Every send moved the thread's tail (isSubmitting, latestMessageId,
latestMessageDepth, and a handleContinue rebound per turn), and the memoized
row context listed all of them as dependencies, so each move handed every row
a new context object and re-rendered the whole thread: three commits of
~12,000 component renders on a 246-message thread, one of them inside the
Enter keypress.
- Keep the row context object stable; the tail values are ref-backed getters
for call-time reads, and rows receive them as props that the row comparator
(isSameTailRelation) only reacts to when the row enters or leaves the tail.
- Fork reads the split target through a getter at click time.
- HoverButtons subscribes to a boolean isSubmitting context and a stable
operations context instead of the combined messages-view context; Fork and
Feedback are memoized so the per-send toolbar pass stops there.
- Make handleContinue referentially stable.
- Cache the Intl formatters behind getMessageTimestamp per locale and clock.
- Add a submit-lag benchmark (e2e/benchmarks-tree) that reports long tasks,
Enter-to-row latency, a CPU profile and per-commit render reasons.
* feat: Resolve GPT Point Releases to Their Family, Add GPT-6.1 Sol
gpt-6.1-sol matched no token, pricing or family key, so agents ran with the
default 32K context and summarized constantly. Point releases now inherit
their family's context, output limit, pricing, reasoning options and native
Responses routing until listed explicitly; GPT-6.1 Sol joins the catalog.
* test: Type the point-release routing helper as ResponsesApiRouting
* 🫸 fix: Keep Historical Artifacts Closed During Regeneration
* 🫸 fix: Restrict Artifact Panel Selection to the Active Response
* 🫸 fix: Keep Failed Regenerations in Their Original Branch
* 🫸 fix: Isolate Artifact Preview Freshness by Response
---------
Co-authored-by: Lia <lia@librechat.ai>
* 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: Resolve SplitText Direction From Its Own Text Under RTL
SplitText wraps every grapheme and every word in its own inline-block box, so
the bidi algorithm never sees the text as one run. Under an RTL document the
boxes lay out right to left, and a Latin greeting or agent name on the landing
page renders mirrored, letter by letter and word by word.
The paragraph and each word box now carry dir="auto", so direction resolves
from the first strong character of the text itself: the paragraph decides word
order and each word decides its own letter order. RTL text still lays out
right to left, and mixed greetings render each part correctly.
Claude-Session: https://claude.ai/code/session_01EhJ8HUVBZ53ArUzgXDnxom
* 🐛 fix: Preserve RTL Greeting Word Order
---------
Co-authored-by: Lia <lia@librechat.ai>
* 🥧 fix: Show Failed Call Fractions and Preview Age
* 🗑️ chore: Remove Unused Failure Labels
* 🕒 test: Verify Live Failure Close Times
* 🧵 fix: Select Accessible Badge Labels Without Nested Ternaries
* 🪆 fix: Show One Failure Pill for Live Phase Groups
---------
Co-authored-by: Lia <lia@librechat.ai>
* ⏩ fix: Clarify Steering Timing and First-Response Fallback
* style: Format First-Response Tooltip Choice
* 💬 fix: Avoid Implying the First Conversation Has Not Started
* fix: Explain the Steer Sooner Enter Override
---------
Co-authored-by: Lia <lia@librechat.ai>
* fix: validate principal config overrides against configSchema
Reject admin config writes whose fields fail configSchema, and drop
invalid fields from stored overrides at merge time so the base value
from librechat.yaml survives.
