Merge pull request #150 from actions/sgoedecke/mock-inference-in-ci

Mock inference in CI

.github/workflows/check-dist.yml

@@ -28,7 +28,7 @@ jobs:
     steps:
       - name: Checkout
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
 
       - name: Setup Node.js
         id: setup-node

.github/workflows/ci.yml

@@ -20,7 +20,7 @@ jobs:
     steps:
       - name: Checkout
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
 
       - name: Setup Node.js
         id: setup-node
@@ -54,22 +54,53 @@ jobs:
     steps:
       - name: Checkout
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version-file: .node-version
+
+      - name: Start Mock Inference Server
+        id: mock-server
+        run: |
+          node script/mock-inference-server.mjs &
+          echo "pid=$!" >> $GITHUB_OUTPUT
+          # Wait for server to be ready
+          for i in {1..10}; do
+            if curl -s http://localhost:3456/health > /dev/null; then
+              echo "Mock server is ready"
+              break
+            fi
+            sleep 1
+          done
 
       - name: Test Local Action
         id: test-action
-        continue-on-error: true
         uses: ./
         with:
           prompt: hello
+          endpoint: http://localhost:3456
         env:
           GITHUB_TOKEN: ${{ github.token }}
 
       - name: Print Output
         id: output
-        continue-on-error: true
         run: echo "${{ steps.test-action.outputs.response }}"
 
+      - name: Verify Output
+        run: |
+          response="${{ steps.test-action.outputs.response }}"
+          if [[ -z "$response" ]]; then
+            echo "Error: No response received"
+            exit 1
+          fi
+          echo "Response received: $response"
+
+      - name: Stop Mock Server
+        if: always()
+        run: kill ${{ steps.mock-server.outputs.pid }} || true
+
   test-action-prompt-file:
     name: GitHub Actions Test with Prompt File
     runs-on: ubuntu-latest
@@ -77,7 +108,26 @@ jobs:
     steps:
       - name: Checkout
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version-file: .node-version
+
+      - name: Start Mock Inference Server
+        id: mock-server
+        run: |
+          node script/mock-inference-server.mjs &
+          echo "pid=$!" >> $GITHUB_OUTPUT
+          # Wait for server to be ready
+          for i in {1..10}; do
+            if curl -s http://localhost:3456/health > /dev/null; then
+              echo "Mock server is ready"
+              break
+            fi
+            sleep 1
+          done
 
       - name: Create Prompt File
         run: echo "hello" > prompt.txt
@@ -87,16 +137,33 @@ jobs:
 
       - name: Test Local Action with Prompt File
         id: test-action-prompt-file
-        continue-on-error: true
         uses: ./
         with:
           prompt-file: prompt.txt
           system-prompt-file: system-prompt.txt
+          endpoint: http://localhost:3456
         env:
           GITHUB_TOKEN: ${{ github.token }}
 
       - name: Print Output
-        continue-on-error: true
         run: |
           echo "Response saved to: ${{ steps.test-action-prompt-file.outputs.response-file }}"
           cat "${{ steps.test-action-prompt-file.outputs.response-file }}"
+
+      - name: Verify Output
+        run: |
+          response_file="${{ steps.test-action-prompt-file.outputs.response-file }}"
+          if [[ ! -f "$response_file" ]]; then
+            echo "Error: Response file not found"
+            exit 1
+          fi
+          content=$(cat "$response_file")
+          if [[ -z "$content" ]]; then
+            echo "Error: Response file is empty"
+            exit 1
+          fi
+          echo "Response file content: $content"
+
+      - name: Stop Mock Server
+        if: always()
+        run: kill ${{ steps.mock-server.outputs.pid }} || true

.github/workflows/codeql-analysis.yml

@@ -30,7 +30,7 @@ jobs:
     steps:
       - name: Checkout
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
 
       - name: Initialize CodeQL
         id: initialize

.github/workflows/licensed.yml

@@ -27,7 +27,7 @@ jobs:
     steps:
       - name: Checkout
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
 
       - name: Setup Node.js
         id: setup-node

.github/workflows/linter.yml

@@ -21,7 +21,7 @@ jobs:
     steps:
       - name: Checkout
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           fetch-depth: 0
 

README.md

@@ -162,6 +162,9 @@ This action now supports **read-only** integration with the GitHub-hosted Model
 Context Protocol (MCP) server, which provides access to GitHub tools like
 repository management, issue tracking, and pull request operations.
 
