keystone-cli 0.5.0 → 0.5.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "keystone-cli",
-  "version": "0.5.0",
+  "version": "0.5.1",
   "description": "A local-first, declarative, agentic workflow orchestrator built on Bun",
   "type": "module",
   "bin": {
@@ -13,7 +13,13 @@
     "lint:fix": "biome check --write .",
     "format": "biome format --write ."
   },
-  "keywords": ["workflow", "orchestrator", "agentic", "automation", "bun"],
+  "keywords": [
+    "workflow",
+    "orchestrator",
+    "agentic",
+    "automation",
+    "bun"
+  ],
   "author": "Mark Hingston",
   "license": "MIT",
   "repository": {
@@ -21,7 +27,12 @@
     "url": "https://github.com/mhingston/keystone-cli.git"
   },
   "homepage": "https://github.com/mhingston/keystone-cli#readme",
-  "files": ["src", "README.md", "LICENSE", "logo.png"],
+  "files": [
+    "src",
+    "README.md",
+    "LICENSE",
+    "logo.png"
+  ],
   "dependencies": {
     "@jsep-plugin/arrow": "^1.0.6",
     "@jsep-plugin/object": "^1.2.2",
@@ -46,4 +57,4 @@
   "engines": {
     "bun": ">=1.0.0"
   }
-}
+}

package/src/runner/llm-executor.test.ts

@@ -132,13 +132,13 @@ describe('llm-executor', () => {
   beforeAll(() => {
     // Mock spawn to avoid actual process creation
     const mockProcess = Object.assign(new EventEmitter(), {
-      stdout: new Readable({ read() {} }),
+      stdout: new Readable({ read() { } }),
       stdin: new Writable({
         write(_chunk, _encoding, cb: (error?: Error | null) => void) {
           cb();
         },
       }),
-      kill: mock(() => {}),
+      kill: mock(() => { }),
     });
     spawnSpy = spyOn(child_process, 'spawn').mockReturnValue(
       mockProcess as unknown as child_process.ChildProcess
@@ -266,7 +266,7 @@ You are a test agent.`;
     expect(result.output).toEqual({ foo: 'bar' });
   });
 
-  it('should throw error if JSON parsing fails for schema', async () => {
+  it('should retry if LLM output fails schema validation', async () => {
     const step: LlmStep = {
       id: 'l1',
       type: 'llm',
@@ -279,7 +279,51 @@ You are a test agent.`;
     const context: ExpressionContext = { inputs: {}, steps: {} };
     const executeStepFn = mock(async () => ({ status: 'success' as const, output: 'ok' }));
 
-    // Mock response with invalid JSON
+    const originalOpenAIChatInner = OpenAIAdapter.prototype.chat;
+    const originalCopilotChatInner = CopilotAdapter.prototype.chat;
+    const originalAnthropicChatInner = AnthropicAdapter.prototype.chat;
+
+    let attempt = 0;
+    const mockChat = mock(async () => {
+      attempt++;
+      if (attempt === 1) {
+        return { message: { role: 'assistant', content: 'Not JSON' } };
+      }
+      return { message: { role: 'assistant', content: '{"success": true}' } };
+    }) as unknown as typeof originalOpenAIChat;
+
+    OpenAIAdapter.prototype.chat = mockChat;
+    CopilotAdapter.prototype.chat = mockChat;
+    AnthropicAdapter.prototype.chat = mockChat;
+
+    const result = await executeLlmStep(
+      step,
+      context,
+      executeStepFn as unknown as (step: Step, context: ExpressionContext) => Promise<StepResult>
+    );
+
+    expect(result.status).toBe('success');
+    expect(result.output).toEqual({ success: true });
+    expect(attempt).toBe(2);
+
+    OpenAIAdapter.prototype.chat = originalOpenAIChatInner;
+    CopilotAdapter.prototype.chat = originalCopilotChatInner;
+    AnthropicAdapter.prototype.chat = originalAnthropicChatInner;
+  });
+
+  it('should fail after max iterations if JSON remains invalid', async () => {
+    const step: LlmStep = {
+      id: 'l1',
+      type: 'llm',
+      agent: 'test-agent',
+      prompt: 'give me invalid json',
+      needs: [],
+      maxIterations: 3,
+      schema: { type: 'object' },
+    };
+    const context: ExpressionContext = { inputs: {}, steps: {} };
+    const executeStepFn = mock(async () => ({ status: 'success' as const, output: 'ok' }));
+
     const originalOpenAIChatInner = OpenAIAdapter.prototype.chat;
     const originalCopilotChatInner = CopilotAdapter.prototype.chat;
     const originalAnthropicChatInner = AnthropicAdapter.prototype.chat;
@@ -298,7 +342,7 @@ You are a test agent.`;
         context,
         executeStepFn as unknown as (step: Step, context: ExpressionContext) => Promise<StepResult>
       )
-    ).rejects.toThrow(/Failed to parse LLM output as JSON/);
+    ).rejects.toThrow('Max ReAct iterations reached');
 
     OpenAIAdapter.prototype.chat = originalOpenAIChatInner;
     CopilotAdapter.prototype.chat = originalCopilotChatInner;
@@ -378,7 +422,7 @@ You are a test agent.`;
       spyOn(client, 'stop').mockReturnValue(undefined);
       return client;
     });
-    const consoleSpy = spyOn(console, 'error').mockImplementation(() => {});
+    const consoleSpy = spyOn(console, 'error').mockImplementation(() => { });
 
     await executeLlmStep(
       step,
@@ -565,7 +609,7 @@ You are a test agent.`;
     };
     const context: ExpressionContext = { inputs: {}, steps: {} };
     const executeStepFn = mock(async () => ({ status: 'success' as const, output: 'ok' }));
-    const consoleSpy = spyOn(console, 'error').mockImplementation(() => {});
+    const consoleSpy = spyOn(console, 'error').mockImplementation(() => { });
 
     await executeLlmStep(
       step,

package/src/runner/llm-executor.ts

@@ -249,9 +249,14 @@ export async function executeLlmStep(
           try {
             output = extractJson(output) as typeof output;
           } catch (e) {
-            throw new Error(
-              `Failed to parse LLM output as JSON matching schema: ${e instanceof Error ? e.message : String(e)}\nOutput: ${output}`
-            );
+            const errorMessage = `Failed to parse LLM output as JSON matching schema: ${e instanceof Error ? e.message : String(e)}`;
+            logger.error(`  ⚠️  ${errorMessage}. Retrying...`);
+
+            messages.push({
+              role: 'user',
+              content: `Error: ${errorMessage}\n\nPlease correct your output to be valid JSON matching the schema.`,
+            });
+            continue;
           }
         }