* fix: Return approved tool and generation error text instead of raw exceptions
* fix: Redact generation error text without the data-schemas helper
* fix: Redact credentials and emails from shown generation errors
* refactor: Move the Wolfram status policy into packages/api
* test: Fix the Wolfram probe constructor and jest summary match
* test: Match the jest pass count through color codes
* fix: Redact credentials a tool input error echoes from its arguments
* fix: Drop free-text fields from typed generation errors under a content policy
* fix: Keep only the code and numeric fields of typed errors under a content policy
* 📎 feat: Copy Chat Attachments into Attached Workspaces
A chat on an attached workspace now copies the user's own attachments into
.librechat/attachments/ before the model runs, through the worker's
import_file operation, and tells the model where they are. A worker without
the operation, a deny file-write policy, or a copy that fails or runs past
its bound leaves a plain note saying the file is not on this machine and why.
* 🧭 fix: Scope Attachment Records by Tenant and Prime Late Subagents
Match the tenant when recording a workspace copy, require write_file so the
folder can always ignore itself, key copies by the live Code API route and
worker, keep a digest when a copy fails so it is not hashed again, stop
waiting on bookkeeping at the deadline, refuse to start on a cancelled turn,
prime lazily initialized subagents through one per-turn primer, and make the
budget configurable as endpoints.agents.workspaceAttachments.timeoutMs.
* 🙈 fix: Copy Attachments Only Where the Folder Ignores Itself
Gate every copy on .librechat/attachments/.gitignore ignoring the whole
folder: create it with *, and accept an existing one only when it already
ignores everything, otherwise report the file as unavailable with how to fix
it. Prime a lazily routed subagent against its own tool context with a
provisioning callback bound to it, and check for a cancelled turn before
arming the deadline timer.
* 🪢 fix: Recheck the Attachment Ignore Rule Each Turn and Share One Request Cap
Drop the process-wide ignore cache and check the folder once per turn, shared
by every agent, rejecting a truncated read. Cap workspace requests per turn
across agents (workspaceAttachments.concurrency, default 3), name converted
images for their stored type, log only safe error metadata, and keep the
bookkeeping write from advancing updatedAt.
* 🔀 chore: Take dev's File Methods Ahead of Merging dev
dev and this branch both add to the same lines of the file methods module.
Take dev's version here so the merge of dev that follows is clean; the next
commit adds recordFileWorkspaceImport back beside dev's new methods.
* 🧩 fix: Ignore Workspace Operations a Newer Worker Advertises
Bring recordFileWorkspaceImport back beside dev's new file methods after the
merge. Drop well-formed operation names this build does not know from a
worker's capabilities, globally and per workspace, instead of rejecting the
whole status, so workers can turn on new operations such as import_file by
default. Known names keep their strict checks, malformed lists are still
refused, and a worker offering only unknown operations offers no workspace
tools rather than failing.
* 🙈 feat: Let Config Profiles Hide librechat.yaml MCP Servers and Remove Their Fields
Config profiles are the `__base__`, role, group and user documents written through `/api/admin/config`. A profile can tombstone `mcpServers.<name>`, and the merge then removes that server from the principal's resolved `mcpConfig`.
* 🛡️ fix: Apply MCP Hiding to Server Instructions, Failed Removals and the Startup Audit
- Server instructions bound for the model now leave out librechat.yaml servers hidden from the acting user, including one a model spec pins. `applyContextToAgent` passes the request user through `formatInstructionsForContext`, which drops the user's hidden set from the user-less catalog it reads.
- A config-tier candidate that removes a librechat.yaml value (a field, or a key of a record such as `headers`) and fails inspection now stands as a failed config-source entry instead of falling back to the librechat.yaml base that still carries the removed value. A failed change that removes nothing still falls back, and an entry with no transport (fields written over a hidden server) keeps the documented librechat.yaml fallback.
- The startup tombstone report reads every tenant's profiles in the system tenant scope, so it also runs from entry points without an outer system scope under strict tenant isolation, and names each profile's tenant.
* 🗂️ fix: Project the Startup MCP Tombstone Report Query and Run It Alongside Connections
* 🔑 fix: Keep Removed OAuth Headers Out of the MCP OAuth Callback
* 📝 docs: Note That Removed OAuth Headers Stay Out of MCP Sign-In
* 🔐 fix: Recheck the MCP OAuth Flow's Server Before Initiation Redirects
OAuth initiation redirected to a flow's stored authorization URL, or started
a new authorization from the flow's stored server URL, without resolving the
server again, so a principal whose profile hid the server after the flow was
created could still be sent to its authorization endpoint. The callback's
active-server check and its OAuth header fallback lived in the CJS route.
resolveMCPOAuthTarget (packages/api/src/mcp/oauth/target.ts) now owns both:
it resolves the server in the principal's current config with a strict read,
refuses a flow admitted for another server, a server the principal can no
longer use, or a redefined one, and selects the headers the flow sends (the
initiation snapshot, else the principal's config). Initiation calls it before
either redirect (404 or 400); the callback calls it before the exchange and
from the token-store fence.
* 📝 docs: Note That a Hidden MCP Server Ends OAuth Sign-Ins Already Started
* 🚩 feat: Publish the yamlMcpHide Admin Config Meta Flag
Profiles can hide librechat.yaml MCP servers and a hidden server is an
access boundary, so GET /api/admin/config/meta now reports
features.yamlMcpHide for the admin panel to gate its hide UI on. The
meta.spec probe checks the behaviour itself: the admin API stores an
mcpServers.<name> tombstone, the merged config drops the server, the
registry withholds it from principal lookups and checkout but not from
app-level lookups, process-backed servers stay, and an unreadable
principal config withholds every hideable server.