+> [!NOTE]
+> The GitHub MCP integration requires a Personal Access Token (PAT) and cannot use the built-in `GITHUB_TOKEN`.
+
 ```yaml
 steps:
   - name: AI Inference with GitHub Tools
@@ -209,7 +212,7 @@ the action:
 | `endpoint`           | The endpoint to use for inference. If you're running this as part of an org, you should probably use the org-specific Models endpoint                         | `https://models.github.ai/inference` |
 | `max-tokens`         | The max number of tokens to generate                                                                                                                          | 200                                  |
 | `enable-github-mcp`  | Enable Model Context Protocol integration with GitHub tools                                                                                                   | `false`                              |
-| `github-mcp-token`   | Token to use for GitHub MCP server (defaults to the main token if not specified). Use a separate PAT for tighter security                                     | `""`                                 |
+| `github-mcp-token`   | Token to use for GitHub MCP server (defaults to the main token if not specified). This must be a PAT in order for MCP to work                                 | `""`                                 |
 
 ## Outputs
 

__tests__/helpers-inference.test.ts

@@ -106,6 +106,8 @@ describe('helpers.ts - inference request building', () => {
         undefined,
         undefined,
         'gpt-4',
+        undefined,
+        undefined,
         100,
         'https://api.test.com',
         'test-token',
@@ -117,6 +119,8 @@ describe('helpers.ts - inference request building', () => {
           {role: 'user', content: 'User message'},
         ],
         modelName: 'gpt-4',
+        temperature: undefined,
+        topP: undefined,
         maxTokens: 100,
         endpoint: 'https://api.test.com',
         token: 'test-token',
@@ -136,6 +140,8 @@ describe('helpers.ts - inference request building', () => {
         'System prompt',
         'User prompt',
         'gpt-4',
+        undefined,
+        undefined,
         100,
         'https://api.test.com',
         'test-token',
@@ -147,6 +153,8 @@ describe('helpers.ts - inference request building', () => {
           {role: 'user', content: 'User prompt'},
         ],
         modelName: 'gpt-4',
+        temperature: undefined,
+        topP: undefined,
         maxTokens: 100,
         endpoint: 'https://api.test.com',
         token: 'test-token',

__tests__/main.test.ts

@@ -75,17 +75,13 @@ vi.mock('fs', () => ({
   writeFileSync: mockWriteFileSync,
 }))
 
-// Mocks for tmp module to control temporary file creation and cleanup
-const mockRemoveCallback = vi.fn()
+// Mocks for tmp module to control temporary file creation
 const mockFileSync = vi.fn().mockReturnValue({
   name: '/secure/temp/dir/modelResponse-abc123.txt',
-  removeCallback: mockRemoveCallback,
 })
-const mockSetGracefulCleanup = vi.fn()
 
 vi.mock('tmp', () => ({
   fileSync: mockFileSync,
-  setGracefulCleanup: mockSetGracefulCleanup,
 }))
 
 // Mock MCP and inference modules
@@ -283,7 +279,7 @@ describe('main.ts', () => {
     expect(mockProcessExit).toHaveBeenCalledWith(1)
   })
 
-  it('creates secure temporary files with proper cleanup', async () => {
+  it('creates temporary files that persist for downstream steps', async () => {
     mockInputs({
       prompt: 'Test prompt',
       'system-prompt': 'You are a test assistant.',
@@ -291,34 +287,16 @@ describe('main.ts', () => {
 
     await run()
 
-    expect(mockSetGracefulCleanup).toHaveBeenCalledOnce()
-
+    // Verify temp file is created with keep: true so it persists
     expect(mockFileSync).toHaveBeenCalledWith({
       prefix: 'modelResponse-',
       postfix: '.txt',
+      keep: true,
     })
 
     expect(core.setOutput).toHaveBeenNthCalledWith(2, 'response-file', '/secure/temp/dir/modelResponse-abc123.txt')
     expect(mockWriteFileSync).toHaveBeenCalledWith('/secure/temp/dir/modelResponse-abc123.txt', 'Hello, user!', 'utf-8')
