Files
gemini-cli/evals/self_diff_review.eval.ts

54 lines
1.6 KiB
TypeScript
Raw Normal View History

feat(core): comprehensive agent self-validation and engineering mandates Major upgrade to the agent's self-validation, safety, and project integrity capabilities through five iterations of system prompt enhancements: Workflow & Quality Mandates: 1. Incremental Validation: Mandates building, linting, and testing after every significant file change to maintain a "green" state. 2. Mandatory Reproduction: Requires creating a failing test case to confirm a bug before fixing, and explicitly verifying the failure (Negative Verification). 3. Test Persistence & Locality: Requires integrating repro cases into the permanent test suite, preferably by amending existing related test files. 4. Script Discovery: Mandates identifying project-specific validation commands from configuration files (package.json, Makefile, etc.). 5. Self-Review: Mandates running `git diff` after every edit, using `--name-only` for large changes to preserve context window tokens. 6. Fast-Path Validation: Prioritizes lightweight checks (e.g., `tsc --noEmit`) for frequent feedback, reserving heavy builds for final verification. 7. Output Verification: Requires checking command output (not just exit codes) to prevent false-positives from empty test runs or hidden warnings. Semantic Integrity & Dependency Safety: 8. Global Usage Discovery: Mandates searching the entire workspace for all usages (via `grep_search`) before modifying exported symbols or APIs. 9. Dependency Integrity: Requires verifying that new imports are explicitly declared in the project's dependency manifest (e.g., package.json). 10. Configuration Sync: Mandates updating build/environment configs (tsconfig, Dockerfile, etc.) to support new file types or entry points. 11. Documentation Sync: Requires searching for and updating documentation references when public APIs or CLI interfaces change. 12. Anti-Silencing Mandate: Prohibits using `any`, `@ts-ignore`, or lint suppressions to resolve validation errors. Diagnostics, Safety & Runtime Verification: 13. Error Grounding: Mandates reading full error logs and stack traces upon failure. Includes Smart Log Navigation to prioritize the tail of large files. 14. Scope Isolation: Instructs the agent to focus only on errors introduced by its changes and ignore unrelated legacy technical debt. 15. Destructive Safety: Mandates a `git status` check before deleting files or modifying critical project configurations. 16. Non-Blocking Smoke Tests: Requires briefly running applications to verify boot stability, using background/timeout strategies for servers. Includes 15 new behavioral evaluations verifying these mandates and updated snapshots in packages/core/src/core/prompts.test.ts.
2026-02-20 14:22:54 -08:00
/**
* @license
* Copyright 2026 Google LLC
* SPDX-License-Identifier: Apache-2.0
*/
import { describe, expect } from 'vitest';
import { evalTest } from './test-helper.js';
describe('Self-Diff Review', () => {
/**
* Verifies that the agent performs a self-review immediately after an edit.
*/
evalTest('USUALLY_PASSES', {
name: 'should review changes immediately after an edit',
files: {
'src/app.ts': 'export const hello = () => "world";',
},
prompt: 'Update src/app.ts to say "hello world" instead of "world".',
assert: async (rig) => {
const toolLogs = rig.readToolLogs();
const editIndex = toolLogs.findIndex(
(log) =>
(log.toolRequest.name === 'replace' ||
log.toolRequest.name === 'write_file') &&
log.toolRequest.args.includes('src/app.ts'),
);
expect(editIndex, 'Agent should have edited src/app.ts').toBeGreaterThan(
-1,
);
// Check for git diff or read_file immediately after the edit
const reviewCall = toolLogs[editIndex + 1];
expect(
reviewCall,
'Agent should have made a call after the edit',
).toBeDefined();
const isReview =
(reviewCall.toolRequest.name === 'run_shell_command' &&
reviewCall.toolRequest.args.includes('git diff')) ||
(reviewCall.toolRequest.name === 'read_file' &&
reviewCall.toolRequest.args.includes('src/app.ts'));
expect(
isReview,
'Agent should have run git diff or read_file immediately after the edit to review its work',
).toBe(true);
},
});
});