* 🧪 test: Align MCP Tombstone Specs With the Portable Config Store
The admin config handlers now write whole documents through
writeConfigDocument, so the end-to-end tombstone spec wires it in place
of the removed upsertConfig/patchConfigFields deps. A reset of a whole
MCP record now also prunes the parents it leaves empty.
* fix: Bound Outbound Calls Made Inside a Request
RAG queries for file search and message context, OCR provider calls,
action HTTP calls, the moderation check, the stored-image fetch for
base64 providers, and Flux result polling had no timeout. A stalled
upstream hung the chat turn, the upload or the tool call indefinitely,
and Flux kept polling forever on statuses that never become Ready.
Each now runs under a budget from a new librechat.yaml timeouts block
with finite defaults (getOutboundTimeoutMs in @librechat/api). Action
calls also receive the run's abort signal, so stopping a run closes the
socket.
* test: Cover a Stalled RAG Query in the Mock Harness
* fix: Share One Flux Budget Across Submit and Polls, Bound the Vertex Token Exchange
* refactor: Poll Flux Results Through pollUntilSettled
* refactor: Leave Flux Polling for a Separate Migration
* fix: Cap Configured Timeouts at the Node Timer Limit
* fix: Bound the Action Token Refresh and the Whole Image Download
* fix: Cap Every Outbound Budget by Wall-Clock Time
Axios's timeout only fires on an idle socket, so a RAG, OCR, moderation,
action or token response that keeps trickling bytes never ended.
outboundRequestBudget pairs each timeout with an AbortSignal deadline, and
action calls combine it with the run's own signal.
* fix: Name the Budget That Ran Out and Give Actions Five Minutes
The actions default rises to five minutes, since action endpoints are
arbitrary third-party APIs and the old minute could cut off legitimately
slow ones after an upgrade.
A budget overrun now fails with an error naming the timeouts key and its
value, whether it was the idle timeout or the wall-clock deadline, and
logAxiosError logs it as a warning so operators know what to raise. A
deadline cancellation is no longer read as a user stop, so an action that
runs out of time reports a timeout instead of looking like a stopped run.
* fix: Hold Outbound Deadlines Weakly So Finished Requests Leave No Timers
* fix: Check the Budget Before Reading an Error Message
* fix: Report a Wrapped Deadline as a Failure, Not a Stop
* fix: Bound the OAuth Code Exchange and Keep the OCR Budget Hint When Protected
* chore: Sort Imports in axios.ts After Rebase
* feat: Defer MCP OAuth Sign-In Until a Deferred Tool Is Used
Adds an oauthPrompt server option (on_start by default, or on_use) and an
operator-declared oauthToolCatalog. With on_use, an operator-configured
OAuth server whose selected agent tools are all deferred no longer prompts
at turn start: tool search offers the declared catalog (or the user's own
cached catalog), and sign-in happens on the first tool call through the
existing in-run OAuth step. Without a declared catalog the turn-start
prompt is kept. Both fields are operator-only.
Related to #16765
* test: Cover MCP OAuth Sign-In on First Deferred Tool Use
* fix: Bound the Declared OAuth Tool Schema to Non-Recursive Types
* refactor: Move MCP OAuth Deferral Orchestration Into packages/api
* refactor: Move the Cached-Catalog OAuth Prompt Branch Into packages/api
* fix: Defer MCP OAuth Only When the Offered Catalog Covers the Selection
* fix: Cover Runtime OAuth, Background Tools, and Tokenless Discovery in Deferral
* fix: Match Legacy Keys, Honor Background Capability, and Reject Unsupported Schema Keywords
* fix: Resolve Wildcard Background Opt-In and Tighten Declared Catalog Matching
* fix: Validate Declared Tool Schema Types and Enums
* fix: Accept Canonical Catalog Keys the Definitions Loader Accepts
* test: Pin the Declared Catalog Requirement for Cached User Catalogs
* fix: Resolve Deferred Tool Options Across MCP Spelling Aliases
* 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
* feat: Look Up Bitbucket Cloud Pull Requests
* fix: Keep the Pull Request Host Out of Overrides and Per-Host Keys
A role, group, user or tenant override can no longer change or remove pullRequests.host, so feature gating and link recording see the operator's host. Links are recorded only from the configured host, and the batch share key and the rate limit cooldown are scoped by host.
* fix: Compare Pull Request Link Hosts Through the Host Resolver
* test: Add Pull Request Scenarios Against a Fake GitHub Enterprise Host
* fix: Keep the Pull Request Provider Out of Overrides and the Batch Share Key
A role, group, user or tenant override can no longer change pullRequests.provider, so feature gating and link recording agree with the lookup, and a batch lookup for one provider is never joined by a request for another.
* fix: Read GitLab History Apart From Open Merge Requests
Merged and closed merge requests are listed on their own, so open fork merge requests on the branch name cannot crowd them out. Rate limit cooldowns are scoped by provider, and a workspace repository with a dot segment is rejected.
* fix: Revalidate Bitbucket Pull Requests and Skip Ones Without a Source Commit
Read a full closed page so a newer closed fork cannot hide an older pull request from the repository, recheck the source branch and fork status on the detail, and treat a pull request without a source commit as unmatchable instead of failing the lookup.