-    expect(mockRemoveCallback).toHaveBeenCalledOnce()
 
     expect(mockProcessExit).toHaveBeenCalledWith(0)
   })
-
-  it('handles cleanup errors gracefully', async () => {
-    mockRemoveCallback.mockImplementationOnce(() => {
-      throw new Error('Cleanup failed')
-    })
-
-    mockInputs({
-      prompt: 'Test prompt',
-      'system-prompt': 'You are a test assistant.',
-    })
-
-    await run()
-
-    expect(mockRemoveCallback).toHaveBeenCalledOnce()
-    expect(core.warning).toHaveBeenCalledWith('Failed to cleanup temporary file: Error: Cleanup failed')
-    expect(mockProcessExit).toHaveBeenCalledWith(0)
-  })
 })

action.yml

@@ -55,7 +55,7 @@ inputs:
     required: false
     default: 'false'
   github-mcp-token:
-    description: The token to use for GitHub MCP server (defaults to GITHUB_TOKEN if not specified)
+    description: The token to use for GitHub MCP server (defaults to the main token if not specified). This must be a PAT for MCP to work.
     required: false
     default: ''
 

dist/index.js

@@ -49496,28 +49496,18 @@ async function simpleInference(request) {
         messages: request.messages,
         max_tokens: request.maxTokens,
         model: request.modelName,
+        temperature: request.temperature,
+        top_p: request.topP,
     };
     // Add response format if specified
     if (request.responseFormat) {
         // eslint-disable-next-line @typescript-eslint/no-explicit-any
         chatCompletionRequest.response_format = request.responseFormat;
     }
-    try {
-        const response = await client.chat.completions.create(chatCompletionRequest);
-        if ('choices' in response) {
-            const modelResponse = response.choices[0]?.message?.content;
-            coreExports.info(`Model response: ${modelResponse || 'No response content'}`);
-            return modelResponse || null;
-        }
-        else {
-            coreExports.error(`Unexpected response format from API: ${JSON.stringify(response)}`);
-            return null;
-        }
-    }
-    catch (error) {
-        coreExports.error(`API error: ${error}`);
-        throw error;
-    }
+    const response = await chatCompletion(client, chatCompletionRequest, 'simpleInference');
+    const modelResponse = response.choices[0]?.message?.content;
+    coreExports.info(`Model response: ${modelResponse || 'No response content'}`);
+    return modelResponse || null;
 }
 /**
  * GitHub MCP-enabled inference with tool execution loop
@@ -49542,6 +49532,8 @@ async function mcpInference(request, githubMcpClient) {
             messages: messages,
             max_tokens: request.maxTokens,
             model: request.modelName,
+            temperature: request.temperature,
+            top_p: request.topP,
         };
         // Add response format if specified (only on final iteration to avoid conflicts with tool calls)
         if (finalMessage && request.responseFormat) {
@@ -49552,10 +49544,7 @@ async function mcpInference(request, githubMcpClient) {
             chatCompletionRequest.tools = githubMcpClient.tools;
         }
         try {
-            const response = await client.chat.completions.create(chatCompletionRequest);
-            if (!('choices' in response)) {
-                throw new Error(`Unexpected response format from API: ${JSON.stringify(response)}`);
-            }
+            const response = await chatCompletion(client, chatCompletionRequest, `mcpInference iteration ${iterationCount}`);
             const assistantMessage = response.choices[0]?.message;
             const modelResponse = assistantMessage?.content;
             const toolCalls = assistantMessage?.tool_calls;
@@ -49567,17 +49556,13 @@ async function mcpInference(request, githubMcpClient) {
             });
             if (!toolCalls || toolCalls.length === 0) {
                 coreExports.info('No tool calls requested, ending GitHub MCP inference loop');
-                // If we have a response format set and we haven't explicitly run one final message iteration,
-                // do another loop with the response format set
