fsSL https://claude.ai/install.sh | bash -s 1.0.93"],
-          "-c",
-          "curl -fsSL https://claude.ai/install.sh | bash -s 1.0.93"
-        ],
         env: { ANTHROPIC_API_KEY: this.apiKey },
         timeout: 12e4,
         // 2 minute timeout
@@ -25761,9 +25758,7 @@ var ClaudeAgent = class {
         onStderr: (chunk) => process.stderr.write(chunk)
       });
       if (result.exitCode !== 0) {
-        throw new Error(
+        throw new Error(`Installation failed with exit code ${result.exitCode}: ${result.stderr}`);
-          `Installation failed with exit code ${result.exitCode}: ${result.stderr}`
-        );
       }
       core2.info("Claude Code installed successfully");
     } catch (error2) {
@@ -25864,10 +25859,7 @@ function processJSONChunk(chunk, agent) {
               ["model", parsedChunk.model],
               ["cwd", parsedChunk.cwd],
               ["permission_mode", parsedChunk.permissionMode],
-              [
+              ["tools", parsedChunk.tools?.length ? `${parsedChunk.tools.length} tools` : "none"],
-                "tools",
-                parsedChunk.tools?.length ? `${parsedChunk.tools.length} tools` : "none"
-              ],
               [
                 "mcp_servers",
                 parsedChunk.mcp_servers?.length ? `${parsedChunk.mcp_servers.length} servers` : "none"
@@ -25888,9 +25880,7 @@ function processJSONChunk(chunk, agent) {
           for (const content of parsedChunk.message.content) {
             if (content.type === "text") {
               if (content.text.trim()) {
-                core2.info(
+                core2.info(boxString(content.text.trim(), { title: "Claude Code" }));
-                  boxString(content.text.trim(), { title: "Claude Code" })
-                );
               }
             } else if (content.type === "tool_use") {
               if (agent) {
@@ -25958,10 +25948,7 @@ function processJSONChunk(chunk, agent) {
         if (parsedChunk.subtype === "success") {
           core2.info(
             tableString([
-              [
+              ["Cost", `$${parsedChunk.total_cost_usd?.toFixed(4) || "0.0000"}`],
-                "Cost",
-                `$${parsedChunk.total_cost_usd?.toFixed(4) || "0.0000"}`
-              ],
               ["Input Tokens", parsedChunk.usage?.input_tokens || 0],
               ["Output Tokens", parsedChunk.usage?.output_tokens || 0],
               ["Duration", `${parsedChunk.duration_ms}ms`],

main.ts

@@ -5,7 +5,7 @@ import { ClaudeAgent } from "./agents";
 export const EXPECTED_INPUTS: string[] = [
   "ANTHROPIC_API_KEY",
   "GITHUB_TOKEN",
-  "GITHUB_INSTALLATION_TOKEN"
+  "GITHUB_INSTALLATION_TOKEN",
 ];
 
 export interface ExecutionInputs {
@@ -27,7 +27,6 @@ export interface MainResult {
   error?: string | undefined;
 }
 
-
 export async function main(params: MainParams): Promise<MainResult> {
   try {
     // Extract inputs from params

package.json

@@ -32,7 +32,7 @@
   },
   "devDependencies": {
     "@types/node": "^20.10.0",
-    "commander": "^14.0.0",
+    "arg": "^5.0.2",
     "esbuild": "^0.25.9",
     "husky": "^9.0.0",
     "typescript": "^5.3.0",

play.ts

@@ -1,7 +1,7 @@
 import { existsSync, readFileSync } from "node:fs";
 import { dirname, extname, join, resolve } from "node:path";
 import { fileURLToPath, pathToFileURL } from "node:url";
-import { Command } from "commander";
+import arg from "arg";
 import { config } from "dotenv";
 import { main } from "./main";
 import { runAct } from "./utils/act";
@@ -10,151 +10,189 @@ import { setupTestRepo } from "./utils/setup";
 // Load environment variables from .env file
 config();
 
-
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = dirname(__filename);
 