* fix: Follow Bitbucket's Next Link When Listing Pull Requests
* fix: Record the Lane Pull Request Without an Aggregation Pipeline Update
* fix: Match an Empty Lane Repository as Unknown
* fix: List Locked GitLab Merge Requests With the Open Ones
* fix: Replace an Expiring Fake Git Host Certificate
* feat: Show the Conversation Pull Request in the Sidebar List
Add a batch lookup route and a sidebar row mark: the state icon with the CI dot, opening the same card beside the row on hover. The row keeps the conversation's own title.
* fix: Address Review Findings on the Sidebar Pull Request Batch and Row Mark
Cap the whole batch with a configurable deadline, make the row mark a keyboard-reachable button that opens the card on focus, refresh rows when the window regains focus through the shared batch, and show a retryable failure state instead of treating a failed first lookup as no pull request.
* fix: Send Sidebar Pull Request Batches One at a Time
Queue every batch request behind the one before it, so an oversized list or an overlapping flush cannot multiply the server's configured lookup concurrency, and bound how long one request may hold the queue so a request that never answers cannot stop the rest.
* fix: Keep the Batch Request Timeout Above the Server Deadline
Default the client's request timeout to just past the longest batchTimeoutSeconds a deployment may configure, so a queued request never starts while the previous request's server-side lookups are still running.
* fix: Keep Batch Lookups Inside the Concurrency Limit and Per-Entry Failures Per Entry
Share one lookup limiter across every batch request so lookups that outlive a request's deadline keep their slot, answer a missing token only for the conversations that needed it, and point the row at its pull request description only while that text is in the page.
* fix: Bound Batch Reads, Scope Concurrency Per Credential, and Gate the Sidebar on the Batch Version
Start the batch deadline when the request arrives and bound the config and lane reads, count lookup slots per credential, advertise a batch route version and ask only a server that has it, take no row space for an absent mark, show a failed refresh over a cached empty answer, keep a malformed response from stalling the request queue, and repair the lane recording tests that still used a mock removed on dev.
* fix: Move Capability Resolution Into the Package and Harden the Batch Route and Queue
Resolve the pull request startup capabilities in packages/api so the route stays wiring only, bound the config and lane reads by the configured deadline, merge identical lanes before taking limiter slots, fall back to the single route when a replica has no batch route, and dispose the batch queue when the signed-in user changes.
* fix: Abort the Batch Fallback With Its Request So the Queue Never Overlaps It
Give the batch fetcher the request's abort signal and pass it to every single-route call, abort a request that times out or is disposed, and stop the fallback workers from starting more calls once it has.
* fix: Forward the Abort Signal to the Batch POST
Pass the request's abort signal through the batch fetcher, the data-provider call and Axios, so a batch request that times out or is disposed is cancelled instead of continuing beside the next one, and settle a request promptly when it is aborted.
* fix: Close the Pull Request Lookup Follow-Ups and Cancel Abandoned Sidebar Rows
* 🏝️ fix: Apply the No-Workspace Decision Wherever an Agent's Code Route Resolves
A conversation set to run without attached workspaces failed mid-turn with
"Choose an attached workspace before using this agent." when its agent's
default machine was attached and the turn loaded a skill or another tool at
execution time. Request-start initialization honoured the decision, but the
execution-time tool loader re-resolved the agent's attached default and then
demanded a workspace selection the conversation intentionally never made.
The per-agent code rule now lives in one typed function,
resolveAgentCodeExecution (packages/api/src/code/agent.ts): it applies the
conversation's no-workspace opt-out, the run-code grant, the stateful
capability and the stateful environment policy, then resolves the base route.
Request-start initialization, lazy subagent metadata and graph members, the
execution-time and definition tool loaders, and the legacy execute_code
loader all go through it, so an opted-out agent lands on the managed default
route with no code tools at every site. Per-call subagent machine choices
offer and accept no attached machine in such a conversation, and a source
invariant test keeps new resolver call sites on the shared rule.
* fix: Pass the Implicit Stateful Route Rollout Into the Shared Code Rule
The per-agent code rule read the implicit stateful route rollout from the environment itself. It now takes it as a parameter, and the request boundaries (initializeAgent and withRequestCodeInputs) resolve it from the deployment configuration, so the rule stays a pure function of its inputs.
* fix: Carry the Initializer's Code Opt-Out Into Tool Definition Loading
initializeAgent hands the tool loaders a partial agent record (id, tools, provider, model) together with the code route it already resolved. Without the agent's route fields the definition loaders could not see a no-workspace opt-out and recomputed code execution as enabled, so tool classification could advertise the programmatic bash tool that execution then refuses. initializeAgent now passes attachedEnvironmentOptOut through every tool loader, and the shared rule keeps an opt-out its caller already decided.
* feat: Record Lane Git State and Serve Conversation Pull Requests
A conversation now stores the branch, head and repository its code lane last reported, written only by the server after a finished bash command and excluded from ordinary reads, generic saves, unset fields and imports. GET /api/convos/:conversationId/pull-request finds the pull request for that branch through an injected GitHub source with a bounded cache, answers null for an absence, and maps failures to stable codes without upstream text. The feature is off unless endpoints.agents.pullRequests is enabled with a token reference.
* fix: Address Review Findings on Pull Request Lookup and Lane Writes
Scope the lookup cache, shared in-flight request and rate limit cooldown to the credential, so one tenant's token never serves another. Ask for open pull requests on their own before falling back to the newest closed one, read every page of check runs up to a bound and never report an incomplete rollup as passing, order lane writes per conversation so an older write cannot land last, and name both required token permissions in the example config.
