* chore: Retire the DALL-E 3 Tool
OpenAI shut down dall-e-3 on May 12, 2026, so the built-in DALL-E 3 tool can no longer generate images on api.openai.com. Remove the tool, its manifest entry, registry definition, and env docs, along with the loadTools functions flag that only existed to register it. The OpenAI Image Tools cover the same use case. Stored agents that still list dalle load without it, and dalle stays in imageGenTools so past images keep rendering.
* test: Cover the Retired DALL-E 3 Tool in Mock Scenarios
* docs: Drop DALL-E from the README Image Generation List
* fix: Price 1-Hour Prompt Cache Writes at the 1-Hour Rate
* test: Cover 1-Hour Cache Write Pricing End to End
* test: Price Cache Scenario on the Configured Mock Model
* fix: Accept a 1-Hour Cache Write Rate in Endpoint Token Configs
* feat: Add Claude Haiku 5.5 Support
Haiku 5.5 (`claude-haiku-5-5`) matched none of the Claude family gates, so
every endpoint treated it as a legacy Claude 4 model: thinking went out as
`enabled` with a budget (400), sampling parameters were sent, thinking blocks
were not bound to the conversation prefix, and it resolved to a 200K context,
64K output and the default token rate.
Add a Haiku version parser and `hasAdaptiveHaikuThinking` (5.5 or later),
and include it in the adaptive thinking, block binding, omitted display,
sampling omission, explicit disabled, effort cap, 1M context and 128K output
gates. Unlike Sonnet 5.5, Haiku 5.5 still accepts `disabled` at effort high or
below, so "thinking off" sends `disabled` and clamps xhigh/max to high. The
settings panel keeps the thinking toggle and hides budget and sampling.
Register it in the Anthropic, Bedrock (`global.anthropic.claude-haiku-5-5`)
and Vertex default model lists, the 1M context map and the 128K output cap,
and price it at $0.10 / $0.50 up to a 100K prompt and $0.50 / $2.50 above,
with matching cache tiers.
Related to #17078, #16911 and #17025.
* test: Cover Claude Haiku 5.5 Parameter Panel
* test: Pin Haiku Scenarios to Model Specs
* fix: Keep Claude Haiku 5.5 Off Regional Vertex Endpoints
happy-dom 20 holds each MutationObserver callback only through a WeakRef, so a garbage collection between two mutations silently unregisters the hook's observer and the removal is never delivered. Under load that GC lands mid-test and the spec reads data-active="true" after the attribute is gone. jsdom keeps registered observers alive the way browsers do.
* fix: Keep Temporary Chats Out of Langfuse
* fix: Gate Forced-Ephemeral and Open Responses Turns From Langfuse
* fix: Treat Legacy and Chat Completion Temporary Turns as Untraced
---------
Co-authored-by: Joel Hirzel <joel.hirzel@eps.ch>
* feat: open durable subagents from background tasks
* fix: isolate subagent focus and discovery from navigation
* fix: retain subagent task identity through navigation
* test: verify task opening without changing the rendered popup
* fix: preserve retained task activity across navigation
* fix: import the core progress key after dev integration
* fix: refresh subagent panel return focus on explicit open
* fix: prefer the subagent panel opener when restoring focus
* test: retain the actual reply stamp across permission grant
* test: scope permission-grant notifications to the target reply
* test: observe the pending reply before granting notification permission
---------
Co-authored-by: Lia <lia@librechat.ai>
* 🧵 test: Cover Harmony Headers in Text Attachments Sent to OpenAI Models on Bedrock
A pasted text/plain attachment is sent inline as file context, so a harmony header such as `<|channel|>analysis` inside it reaches every later turn. OpenAI models on Bedrock answer that request with an event stream that closes before any event, which crashed the turn in @librechat/agents 4.0.4 with `Cannot use 'in' operator to search for 'tool_calls' in undefined`.
Drive LibreChat's own file-context formatting through the SDK's Bedrock model with a recorded response and assert the header reaches an OpenAI model neutralized and other model families unchanged.
* 📦 chore: Bump @librechat/agents to v4.0.5
* 📦 chore: Regenerate the Lockfile with npm 11 to Keep libc Constraints
* perf: Debounce People Picker Search
* perf: Index-Bound Word-Prefix Search for Users and Groups
Principal search and admin user search matched name, email and username
with case-insensitive regexes, which scan the whole users collection.
Users and groups now store lowercase, accent-folded token arrays derived
from each searchable field, indexed with tenantId, and searches match
anchored token prefixes. Schema middleware keeps the tokens in sync on
save, insertMany and update queries. Documents written before the tokens
existed stay findable through a fallback branch until
npm run migrate:search-tokens backfills them; startup warns while any
remain.