                 if (request.responseFormat && !finalMessage) {
                     coreExports.info('Making one more MCP loop with the requested response format...');
-                    // Add a user message requesting JSON format and try again
                     messages.push({
                         role: 'user',
                         content: `Please provide your response in the exact ${request.responseFormat.type} format specified.`,
                     });
                     finalMessage = true;
-                    // Continue the loop to get a properly formatted response
                     continue;
                 }
                 else {
@@ -49585,9 +49570,7 @@ async function mcpInference(request, githubMcpClient) {
                 }
             }
             coreExports.info(`Model requested ${toolCalls.length} tool calls`);
-            // Execute all tool calls via GitHub MCP
             const toolResults = await executeToolCalls(githubMcpClient.client, toolCalls);
-            // Add tool results to the conversation
             messages.push(...toolResults);
             coreExports.info('Tool results added, continuing conversation...');
         }
@@ -49604,6 +49587,38 @@ async function mcpInference(request, githubMcpClient) {
         .find(msg => msg.role === 'assistant');
     return lastAssistantMessage?.content || null;
 }
+/**
+ * Wrapper around OpenAI chat.completions.create with defensive handling for cases where
+ * the SDK returns a raw string (e.g., unexpected content-type or streaming body) instead of
+ * a parsed object. Ensures an object with a 'choices' array is returned or throws a descriptive error.
+ */
+async function chatCompletion(client, params, context) {
+    try {
+        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+        let response = await client.chat.completions.create(params);
+        coreExports.debug(`${context}: raw response typeof=${typeof response}`);
+        if (typeof response === 'string') {
+            // Attempt to parse if we unexpectedly received a string
+            try {
+                response = JSON.parse(response);
+            }
+            catch (e) {
+                const preview = response.slice(0, 400);
+                throw new Error(`${context}: Chat completion response was a string and not valid JSON (${e.message}). Preview: ${preview}`);
+            }
+        }
+        if (!response || typeof response !== 'object' || !('choices' in response)) {
+            const preview = JSON.stringify(response)?.slice(0, 800);
+            throw new Error(`${context}: Unexpected response shape (no choices). Preview: ${preview}`);
+        }
+        return response;
+    }
+    catch (err) {
+        // Re-throw after logging for upstream handling
+        coreExports.error(`${context}: chatCompletion failed: ${err}`);
+        throw err;
+    }
+}
 
 /**
  * Helper function to load content from a file or use fallback input
@@ -49674,12 +49689,14 @@ function buildResponseFormat(promptConfig) {
 /**
  * Build complete InferenceRequest from prompt config and inputs
  */
-function buildInferenceRequest(promptConfig, systemPrompt, prompt, modelName, maxTokens, endpoint, token) {
+function buildInferenceRequest(promptConfig, systemPrompt, prompt, modelName, temperature, topP, maxTokens, endpoint, token) {
     const messages = buildMessages(promptConfig, systemPrompt, prompt);
     const responseFormat = buildResponseFormat(promptConfig);
     return {
         messages,
         modelName,
+        temperature,
+        topP,
         maxTokens,
         endpoint,
         token,
@@ -52561,10 +52578,8 @@ function loadPromptFile(filePath, templateVariables = {}) {
         throw new Error(`Prompt file not found: ${filePath}`);
     }
     const fileContent = fs.readFileSync(filePath, 'utf-8');
-    // Apply template variable substitution
-    const processedContent = replaceTemplateVariables(fileContent, templateVariables);
     try {
-        const config = load(processedContent);
+        const config = load(fileContent);