* fix: Address Second Round Review Findings on Pull Request Lookup
Clear the stored lane when the conversation moves to another workspace or leaves attached execution, so the previous workspace's pull request is never served. Fence the lane write on the time it was reported, so an older report from another replica cannot replace a newer one. Honor GitHub's reset window up to an hour instead of ten minutes, and expose the request timeout, whole lookup timeout and check run page limit as endpoints.agents.pullRequests settings with the previous values as defaults.
* fix: Authorize Pull Request Repositories and Fence Lane Writes
Require an administrator allowlist of repositories before the server token is used, since a worker reports its own repository. Match a pull request to the commit the conversation last ran at, so a reused branch name does not return another pull request. Fence the lane write on the workspace the command ran in, so a write queued before a move cannot bring the old lane back, and record a subagent thread's lane on the visible root conversation. Key the lookup cache and shared request by the recorded commit and the lookup policy, and document the worker flag in the example config.
* fix: Order Lane Reports by a Database Sequence and Tighten Head Matching
Replace wall clock ordering of lane reports with a per conversation counter reserved in the database when the command settles, so replica clock skew cannot reject a newer report. Treat a pull request whose head is behind the recorded commit as no match, and expose the candidate and comparison bounds of the history search as settings. Read the app config and the owner scoped lane together, inject the HTTP transport from the route instead of falling back to a global, skip a report identical to the last one written, and name the Contents permission that commit comparison needs.
* fix: Queue Lane Reports by the Visible Conversation and Read Checks for the Current Head
Order lane report reservations by the conversation the write lands on, so two subagent threads of one conversation share a chain, and write every report instead of skipping a repeat, since another recorder may have changed the lane in between. Read a pull request's detail before its check runs and use the head it reports, so line counts, mergeability and checks describe one commit. Treat an expired temporary chat as gone when reserving, writing and reading the lane, and make the shared cache capacity a setting.
* fix: Fence Lane Reports by Attachment Epoch and Place Recorders Before Reporting
Strip dotted descendants of lane fields from saves and imports, partition the pull request cache per credential, resolve the lane write target when the tool is created, and fence reports with a codeAttachmentEpoch that advances only on an owner move or detach.
* fix: Fence Subagent Lane Reports by the Root Epoch and Cover Forks and Moved Heads
Read the visible root's attachment epoch for a subagent report, gate lane recording on pullRequests.enabled, make the credential partition bound a config field, revalidate the detail head against the recorded commit, and find fork pull requests through the recorded commit.
* fix: Keep the Credential Cache Bound Deployment Wide and Budget Head Revalidation
Apply the largest configured credential bound to the shared cache, count cache hits as use, spend the comparison budget on a moved head, and fix the data-provider config default test.
* fix: Cool Down Headerless Secondary Rate Limits
A 403 whose message names a secondary rate limit now starts the credential cooldown for at least a minute instead of being cached as a ten second failure.
* 🧬 fix: Default Subagents to Their Parent's Code Machine and Workspace
A subagent resolved its code machine from its own saved fields only, so
when a chat routed the parent to an allowlisted machine the subagent fell
back to its own default and whatever workspace the conversation had
selected there, or failed the parent's whole turn when none existed.
Subagents now inherit the parent's resolved attached machine when it is
their own default or on their allowlist where per-chat choice applies,
the principal can use it, the sealed decision already selected it, and no
selection names the subagent as an explicit owner. Inheritance is derived
per request from the sealed decision and current agent configuration and
fed to every routing site; it is never persisted, so existing chats and
older clients get it without a decision change. The composer mirrors it.
A subagent whose machine has no usable workspace is advertised to the
parent as unavailable instead of failing the turn; a locked decision
stays fatal.
* fix: Inherit Across Every Spawning Path, Graph Members and Deferred Code Tools
A subagent now inherits only when every agent that can spawn it, at any
depth, runs on the same machine; spawn cycles keep their own routes.
Subagent graph members inherit like explicit subagents. The deferred
execute_code loader passes inheritance to its fallback resolver. Subagent
reads and VIEW checks move to a packages/api loader shared by inheritance,
lazy descriptors and graph members. The composer lets a draft's subagent
follow its parent's machine before a workspace is picked, and keeps an
inheriting subagent's former selection without asking for a move.
* fix: Prune Spawn Back-Edges, Gate the Composer on an Inheritance Capability
Spawning parents follow the run's path semantics, so a back-edge the
descriptor tree prunes no longer blocks inheritance, and agents that spawn
each other follow their single outside parent machine when it routes them
all there. Only admitted subagents count toward the walk's node limit.
Graph members re-read and re-check VIEW when they initialize. The composer
echoes an inheriting subagent's former selection without gating on that
machine, and mirrors inheritance only when the API advertises
codeWorkspaceInheritanceVersion; approval modes keep intersecting every
machine an agent could use.
* fix: Settle Each Mutually Spawning Group on Its Own
Independent spawn cycles no longer merge into one group, so a pair under
one parent follows that parent even when another pair follows a different
machine. Also types the data-provider spec for its ES5 target.
* fix: Bound the Inheritance Walk, Normalize Parallel Roots, Move Subagent Availability to TS
The walk reads in batches no larger than the remaining node budget, stops
once the limit is reached, and runs through the run's subagent load limiter.