* perf: Build Search Token Indexes in the Backfill Migration
* fix: Reconcile Stale Search Tokens and Close Write-Path Gaps
The backfill now compares every document's tokens with its source fields and
repairs stale ones, so writes from an older server during a rolling deploy are
recoverable by rerunning it. Undefined assignments no longer clear tokens,
upsert inserts initialize every token field, every query word is required,
relevance scoring folds accents like the filter does, and the startup check
counts both collections in parallel and owns its error logging.
* perf: Drop the Second People Picker Debounce
SearchPicker already debounces query changes by 500ms; the extra 300ms delay
in UnifiedPeopleSearch stacked on top of it.
* fix: Match Nothing for Queries Over the Word Cap
Dropping the query word cap let one request build an unbounded filter. Restore
a 16 word cap, but a longer query now matches nothing instead of ignoring its
extra words, so a cap never produces a false positive.
* test: Cover Word-Prefix People Search and the Token Backfill End to End
* fix: Tokenize Fields an Upsert Copies From Its Filter
MongoDB copies equality conditions from an upsert's filter into the inserted
document, so their token fields now derive from that value instead of being
initialized empty.
* fix: Rank Accent-Folded User Matches Like the Search Filter
* test: Run the Token Backfill Without Blocking the Playwright Worker
* perf: Index Search Tokens Tenant First for Scoped Searches and Ignore Undefined Unsets
* test: Skip the Module-Scope Index Rebuild When the Scenario Runs the Backfill
* fix: Exit the Search Token Migration When Redis Clients Stay Open
* perf: Probe for Missing Search Tokens at Startup Instead of Counting Them
* test: Explain the Startup Probe the Warning Actually Issues
* fix: Keep Search Token Fields Out of Serialized Users and Groups
* fix: Keep the Admin Search Legacy Fallback Anchored at the Field Start
* fix: Score Groups Only on the Fields Search Matches
* perf: Start the Search Token Probe Alongside the Permission Checks
* fix: Warn at Startup When the Search Token Indexes Were Never Built
* fix: Score People Search Matches With the Filter's Word Separators
* perf: Check the Token Indexes Before Probing for Unmigrated Documents
* test: Drive the Failed Startup Probe Past the Index Check
* fix: Score People Search Matches Independently of Word Order
A backend jest worker keeps every module and transform it has loaded for its whole life and grows past three gigabytes on a packages/api shard. Share a 1GB workerIdleMemoryLimit from config/jest.workers.cjs and apply it to api, packages/api and data-schemas, as the client config already does.
* test: Spec the First-Sentence Diagnostic Comparison in the Suppression Runner
* test: Copy the Diagnostics Module into the Suppression Spec Synthetic Roots
* test: Copy the Diagnostic Comparison Spec into the Suppression Synthetic Roots
* refactor: Drop Queue Reset Helper From The Client Queue Module
* refactor: Drop Unused Shortcut Getters And Subagent Batch Wrappers
* refactor: Move HITL Approval Payload Builder To Test Fixtures And Drop Dead Helpers
* refactor: Remove Unused Agent Helpers and Superseded Wrappers
* refactor: Remove Dead Auth, Middleware, Protection and Prompt Helpers
* refactor: Remove Dead MCP, File, Skill and Endpoint Code
* refactor: Remove Dead Stream, Telemetry, Utility and Type Exports
* chore: Drop Unused packages/api Dependencies From bun.lock
* refactor: Keep Published Compatibility and Deprecated packages/api Exports
Bump @modelcontextprotocol/sdk from ^1.30.0 to ^1.32.1 in api and
packages/api. 1.31.0 binds stored OAuth credentials to the issuing
authorization server; LibreChat does not use the affected auth()/
authProvider path and already pins refreshes to the stored token endpoint.
1.32.0 makes the client transports follow same-origin redirects
themselves, which would re-issue requests the connection fetch declined
(hop cap, reauthorized tools/call). Set redirectPolicy: 'follow' so
redirects stay with the connection fetch, and cover both transports with
real-SDK tests.
Plan the complete mock suite's shards from recorded per-file durations instead of Playwright's equal-count split, widen the memory lanes from six to ten, and give the nightly full suite ten shards per store so it fits its 30-minute job timeout again.
* 🔐 fix: Skip Reasoning Labels for Encrypted Reasoning
OpenAI models on Bedrock (and Claude's redacted thinking on Bedrock) return
reasoning as encrypted `redactedContent` bytes. The current agents SDK
base64-encodes those bytes into the reasoning stream, so they arrive as THINK
text: the reasoning-label wiring summarized the blob ("Inspecting an encoded
data snapshot"), activity and phase labels quoted it as reasoning evidence, and
the UI rendered it as thinking.
