librechat-mnemonic AGENTS.md

Agent Guidelines

Workflow

1. Read before writing. Understand the existing code, tests, and architecture before making changes. Start with README.md, then the files relevant to the task.
2. Make focused changes. Do not reformat unrelated code, rename variables, or "improve" things outside the scope of the task.
3. Add or update tests. Every behavioural change should have corresponding test coverage. Run npm test before committing.
4. Run validation. Always run npm test and npm run typecheck before pushing. If either fails, fix the issue before committing.
5. Commit checkpoints. For longer tasks, commit and push coherent checkpoints rather than waiting until the end. Each commit should leave the codebase in a working state.
6. Do not merge. Create or update a pull request targeting main. Let a human review and merge.

Code Style

• Follow the existing patterns in the codebase. Do not introduce new dependencies without justification.
• Keep comments concise and explain why, not what.
• Use the existing logger (pino) for all logging. Never console.log.
• Environment-driven configuration via zod schemas. No config files.

Testing

• Tests use vitest. Run npm test for the full suite.
• Mock external dependencies (mnemonic, MongoDB, HTTP) — never hit real services in tests.
• Each test file should be self-contained with its own fixtures.

Architecture Notes

• src/proxy/handler.ts — the request path; where memory recall and injection happen.
• src/memory/service.ts — the shared core for both the proxy and MCP endpoint.
• src/mnemonic/client.ts — the MCP client with read/write queue separation.
• src/cache.ts — TTL cache with hit/miss tracking.
• src/telemetry.ts — Langfuse telemetry (noop when credentials are missing).
• src/config.ts — all configuration, env-driven via zod.