AGENTS.md is the repository's contributor guidance. Backend code lives in packages/api
(TypeScript) and packages/data-schemas (database methods); api is legacy Express wiring.
Shared API types and services live in packages/data-provider, and the React app lives in
client with shared primitives in packages/client.
Normally branch off dev and target dev; gh pr create defaults to main, so pass --base dev
explicitly. main is the released branch, kept as a fast-forward of dev and synced as-is — never
open a backport pull request to main, because anything merged to dev reaches it at the next
sync. Pull requests opened against main are retargeted automatically.
Maintainer-directed canary exception: Experimental work, or a PR ready to merge after review but
not yet suitable for the next main sync, may instead target canary. The maintainer may choose
this before work starts or while reviewing a stale PR. For new canary work, branch from the current
origin/canary and pass --base canary explicitly; do not silently retarget an existing PR or
promote canary code to dev. If a stale PR is redirected to canary, first check its base, diff and
reviewed head against current canary; coordinate any rebase or new PR with the maintainer. A PR
explicitly based on canary stays there; the main-to-dev retarget workflow does not move it.
Still link related issues in the PR (for example, Related to #N) so the work remains
traceable. Fixes #N does not close an issue on a dev or canary merge — GitHub honors closing
keywords only on the default branch. Close resolved issues by hand after merging. Worktrees share
one stash stack, so never use a bare git stash pop. Prefer a WIP commit; if a stash is necessary,
apply the specific tagged entry. The target: main label and release-bound upstream branches are
exempt from the main-to-dev retarget workflow; do not use either to backport ordinary work.
Write the description for a reader who has not followed the branch: what breaks, what triggers it,
how it behaves after the change, then one or two views of the mechanism — a focused diff, a call
tree, a shallow file tree, or a Mermaid sequence. Keep only what the change carries, and describe
the code as it stands rather than narrating earlier commits or review rounds. Naming the merged
pull request that caused the bug is not the same thing; that is history the reader needs. The
formats and examples live in .github/pull_request_template.md.
Read the inline review threads themselves — a summary comment or notification list omits findings.
Audit each one against the current code, fix what is valid, and reject what is obsolete in a reply
that says why. After each round: focused tests, npx tsc --noEmit in every workspace you changed,
push, then request the next review naming the pull request's exact remote head — a clean review of
an earlier head says nothing about what you just pushed, and CI runs on its own clock. Reply on
each resolved thread with the resolving commit and evidence. After two actionable rounds, stop
patching thread by thread and read the subsystem by invariant instead: identity, authorization,
persistence, retry, cancellation, cleanup and mixed-version behavior where relevant. Which reviewer
and what phrase triggers it will change; that the review must cover the exact pushed head will not.
A clean review is one completion signal, not the definition of done. Ship the observable experience — loading, empty, success, failure, cancellation, retry, restored session — with strings localized, accessibility intact, defaults and stored data preserved, and no backend capability left without a frontend entry point. Keep fixes small and test the missed behavior. Report the pushed head, what you ran locally, CI state, the review result at that head, any finding you rejected with the reasoning, and checks you could not run.
For startup, auth, config, file, or message-loading changes, avoid serial database
reads and reuse loaded request data. Run npm run lighthouse before completion:
the CI lane adds 250 ms per Mongo query and checks the visible conversation's LCP.
See budgets, reproduction and failure diagnosis.
A green build is not a typecheck: packages/api, packages/client and packages/data-schemas build
with tsdown, which emits without checking types. Run npx tsc --noEmit in the workspace you
changed. packages/client excludes *.spec.ts(x) and *.test.ts(x) from typechecking entirely.
npm run sort-imports with no arguments rewrites every source root — pass the paths you touched.
Run npm run static-checks -- --against origin/dev to reproduce the PR's static checks; use
npm run static-checks:full for the slower gates. Fix all formatting, lint, and TypeScript
warnings/errors in the code you change.
All new backend behavior belongs in TypeScript under packages/api; database-specific shared logic
belongs in packages/data-schemas, and frontend/backend shared API logic belongs in
packages/data-provider. Build data-provider from the project root with npm run build:data-provider.
/api holds wiring, not behavior. When a change would add logic to a CJS file there — a branch, a
helper, a validation step, a service call — the logic belongs in packages/api, and the JS file
keeps requires, route registration and the call into the TS module (MCPRequestContext.js is the
shape, thirteen lines of re-export). "Minimum" means how much behavior /api gains, not how small
the diff is, and the rule applies to editing existing CJS, which is the common case.
Database contracts belong to packages/data-schemas. Keep Mongoose types (FilterQuery,
Types.ObjectId, Document) out of exported signatures in packages/api, packages/data-provider
and client, because they make the storage engine part of that module's public API. Take and return
plain typed objects and express the query behind a data-schemas method. The boundary already leaks
across packages/api, so stop widening it rather than rewriting what exists; the client carries none
of it and must stay that way.
New levers ship configurable: a limit, timeout, toggle or capability introduced in code earns a field
on configSchema (packages/data-provider/src/config.ts) so it can be set in librechat.yaml, with
a default that reproduces today's behavior. Hard-coded constants and env-only switches need a reason.
Modules take their dependencies rather than reaching for them: code in packages/api receives its
config, database methods and clients from the caller, the way createModels(mongoose) receives the
app's connection, instead of importing app singletons or reading global state. Integrations (provider
SDKs, storage backends, vector stores, OAuth servers) arrive through an interface the caller
supplies, so a second implementation is a new argument instead of a new branch. The static singletons
under packages/api/src/mcp are the shape to stop extending, not a pattern to copy. This is the
backend half of client state ownership: pass it in, do not reach for it.