Add a shared `stripOpaqueReasoning` / `isOpaqueReasoning` predicate in
data-provider that recognizes long unbroken base64/base64url runs (Bedrock
redactedContent, Anthropic redacted_thinking data, OpenAI encrypted_content,
Gemini thought signatures) and keep them out of every label path:
- reasoning labels drop encoded deltas, ignore persisted payloads on resume,
and re-check the accumulated text before a billed label call
- activity-label and phase-summary evidence carries readable excerpts only;
phase anchoring still uses the raw excerpt
- the THINK renderer, grouped reasoning rows and the live header show a
neutral "Reasoning (encrypted)" marker instead of the payload, and ignore
labels previously generated from it
Readable reasoning on every provider, including readable text mixed with an
encoded payload in the same step, is still labeled.
* fix: Classify streamed and persisted payloads by readable content
- isOpaqueReasoning / stripOpaqueReasoning ignore the persisted `<think>`
wrapper, so stored encrypted reasoning renders as the encrypted marker and
never reaches label evidence as tags-only "text"
- add advanceOpaqueReasoning, a streaming cursor that holds a trailing
payload-alphabet run until it settles as a word or grows into a payload,
so a payload split across sub-threshold deltas is dropped as it streams
- reasoning labels count only settled readable characters plus a word-sized
pending tail, so ciphertext fragments never advance the minChars or
updateChars gates or trigger extra billed label calls
* fix: Show readable reasoning around payloads and settle words before counting
- the reasoning disclosures, grouped rows, live header and conversation
export render and copy the readable text of a THINK part with encoded
payloads removed; the encrypted marker stays for parts with nothing readable
- reasoning labels count and send only settled readable text: an undecided
trailing run never advances minChars/updateChars or reaches the prompt,
whatever the configured thresholds, and a short final word settles when the
step closes
* fix: Classify reasoning before every bound and across resume
- once a run is classified as a payload, it continues through the next
non-payload character even when that arrives in the same delta as readable
text, so no tail fragment of the payload is released as readable
- add sliceReasoningTail: the activity-phase rolling reasoning window (and its
snapshot/restore bounds) never starts inside an unbroken run, so a payload cut
by the window cannot leave a fragment too short to classify; the window stays
a substring of the raw THINK text, so phase anchoring is unchanged
- a resumed reasoning step replays its persisted text through the stream
filter and keeps the resulting cursor, so a payload prefix that was still
undecided at the snapshot is dropped once the continuation classifies it
* fix: Classify phase evidence while it streams and mask payloads while pending
- activity phases classify each reasoning delta with the stream cursor and keep
a readable rolling window beside the raw anchor window; tracked activities
carry readableExcerpts as prompt evidence, so a window cut into a payload can
never forward a fragment (older snapshots fall back to stripping); the
cursor-free sliceReasoningTail is removed
- add readableReasoning / flushOpaqueReasoning: while reasoning streams, the
disclosures and live header hold back a trailing run that could still grow
into a payload instead of showing its first characters
- the message editor offers no field for payload-only reasoning and shows only
the readable text of mixed reasoning; an unchanged part is never written, so
its stored payload stays intact
* fix: Replay legacy phase snapshots through the reasoning stream filter
A pending-reasoning lane restored from a snapshot without a readable window
or cursor is replayed through advanceOpaqueReasoning instead of stripped, so a
payload prefix that was still undecided at the snapshot stays in the cursor
and is dropped once the continuation classifies it.
* fix: Drop the cut leading run of a full legacy phase window
A legacy pending-reasoning window that filled the 600-char bound may have been
cut inside a payload, so its leading payload-alphabet run is dropped before the
replay; a shorter window is the whole lane and is replayed as is.
* fix: Keep payload-bearing reasoning read-only and drop payload-derived titles on resume
- reasoning that carries an encoded payload is not offered in the message
editor: the edit endpoint replaces the whole value and the payload cannot be
positioned in an edited draft, so editing would silently delete it
- a resumed reasoning step whose persisted text has no readable reasoning
drops its existing title (which could only describe the payload) instead of
carrying it forward where a short readable continuation would revive it
* fix: Reset a resumed payload-derived title on the client, only for payload-only steps
- a resumed step drops its title only when its persisted reasoning is entirely
an encoded payload, so a short readable token (still undecided for the
stream cursor) keeps a valid title
- dropping it emits a same-step on_reasoning_label reset, which the client
already applies, so a resumed client clears the snapshot title too
* fix: release OpenID delivery lease when the client disconnected before delivery
A refresh whose client disconnects before the response is sent claims the
delivery lease and then arms finish/close listeners on a response that has
already closed. Neither event fires again, so the lease is never released and
every later refresh for the same generation polls until the 30-second lease
expires.
Arm the listeners only while the response is open; otherwise release the lease
directly after the send.
* fix: defer the silent refresh until a prerendered page is shown
A prerendered page runs the startup silent refresh before the user sees it.
When the browser discards the prerender, the refresh is aborted mid-flight
after the server has already rotated the refresh token, and the page the user
actually opens has to recover from a session it never received.
Wait for prerenderingchange before refreshing so a prerender has no effect on
the session.
* refactor: move OpenID response delivery into packages/api