Roots are keyed by saved agent ID, so a parallel root such as id____1 can
never receive an inheritance entry. Resolving a lazy subagent's code route,
reporting it unavailable, and guarding its descriptor now live in
packages/api (resolveSubagentCodeAvailability, guardUnavailableSubagent).
* fix: Keep Missing Subagent Loads Parallel Near the Node Limit
Batches never shrink below eight IDs, so missing or unviewable subagents
near the limit still load in parallel under the injected limiter, and at
most a handful of reads land past the limit.
* fix: Separate workspace admission and transport budgets
* fix: Validate workspace run reserves and retry test types
---------
Co-authored-by: Lia <lia@librechat.ai>
* 📝 docs: Describe Attached-Workspace Lanes, Sandbox Scratch, and Skill Files in Tool Descriptions
- Lane-mode `cwd` now states that any call outside `.worktrees/<name>` (with or
without cwd, and file tools on such paths) is checkout-wide: it waits for
running worktree calls and blocks new ones until it finishes, so checkout-wide
calls are reserved for creating, pruning or removing worktrees.
- `command`/`cwd` state that the command starts in cwd, so the model does not
also cd into it; the lane-mode cd hint says the call ran checkout-wide.
- Workers advertising the native SRT sandbox profile (`anthropic-srt`,
`anthropic-srt:<preset>`) get a bash bullet that / and /tmp are read-only and
scratch files belong in $TMPDIR or the workspace. Carried as
`codeWorkspace.nativeSandbox` into tool definitions and the bash tool.
- The attached skill-aware `read_file` description says skill files are not on
the attached machine and must be written into the workspace to run there.
- Tightened the attached bash description so the native variant stays within
the 1024-character tool description limit.
* 📝 fix: Keep the Skill-Copy Warning on the Command-Only Workspace's Skill read_file
A workspace that offers execute_command but not read_file falls back to the
skill-only read_file definition. It now carries the same sentence as the
attached skill-aware definition, so the model knows bash_tool cannot reach
skill files by path there.
* 📝 fix: Name bash_tool in the Skill-File Warning Only When It Is Registered
The skill-copy sentence on attached read_file definitions now depends on the
attached bash_tool being registered beside it (includeBash and an advertised
execute_command), for the attached skill-aware definition, its custom
line-limit variant, and the skill-only fallback. Workspaces without commands
keep the plain skill description.
* feat: Route Linked Worktree Requests into Per-Worktree Lanes
Workers that advertise the git_linked_worktree workspace scope run each
.worktrees/<name> of a registered checkout in its own scheduling lane.
Accept the scope in worker status, map worktree-scoped paths and command
working directories onto the worktree field, and restore the prefix on
returned paths so tool output stays relative to the checkout.
* fix: Advertise Worktree cwd Routing in Model-Facing Bash Definitions
* refactor: Normalize the Lane Flag in TypeScript, Pass It Through /api Unchanged
* feat: Gate Linked Worktree Lanes Behind an Environment configSchema Toggle
* test: Give the Lane Toggle Fixture a Default Environment
* ⏳ fix: Keep BYOM tools queued across capacity windows
* ⏳ fix: Configure the admission budget and refresh queued credentials
Addresses the Codex review on #16060:
- Add `limits.maxQueueWaitMs` to the code environment config schema so a
deployment can shorten or disable the five-minute admission budget from
`librechat.yaml`; omission keeps the merged behavior, `0` surfaces the
first capacity expiry to the model. Resolved through
`resolveAttachedWorkspaceQueueWaitMs` and passed by ToolService, matching
how `maxCommandTimeoutMs` already flows.
- Accept an auth-header supplier in `executeWorkspaceTool` and mint it per
admission attempt, so a call queued past the Code API token TTL presents
fresh credentials instead of failing permanently with 401. The attached
Bash tool and the workspace file operations now pass suppliers.
- Recheck the deadline after a clamped wait and throw the saved rejection
rather than opening an admission window that could still be admitted
once the budget is spent.
* fix: Complete workspace admission wiring and caller regressions
* fix: Share protected edit retry budgets and schema limits
---------
Co-authored-by: Lia <lia@librechat.ai>
* fix: Enable programmatic Bash in selected attached workspaces
* Centralize project PTC construction and preserve recognized capabilities
* Prevent stale list refreshes from restoring removed pins
* 🔎 fix: Enforce WEB_SEARCH Role Permission on Agents
Squash of #15822 onto dev, preserving authorship:
- 🔎 fix: Enforce WEB_SEARCH Role Permission on Agents
- 🧪 test: Cover the Runtime Tool Loader's Web Search Gate
- 🔒 fix: Gate Web Search on the OpenAI-Compatible Embedder Surface
- 🩹 fix: Authorize Web Search by the Execution Context's User
- 📝 docs: Note the Grant Memo's One-Subject-Per-Request Assumption
- 🧹 fix: Deny Web Search Before an Endpoint Default Can Grant It
- 🙈 fix: Withhold the Web Search Model Parameter From a Denied Role
- 🧱 fix: Strip Provider-Native Search an Endpoint Re-Enables After Denial
- 🔌 fix: Close the OpenRouter Plugin Path and Stop Pruning Hidden Parameters
* 🧭 fix: Gate Provider-Native Web Search on the Built Config
The web search role gate ran before the provider builder and only when the
agent's stored `web_search` was not `false`, so an endpoint's `addParams`
re-enabled native search for a denied role, and every OpenAI-compatible and
Responses request read the role whether or not search was in play.
