⚙️ Config v1.3.19
Configuration changes since Config v1.3.18, included in LibreChat v0.8.9. Set the top-level version in librechat.yaml to 1.3.19 after reviewing these changes.
| Area | Changes and defaults |
|---|---|
| Remote bridge upgrade | Upgrade host and worker together: transport capability 2 is mandatory; host tools also require capability 2. Saved bindings remain protocol 1. Optional remote.recovery.routes maps stable worker owner IDs to allowlisted direct URLs. Final snapshot reassembly is bounded by maxBufferedBytes. |
| Native Google rates | endpoints.google.tokenConfig.<model> requires non-negative prompt, completion, and cacheRead rates in USD per million tokens. Explicit zero rates apply to LibreChat billing and optional context-cost display; omitted models keep built-in pricing. |
| SDK runtimes | Optional endpoints.agents.sdkRuntimes.<id> registrations select a local module or remote WebSocket runtime; roles default to ADMIN. Recovery is off unless configured. Remote defaults: connect 10000 ms, RPC 60000 ms, run 600000 ms, frames 1048576 bytes, buffering 4194304 bytes. Recovery defaults when enabled: disconnect 15000 ms, retry 250 ms, journal 67108864 bytes. |
| SDK rollout | Upgrade all replicas before binding Agents. Remote workers require a resolved secret of at least 32 characters and independent identity authorization. Recovery does not restart an in-flight graph after worker loss; worker journal retention and purge are operator-owned. |
| Background child streams | endpoints.agents.subagentActivity.generationJobs: true; generationContentBytes: 4194304 (0–8388608). The generation stream observes the child and is separate from task execution/cancellation. |
| Config cache | Deployment-owned configCache.maxBytes: 16777216 and maxEntries: 1000 bound local serialized APP_CONFIG retention. Eviction does not change effective config or permissions. |
| Agent attribution | {{LIBRECHAT_AGENT_ID}} identifies the primary persisted Agent in provider headers and the executing tool's owner in operator-defined MCP headers; plain/ephemeral chats have an empty value. |
| Theme backdrop | interface.theme.appearance.welcomeBackdrop accepts none or dots. A backdrop appears only when the build provides it; standard LibreChat keeps a plain surface without a registered implementation. |
| Background Responses | endpoints.agents.remoteApi.background.enabled: false; acknowledgementTimeoutMs: 10000 (100–60000 ms). API-key auth and transaction-capable storage are required. |
| Slack Channels | endpoints.agents.channels.enabled: false; transport: socket_mode; nativeSessions: false; lifecycleReactions: false; signatureToleranceSeconds: 300; socketReconnectDelayMs: 5000. Shared workspace deployment uses webhook transport and UI-managed app credentials. |
| Shared Slack apps | channels.slackApp.enabled: false; flowLifetimeMs: 600000; requestTimeoutMs: 10000; maxResponseBytes: 65536. Gateway defaults: ingress concurrency 32, worker concurrency 4, discovery batch 32, polling 1000 ms. |
| Channel environment and deployment | CHANNEL_SERVICE_TOKEN (at least 32 characters), CHANNEL_ID, CHANNEL_TENANT_ID, Socket Mode SLACK_APP_TOKEN, shared setup SLACK_WORKSPACES_ENABLED/SLACK_GATEWAY_URL, CHANNEL_GATEWAY_PORT, and release-matched LIBRECHAT_IMAGE_TAG. Optional Compose overlays and Helm gateway deployment do not enable themselves. |
| Tool deferral | modelSpecs.list[].deferLoading: omitted/false, true for all eligible tools, or a list of equipped tool names/capabilities. Saved Agents use tool_options.<tool>.defer_loading. Requires deferred_tools; orchestration tools remain loaded. |
| Prompt instructions | endpoints.agents.instructions.langfuse.enabled: false, separate from Langfuse prompt-catalog synchronization. |
| Browser limits | clientLimits: refresh 30000, TTS 60000, SharePoint 300000, CloudFront 15000 ms; upload concurrency 3. |
| Server budgets | timeouts.directoryMs: 30000, toolsMs: 60000, speechMs: 300000, imageGenerationMs: 300000; imageFetchMs: 30000 now also bounds blob-storage reads. |
| Lookup concurrency | permissionPrincipalLookups: 4, connectedAgentLoads: 4, mcpCatalogLookups: 4, skillNameLookups: 8, skillFileInspections: 4, resumeAgentLoads: 4; each accepts 1–32. |