This section applies to backend services (packages/api) and database methods/repositories
(packages/data-schemas), not to React components. For new or substantially changed code in
those layers, choose failure behavior at the contract boundary instead of treating every
unsuccessful outcome as an exception:
- Return a plain value on success. Use
nullonly for a documented absence (for example, a successful lookup with no record), never to hide a query failure. - For an expected failure the caller can handle without aborting the operation, prefer a typed
discriminated result such as
{ ok: true; value: T } | { ok: false; error: { code: string; message?: string } }. Keep an existing domain-specific result shape when changing it would break callers; do not introduce interchangeableok,valid, and bare{ message }contracts in the same service. Codes should be stable, machine-readable identifiers when the caller needs to distinguish failures. Internal validation helpers may use the existing local pattern. - Throw for unexpected operational failures (database, provider, network) and violated invariants.
Use a typed/coded error when a boundary must recognize a particular failure. Catch where you can
recover or translate it; otherwise let it reach the owning request or job boundary. Do not catch
an exception and return a truthy record,
null, or{ message }in its place: callers can then mistake an outage for data or send it as an HTTP 200. Keep separate migrations for existing contracts, with their callers and tests, rather than changing wire shapes incidentally. - At the HTTP, stream, or job boundary, map failures to the existing status and response contract.
Only expose intentionally approved codes and helpful user-safe information. Error disclosure is
a security boundary: never forward raw exceptions, stacks, query text, credentials, request
bodies, headers, provider payloads, or arbitrary
error.messageto a client or persisted user-visible status. Sanitize diagnostics and log only what the path can safely retain. For errors that may echo submitted content, prefergetSafeErrorMetadataover error text;packages/api/src/utils/errors.tsdocuments when redacted error text is appropriate. - Localization is a UI responsibility, not a database-method return shape. When a failure reaches
a user-facing surface, expose a safe, stable code the UI can map to localized, actionable copy
and a fallback. See
client/src/components/Messages/Content/Error/registry.tsandclient/src/components/Chat/Trace/Viewer.tsx. Keep any intentionally disclosed provider detail behind its existing protection policy rather than treating arbitrary upstream text as safe.
Examples of local contracts, not a mandate to copy their shapes: packages/api/src/admin/auditLog.ts
uses a parsing result, packages/api/src/agents/openai/service.ts validates before mapping to an
OpenAI-compatible response, and packages/api/src/skills/sync/errors.ts names a sync failure by
stable code. packages/data-schemas/src/methods/prompt.ts currently returns { message } on query
failure; do not extend that pattern. Test absence versus failure, the boundary's non-success response,
and that sensitive diagnostics cannot reach a user-visible surface.
For frontend work, compose existing @librechat/client primitives and variants before adding
feature-local styles. Use semantic theme/Tailwind roles for color and shared appearance; do not
introduce raw palette utilities, hard-coded colors, or arbitrary theme CSS. If the system cannot
express a reusable design need, deepen the shared primitive or versioned theme-token registry
instead of copying classes into a feature. Theme definitions select semantic colors and appearance,
not selectors, app behavior, or alternate layouts. Keep genuine layout and behavior local, and
explain why any new custom CSS cannot use shared primitives. Support light/dark and reduced motion;
preserve defaults and test a deliberately different reference theme for reusable new variants.
When adding or changing code that mutates user documents, invalidate the auth user document cache
for affected users, including bulk role and user mutations. Otherwise burst-cached JWT requests can
serve a stale req.user until the cache expires.
The client is migrating from Recoil to Jotai. New state is always Jotai, even in a file that already
imports Recoil; many files import both, so mixed imports say nothing about which to use. For existing
state the unit of conversion is one atom plus every file that reads or writes it, because an atom
cannot be half converted — convert the areas you touch, not the whole store.
Split by ownership: state a feature both writes and reads is feature-owned, so convert it to Jotai
and keep it inside the feature; app-global preferences and shell state a feature merely consumes
(maximizeChatSpace, showScrollButton, enterToSend, artifact visibility) must be passed in
through props or a small host-supplied context rather than reached for through ~/store; when a
consumer sits outside the feature you are changing, leave that atom on Recoil and pass it in. Passing
them in is what lets a feature move to its own workspace later without a rewrite, and it keeps the
Jotai conversion scoped to the state a feature owns. For persisted atoms, use the helpers in
client/src/store/jotai-utils.ts.
Use short, single-word file names when possible; group related modules in a single-word directory.
Prefer flat early returns, pure functions, and explicit types. Avoid any, broad unknown/casts,
duplicated types, and dynamic imports. Extract genuinely repeated logic instead of copying it;
write comments only for non-obvious behavior or public API contracts.
Imports have three sections: package values (shortest line first, with react first), import type
(longest line first, package types before local types), and local/project values (longest line first).
Use standalone import type { ... }, not inline type within a value import. Run the scoped
import sorter on files you change.
Avoid extra passes over shared message arrays and unnecessary allocations. For startup and request paths, reuse already-loaded user/config data, avoid serial database reads, and start independent reads in parallel. Do not weaken authorization or tenant checks or write a response before those checks succeed.
Use useLocalize() for all visible copy and update only English keys in
client/src/locales/en/translation.json. Use semantic HTML, keyboard behavior, and ARIA labels.
Use React Query for API interactions and invalidate related queries after mutations; define query
and mutation keys in packages/data-provider/src/keys.ts. Put shared endpoint definitions, types,
and data services in packages/data-provider, and encode dynamic URL parameters. Cover loading,
success, and failure states in focused component tests.
Run focused Jest tests from the owning workspace, not the entire monorepo. Prefer real logic and
spies over mocked internals; use mongodb-memory-server for database queries and the real MCP SDK
for MCP behavior. Mock only external HTTP APIs or uncontrollable services. For frontend tests use
the component's colocated __tests__ and test/layout-test-utils where applicable.