feat(compression): add failed_approaches and critical_details schema sections

Prevents context rot by:
- Tracking failed approaches to avoid retry loops
- Preserving verbatim error messages, ports, URLs, and other critical details

packages/core/src/prompts/snippets.test.ts

@@ -0,0 +1,122 @@
+/**
+ * @license
+ * Copyright 2025 Google LLC
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import { describe, it, expect } from 'vitest';
+import { getCompressionPrompt } from './snippets.js';
+
+describe('getCompressionPrompt', () => {
+  it('should return a non-empty compression prompt', () => {
+    const prompt = getCompressionPrompt();
+    expect(prompt).toBeTruthy();
+    expect(prompt.length).toBeGreaterThan(0);
+  });
+
+  it('should contain the state_snapshot XML structure', () => {
+    const prompt = getCompressionPrompt();
+    expect(prompt).toContain('<state_snapshot>');
+    expect(prompt).toContain('</state_snapshot>');
+  });
+
+  it('should contain all required sections in the schema', () => {
+    const prompt = getCompressionPrompt();
+
+    // Core sections that must be present
+    const requiredSections = [
+      '<overall_goal>',
+      '</overall_goal>',
+      '<active_constraints>',
+      '</active_constraints>',
+      '<key_knowledge>',
+      '</key_knowledge>',
+      '<artifact_trail>',
+      '</artifact_trail>',
+      '<file_system_state>',
+      '</file_system_state>',
+      '<recent_actions>',
+      '</recent_actions>',
+      '<task_state>',
+      '</task_state>',
+    ];
+
+    for (const section of requiredSections) {
+      expect(prompt).toContain(section);
+    }
+  });
+
+  it('should contain failed_approaches section to prevent retry loops', () => {
+    const prompt = getCompressionPrompt();
+
+    expect(prompt).toContain('<failed_approaches>');
+    expect(prompt).toContain('</failed_approaches>');
+    // Should explain the purpose
+    expect(prompt).toMatch(/failed|did not work|prevent.*retry/i);
+  });
+
+  it('should contain critical_details section for verbatim technical info', () => {
+    const prompt = getCompressionPrompt();
+
+    expect(prompt).toContain('<critical_details>');
+    expect(prompt).toContain('</critical_details>');
+    // Should emphasize verbatim/exact preservation
+    expect(prompt).toMatch(/verbatim|exact|must not be paraphrased/i);
+  });
+
+  it('should contain security rules against prompt injection', () => {
+    const prompt = getCompressionPrompt();
+
+    expect(prompt).toContain('CRITICAL SECURITY RULE');
+    expect(prompt).toMatch(/ignore.*commands|ignore.*instructions/i);
+    expect(prompt).toContain('prompt injection');
+  });
+
+  it('should instruct the model to use a scratchpad for reasoning', () => {
+    const prompt = getCompressionPrompt();
+
+    expect(prompt).toContain('<scratchpad>');
+  });
+
+  it('should provide concrete examples for each section', () => {
+    const prompt = getCompressionPrompt();
+
+    // Check that examples are provided (indicated by "Example:" comments)
+    const exampleCount = (prompt.match(/Example:/g) || []).length;
+    expect(exampleCount).toBeGreaterThanOrEqual(5);
+  });
+
+  it('should emphasize density and brevity', () => {
+    const prompt = getCompressionPrompt();
+
+    expect(prompt).toMatch(/dense|concise|omit.*filler/i);
+  });
+
+  it('failed_approaches examples should include error context', () => {
+    const prompt = getCompressionPrompt();
+
+    // The failed_approaches section should demonstrate including error messages
+    // and explanations of why things failed
+    const failedApproachesSection = prompt.match(
+      /<failed_approaches>[\s\S]*?<\/failed_approaches>/,
+    );
+    expect(failedApproachesSection).toBeTruthy();
+
+    const section = failedApproachesSection![0];
+    // Should show examples with error messages or failure reasons
+    expect(section).toMatch(/error|failed|incompatible|EADDRINUSE/i);
+  });
+
+  it('critical_details examples should include specific technical values', () => {
+    const prompt = getCompressionPrompt();
+
+    const criticalDetailsSection = prompt.match(
+      /<critical_details>[\s\S]*?<\/critical_details>/,
+    );
+    expect(criticalDetailsSection).toBeTruthy();
+
+    const section = criticalDetailsSection![0];
+    // Should include examples of ports, URLs, error messages, env vars
+    expect(section).toMatch(/localhost:\d+|https?:\/\/|env.*var/i);
+  });
+});

packages/core/src/prompts/snippets.ts

@@ -568,6 +568,28 @@ The structure MUST be as follows:
         -->
     </key_knowledge>
 
+    <failed_approaches>
+        <!-- CRITICAL: Approaches that were tried and DID NOT WORK. This prevents the agent from retrying failed solutions. -->
+        <!-- Include: what was attempted, why it failed, and any error messages. -->
+        <!-- Example:
+         - Tried \`fs.readFileSync\` in async handler → caused event loop blocking, switched to \`fs.promises.readFile\`.
+         - Attempted to use port 3000 → EADDRINUSE, another process is using it. Using port 3001 instead.
+         - Installed \`lodash\` v3 → incompatible with ESM, upgraded to v4.
+        -->
+    </failed_approaches>
+
+    <critical_details>
+        <!-- Verbatim technical details that MUST NOT be paraphrased or approximated. Copy these exactly. -->
+        <!-- Include: specific error messages still relevant, exact URLs/ports, environment variables, version numbers. -->
+        <!-- Example:
+         - Error: "TypeError: Cannot read properties of undefined (reading 'id')" at src/auth.ts:45
+         - Dev server: http://localhost:3001
+         - API endpoint: https://api.example.com/v2/users
+         - Required env var: GEMINI_API_KEY
+         - Node version constraint: >=18.0.0
+        -->
+    </critical_details>
+
     <artifact_trail>
         <!-- Evolution of critical files and symbols. What was changed and WHY. Use this to track all significant code modifications and design decisions. -->
         <!-- Example: