* 🥊 test: Prove MCP Refresh Coordination Across Replica Processes
Fork real replica processes with production Mongo, Redis, encryption and OAuth adapters, and assert one redemption per rotation, peer adoption, and recovery after the authorizing replica is killed. A negative control with coordination disabled observes two redemptions, so a green coordinated run cannot pass on timing alone.
Adds a refreshGate hook to the shared OAuth test provider so concurrent refreshes can be held open. No production code changes.
* test: Establish Replica Overlap Instead of Assuming It
The coordinated assertion counted provider redemptions after a fixed delay, so a peer that arrived once the winner had already persisted would read a fresh credential, never contend, and still leave one redemption: the test would have passed while proving nothing. Both replicas now report entering the refresh path on a credential they observed expired, and getTokens own onRefreshSuccess/onTokensAdopted hooks report which side each took, so the pair (one redeemed, one adopted) is asserted rather than inferred from timing.
A worker that never signalled ready was also left running with live Mongo and Redis connections, which would outlive the suite and disturb the rest of the integration lane; boot failures now reap the child and report the workers own initialization error.
---------
Co-authored-by: Lia <lia@librechat.ai>
Classify HTTP 408, 429, and 5xx before legacy body-message classifiers can invalidate client registration or start consent. Preserve typed temporary failures across refresh storage and verify retry with the existing grant against the real SDK test provider. Keep permanent grant errors and existing UI retry behavior intact.
Co-authored-by: Lia <lia@librechat.ai>
* ✂️ fix: Name Tool Rounds From the Trace, and Match a Cut Response From Its End
* 🧵 fix: Join a Cut Model Call's Round Only When One Is Waiting
* 🏷️ fix: Read Labels From the End Only for the Response the Limit Cut
* 🪪 fix: Say a Tool Round Came From the Conversation Only When It Did
* 🧾 fix: Carry Where a Tool Round's Calls Came From
* refactor(redis): import script cache proposal from #16069
Squashed copy of PR #16069 at 7e994194c2, rebased onto current dev without changing the original branch.
Co-authored-by: Marco Beretta <81851188+berry-13@users.noreply.github.com>
* refactor(redis): cache independent claims without queueing streams
---------
Co-authored-by: Lia <lia@librechat.ai>
Co-authored-by: Marco Beretta <81851188+berry-13@users.noreply.github.com>
* fix: give same-batch code calls the files a skill just loaded
A batch that carries a skill call next to execute_code/bash ran both
against the code-session snapshot the graph took when it planned the
batch, so the sandbox had no /mnt/data/skills/<name>/ for that turn.
Code calls in a mixed batch now wait on the skill calls and merge their
uploaded refs into their own context before dispatch.
* test: pin the spec's own file ref kind so CodeEnvFile stays assignable
---------
Co-authored-by: Lia <lia@librechat.ai>
* 🥢 fix: Exclude Prompt Inputs From Terminal Events
Terminal events carried prompt-building inputs into every place a FINAL is
kept: the runtime cache, the durable Redis job hash, Pub/Sub publication and
late/cross-replica replay. Neither terminal producer projected the payload, so
a 706 KiB attachment-heavy FINAL was stored and published at full size.
sanitizeMessageForTransmit removed fileContext and files[].text from
requestMessage at the controller, but never image_urls, and responseMessage
was not sanitized at all.
projectTerminalEvent is one shared, schema-aware projection applied at both
GenerationJobManager.emitDone() and publishTerminalClaim(), and on read where
a stored record may predate it. It excludes fileContext, image_urls and
embedded file bodies from requestMessage, responseMessage and every
runMessages entry, while preserving message content, attachment references and
display metadata, and the terminal/reconciliation protocol fields.
publishTerminalClaim compared publicationEvent !== finalEvent to decide
persistenceFailed. Projection allocates a new object, so that identity check
now compares against the projected intendedEvent; a successfully persisted
FINAL would otherwise have been reported as a persistence failure.
* fix: Preserve Output Attachment Text and Project Live Terminal Frames
Addresses three review findings at 9e5f6f8ce.
Output attachment text was excluded as if it were a prompt input. `files`
carries the user's uploads, whose `text` is the body extracted for the prompt,
but `attachments` holds resolved artifactPromises that BaseClient assigns onto
responseMessage, so its `text` is model-generated output rendered inline by
TextAttachment from `file.text ?? ''`. Excluding it was also unrecoverable:
useAttachmentPreviewSync polls the preview endpoint only while status is
'pending', and an absent status reads as 'ready', so a late or cross-replica
subscriber whose only source is the stored FINAL rendered a blank preview.
Each collection now carries its own exclusion set; storage bookkeeping (_id,
__v) is still excluded from both.
A live Pub/Sub FINAL from an older replica bypassed projection. The store-read
paths were covered, but during a rolling deploy an old generation owner
publishes an unprojected frame that a new subscriber cached on its runtime and
forwarded to the browser. queueDone is the one choke point every terminal
delivery to a subscriber passes through, so the projection now runs there.
The declared return type was the input type T, so the exported projected
contract was never enforced. FinalMessageFields carries an index signature,
which makes Omit alone a no-op, so ProjectedMessageFields now maps the
transient fields to `?: never` and overloads return the projected contract for
a FinalEvent, the exact type for a non-terminal event, and ServerSentEvent for
the union every manager call site passes.
* fix: Spell Out Transient File Fields for isolatedDeclarations
tsdown builds packages/api with --isolatedDeclarations, which rejects a spread
element in an inferred array type (TS9018), so composing TRANSIENT_FILE_FIELDS
from a shared bookkeeping tuple broke Build packages and every lane that
depends on a built package. tsc --noEmit does not use that flag and passed,
which is why this reached CI.
Both tuples are now spelled out, with a comment recording the constraint.
---------
Co-authored-by: Lia <lia@librechat.ai>
* 🥖 fix: Refresh MCP Credentials That Arrive Already Expired
storeTokens replaced a stated access-token lifetime with the 365-day default whenever it had already elapsed, so a credential that arrived expired was recorded as valid for a year. getTokens decides to refresh from that record, so it never did: recovery waited for the resource server to reject a request, and on connection establishment a second rejection of the refreshed tokens starts interactive OAuth instead.