The gate now reads the builder's output: when it carries a native search tool
or OpenRouter's `web` plugin, the `WEB_SEARCH` grant is resolved and the search
stripped on denial. The routes that reach `initializeAgent` with `runtime` and
no `req` pass a lazy `resolveWebSearchGrant` that joins their request-memoized
grants, forwarded through agent discovery to handoffs and subagents. The
embedder surface wires it whenever `getRoleByName` is supplied, independent of
`appConfig`.
---------
Co-authored-by: TomasPalsson <tomas2212042710@gmail.com>
* feat: add run-scoped file sharing for subagents
Delegate current-message files through authorized execution catalogs and publish selected immutable code-output versions through the existing file pipeline.
Requires the companion agents SDK context adapter before the dependency can be released and pinned.
* fix: preserve child tool artifacts and verify file-sharing lifecycle
Keep citations and other non-sandbox artifacts on their existing delivery path while retaining private code outputs and inspected images until publication.
Cover native and text delivery, nested and sibling authorization, concurrent runs, cancellation cleanup, and checkpoint recovery through the browser harness.
* test: align file-sharing fixtures and integration cleanup
* test: repair LibreOffice conversion fixtures
* test: verify deployment skill bytes across checkout formats
* fix: align subagent file budgets and refresh boundaries
* chore: update agents SDK to 3.8.6
* refactor: move code output processing into API workspace
* fix: enforce shared file budgets for parents and teams
* fix: declare the shared file encoder interface
* fix: enforce FILE_SEARCH role permission server-side
The FILE_SEARCH role permission is stored, served by GET /api/roles/:name and
settable through the admin API, but nothing ever checks it.
PermissionTypes.FILE_SEARCH occurs zero times under api/server; the three
occurrences in @librechat/api are all in the interface-to-role sync, which
writes the permission rather than checking it.
Measured on v0.8.8-rc1 and confirmed unchanged in rc2: a user whose role has
FILE_SEARCH.USE = false can still upload a document with
tool_resource=file_search and gets 200 with "embedded": true. The same user
calling execute_code correctly gets 403.
Two gates, both mirroring toolAccessPermType in ~/server/controllers/tools.js:
* Upload (api/server/routes/files/files.js): a tool-resource-to-permission map
checked before any branching, so it also covers the assistants path. Returns
403 with the same log line as the RUN_CODE gate. execute_code is included
because it had the same hole: the tool CALL was gated, the upload was not.
* Tool loading (api/app/clients/tools/util/handleTools.js): checked outside the
lazy loader so a denied user never gets the tool equipped, rather than one
that fails when called. Fails closed if the permission check itself throws.
All four required imports were already present in that file, next to an
existing FILE_CITATIONS check.
Both changes live in /api (JavaScript) rather than /packages/api (TypeScript)
as CLAUDE.md prefers for new backend code: each mirrors an existing pattern in
exactly these files, and a second permission gate in another language and
package would be harder to keep in step with the first.
Tests: api/app/clients/tools/util/handleTools.fileSearchPermission.test.js
covers permit, deny, the log line, and the fail-closed path.
* fixup! fix: enforce FILE_SEARCH role permission server-side
* 🔐 fix: Gate `file_search` and `execute_code` on Role Permissions
The `FILE_SEARCH` gate added in the previous commit lives in
`handleTools.loadTools`, which `loadAgentTools` only reaches when
`definitionsOnly` is false. That default is false on the chat path
(`Endpoints/agents/initialize.js`) but true on the responses and
OpenAI-compatible controllers, so on those paths a denied user still got
the tool definition advertised to the model and their files primed by
`primeSearchFiles`, and the call failed later in `loadToolsForExecution` —
the "gets one that fails when called" outcome the gate set out to avoid.
Move the decision to the capability filters that both loaders already run,
mirroring how `canUseMCP` resolves the `MCP_SERVERS` permission ahead of the
synchronous filter. `AgentCapabilities` stays the instance-wide deployment
switch; the role grant is a second condition a tool has to clear. Because
`hasFileSearch`/`hasExecuteCode` derive from the filtered list, priming is
skipped for a denied user too.
`RUN_CODE` had the same hole: `toolAccessPermType` in
`~/server/controllers/tools.js` gates only `POST /tools/:toolId/call`, not
the agent run, so a role without `RUN_CODE.USE` could still execute code
through an agent. Both tools go through the same map.
Checks use `checkAccessWithRequestCache` and run only for tools whose
capability is already enabled, so a run costs at most one role read. A check
that throws denies the tool.
Also formats the test file added in the previous commit, which Prettier
rejects as-is.
* 🔐 fix: Close the Remaining `RUN_CODE` and `FILE_SEARCH` Bypasses
Codex review of the previous commit found four ways past the gates. All four
share a cause: the permission map was copied per call site, so every boundary
that was not copied into stayed open.
Moves the maps and the check into `packages/api/src/tools/rolePermissions.ts`,
per CLAUDE.md ("all new backend code must be TypeScript in `/packages/api`",
"keep `/api` changes to the absolute minimum"), and points the four `/api`
boundaries at it.
- `handleTools.loadTools` now gates every tool in `toolRolePermissions`, not
just `file_search`. It is the shared boundary the Assistants required-action
flow crosses via `processRequiredActions`, which never passes the agent
capability filter — so a legacy Assistant with a raw function named
`execute_code` ran arbitrary code for a role denied `RUN_CODE`.
- The check is now request-cached (`checkAccessWithRequestCache`). The previous
`checkAccess` call omitted `req`, so the agent path resolved the same grant
twice and issued a second serial `getRoleByName` during chat startup, against
the rule in AGENTS.md.