         if (!config.messages || !Array.isArray(config.messages)) {
             throw new Error('Prompt file must contain a "messages" array');
         }
@@ -52577,6 +52592,13 @@ function loadPromptFile(filePath, templateVariables = {}) {
                 throw new Error(`Invalid message role: ${message.role}`);
             }
         }
+        // Prepare messages by replacing template variables with actual content
+        config.messages = config.messages.map(msg => {
+            return {
+                ...msg,
+                content: replaceTemplateVariables(msg.content, templateVariables),
+            };
+        });
         return config;
     }
     catch (error) {
@@ -52596,9 +52618,6 @@ function isPromptYamlFile(filePath) {
  * @returns Resolves when the action is complete.
  */
 async function run() {
-    let responseFile = null;
-    // Set up graceful cleanup for temporary files on process exit
-    tmpExports.setGracefulCleanup();
     try {
         const promptFilePath = coreExports.getInput('prompt-file');
         const inputVariables = coreExports.getInput('input');
@@ -52624,7 +52643,10 @@ async function run() {
         }
         // Get common parameters
         const modelName = promptConfig?.model || coreExports.getInput('model');
-        const maxTokens = parseInt(coreExports.getInput('max-tokens'), 10);
+        let maxTokens = promptConfig?.modelParameters?.maxTokens ?? coreExports.getInput('max-tokens');
+        if (typeof maxTokens === 'string') {
+            maxTokens = parseInt(maxTokens, 10);
+        }
         const token = process.env['GITHUB_TOKEN'] || coreExports.getInput('token');
         if (token === undefined) {
             throw new Error('GITHUB_TOKEN is not set');
@@ -52633,7 +52655,7 @@ async function run() {
         const githubMcpToken = coreExports.getInput('github-mcp-token') || token;
         const endpoint = coreExports.getInput('endpoint');
         // Build the inference request with pre-processed messages and response format
-        const inferenceRequest = buildInferenceRequest(promptConfig, systemPrompt, prompt, modelName, maxTokens, endpoint, token);
+        const inferenceRequest = buildInferenceRequest(promptConfig, systemPrompt, prompt, modelName, promptConfig?.modelParameters?.temperature, promptConfig?.modelParameters?.topP, maxTokens, endpoint, token);
         const enableMcp = coreExports.getBooleanInput('enable-github-mcp') || false;
         let modelResponse = null;
         if (enableMcp) {
@@ -52650,10 +52672,13 @@ async function run() {
             modelResponse = await simpleInference(inferenceRequest);
         }
         coreExports.setOutput('response', modelResponse || '');
-        // Create a secure temporary file instead of using the temp directory directly
-        responseFile = tmpExports.fileSync({
+        // Create a temporary file for the response that persists for downstream steps.
+        // We use keep: true to prevent automatic cleanup - the file will be cleaned up
+        // by the runner when the job completes.
+        const responseFile = tmpExports.fileSync({
             prefix: 'modelResponse-',
             postfix: '.txt',
+            keep: true,
         });
         coreExports.setOutput('response-file', responseFile.name);
         if (modelResponse && modelResponse !== '') {
@@ -52670,18 +52695,6 @@ async function run() {
         // Force exit to prevent hanging on open connections
         process.exit(1);
     }
-    finally {
-        // Explicit cleanup of temporary file if it was created
-        if (responseFile) {
-            try {
-                responseFile.removeCallback();
-            }
-            catch (cleanupError) {
-                // Log cleanup errors but don't fail the action
-                coreExports.warning(`Failed to cleanup temporary file: ${cleanupError}`);
-            }
-        }
-    }
     // Force exit to prevent hanging on open connections
     process.exit(0);
 }

eslint.config.mjs

@@ -19,7 +19,7 @@ const compat = new FlatCompat({
 
 export default [
   {