A lifetime stated by the response is now persisted as it stands, including when it has elapsed, so the next read refreshes it silently. A JWT exp keeps its existing leniency because it comes from the provider clock, where skew can make a live credential look expired.
* style: Sort the jsonwebtoken Import in the Expiry Suite
Package imports are ordered by line length, so `keyv` precedes `jsonwebtoken`. Fixes the Static checks import-order gate.
* fix: Renew Expired OAuth Results Before Connection Readiness
Preserve finite zero lifetimes at exchange and storage. Both callback consumers re-read expired flow results through coordinated refresh after crossing and releasing the callback persistence fence. Cover the real SDK callback transaction, concurrent token waiter, zero-expiry updates, retryable failures, bounded recovery, cancellation, and the unchanged fast path.
---------
Co-authored-by: Lia <lia@librechat.ai>
* 🐣 feat: Invoke Newly Authored Skills In the Same Run
* test: Pin the SDK skill tool text the authoring variant rewrites
* fix: Replace a stale skill definition instead of suppressing the duplicate
* fix: Pin Skill Invocation to the Doc Authored In the Run
---------
Co-authored-by: Lia <lia@librechat.ai>
* 🗂️ fix: Keep a Running Attachment Chat in the Sidebar
Since #15694, a turn with model-bound attachments defers its user-message
write until the model admits the payload. That write was also what created
a new conversation's row, so for the whole first model call, which with a
long generated script or extended thinking can run for minutes, the
conversation did not exist. Any conversation-list refetch in that window
(starting another chat, deleting one, reloading) dropped the running chat
from the sidebar until its run finished, and GET /api/convos/:id answered
404 for it.
When attachments are the only reason to defer, the new conversation's row
is now written as the run starts, carrying the same fields the message
write would set but no message. The message itself stays deferred until
admission, and its write waits for a seed still in flight so the two
upserts never race. A content policy still holds back every write, the
row included, as #14425 intended.
The conversation half of saveMessageToDatabase moves into
saveConversationToDatabase so the seed and the message write share one
path.
* 🧱 refactor: Own Turn Conversation Writes in packages/api
The conversation half of a turn's writes (the row a message save
upserts, the retention context it stamps, and the new seed ahead of a
deferred first message) moves into `conversations/save.ts`, which takes
`getConvo`/`saveConvo` from the caller. BaseClient keeps only the
wiring: the retention lookup that must stay synchronous for Stop, the
message save, and the calls into the module.
The seed passes an empty `appendMessageIds`, which tells `saveConvo`
the row holds no messages yet, so it no longer reads the message list
before the upsert. It also resolves the conversation once and leaves it
on the request, so the deferred message save reuses it instead of
looking the conversation up again.
save.spec.ts drives the real data-schemas methods against an in-memory
MongoDB: the seed creates an empty row without reading messages, the
message save appends to it without a second lookup, an existing chat and
a subagent thread are left alone, a temporary chat keeps its retention,
and a failed lookup settles instead of rejecting.
Claude accepts base64 documents only for PDF. Through an OpenAI-compatible
gateway, a textual file was sent as an OpenAI `file` part with its own media
type, which the gateway maps to a base64 document that Claude rejects with a
400 on every later turn. Send it as a text part with the decoded contents,
using the same Claude detection as the document filter.
Fixes#16054
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix: Stop Sealing a Code Environment Decision a Chat Never Made
Switching an existing chat to a coding agent left the composer unable to
send. The chat had no workspace to select, Send stayed disabled, and the
workspace chip was a read-only status with a recovery hint.
Both layers treated "no decision recorded" as "decided to run without an
attached environment". `readPersistedDecision` documented that a stored
conversation always carries a decision and inferred `without_attached`
whenever `codeEnvironmentMode` and `codeWorkspaces` were both absent,
which is the ordinary state of a chat whose turns never involved a
code-capable agent. `resolveConversationCodeEnvironmentDecision` then
sealed that inference, so the first workspace its owner picked came back
as `locked`. The client mirrored the same premise: `locked` was any saved
chat, so `useCodeWorkspace` withheld agent defaults, a remembered
selection, and a sole workspace, and `CodeWorkspaceMenu` rendered its
non-interactive locked branch.
Only a recorded decision is sealed now. A conversation holding neither
field is undecided, so its first coding turn establishes the decision,
and `resolvePersistableCodeEnvironmentDecision` records that decision
with its selections; writing the mode alone would leave `attached`
without the selections the next turn validates. On the client the same
predicate drives `locked`, and the selection defaults that already
applied to a new chat now apply to any chat that has not decided.
An explicit `without_attached` stays sealed, a legacy row still infers
`attached` from the selections it stores, and a sealed decision is still
replaced only by its owner's move.
* fix: Record the first code environment decision on an existing chat
A saved chat that holds no decision may record the one a run establishes,
but every caller persists through `saveConvo`, which carries those fields
only in `$setOnInsert`. On a row that already exists the write was dropped,
so the choice never survived a reload and a later turn could resolve a
different workspace.
`saveConvo` now fills a whole decision onto a row that holds none, by
compare-and-set on that absence: a writer that decided first keeps its
decision, a chat that already holds one is untouched, and bare selections
never seed a row.
* fix: Record the decision a failed turn of a saved chat ran under
`saveErrorTurn` built its conversation fields only when it was seeding a new
row, so an existing chat kept the stored error turn without the decision that
turn ran under: a retry could then pick a different workspace than the failure
already recorded. The validated decision now travels on both branches, where
the conditional write in `saveConvo` keeps it from replacing a stored one.
* refactor: Resolve the error turn's decision fields in `packages/api`
The error-turn save derived its own decision fields, putting persistence
policy in a CJS controller where `/api` is meant to hold wiring only. It now
calls `resolvePersistableCodeEnvironmentDecision`, the resolver the streaming
saves already use, so one rule decides what a turn may record: a chat that
holds a decision keeps it and receives no decision fields at all.
* fix: Keep the legacy code environment lock until the protocol is advertised
The composer unlocked a saved chat that holds no decision without consulting
`codeEnvironmentDecisionVersion`, so during the rollout window it could attach
an agent default and submit it to a replica that still reads a field-less row
as a sealed `without_attached` and rejects the turn as `locked`.
The unlock now waits for the deployment to advertise the protocol, which makes
the hook identical to its previous behavior while the flag is unset. Nothing is
lost by waiting: the unmade decision is only reported once that flag is on.
---------
Co-authored-by: Lia <lia@librechat.ai>
* 🫗 fix: Roll Back Incomplete Skill Archive Imports
A .zip or .skill import created the skill first and persisted its bundled
files one by one, keeping the skill and answering 201 when a file failed
validation, a size limit, storage, or its database row. The client never
read the internal `_importSummary`, so it showed "Skill created" and
navigated to a skill that silently lacked the resources SKILL.md names.
Import is now atomic: any failed bundled file rolls back the skill, its
file rows, the ACL grant, and the blobs already written to storage, and
the request answers 422 `skill_import_incomplete` with the failed paths
and reasons. The upload dialog stays open and lists them.
* fix: Localize Import Failure Reasons and Report Every Lost File
Addresses the first review round on the archive import rollback.
Failure reasons are now stable codes (`invalid_path`, `file_too_large`,
`archive_too_large`, `archive_entry_changed`, `persistence_failed`) that
the client localizes, so a non-English locale no longer sees English text
and storage or driver messages stay in the server log.
The decompression budget stops the scan mid-archive, and the entries it
never reached were missing from the response, telling the user they had
seen the whole failure set. Those paths are now reported too.
Rollback deletes stored blobs only once `deleteSkill` confirms the skill
is gone; otherwise the blobs stay and the request answers 500
`skill_import_rollback_failed`, because a surviving skill pointing at
deleted files is worse than unreferenced storage.
The failure list scrolls instead of overflowing the dialog, and a
response that resolves after the dialog was dismissed no longer
repopulates it for the next session.
* fix: Ignore Late Skill Import Successes
* fix: Make Skill Import Rollback Retryable
* fix: Propagate Incomplete Skill Cleanup
* fix: Settle Rollback Blob Cleanup
* test: Type Partial Cleanup Fixture
* fix: Preserve Settled Skill Cleanup Steps
* fix: Export Skill Cleanup Step Type
* fix: Reconcile Partially Deleted Skills
* fix: Reconcile Skill Rollback Caches
* style: format cleanup reconciliation changes
* fix: complete skill rollback cleanup
* chore: leave generated locale catalogs unchanged
* fix: preserve synced skill blob cleanup
* test: type synced blob cleanup mock
* fix: Validate Archive Paths and Bound Rollback Cleanup
---------
Co-authored-by: Lia <lia@librechat.ai>
Co-authored-by: Danny Avila <danny@librechat.ai>
* ⏳ 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>
When an MCP server's resource endpoint rejects the access tokens its authorization server keeps issuing, a connection being established silently refreshed on every 401 and retried with a token that was rejected again. Once the connect attempts ran out, the failure surfaced as an ordinary connect error, so reinitialize reported `oauthRequired: false`, agent tool loading never showed a sign-in prompt, and agents whose MCP tools all came from that server failed closed.
The establishment handler in `MCPConnectionFactory` now remembers the tokens its silent refresh issued. When the server rejects that same credential set again while those tokens are still the stored credential, it starts interactive OAuth instead of refreshing, which issues the authorization URL through `oauthStart` so tool loading prompts for sign-in. A credential another request stored in the meantime, or a different credential set, keeps the existing refresh path.
Since #15694 bounded each turn's attachments across history, a thread whose
history counts more than `fileLimit` model-bound files is refused on every turn,
including turns that attach nothing. Two kinds of file reached that count that
never belonged in the prompt.
Code outputs: priming clears an expired sandbox reference on the turn's copy of
the record so the file is re-provisioned, and a route-less record without a
reference was classified as prompt content, so it counted toward the limit and
could be encoded as media. Code outputs now stay tool-owned regardless of
reference liveness, through one predicate shared by admission, BaseClient
delivery and the child run-file encoder.
Tool-routed spreadsheets: with `textFallbackWithoutTools`, #16027 delivered the
fallback text until a tool held a copy. Run Code receives its copy only on its
first call, which a refused turn never makes, so the text counted on every later
turn. An enabled Run Code is a reader again for the types it can read; File
Search still reads only what its store holds.
The agent tool-resource branch of DELETE /files removed a file's association
from the agent and answered 200 without reaching processDeleteRequest, so the
bytes on disk and the RAG chunks of a file the caller owned survived the delete.
A replaced File Search document kept answering from its old chunks (#15930).
An attached file now goes through the full delete pass, but only when nothing
else holds it: a file the caller does not own, and a file any other
(agent, tool_resource) pair still references, are unlinked and left whole.
Nothing is unlinked until the destroy it depends on has succeeded.
- packages/api/src/files/deletion.ts: deleteAgentResourceFiles owns the
operation with its db methods and delete pass injected.
- packages/data-schemas: getSharedResourceFileIds answers the last-reference
question, excluding the pair being removed by the agent's _id.
- Local/crud.js: unlinkFile propagates non-ENOENT, so a file whose bytes
survive is no longer reported as deleted.
Co-authored-by: Mohammed Alshyakh <mohmedmm@users.noreply.github.com>
* fix: Resolve live OpenID bearer before the MCP early domain gate
* fix: Drop unread credentials from the MCP early domain gate
The early domain gate resolves the whole server config through
`processMCPEnv`, but decides from the URL alone. A credential placeholder in
any other field therefore raised `OpenIDReauthRequiredError` from a stale
request-time OpenID snapshot, before the connection path could refresh that
bearer, and the tool was dropped from the agent's toolset.
`buildMCPDomainValidationConfig` narrows the config to what the decision reads,
so the gate needs no live credential. A URL placeholder still fails closed, and
the argument is never mutated, so direct-bearer recovery keeps the placeholder
it knows how to refresh.
Replaces the per-tool `upstreamTokenProvider` call the gate would otherwise
make (1 + N per request per server) and covers every credential-bearing field,
not only `Authorization`.
Co-authored-by: Artyom Bogachenko <SpectralOne@users.noreply.github.com>
---------
Co-authored-by: Lia <lia@librechat.ai>
Co-authored-by: Artyom Bogachenko <SpectralOne@users.noreply.github.com>
* fix: Serialize MCP OAuth token refresh across replicas
`MCPTokenStorage.inflightRefreshes` coalesces refresh-token redemptions
inside one Node process. Behind a load balancer without session affinity,
one user's concurrent requests land on different replicas, each reads the
same not-yet-rotated refresh token and redeems it. RFC 9700 servers treat
the second redemption as replay and revoke the whole grant family, so the
user is asked to authorize the MCP server again.
`forceRefreshTokens` now takes a cross-replica refresh flight before
redeeming, using the same `FlowStateManager.acquireLease` primitive the
OAuth teardown fence uses under a distinct key. A replica that waited for
the flight adopts the tokens the holder rotated instead of redeeming
again; when the flight is still held after the wait window it falls back
to the unfenced redemption every earlier release performed.
* fix: Redeem unfenced when the refresh flight lease store fails
* fix: Fence MCP refresh waiters instead of redeeming unfenced
A waiter that could not take the cross-replica refresh flight fell through to
an unfenced redemption after 10s, while the holder's own stale abort does not
fire until 60s. A refresh taking 11 to 60 seconds therefore still let two
replicas redeem one refresh token, the replay this fence exists to prevent.
The waiter now polls until it either adopts the tokens the holder rotated or
acquires the flight, and fails as MCPTokenRefreshUnavailableError rather than
redeeming beside a live holder. `getTokens` callers defer connection recovery
on that error, and the stored credential is left intact for a later attempt.
The flight is keyed by the stored credential (tenant, user, server name)
rather than the caller's OAuth binding digest, so a rolling config change
cannot hand two replicas different locks over one stored token. The wait is
an operator lever, `oauthRefreshWaitTimeout`, clamped to half the stale
window. An adoption read that fails no longer gives up the held flight.
* fix: Annotate derived refresh-flight constants for isolatedDeclarations
* fix: Recapture the credential generation a peer published on adoption
* fix: Hold the MCP refresh flight until redemption settles
Four corrections to the cross-replica refresh flight.
The flight lease equalled the window that aborts a stalled redemption, on
the claim that an expired flight could never belong to a redemption still
able to reach the token endpoint. Aborting proves no such thing: the
request may have been processed with its response lost, and a stalled
event loop can delay the abort past its own deadline. The lease now
outlives the abort, so a peer cannot redeem a credential the provider has
already rotated. A live replica still releases on settle, so the margin is
paid only by one that died holding the flight.
The credential snapshot was taken after the first failed acquisition, so a
holder that stored and released in that gap was snapshotted post-rotation:
the next attempt saw an unchanged record and redeemed the credential it
should have adopted. It is taken before the first attempt now.
`oauthRefreshWaitTimeout` accepted zero while the runtime mapped every
non-positive value to the default, so the config validated and then
behaved contrary to its value. Zero is rejected at load.
Adoption announced a second credential change through
`handleOAuthRefreshSuccess`, whose `onOAuthCredentialsChanged` advances
authorization state after persistence this replica did not perform. That
retired the generation recaptured beside it and fenced the build against
its own tool publication. Adoption now updates the local token-flow cache
and recaptures, without announcing.
* fix: Check every held MCP refresh flight for a peer's rotation
Four corrections, two of them consequences of moving the credential
observation ahead of the first lease attempt.
That move created a storage read inside `beginRefreshFlight`, and its
failure reached a handler written for a lease-store outage, so a transient
read error became an unfenced redemption beside a live peer. The flight
now handles its own reads: losing the observation costs adoption and
nothing else, and only the lease store failing may redeem unfenced.
The observation was also never compared when the first acquisition
succeeded. A peer that rotated and released before this replica contended
left an acquisition that looked uncontended, and its fresh credential was
redeemed a second time. Every acquisition now runs one rotation check,
which also collapses two code paths into one.
`getTokens` already loads the refresh record to decide a refresh is
needed, so it is passed on as the observation baseline instead of read
again, and the read taken under the flight is reused as the credential
redeemed. Two reads on a latency-counted path where there were three.
`runSilentRefresh` collapsed contention into null, sending the 401 path to
interactive OAuth, so a slow peer prompted the user to authorize a server
whose credential was about to be valid. The retryable outcome now travels
through the silent-refresh layers and the connection defers, matched by
name as well as identity because these errors cross the package boundary.
* chore: Regenerate the Agents OpenAPI spec for repositoryInstructions
The drift check added by #15928 fails on dev: #16008 added
`repositoryInstructions` to the agent schema without regenerating the
committed spec, so every branch merging dev inherits the failure. This is
the remedy the check itself prescribes and carries no hand-written change.
Drop this commit if dev regenerates first.
* fix: preserve MCP refresh outcomes across replica boundaries
* fix: keep MCP adoption fenced through publication
* fix: anchor MCP refresh adoption to rejected credentials
* fix: Complete OAuth Adoption Adapters and Isolate Legacy Flow Readers
* fix: Preserve OAuth Request Identity Through Recovery and Discovery
* fix: Fence OAuth Adoption Against All Credential Writers
* fix: Gate Coordinated OAuth Rollout and Preserve Unauthenticated Identity
* fix: Preserve OAuth Coalescing Across the Staged Rollout
* style: Sort OAuth Timeout Constant Import
* fix: Invalidate Both OAuth Token Flow Protocols on Rotation
* test: Complete the Rollback Token Flow Fixture
---------
Co-authored-by: Lia <lia@librechat.ai>
Co-authored-by: Danny Avila <danny@librechat.ai>
* 🗣️ fix: Keep Provider Error Text on Unclassified Agent Failures
An upstream failure LangChain does not classify discarded the provider's own
message and answered with the generic upstream sentence plus a status, so a
gateway or privacy-proxy rejection lost the only account of what happened.
The provider text now rides along in the persisted payload as `message`,
withheld only where a content policy inspects the traffic — the same condition
`getUserFacingProviderError` and `getUserFacingRequestError` already decide by.
* fix: Harden Provider Error Detail Retention
---------
Co-authored-by: Danny Avila <danny@librechat.ai>
* 📭 fix: Deliver Attachment Text Until a File Tool Holds It
A `none`-routed upload had its extracted text withheld whenever a file tool was
enabled, on the assumption the tool would serve the file. A plain chat with the
File Search toggle runs an ephemeral agent, so the upload is filed under no tool
resource and never embedded: the text was withheld for a vector store that never
received the file, and the attachment reached nothing.
Withholding now requires the record to show the tool holds the file, which is the
same evidence deferred provisioning reads before queueing it. `resources.ts`
delegates to that shared predicate so the two cannot disagree again.
* test: Give the Code-Running Primary Agent a Sandbox Copy
---------
Co-authored-by: Lia <lia@librechat.ai>
* 🪜 feat: Trace Viewer Steps, Sequence Scale and Previews
* 🪜 fix: split tool-only rounds by run step; preview only a step's own roots
* 🪜 fix: keep windows and previews on the records they describe across pages, modes and failed wrappers
* 🪜 fix: keep step folds and previews honest across page boundaries and failed turns
* 🪜 fix: causal tie-breaks, final-only previews, span-shaped title runs, and truncating long names
* 🪜 fix: index previews in one pass, split agent handoffs, withhold parallel lanes
* 🪜 fix: skip the holes a streaming message leaves in its content
* 🪜 fix: reserve a round for compaction summaries; keep parallel-lane tools with their own model call
* 🪜 fix: record a handoff that starts with reasoning; look lanes up in a map
* 🪜 fix: per-lane leading steps, wrappers for turns without work yet, no final alignment for runs that ended early
* 🪜 fix: append to a step's per-tool bucket instead of copying it per call
* fix: Preserve Trace Preview Alignment Across Run Boundaries
* perf: Cache Trace Ancestors and Lane Branches
* 📪 feat: Send Per-Request MCP Headers Without Hiding the Tool Catalog
A `{{LIBRECHAT_BODY_*}}` placeholder in an MCP server's `headers` marks the
server request-scoped, and catalog discovery has no conversation to resolve it
against, so LibreChat skipped tool discovery and the server appeared with no
individual tools.
Operators can now declare those headers separately in `requestHeaders`, which
discovery omits and chat turns resolve and merge over `headers`.
* fix: normalize requestHeaders at each resolution entry point
Registering the new field in every consumer that enumerates header-bearing
config fields left five gaps: startup inspection, case-insensitive override,
the admin-configurable comparison surface, Graph token preprocessing, and
direct-bearer recovery.
Fold `requestHeaders` into `headers` at the entry of each resolution pipeline
instead, so Graph preprocessing, direct-bearer detection, `processMCPEnv` and
the transports keep reading one header map and never learn a second exists.
Catalog paths strip it at the same boundary, startup inspection included.
* fix: keep servers with chat-only headers off the shared app connection
A static `requestHeaders` map left a server eligible for app-level sharing, and
`ConnectionsRepository.loadConnection` handed the raw config to the factory, so
the shared session's own `initialize` and `tools/list` carried headers
documented as chat-only — and every later catalog read reused that session.
One session cannot serve both sides of the field. Exclude such configs from
`canUseAppConnection`, the single predicate both `connectAppServers` and the
repository's own app-level gate already consult, so the chat turn keeps its
headers on a user-scoped connection while catalog work never sees them.
* fix: normalize requestHeaders before the manager's bearer decisions
`createUserConnectionInternal` reads `usesDirectOpenIDBearerRecovery` and
`resolveDirectOpenIDBearerConfig` on the config it receives, while the merge sat
in `resolveRuntimeConfig`, a leaf helper below those calls. An `Authorization`
template declared in `requestHeaders` therefore never entered direct-bearer
mode: a `tools/list` 401 skipped the force-refresh and the connection was
returned with a stale bearer.
Normalize at the birth of that pipeline's config instead, ahead of all four
decisions, and make `getMCPAppToolsPublicationGeneration` normalize before
hashing. Connection paths hold a merged config while cache paths hold the
declared one, so without a canonical hash the two would address different
catalog slices.
* fix: Preserve MCP request header ownership across lifecycle boundaries
* fix: Narrow remote header identity before rebuilding config
* fix: Preserve request header precedence over catalog API keys
* docs: Synchronize generated agent API schemas
* fix: Ignore shadowed generated MCP user key requirements
---------
Co-authored-by: Lia <lia@librechat.ai>
Co-authored-by: Danny Avila <danny@librechat.ai>
Regenerates packages/api/openapi/agents.openapi.json from the code so
openapi:check passes again. #16008 added repositoryInstructions to the
agent schema and merged before #15928, whose committed spec was
generated from a branch that predated the field, leaving dev red.
Co-authored-by: Lia <lia@librechat.ai>
* 🔌 feat: OpenAPI spec and docs for the public Agents API
* 🔧 fix: address review on Agents API OpenAPI spec and docs:
- adapter: mark JSON request bodies as required
- openapi.js: move gate, config read and spec serving into packages/api
as an injected router factory; leave /api as wiring only
- docs: compute URLs in-browser and use a relative servers url so docs
work under a base path (e.g. /chat)
- build: copy the generated spec into dist so the multi-stage image ships it
- agents: reuse the enforced list envelope for the docs so the AgentList
schema cannot drift from the runtime cursor and strictness
* 🔧 fix: align Agents API OpenAPI contract with enforced runtime
Second review round on the public Agents API spec. Three cases where the
published contract described something the server does not do, so generated
clients would fail against the live API:
- Upload form: document the `purpose` field (enum: file_search, execute_code,
context) that POST /agents/{id}/files actually requires, instead of the
internal-only `tool_resource`; requests following the old form were rejected
with 400 because `purpose` was absent.
- Auth: drop the `apiKeyBearer` scheme from the management and skill endpoints.
They gate on OIDC/M2M via requireAgentManagementAuth and deliberately reject
the API-key fallback; the API key belongs to the out-of-scope
OpenAI-compatible endpoints and returns when those are documented.
- Skill update: document the 409 conflict PATCH /skills/{id} returns on a stale
expectedVersion. The agent update path collapses its internal 409 to 400, so
it correctly keeps no 409.
* 🔧 fix: serve docs on experimental server and document the 401
Third review round on the public Agents API spec.
- Experimental server: mount routes.openapi in api/server/experimental.js before its /api 404 handler. Without it, npm run backend:experimental returns 404 for /api/docs and /api/openapi.json even when openapi.enabled is true, because that server never mounted the router.
- Auth failure: document the 401 that requireAgentManagementAuth returns on a missing, expired, malformed, or unbound token. Its body is { "error": "Unauthorized" }, a flatter shape than the management error schema, so add a dedicated UnauthorizedError schema and a shared 401 response to the agent and skill contracts rather than reusing the standard error schema.