- `codeExecutionEnabled` is gated by the role result at all three loaders, not
only the `execute_code` entry in the filtered list. It drives tool
classification and the programmatic bash tool, so an agent pairing
`execute_code` with an MCP `code_execution` tool still advertised and
instantiated `run_tools_with_bash` for a denied role.
- The upload map covers `EToolResources.code_interpreter`, which the Assistants
builder posts instead of `execute_code`, and native `code_interpreter` /
`file_search` tools are dropped when an assistant is created or updated.
Those run inside the provider and reach neither loader, so configuration time
is the only place they can be gated.
Renames `handleTools.fileSearchPermission.test.js` to
`handleTools.rolePermissions.test.js` — it now covers both tools.
* 🔐 fix: Gate the Image Upload Route and the v1 Assistant Writers
Second Codex round found the same shape again: boundaries the map had not been
wired into. Adds two shared helpers next to the maps in
`packages/api/src/tools/rolePermissions.ts` so a new upload handler or assistant
writer gets the check by calling one function.
- `/files/images` accepts `tool_resource` and routes agent uploads to
`processAgentFileUpload` on its own. The Code Files UI sends images there, so
every image was a way around the upload boundary for a role denied
`RUN_CODE`. Both upload routes now call
`checkToolResourceUploadPermission`.
- That helper passes `req`, so the role read joins the request cache. The
previous `checkAccess` call omitted it, and an allowed Assistants upload goes
on to `processFileUpload` → `addResourceFileId` → `updateAssistant`, whose own
check does pass `req` — two serial lookups per upload, against AGENTS.md.
- `/assistants/v1` is still mounted and its `createAssistant` / `patchAssistant`
write native tools straight to the provider. Both now filter through
`resolveAssistantToolPermissions`, which v2 also uses in place of its local
copy.
Run-time enforcement for native assistant tools is deliberately not here: it
means overriding `body.tools` at `createRun` in both chat controllers, which
changes run semantics, and is better argued on its own.
* 🔐 fix: Gate Legacy `retrieval` and the Code-Environment File Tools
Third Codex round, same shape twice more.
- The v1 assistant builder submits `{ type: 'retrieval' }` where v2 submits
`file_search` (`AssistantPanel.tsx:203`). Only the v2 spelling was mapped, so
`resolveAssistantToolPermissions` treated legacy retrieval as ungated and both
v1 writers preserved it for a role denied `FILE_SEARCH`. Both spellings now
answer to the same grant.
- `codeEnvAvailable` in the agent initializer was capability-only, and
`initializeAgent` rebuilds `bash_tool`, `read_file` and the workspace file
tools from it — after the tool loader has already dropped `execute_code` for a
denied role. The bash gate stops command execution, but those handlers still
read, search, create and edit files in an attached code environment. The flag
now carries the role grant, and short-circuits before the role read when the
deployment has the capability off.
* 🔐 fix: Resolve Tool Grants Once Per Request and Pair Every Capability Gate
Four review rounds found thirteen issues, all one class: a boundary the
permission check had not been copied into. Rounds 2 and 3 each found gaps
created by the previous round's fix, and a manual sweep of tool constructors
and upload routes still missed the `codeEnvAvailable` family entirely — because
the thing worth enumerating is not "paths that reach a tool" but "reads of the
capability".
There are 28 such reads. This replaces per-site permission checks with one
resolution per request, pairs every read that is a gate, and pins the list so a
new one cannot be added silently.
- `resolveToolRoleGrants` resolves `RUN_CODE` and `FILE_SEARCH` together and
memoizes on the request. Both perf findings disappear by construction: gates
share one role read instead of each issuing its own.
- Newly paired gates: `codeEnvAvailable` in the OpenAI-compatible, Responses and
memory-agent initializers (`initializeAgent` rebuilds `bash_tool`, `read_file`
and the workspace file tools from it, and hands their handlers the code
environment); the capability checks in agent upload processing and in
agent-management upload purposes; and the library chat-completion service,
behind an optional `getRoleByName` dependency so embedders are not broken.
- The v1 Knowledge upload posts `assistant_id` with no `tool_resource`, so there
is no resource to authorize. It now reads the assistant's own native tools and
requires their grants, instead of treating a missing resource as unrestricted.
- The startup role read moved off the critical path into the existing
`Promise.all`, and the direct tool-call check passes `req` so the shared
loader gate reuses its result rather than repeating the lookup.
- `toolCapabilityGates.spec.js` pins every file that reads either capability
with a count and a reason. A new read fails the suite until it is classified,
so the next boundary announces itself instead of waiting for a reviewer.
* 🔐 fix: Abort Assistant Writes on Role-Lookup Failure, Authorize Legacy Uploads First
Round 5 found no missed boundary — the inventory held. All three findings are
about how the gates behave.
- Failing closed is right where a denial blocks the operation, and wrong where a
denial instead filters a payload the caller persists. A transient role-store
outage during an assistant create or patch was silently stripping
`code_interpreter` / `file_search` / `retrieval` and saving the result, turning
a momentary failure into permanent config loss. `checkToolRolePermission`
gains `throwOnError`, and `resolveAssistantToolPermissions` uses it: every one
of its callers either persists the filtered list or rejects the request, so a
lookup failure has to propagate.
- The legacy-assistant upload gate ran after `sanitizedUploadFn` had already
pushed the file to the provider, so a denied role left an untracked remote file
and got a 500 for what is a 403. Moved to the route, next to the tool-resource
gate, before any bytes are sent.