* fix: require the merge key on partial custom endpoint overrides
* test: register a dedicated user for config override scenarios
* test: log in once per worker in config override scenarios
* fix: strip invalid overrides under dotted record keys
* fix: ignore a stored overrides document that is not an object
* fix: reject indexed patches into merged-by-name override arrays
* fix: match union options by shape, apply supplied-value refinements, reject paths past scalars
* refactor: validate config overrides on the merged result instead of the patch
* test: cover record refinements on written config fields
* fix: validate config field writes on top of the stored override
* fix: attribute id-addressed issues to their item and keep valid sections after repairs
* fix: judge field writes by what they change, over the principal's tombstoned base
* fix: validate whole-document config writes on the principal's tombstoned base
* fix: report config override issues as stable codes instead of schema messages
* fix: reject indexed writes into keyed arrays and keep filtered MCP servers out of validation state
* fix: judge each override layer only on the values it supplies
* fix: reject merged-by-key array items that are not objects
* fix: report repeated merge keys on the items the keyed merge discards
* fix: reject keys a union option drops when another option defines them
* 🔲 fix: Anchor Quote Popup to Visible Selection
* 🔲 fix: Keep Cross-Container Quotes Anchored to Visible Text
* 🔲 perf: Bound Quote Geometry Walk to Selected Text
---------
Co-authored-by: Lia <lia@librechat.ai>
Sonnet 5.5 (`claude-sonnet-5-5`) rejects `thinking.type: disabled` and takes
`between_tools` as its lowest thinking setting, binds thinking blocks to the
conversation prefix, caps effort at `high` while thinking is off, rejects
sampling parameters, supports the `updates` thinking display, and ships with
128k output / 1M context at Sonnet 5 pricing.
Family gates in data-provider now use "5.5 or later" semantics so Opus 5.6+,
Sonnet 5.6+ and the Mythos-class line (Fable) inherit the newest contract
instead of falling back to the Opus 5 / Sonnet 5 paths.
* 🔒 fix: Guard 2FA Deep-Link Redirects Against Blocked Session Storage
The post-login redirect store was the one session-storage consumer in the
client that reached for it bare-handed: a context where storage access
throws crashed the login screen's persist effect and the post-challenge
boot instead of dropping the destination. Route every access through
throwing-safe helpers so a blocked store costs the deep link, never the
sign-in, and pin the ordinary 2FA challenge's deep-link survival and
consumption with mock-harness scenarios plus blocked-storage unit tests.
The challenge screen's code inputs also gain accessible names.
* ♿ fix: Name the 2FA Verify Button by Its Visible Text
The challenge submit button carried aria-label 'Continue' over the visible
text 'Verify', so its accessible name did not contain the visible label
(WCAG 2.5.3). The visible text names the button on its own; the label is
dropped.
* 🧪 test: Keep the 2FA Mock Harness From Re-Declaring Destinations
The shell's authenticated queries 401 against the real backend under the
stand-in bearer, and the auth-recovery login bounce carries the current URL
as redirect_to, re-persisting the deep link mid-test. Answer those queries
empty and leave /api/config real.
* 🧪 test: Answer the Shell's Empty Shapes in the 2FA Deep-Link Harness
The composer and its selectors iterate their query results, so the stand-in
session's shell queries need per-endpoint empty payloads rather than a
blank object; the destination's marker rides the query string so no
conversation has to exist for the landing to hold; and the second sign-in
declares its own destination, which is the deterministic form of
consumption-across-sign-ins.
* 🧪 test: Assert the 2FA Shell Across Viewports
The account button hides in the mobile drawer and the drawer never
satisfies a pointer click's stability check, so the shell assertion polls
for whichever marker the viewport shows and the session ends through an
in-page logout request, which the route mocks can actually see.
* 🧪 test: Locate the Login Password Field by Label
* 🧪 test: Let the CI Storage Polyfill Replace the Blocked Store
The CI-mode build ships a storage polyfill that answers a broken
sessionStorage by installing its own in-memory shim with
defineProperty, which crashed against a non-configurable blocker and
took the app boot with it. The blocker stays configurable: the shim is
per-document, so the challenge's document swap still empties it and the
landing contract is the same either way.
Route modules build their rate limiters while they load, and both server
entries required them before startup checks copied rateLimits from
librechat.yaml into the environment. The conversationsImport, tts, stt
and route-scoped fileUploads budgets therefore ignored the yaml unless
the matching environment variable was also set. Require the routes only
after performStartupChecks in index.js and experimental.js.
* 📡 fix: Explain Model Streams That Die Mid-Response
* 📡 fix: Bound Model Stream Failures to Response Bodies
---------
Co-authored-by: Lia <lia@librechat.ai>