* 🔧 fix: document skill-update, auth-conflict, upload-limit, and watch-copy gaps
* 🔧 fix: added error shapes for permission denied and internal server errors in API responses docs
* fix: cover the auth-layer 500 shape and pin the account-deletion code
* fix: update OpenApiRouterDeps documentation and simplify request handling in isEnabled function
* 🔧 fix: gate the Swagger asset mount behind the OpenAPI toggle
* 🔧 fix: accept the malformed-JSON 400 envelope in the contract
* fix: accept the tenant-isolation 403 envelope in the contract
* fix: add description for content property in SkillFileUpdateRequest to clarify byte limit
* 🧼 fix: Strip Cache-Bust Query Before Local Vision Encode
Reused code-interpreter images persist filepath with a ?v= suffix.
prepareImagesLocal now strips it before disk encode, matching crud/share.
Co-authored-by: Zsanz3 <Zsanz3@users.noreply.github.com>
* 🧪 test: Isolate prepareImagesLocal query-strip coverage
Mock sharp and resize so the local vision-encode test does not load
the rest of the file-strategy graph.
Co-authored-by: Zsanz3 <Zsanz3@users.noreply.github.com>
* 🪒 fix: Strip Cache-Bust Suffixes From Every Local File Read
Reused code-interpreter outputs persist a `?v=<timestamp>` suffix on the file
document's filepath, and local storage resolves that field into a filesystem path.
prepareImagesLocal already handled it after the cherry-picked commits; getLocalFileStream
did not, so every download-stream consumer (the download route, provisioning, agent and
skill file reads) still hit ENOENT for a regenerated image.
Move the strip into one documented helper and call it from all three local read paths.
* 🧱 refactor: Move Cache-Bust Stripping Into the Storage Package
`/api` holds wiring, not behavior, so `stripCacheBust` moves out of a new CJS helper and
into `packages/api/src/storage/path.ts`, beside `resolveDownloadPath` — the function that
already decides what a stored filepath means to a download. Its doc comment now states why
the two differ: a remote strategy's query string can carry a presigned signature, so only
the local paths may strip.
The local storage modules keep just the call into it. Semantics are covered by the storage
package's own test; the `/api` specs pin the wiring.
Addresses the Codex P1 finding on 5f8dbf3cf.
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Zsanz3 <Zsanz3@users.noreply.github.com>
Co-authored-by: Lia <lia@librechat.ai>
* 🎚️ fix: Allow Opting Out of the MCP OAuth `resource` Parameter
Adds `mcpServers.<name>.oauth.send_resource_parameter`, defaulting to true, so
operators whose authorization server rejects RFC 8707 `resource` can complete
the MCP OAuth flow. Suppression applies at send time only: Protected Resource
Metadata is still discovered, still validated against the server URL
(RFC 9728 §3.3) and still recorded on the stored client binding.
* 🧹 chore: Drop Accidentally Tracked node_modules Symlinks
Worktree-local symlinks were committed because .gitignore matches node_modules/ as a directory, not as a symlink of that name. They pointed outside the repository, so CI dependency installation failed and every downstream check with it.
* 🧪 test: Cover the MCP OAuth `resource` Opt-Out at Token Exchange
The authorization_code exchange reads the decision from flow metadata rather than live config, so a code obtained without `resource` is never exchanged with one. Covers both the opted-out flow and a flow initiated before the field existed, which must keep sending the parameter.
* 💅 style: Apply Prettier to the Token Exchange Test
* 🔒 fix: Make the MCP OAuth `resource` Opt-Out Actually Take Effect
Addresses three Codex findings on the opt-out:
- `startAuthorization` copies the configured `authorization_url` verbatim, so a
`resource` an admin left in that URL survived into the request and Entra kept
rejecting the flow. Opting out now deletes the parameter on both authorize
paths instead of only declining to add one.
- `.omit()` let a user-managed submission carrying `send_resource_parameter`
parse successfully while discarding the field, so the opt-out was silently
ignored. It is now `z.never().optional()`, matching the documented contract
and the `audience` precedent.
- A pending flow built with `resource` could be replayed for the full pending
window after an operator set the option, reissuing the request they had just
reconfigured away from. `isCurrentServerOAuthFlow` now compares the captured
decision with live config. Stored tokens are unaffected, so changing the
option still does not force re-authentication.
* 🧯 fix: Suppress Inherited MCP OAuth `resource` and Stale Replays
Addresses two further Codex findings on the opt-out:
- The MCP SDK uses `token_endpoint` verbatim and the refresh paths post to the
resolved token URL as-is, so a `resource` an admin left in `token_url` (or one
on a discovered token endpoint) still reached the provider. The outbound URL is
now sanitized at the exchange and in all three refresh paths. Stored metadata
keeps its configured form so the client binding still matches.
- Pending-flow replay was only guarded inside MCPConnectionFactory, so the
browser initiate endpoint could still redirect to the stored URL. The shared
pending helper now takes optional live config and refuses a replay whose
captured decision no longer matches; the route passes it.
* 🪤 fix: Scope the MCP OAuth `resource` Opt-Out to Paths It Can Serve Correctly
The browser initiate route compared the pending flow against `resolveConfigServers`,
which deliberately omits unmodified YAML servers (see `mcp/context.ts`). For the very
config this option targets that read a captured `false` against `undefined`, reported
drift on every initiation, and then returned 400 because factory-written flow states
carry no `oauth` property to rebuild from. Reverted rather than patched: doing it right
needs effective-config resolution behind a `packages/api` helper, per the `/api`
wiring-only rule, and this sandbox cannot run that route suite.
Retained and corrected:
- The pending-flow guard no longer rides on `isCurrentServerOAuthFlow`, which also
gates reuse of a recent COMPLETED flow. It was invalidating already-issued tokens,
contradicting the promise that changing this option never forces re-authentication.
Now a PENDING-only predicate at the two replay sites.
- `ServerConfigsDB.sanitizeUserManagedOAuthConfig` strips `send_resource_parameter`, so
a DB-backed or imported user config cannot activate the admin-only opt-out.
---------
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: Restore OpenID Refresh After the Session Store TTL Expires
express-session raises `failed to load session` only when the store holds no
record for the session id. #15655 treated that as terminal, so a refresh whose
cookie refresh token was still valid ended in 401 OPENID_SESSION_MISSING once
the OpenID session store TTL (15 minutes by default) elapsed before the IdP
access token (~1 hour with Entra ID).