- `resendFiles` hydration derived its resource set from unfiltered `agent.tools`,
so a denied role still paid the thread walk and code-file reads for a tool
about to be dropped. The `execute_code` half now keys off
`effectiveCodeEnvAvailable`, which already carries the grant.
The `file_search` half of that last one needs the grant threaded into
`initializeAgent`, which is plumbing in the most delicate file here for a latency
win on an already-correct path — deliberately left for a follow-up.
* 🧪 fix: Declare `resolveToolRoleGrants` in the Upload Spec's Module Stub
`Tests: api (shard 2/3)` failed on `a0b38c3`: `process.spec.js` stubs
`@librechat/api` wholesale, so the `resolveToolRoleGrants` call added to the
upload gate resolved to undefined and threw. Five failures in
`processAgentFileUpload`, plus one cascading `processFileURL` assertion from the
same suite.
The gate was verified through its own spec and the routes above it, but not
through the spec of the file it lives in — `process.spec.js` was never run
locally. Every spec belonging to a file this PR touches now has been, and the
mongodb-backed ones that still cannot run here were checked statically for the
same stub shape: all but `process.integration.spec.js` spread `requireActual`,
and that one never reaches these paths.
* ⚡ fix: Skip the Role Lookup When Code Execution Is Disabled
The browser initializer already guards this; the OpenAI-compatible and Responses
controllers did not, so every request on those endpoints issued a role read whose
result the capability short-circuit then discarded. Both now start the lookup
only when the deployment has `execute_code` enabled, matching the third.
* ⚡ fix: Drop Three Redundant Reads on the Gated Paths
- The memory initializer resolved grants unconditionally, so a deployment
without `execute_code` paid a role read for a flag that could only be false.
Guarded like the three chat initializers.
- Agent-management upload purposes resolved grants before `checkCapability`, so
a purpose the deployment has switched off paid a role read on the way to being
rejected. Capability first, then the grant.
- The legacy-assistant preflight builds an OpenAI client to read the assistant's
tools, and `processFileUpload` then built a second one — re-reading the user's
key expiry and values. The preflight now hands its client on, and processing
builds one only when it wasn't given one.
---------
Co-authored-by: Paul <200737214+SSIG-IT@users.noreply.github.com>
* ♻️ refactor: Take Typed Criteria, Not Mongo Filters, for Actions and Assistants
`createActionMethods` and `createAssistantMethods` accepted `FilterQuery<T>`,
putting Mongoose's query language in the package's public surface: every caller
wrote raw Mongo, and no engine other than Mongo could satisfy the signatures.
Both now take domain criteria — `ActionQuery` and `AssistantQuery` — naming
concepts rather than stored fields, with an array meaning "any of these". A
per-domain field map translates them, so `FilterQuery` survives only inside the
two translators.
`buildFilter` throws on a criterion the field map does not cover. Callers in
`api/` are JavaScript and get no compile-time checking, and the server sets
`strictQuery`, under which a silently dropped criterion widens the filter
instead of narrowing it — an unscoped `findOne` returns an arbitrary document
rather than none. Failing closed is the only safe default.
`AssistantQuery.avatarFilepath` covers the avatar-authorization lookup, which
queries a nested path no plain field name reaches.
No behaviour change: every migrated call site produces the same filter as before.
* 🧪 test: Update Avatar Authorization Assertion to Typed Criteria
`validateImages.spec.js` asserted the old `{ 'avatar.filepath': { $in: [...] } }`
call shape for `getAssistant`. The earlier sweep matched on `action_id`,
`agent_id` and `assistant_id`, none of which appear in a dotted avatar path, so
this one assertion was missed.
The `getAgent` assertions in the same file keep the filter shape — that method
still takes a `FilterQuery`.
* 🔒 fix: Reject Unknown Criteria Before Omitting Undefined Ones
`buildFilter` skipped a criterion whose value was `undefined` before checking it
against the field map, so a JavaScript caller misspelling a key that happened to
carry `undefined` — `deleteActions({ agent_id: maybeId })` — produced `{}` and
reached `deleteMany` unscoped. That is the exact failure the guard exists to
prevent, defeated by the order of the two checks.
Validate every key first, then apply the undefined-omission rule. The lookup is
an own-property check so inherited names cannot resolve through the field map's
prototype: `{ toString: 'x' }` previously found `Function.prototype.toString`,
passed the truthiness test, and wrote a garbage filter key instead of throwing.
A recognized criterion left `undefined` is still omitted, so optional parameters
keep working.
* 🔒 fix: Reject Query Fragments as Criteria and Keep the Translator Internal
Three review findings on the criteria translator, all of the same shape: the
guard was narrower than the surface it protects.
Criterion values are now validated as scalars or lists of scalars. `matchAny`
copied any non-array through unchanged, so a JavaScript caller could pass
`{ agentId: { $ne: null } }` and have the operator land verbatim in the filter
— the fail-open case the field map exists to close. `null` is rejected too; it
would have matched documents missing the field entirely.
`matchAny`, `buildFilter` and `FieldMap` are no longer re-exported from the
package barrel. They emit Mongo, and shipping them as public API undercut the
point of the change; they were reachable as `require('@librechat/data-schemas')
.buildFilter`. `OneOrMany` moves to `types/query` because the domain query types
reference it and it names no engine — which also drops the `types` -> `utils`
import.
`loadActionSets` names `ActionQuery` through `import()` so the JSDoc type
actually resolves.