-    ignores: ['**/coverage', '**/dist', '**/linter', '**/node_modules'],
+    ignores: ['**/coverage', '**/dist', '**/linter', '**/node_modules', 'script/**'],
   },
   ...compat.extends(
     'eslint:recommended',

script/mock-inference-server.mjs

@@ -0,0 +1,71 @@
+#!/usr/bin/env node
+/**
+ * A simple mock OpenAI-compatible inference server for CI testing.
+ * This returns predictable responses without needing real API credentials.
+ */
+
+import http from 'http'
+
+const PORT = process.env.MOCK_SERVER_PORT || 3456
+
+const server = http.createServer((req, res) => {
+  let body = ''
+
+  req.on('data', chunk => {
+    body += chunk.toString()
+  })
+
+  req.on('end', () => {
+    console.log(`[Mock Server] ${req.method} ${req.url}`)
+
+    // Handle chat completions endpoint
+    if (req.url === '/chat/completions' && req.method === 'POST') {
+      const request = JSON.parse(body)
+      const userMessage = request.messages?.find(m => m.role === 'user')?.content || 'No prompt'
+
+      const response = {
+        id: 'mock-completion-id',
+        object: 'chat.completion',
+        created: Date.now(),
+        model: request.model || 'mock-model',
+        choices: [
+          {
+            index: 0,
+            message: {
+              role: 'assistant',
+              content: `Mock response to: "${userMessage.slice(0, 50)}..."`,
+            },
+            finish_reason: 'stop',
+          },
+        ],
+        usage: {
+          prompt_tokens: 10,
+          completion_tokens: 20,
+          total_tokens: 30,
+        },
+      }
+
+      res.writeHead(200, {'Content-Type': 'application/json'})
+      res.end(JSON.stringify(response))
+      return
+    }
+
+    // Health check endpoint
+    if (req.url === '/health' || req.url === '/') {
+      res.writeHead(200, {'Content-Type': 'application/json'})
+      res.end(JSON.stringify({status: 'ok'}))
+      return
+    }
+
+    // 404 for unknown routes
+    res.writeHead(404, {'Content-Type': 'application/json'})
+    res.end(JSON.stringify({error: 'Not found'}))
+  })
+})
+
+server.listen(PORT, () => {
+  console.log(`[Mock Server] Listening on http://localhost:${PORT}`)
+  console.log('[Mock Server] Endpoints:')
+  console.log('  POST /chat/completions - Mock chat completion')
+  console.log('  GET  /health           - Health check')
+})

src/helpers.ts

@@ -82,6 +82,8 @@ export function buildInferenceRequest(
   systemPrompt: string | undefined,
   prompt: string | undefined,
   modelName: string,
+  temperature: number | undefined,
+  topP: number | undefined,
   maxTokens: number,
   endpoint: string,
   token: string,
@@ -92,6 +94,8 @@ export function buildInferenceRequest(
   return {
     messages,
     modelName,
+    temperature,
+    topP,
     maxTokens,
     endpoint,
     token,

src/inference.ts

@@ -15,6 +15,8 @@ export interface InferenceRequest {
   maxTokens: number
   endpoint: string
   token: string
+  temperature?: number
+  topP?: number
   responseFormat?: {type: 'json_schema'; json_schema: unknown} // Processed response format for the API
 }
 
@@ -45,6 +47,8 @@ export async function simpleInference(request: InferenceRequest): Promise<string
     messages: request.messages as OpenAI.Chat.Completions.ChatCompletionMessageParam[],
     max_tokens: request.maxTokens,
     model: request.modelName,
+    temperature: request.temperature,
+    top_p: request.topP,
   }
 
   // Add response format if specified
@@ -53,21 +57,10 @@ export async function simpleInference(request: InferenceRequest): Promise<string
     chatCompletionRequest.response_format = request.responseFormat as any
   }
 
-  try {
-    const response = await client.chat.completions.create(chatCompletionRequest)
-
-    if ('choices' in response) {
-      const modelResponse = response.choices[0]?.message?.content
-      core.info(`Model response: ${modelResponse || 'No response content'}`)
-      return modelResponse || null
-    } else {
-      core.error(`Unexpected response format from API: ${JSON.stringify(response)}`)
-      return null
-    }
-  } catch (error) {
-    core.error(`API error: ${error}`)