Reload the persisted record through one tolerant helper: an absent record is an
empty session, not a session that advanced past this result, so publication
seeds a new record. Store outages still propagate, and logout still fails a
retired refresh token through its durable revoked publication flight.
* 🧪 test: Pin the logout fence around the tolerated session reload
Tolerating an absent session record leaves the durable revoked publication
flight as the only thing between a logged-out refresh token and a resurrected
session, so cover both of its checks with an expired record: the tombstone read
that precedes the reload, and the completion that follows it.
Drop a needless optional chain in publishResolvedSessionTokens, whose `req`
parameter is not optional.
---------
Co-authored-by: Lia <lia@librechat.ai>
* 🧷 feat: Carry Ask User Answers Past Compaction
The user's answers to `ask_user_question` lived only in the tool result that
received them. A checkpoint summary, the tool-output pruner or a full context
window could remove that message, and the agent then asked again or acted on
a summary's paraphrase.
Every answered ask on the conversation branch is now rendered verbatim into
the run's dynamic instructions, on the first request and on every resume,
where the answer just given is read from the seeded content. Older answers
drop first past a token budget; the newest set is always kept.
New lever: `endpoints.agents.askUserQuestion.retainedAnswers` (`enabled`,
`maxTokens`), on by default with a 4096-token budget.
* 🧷 feat: Carry Ask User Answers Past Compaction
The user's answers to `ask_user_question` lived only in the tool result that
received them. A checkpoint summary, the tool-output pruner or a full context
window could remove that message, and the agent then asked again or acted on
a summary's paraphrase.
Every answered ask on the conversation branch is now quoted verbatim at the
top of the current user turn's prompt copy, in the user's own role: rebuilt
from the stored rows every turn, never persisted, excluded from the memory
copy, and counted in the prompt total with the active encoding. A warm
event-actor turn reads the branch on demand. Older answers drop first past a
token budget that includes separators; the newest set is always kept.
New lever: `endpoints.agents.askUserQuestion.retainedAnswers` (`enabled`,
`maxTokens`), on by default with a 4096-token budget.
* 🧷 refactor: Read The Warm Branch Inside The Module
The warm event-actor read lived as a conditional and a `getMessages` call in
the CJS client. The typed module now takes `historyLoaded`, the identity and
the query as dependencies, decides when to read, and owns the failure policy
for both the read and the block itself; the client passes plumbing only.
* 🧷 fix: Complete The Branch Behind A Summary Cut
`loadHistory` already stops its walk at a checkpoint summary, so the rows
`buildMessages` receives never held the answers this feature carries. The
module now detects a branch that stops short of its root, whatever cut it,
and completes it through the injected stored-row query, stored rows winning.
The block goes on the latest user-authored row so a continued generation
keeps it, and the render measures the whole block before reserving the
omission note, so a set is only dropped when something truly has to go.
* 🧷 fix: Reuse Loaded History And Read Every Ask Stamp
A post-compaction turn read the conversation twice: once in `loadHistory`,
which cuts at the checkpoint summary, and again in the module to complete
the branch. The loader now keeps the rows it read and the module prefers
them, querying only for a turn that made no history read.
Two stored shapes were missed: a pause record whose `args` is the bare
question string, now read as the legacy question, and a call whose input
failed validation, whose persisted error output is no longer quoted as an
answer.
* 🧷 fix: Keep The Block Off The Stored Row And Out Of Memory Copies
Array-form content is replaced instead of edited through the array the
stored row shares, the block's cost is measured as two fresh counts of the
same copy rather than against a possibly calibrated stored count, the
memory copy is only built when memory extraction will read it, the loaded
rows reach the agents client through a hook and are released once the block
exists, and the tool-call-id merge is gone: rows are unique by message id
and providers may reuse ids across turns.
* 🧷 refactor: Apply The Answers Block Inside The Module
The placement lookup, the prepend, the count delta and the memory-copy
decision lived in the client after the context kickoff. They now live in
`applyRetainedAnswers`, which receives the rows, the prompt copies, the
count map, the counter and the memory dependency; the client keeps one
call and passes plumbing.
* fix: Apply Retained Answers After Summary Replay 🧷
* perf: Count Retained Context Exactly Within Its Budget 🧮
* fix: Preserve Retained Counts Across Checkpoint Recounts 🧾
* fix: Isolate Retained Context Counting and Provenance 🧵
* fix: Replace Warm Retained Context and Verify Answer Provenance 🪪
* fix: Use Workspace-Compatible Array Traversal 🧭
* fix: Refresh Retained Context at the Warm Turn Boundary 🧲
* 🪂 feat: Opt-In Text Fallback for Files No Tool Can Read
An upload routed to tools (`llmDeliveryPath: none`) reaches nothing when the
agent handling the turn runs neither Run Code nor File Search, or runs one that
cannot read the type. With `fileConfig.endpoints.<endpoint>.textFallbackWithoutTools`
(also accepted at the top level and inherited from `default`), such a turn
delivers the text extracted at upload instead. The setting is off by default,
which keeps today's behavior.
- Upload: when the setting is on and an inferred `none` route finds no reading
tool, the built-in document parser or native text reader stores the text on
the file, best effort, under the extracted-text size cap and content policy.
- Turn: `resolveTurnLLMDeliveryPath` re-resolves each inferred route against
the tools the agent runs, and `applyTurnTextFallback` marks copies of the
records a turn loads, so endpoint filtering, limits, inspection, history
replay and steer media all see the text the turn delivers. The stored route
is never rewritten.
- Explicit destinations, records predating routing, and turns whose tools are
unknown are left as they are.
* 🧪 test: Resolve Custom Endpoint Fallback Under the Endpoint Name
Uploads route a custom endpoint agent by its saved provider, the endpoint name. initializeAgent marks turn copies before getProviderConfig swaps the provider for the backing client, and the endpoint keeps the name afterward, so the client and child encoders resolve the opt-in under the same endpoint. Cover both the initialization path and the post-initialization agent shape.