-async function loadPrompt(filePath: string): Promise<string> {
+export async function run(
-  const ext = extname(filePath).toLowerCase();
+  prompt: string,
-
+  options: { act?: boolean } = {}
-  // Try to resolve the file path
+): Promise<{ success: boolean; output?: string | undefined; error?: string | undefined }> {
-  let resolvedPath: string;
-
-  // First try as fixtures path
-  const fixturesPath = join(__dirname, "fixtures", filePath);
-  if (existsSync(fixturesPath)) {
-    resolvedPath = fixturesPath;
-  } else if (existsSync(filePath)) {
-    resolvedPath = resolve(filePath);
-  } else {
-    throw new Error(`File not found: ${filePath}`);
-  }
-
-  switch (ext) {
-    case ".txt": {
-      // Plain text - pass directly as prompt
-      return readFileSync(resolvedPath, "utf8").trim();
-    }
-
-    case ".json": {
-      // JSON - stringify and pass as prompt
-      const content = readFileSync(resolvedPath, "utf8");
-      const parsed = JSON.parse(content);
-      return JSON.stringify(parsed, null, 2);
-    }
-
-    case ".ts": {
-      // TypeScript - dynamic import and stringify default export
-      const fileUrl = pathToFileURL(resolvedPath).href;
-      const module = await import(fileUrl);
-
-      if (!module.default) {
-        throw new Error(`TypeScript file ${filePath} must have a default export`);
-      }
-
-      // If it's a string, use it directly
-      if (typeof module.default === "string") {
-        return module.default;
-      }
-
-      // If it's a MainParams object with a prompt field, extract the prompt
-      if (typeof module.default === "object" && module.default.prompt) {
-        return module.default.prompt;
-      }
-
-      // Otherwise stringify it
-      return JSON.stringify(module.default, null, 2);
-    }
-
-    default:
-      throw new Error(`Unsupported file type: ${ext}. Supported types: .txt, .json, .ts`);
-  }
-}
-
-async function runPlay(filePath: string, options: { act?: boolean }): Promise<void> {
   try {
-    // Load the prompt from the specified file
-    const prompt = await loadPrompt(filePath);
-
     if (options.act) {
       // Use Docker/act to run the action
       console.log("🐳 Running with Docker/act...");
       runAct(prompt);
+      return { success: true };
+    }
+
+    // Setup test repository and run directly
+    const tempDir = join(process.cwd(), ".temp");
+    setupTestRepo({ tempDir, forceClean: true });
+
+    // Change to the temp directory
+    const originalCwd = process.cwd();
+    process.chdir(tempDir);
+
+    console.log("🚀 Running action with prompt...");
+    console.log("─".repeat(50));
+    console.log("Prompt:");
+    console.log(prompt);
+    console.log("─".repeat(50));
+
+    // Set environment variables from our .env for the action to use
+    const { EXPECTED_INPUTS } = await import("./main");
+    EXPECTED_INPUTS.forEach((inputName) => {
+      const value = process.env[inputName];
+      if (value) {
+        process.env[`INPUT_${inputName.toLowerCase()}`] = value;
+      }
+    });
+
+    // Run main with the new params structure
+    const inputs: any = {
+      prompt,
+      anthropic_api_key: process.env.ANTHROPIC_API_KEY || "",
+    };
+
+    // Add optional properties only if they exist
+    if (process.env.GITHUB_TOKEN) {
+      inputs.github_token = process.env.GITHUB_TOKEN;
+    }
+
+    if (process.env.GITHUB_INSTALLATION_TOKEN) {
+      inputs.github_installation_token = process.env.GITHUB_INSTALLATION_TOKEN;
+    }
+
+    const result = await main({
+      inputs,
+      env: process.env as Record<string, string>,
+      cwd: process.cwd(),
+    });
+
+    // Change back to original directory
+    process.chdir(originalCwd);
+
+    if (result.success) {
+      console.log("✅ Action completed successfully");
+      if (result.output) {
+        console.log("Output:", result.output);
+      }
+      return { success: true, output: result.output || undefined, error: undefined };
     } else {
-      // Setup test repository and run directly
+      console.error("❌ Action failed:", result.error);
-      const tempDir = join(process.cwd(), ".temp");
+      return { success: false, error: result.error || undefined, output: undefined };
-      setupTestRepo({ tempDir, forceClean: true });
+    }
+  } catch (error) {
+    const errorMessage = (error as Error).message;
+    console.error("❌ Error:", errorMessage);
+    return { success: false, error: errorMessage, output: undefined };
+  }
+}
 
-      // Change to the temp directory
+// CLI execution when run directly
-      process.chdir(tempDir);
+if (import.meta.url === `file://${process.argv[1]}`) {
+  const args = arg({
+    "--help": Boolean,
+    "--act": Boolean,
+    "--raw": String,
+    "-h": "--help",
+  });
 
-      console.log("🚀 Running test in .temp directory...");
+  if (args["--help"]) {
-      console.log("─".repeat(50));
+    console.log(`
-      console.log(`Prompt from ${filePath}:`);
+Usage: tsx play.ts [file] [options]
-      console.log(prompt);
-      console.log("─".repeat(50));
 