-    throw error
-  }
+  const response = await chatCompletion(client, chatCompletionRequest, 'simpleInference')
+  const modelResponse = response.choices[0]?.message?.content
+  core.info(`Model response: ${modelResponse || 'No response content'}`)
+  return modelResponse || null
 }
 
 /**
@@ -101,6 +94,8 @@ export async function mcpInference(
       messages: messages as OpenAI.Chat.Completions.ChatCompletionMessageParam[],
       max_tokens: request.maxTokens,
       model: request.modelName,
+      temperature: request.temperature,
+      top_p: request.topP,
     }
 
     // Add response format if specified (only on final iteration to avoid conflicts with tool calls)
@@ -112,11 +107,7 @@ export async function mcpInference(
     }
 
     try {
-      const response = await client.chat.completions.create(chatCompletionRequest)
-
-      if (!('choices' in response)) {
-        throw new Error(`Unexpected response format from API: ${JSON.stringify(response)}`)
-      }
+      const response = await chatCompletion(client, chatCompletionRequest, `mcpInference iteration ${iterationCount}`)
 
       const assistantMessage = response.choices[0]?.message
       const modelResponse = assistantMessage?.content
@@ -133,20 +124,13 @@ export async function mcpInference(
       if (!toolCalls || toolCalls.length === 0) {
         core.info('No tool calls requested, ending GitHub MCP inference loop')
 
-        // If we have a response format set and we haven't explicitly run one final message iteration,
-        // do another loop with the response format set
         if (request.responseFormat && !finalMessage) {
           core.info('Making one more MCP loop with the requested response format...')
-
-          // Add a user message requesting JSON format and try again
           messages.push({
             role: 'user',
             content: `Please provide your response in the exact ${request.responseFormat.type} format specified.`,
           })
-
           finalMessage = true
-
-          // Continue the loop to get a properly formatted response
           continue
         } else {
           return modelResponse || null
@@ -154,13 +138,8 @@ export async function mcpInference(
       }
 
       core.info(`Model requested ${toolCalls.length} tool calls`)
-
-      // Execute all tool calls via GitHub MCP
       const toolResults = await executeToolCalls(githubMcpClient.client, toolCalls as ToolCall[])
-
-      // Add tool results to the conversation
       messages.push(...toolResults)
-
       core.info('Tool results added, continuing conversation...')
     } catch (error) {
       core.error(`OpenAI API error: ${error}`)
@@ -178,3 +157,43 @@ export async function mcpInference(
 
   return lastAssistantMessage?.content || null
 }
+
+/**
+ * Wrapper around OpenAI chat.completions.create with defensive handling for cases where
+ * the SDK returns a raw string (e.g., unexpected content-type or streaming body) instead of
+ * a parsed object. Ensures an object with a 'choices' array is returned or throws a descriptive error.
+ */
+async function chatCompletion(
+  client: OpenAI,
+  params: OpenAI.Chat.Completions.ChatCompletionCreateParams,
+  context: string,
+): Promise<OpenAI.Chat.Completions.ChatCompletion> {
+  try {
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    let response: any = await client.chat.completions.create(params)
+    core.debug(`${context}: raw response typeof=${typeof response}`)
+
+    if (typeof response === 'string') {
+      // Attempt to parse if we unexpectedly received a string
+      try {
+        response = JSON.parse(response)
+      } catch (e) {
+        const preview = response.slice(0, 400)
+        throw new Error(
+          `${context}: Chat completion response was a string and not valid JSON (${(e as Error).message}). Preview: ${preview}`,
+        )
+      }
+    }
+
+    if (!response || typeof response !== 'object' || !('choices' in response)) {
+      const preview = JSON.stringify(response)?.slice(0, 800)
+      throw new Error(`${context}: Unexpected response shape (no choices). Preview: ${preview}`)