* ♻️ refactor: Share One Extracted-Text Storage Cap Across Uploads
The fallback defined its own copy of the 15 MiB cap context uploads already apply to extracted text. Export MAX_STORED_EXTRACTED_TEXT_BYTES from files/extract and use it in both places, documenting that it bounds what a MongoDB file document can hold rather than operator policy.
* 🧭 fix: Keep Fallback Text Wherever a Later Turn Can Deliver It
Extraction skipped uploads filed under a reading tool, so a handoff agent
without that tool, or the same agent after its tools or grants changed, had
no text to fall back to. Store fallback text for every inferred route on a
message attachment; files kept on an agent's tool resources never reach a
prompt and still skip it.
Turn marking only ran while the fallback was enabled, so a stored tool-routed
record whose type the endpoint now routes to text stayed marked `none` and
`extractFileContext` skipped it. applyTurnTextDelivery (renamed from
applyTurnTextFallback) now marks any stored `none` record with text that the
turn resolves to text, and returns early when no such record is loaded.
* 🚦 fix: Admit Every Attachment by the Route Its Turn Delivers
A record upload routed to tools that the current endpoint sends to the
provider stayed marked `none` on the turn copy, so admission skipped it while
BaseClient encoded its bytes, bypassing attachment count and size limits.
applyTurnDelivery (renamed from applyTurnTextDelivery) now gives every record
whose route upload inferred the route the turn resolves, before endpoint
filtering, limits and inspection, matching the run-file encoder.
hasInferredLLMDeliveryPath shares the inferred-route rule with the resolver.
Fallback extraction selected the document parser inside the CJS upload path.
resolveUploadFallbackText now takes the upload and picks parseDocument or
parseTextNative itself, with the extractors injectable, so process.js passes
request data only. Its tests run the real parser against the xlsx fixture and
the real native reader against temporary files.
* 🎚️ fix: Read Custom Endpoint Dialect From Config on Both Sides of Init
initializeAgent materializes turn routes before getProviderConfig swaps the
saved custom endpoint name in agent.provider for its backing client, so
routing read that name as a non-OpenAI dialect. Media a custom endpoint opted
into could then be materialized off the provider path, skipping admission,
while the finalized client still sent the bytes. resolveAgentDeliveryRouting
now reads the dialect with getCustomEndpointProvider, as upload does, which
holds before and after the swap.
The upload fallback plan now keys on the chosen-destination marker the turn
resolver reads, so an upload recorded as chosen, including every legacy
chooser upload, never pays for text a turn cannot deliver.
* 🛡️ fix: Apply Turn Routes Only While the Record Stays Admitted
Resolution while initializing and at delivery can disagree: delivery reads a
Responses API choice only the finished client config holds. Applying a route
that takes a record out of admission, off the model path or to text it never
stored, would then skip the limits for a file the client still sends.
applyTurnDelivery now applies a route only while the record stays model-bound;
admitting a record the turn leaves out cannot slip past a limit.
* 🕹️ feat: Run Declared Project Actions on Attached Workers (#15943)
* feat: use declared attached project actions
* Complete named action definitions and approval previews
* Sort named environment imports
* Sort native environment fixture imports
* 🔓 fix: Keep Code Approval Mode Selectable Mid-Run (#15938)
* 🔓 fix: Keep Code Approval Mode Selectable Mid-Run
The composer's code approval mode selector was disabled while a run streamed, and a mode picked during a run was reverted when the final or abort event merged the server conversation back. The selector now stays usable in flight and both merges retain a mid-run selection, which the next send carries.
* 🔓 fix: Retain Mid-Run Approval Mode Across Cache and Recovery
Stamp the live approval mode onto the submission conversation at send so the retention baseline matches what was read, keep a retained mode in the conversation query cache on final and cancel, and rebuild failed or aborted conversations from a preset that carries the mid-run pick.
* 🔓 fix: Scope Retained Approval Mode to the Submitted Conversation
Retention now requires the live conversation to be the one the run submitted, or the id the server assigned to a new chat, so navigating elsewhere mid-run cannot write that conversation's mode back. A single recovery helper rebuilds failed or aborted conversations from the retained preset and patches the detail cache for every recovery site.
* 🔓 refactor: Keep the Local Approval Mode on Every Server Merge
Replace per-path retention with one rule: a locally picked code approval mode is newer than any server copy of the same conversation, so the final and abort merges and the error recovery preset keep it. The selector writes the pick into the conversation's detail cache, which navigation rebuilds from, so the pick is conversation-scoped without reading the index-global atom from the SSE handlers.
* 🔓 fix: Prefer the Live Conversation When Caching the Final Merge
The final handler's detail-cache merge now takes the local approval mode from the open conversation when it is the one that finished, falling back to the cached record only for a conversation that is no longer on screen.
* 🔓 fix: Keep the Local Approval Mode Through Settled-Start Reconciliation
The resumable transport's settled-start reconciliation and replacement handoff wrote a fetched conversation straight into conversation state and the detail cache. Both now apply the same local-mode rule as the event handlers, treating a new chat's pending id as the settled conversation's own.
* 🔓 fix: Seed the Detail Record When a Mode Is Picked
A conversation that has its id but no detail record yet now gets one from the live conversation when a mode is picked, the same key the resumable transport seeds optimistically, so recovery and navigation find the pick. A chat with no id yet still keeps the pick in conversation state alone.
* ⬆️ chore: bump agents to 3.8.7 (#15947)
* 🧭 refactor: Settle Turn Delivery Routing Once in Agent Initialization
`initializeAgent` resolved the provider and its client options only after the
turn's files were loaded and admitted, and the two delivery readers each rebuilt
the attachment routing from the agent object at their own moment: `BaseClient`
took the custom-endpoint dialect from the already-swapped `agent.provider`, the
run-file encoder from whatever child config it was handed.
Move `getProviderConfig`/`getOptions` ahead of file discovery, where nothing in
between fed them, and settle one `deliveryRouting` value with every input final:
the file policy under the endpoint's own name, the dialect its config declares,
the Responses API decision the model call uses, and the transcription setting.
`InitializedAgent`, the child encoder and `BaseClient` consume that value;
`resolveTurnLLMDeliveryPath` is the one place a stored route is resolved again.
* fix: Preserve attachment reachability across handoffs and late steers
* test: Type the untrusted attachment reference fixture
---------
Co-authored-by: Ravi Kumar L <upman@users.noreply.github.com>