-      // Set environment variables from our .env for the action to use
+Test the Pullfrog action with various prompts.
-      const { EXPECTED_INPUTS } = await import("./main");
+
-      EXPECTED_INPUTS.forEach((inputName) => {
+Arguments:
-        const value = process.env[inputName];
+  file                    Prompt file to use (.txt, .json, or .ts) [default: fixtures/basic.txt]
-        if (value) {
+
-          process.env[`INPUT_${inputName.toLowerCase()}`] = value;
+Options:
+  --act                   Use Docker/act to run the action instead of running directly
+  --raw [prompt]          Use raw string as prompt instead of loading from file
+  -h, --help              Show this help message
+
+Examples:
+  tsx play.ts                        # Use default fixture
+  tsx play.ts fixtures/basic.txt     # Use specific text file
+  tsx play.ts custom.json            # Use JSON file
+  tsx play.ts --act fixtures/test.ts # Use TypeScript file with Docker/act
+  tsx play.ts --raw "Hello world"    # Use raw string as prompt
+    `);
+    process.exit(0);
+  }
+
+  let prompt: string;
+
+  if (args["--raw"]) {
+    // Use raw prompt string
+    prompt = args["--raw"];
+  } else {
+    // Load prompt from file
+    const filePath = args._[0] || "fixtures/basic.txt";
+    const ext = extname(filePath).toLowerCase();
+    let resolvedPath: string;
+
+    // First try as fixtures path
+    const fixturesPath = join(__dirname, "fixtures", filePath);
+    if (existsSync(fixturesPath)) {
+      resolvedPath = fixturesPath;
+    } else if (existsSync(filePath)) {
+      resolvedPath = resolve(filePath);
+    } else {
+      throw new Error(`File not found: ${filePath}`);
+    }
+
+    switch (ext) {
+      case ".txt": {
+        // Plain text - pass directly as prompt
+        prompt = readFileSync(resolvedPath, "utf8").trim();
+        break;
+      }
+
+      case ".json": {
+        // JSON - stringify and pass as prompt
+        const content = readFileSync(resolvedPath, "utf8");
+        const parsed = JSON.parse(content);
+        prompt = JSON.stringify(parsed, null, 2);
+        break;
+      }
+
+      case ".ts": {
+        // TypeScript - dynamic import and stringify default export
+        const fileUrl = pathToFileURL(resolvedPath).href;
+        const module = await import(fileUrl);
+
+        if (!module.default) {
+          throw new Error(`TypeScript file ${filePath} must have a default export`);
         }
-      });
 
-      // Run main with the new params structure
+        // If it's a string, use it directly
-      const inputs: any = {
+        if (typeof module.default === "string") {
-        prompt,
+          prompt = module.default;
-        anthropic_api_key: process.env.ANTHROPIC_API_KEY || "",
+        } else if (typeof module.default === "object" && module.default.prompt) {
-      };
+          // If it's a MainParams object with a prompt field, extract the prompt
-      
+          prompt = module.default.prompt;
-      // Add optional properties only if they exist
+        } else {
-      if (process.env.GITHUB_TOKEN) {
+          // Otherwise stringify it
-        inputs.github_token = process.env.GITHUB_TOKEN;
+          prompt = JSON.stringify(module.default, null, 2);
-      }
-      
-      if (process.env.GITHUB_INSTALLATION_TOKEN) {
-        inputs.github_installation_token = process.env.GITHUB_INSTALLATION_TOKEN;
-      }
-
-      const result = await main({
-        inputs,
-        env: process.env as Record<string, string>,
-        cwd: process.cwd(),
-      });
-
-      if (result.success) {
-        console.log("✅ Test completed successfully");
-        if (result.output) {
-          console.log("Output:", result.output);
         }
-      } else {
+        break;
-        console.error("❌ Test failed:", result.error);
-        process.exit(1);
       }
+
+      default:
+        throw new Error(`Unsupported file type: ${ext}. Supported types: .txt, .json, .ts`);
+    }
+  }
+
+  try {
+    const result = await run(prompt, { act: args["--act"] || false });
+
+    if (!result.success) {
+      process.exit(1);
     }
   } catch (error) {
     console.error("❌ Error:", (error as Error).message);
     process.exit(1);
   }
 }
-
-// Set up CLI
-const program = new Command();
-
-program
-  .name("play")
-  .description("Test the Pullfrog action with various prompts")
-  .version("1.0.0")
-  .argument("[file]", "Prompt file to use (.txt, .json, or .ts)", "fixtures/basic.txt")
-  .option("--act", "Use Docker/act to run the action instead of running directly")
-  .action(async (file: string, options: { act?: boolean }) => {
-    await runPlay(file, options);
-  });
-
-// Parse arguments and run
-program.parseAsync(process.argv).catch((error) => {
-  console.error("Fatal error:", error);
-  process.exit(1);
-});