| Spreadsheet parsing | spreadsheetParsing.timeoutMs: 5000 (100–60000 ms); maxConcurrent: 2 (1–8 per API process). |
| Files | fileConfig.deletionRetryLimit: 5 persists discarded-upload retry counts across reloads. fileConfig.stt.supportedMimeTypes now takes effect; configure a speech provider separately. |
| Pagination and logs | messages.maxPageSize: 1000 (1–10000); violationLog.maxEntries: 100; mcpSettings.oauthFailureRetentionMs: 86400000 (up to 30 days). |
| Interface | landingFooter: true now places the footer at page bottom; composer restores under-composer placement. messageAuthor: always also accepts whenDifferent. analyticsConsent: true may be disabled by deployments that handle consent themselves. |
| Telemetry | telemetry.stackTraceLimit accepts 1–1000 frames; omitted, preserves the process default. Service-version precedence is OTEL_SERVICE_VERSION → BUILD_COMMIT → package version. Logs and request spans carry a config digest; tool lifecycle logs omit payloads. |
| Optional favorites storage | persistence.favorites: mongo by default; postgres requires PostgreSQL 16, explicit schema migration, offline import and verification, and coordinated cutover. This does not replace MongoDB or the RAG database. |
| Override ownership | Storage, browser limits, telemetry, management API, shared Slack-app configuration, and other operator-owned fields cannot be activated through stored Base/role/group/user overrides. Inspect applied/skipped config reports. |
| Account recovery | Password-reset credentials are emailed only; they are never displayed as a fallback. Configure email delivery. |
Persistent references: config structure, file configuration, model specifications, and admin configuration.
Startup index compatibility
Background-delivery polling uses the shorter trigger_delivery_background_poll name for new indexes. Equivalent existing indexes are reused without dropping or renaming them, including their effective collation. No new operator setting or manual polling-index migration is required. This is distinct from the MCP reservation migration below; see startup compatibility guidance.
MCP name reservation migration
Tenant-aware MCP name reservations require a unique { name, tenantId } index. Legacy tenantless reservations remain global tombstones and must not be deleted or reassigned. For the targeted online procedure, block reservation/account-erasure writers, upgrade every writer, run migrate:mcp-reservations:dry-run and migrate:mcp-reservations -- --writers-upgraded, verify the replacement index, then reopen writes. Chat can stay available. See the full migration procedure; the broad tenant-index migration still requires all writers stopped.
Remote provider-text adapter
SDK 4.2.7 adds a programmatic hostServices.providerText adapter with negotiated capability providerText: 1. It requires a trusted host policy, retained admission quotas, and a worker factory that preserves the supplied protection binding; it is not a YAML opt-in. Saved-Agent model-bound content/PII restrictions remain. See SDK protection and compatibility.
Redis outage recovery
REDIS_RETRY_MAX_ATTEMPTS now defaults to 0 (unlimited), replacing 10. Explicit positive retry caps are preserved. Reconnect delays remain bounded by REDIS_RETRY_MAX_DELAY (default 3000 ms). Subscribers restore registrations without automatically replaying uncertain application writes or restarting failed Agent runs. See Redis recovery.
Google Hydrogen
gemini-4-hydrogen joins the default Google model catalog. Explicit GOOGLE_MODELS or endpoints.google.models lists must include it to expose it. Hydrogen supports LOW, MEDIUM, HIGH (default), and XHIGH thinking; MINIMAL becomes LOW. The 1,048,576-token shared window reserves up to 65,536 tokens for output, including thinking. Native Google rate overrides also apply to native Vertex AI; see Google setup.
Browser telemetry privacy
Trace export strips credential headers and omits request and response bodies for authentication and credential endpoints. Ordinary request bodies remain available with advanced capture. If capture was enabled, review retained telemetry and collector access, retire exposed tokens or keys, and remove affected data under your retention policy. See upgrade guidance.