+    }
+
+    return response as OpenAI.Chat.Completions.ChatCompletion
+  } catch (err) {
+    // Re-throw after logging for upstream handling
+    core.error(`${context}: chatCompletion failed: ${err}`)
+    throw err
+  }
+}

src/main.ts

@@ -18,11 +18,6 @@ import {
  * @returns Resolves when the action is complete.
  */
 export async function run(): Promise<void> {
-  let responseFile: tmp.FileResult | null = null
-
-  // Set up graceful cleanup for temporary files on process exit
-  tmp.setGracefulCleanup()
-
   try {
     const promptFilePath = core.getInput('prompt-file')
     const inputVariables = core.getInput('input')
@@ -53,7 +48,11 @@ export async function run(): Promise<void> {
 
     // Get common parameters
     const modelName = promptConfig?.model || core.getInput('model')
-    const maxTokens = parseInt(core.getInput('max-tokens'), 10)
+    let maxTokens = promptConfig?.modelParameters?.maxTokens ?? core.getInput('max-tokens')
+
+    if (typeof maxTokens === 'string') {
+      maxTokens = parseInt(maxTokens, 10)
+    }
 
     const token = process.env['GITHUB_TOKEN'] || core.getInput('token')
     if (token === undefined) {
@@ -71,6 +70,8 @@ export async function run(): Promise<void> {
       systemPrompt,
       prompt,
       modelName,
+      promptConfig?.modelParameters?.temperature,
+      promptConfig?.modelParameters?.topP,
       maxTokens,
       endpoint,
       token,
@@ -95,10 +96,13 @@ export async function run(): Promise<void> {
 
     core.setOutput('response', modelResponse || '')
 
-    // Create a secure temporary file instead of using the temp directory directly
-    responseFile = tmp.fileSync({
+    // Create a temporary file for the response that persists for downstream steps.
+    // We use keep: true to prevent automatic cleanup - the file will be cleaned up
+    // by the runner when the job completes.
+    const responseFile = tmp.fileSync({
       prefix: 'modelResponse-',
       postfix: '.txt',
+      keep: true,
     })
 
     core.setOutput('response-file', responseFile.name)
@@ -114,16 +118,6 @@ export async function run(): Promise<void> {
     }
     // Force exit to prevent hanging on open connections
     process.exit(1)
-  } finally {
-    // Explicit cleanup of temporary file if it was created
-    if (responseFile) {
-      try {
-        responseFile.removeCallback()
-      } catch (cleanupError) {
-        // Log cleanup errors but don't fail the action
-        core.warning(`Failed to cleanup temporary file: ${cleanupError}`)
-      }
-    }
   }
 
   // Force exit to prevent hanging on open connections

src/prompt.ts

@@ -7,9 +7,16 @@ export interface PromptMessage {
   content: string
 }
 
+export interface ModelParameters {
+  maxTokens?: number
+  temperature?: number
+  topP?: number
+}
+
 export interface PromptConfig {
   messages: PromptMessage[]
   model?: string
+  modelParameters?: ModelParameters
   responseFormat?: 'text' | 'json_schema'
   jsonSchema?: string
 }
@@ -101,11 +108,8 @@ export function loadPromptFile(filePath: string, templateVariables: TemplateVari
 
   const fileContent = fs.readFileSync(filePath, 'utf-8')
 
-  // Apply template variable substitution
-  const processedContent = replaceTemplateVariables(fileContent, templateVariables)
-
   try {
-    const config = yaml.load(processedContent) as PromptConfig
+    const config = yaml.load(fileContent) as PromptConfig
 
     if (!config.messages || !Array.isArray(config.messages)) {
       throw new Error('Prompt file must contain a "messages" array')
@@ -121,6 +125,14 @@ export function loadPromptFile(filePath: string, templateVariables: TemplateVari
       }
     }
 
+    // Prepare messages by replacing template variables with actual content
+    config.messages = config.messages.map(msg => {
+      return {
+        ...msg,
+        content: replaceTemplateVariables(msg.content, templateVariables),
+      }
+    })
+
     return config
   } catch (error) {
     throw new Error(`Failed to parse prompt file: ${error instanceof Error ? error.message : 'Unknown error'}`)