Opening a long conversation built a toolbar for every message as its rows
mounted: the copy, edit, branch, feedback, read-aloud and rerun buttons
that show on hover or under the latest message. That was close to half of
what opening a 246-row thread cost the main thread, and most of it in a
thread four times as long.
A row a long thread adds after its first commit now leaves its toolbar out
until the thread's row observer finds it within a screen of the viewport,
then keeps it. The footer holds the toolbar's height throughout. The rows
a thread opens on, the latest message, short threads and rows outside the
main thread view are unchanged.
The copy-last-response and edit-last-message shortcuts reach their row
first, and screenshot export mounts every toolbar before it clones.
Adds e2e/benchmarks-fade/open.perf.spec.ts, which traces a reload, a switch
from the sidebar and a switch while a reply is in flight.
* 🎯 feat: Let @ Pick the Model of a Conversation Open Beside the Chat
* 🎚️ fix: Hand the Modular Chat Preference to the Mention List Instead of Reading the Store
* 🚪 feat: Show the Session Rail When Work Starts Running
The rail now shows itself when work first starts running in a conversation,
on screens wide enough to dock it, so a reader does not have to notice the dot
on the header control. It takes no focus and never opens as the small-screen
sheet. Showing itself is not the reader's standing choice: it is not
remembered and ends with the conversation. Closing the rail is an answer, and
it is not shown again for that conversation, whatever starts later.
interface.sessionRailAutoOpen (default true) turns it off.
* 🚪 feat: Let a Reader Keep the Session Rail From Showing Itself
A rail that shows itself for every conversation's work is one a reader who
does not want it has to close every time. Two ways out, both the reader's:
Closing the rail is now remembered for that conversation in local storage, so
it is not shown again there on this visit or after coming back to it. The
memory is the last hundred conversations; opening the rail by hand takes a
conversation off it.
Settings > General gains "Open the session rail when work starts", on by
default and offered only where the deployment has the rail and lets it show
itself.
* ⏱️ fix: Time a Finished Subagent From Its Dispatch, Not From Its Reply
Every finished subagent read "Completed 0s" in the session rail and the
background tasks popover. A child task is two records, its brief and its
reply, and the parent's index reports the time of the newest one: the
dispatch while the task runs, the end once it has replied. The client took
that for the start, so a finished task's duration was its end minus its end.
The index now also reports startedAt, the brief's time, read in the same two
queries. The client times a finished task from startedAt to its reply, and
shows no duration rather than a wrong one when the start is not known (an
older server, or a brief outside the rows the index read).
* ⏱️ fix: End a Finished Subagent at Its Thread's Timestamp, Not Its Reply's
A child that pauses for an approval writes its reply at the pause and updates
that same record when it resumes and ends. Taking the reply's creation time
for the end left out everything after the approval, and aged a task that had
just finished out of the list if it had waited over an hour for one. The end
is the thread's timestamp again, as it was; only the start is new.
The dialog lists a server's tools once, for all agents. With the
whole-server item ticked and a named tool left out, the agent that
selects the whole server could call that tool while an agent that names
it was refused, and the tool showed as unticked under copy that says
anything unticked is refused.
Ticking the whole-server item now ticks every named tool on the server
and locks those boxes; unticking it leaves them ticked for the owner to
choose from. The item sits above the tick-all box instead of under it,
and its label wraps at word boundaries on a narrow screen.
* 🗄️ feat: Let a Schedule's Owner Allow Every Tool on a Declared MCP Server
An agent may select a declared server as a whole instead of naming its tools. An
owner authorizing tool by tool had nothing to tick for that, so those tools were
left out unless they allowed the whole agent. The selection is now one item for the
server, and it stands for every tool there, including ones added later.
* 📌 fix: Keep Pins, Approvals and Change Warnings Working Under a Whole-Server Grant
Renewing a grant of the whole server dropped the pins of the tools used under it,
since none is named in it. An agent granted the whole server was not paused for a
tool the owner approves each time. And the dialog showed no warning, and left the
item ticked, when a tool used under it had changed.
* 🙋 fix: Keep a Stored Approval for a Tool the Whole-Server Item Still Covers
A question about a named tool was dropped on renewal once no agent named the tool,
even where an agent still reached it through a grant of the whole server. The tool
then ran unattended. It is kept while anything on the server still covers it, by
the service and in what the dialog sends.
* 🧾 fix: Keep Every Whole-Server Warning, Question and Stored Choice in View
A changed tool on a server now unticks the whole-server item as well as
the tool's own box, so the owner looks at every change the item covers.
A question the grant holds about a tool with no ticked box of its own
gets a row under the whole-server item, where the owner can keep it or
stop it. Tools a stored whole-server grant already allows start ticked
once the agents name them.
* fix: Drop the Agent Icon From the Greeting Welcome Heading
* docs: Describe the Greeting Heading Without Its Icon in the Example Config
* test: Cover the Greeting Heading Without the Agent Icon in E2E Scenarios
* 🔌 feat: Let a Schedule's Owner Allow Tools on an MCP Server the Operator Did Not Declare
An owner who authorizes a schedule tool by tool could tick only tools on servers
declared under mcpConsent.resources. A local server, one they added, or one the
operator never declared was covered only by a grant of the whole agent. Its tools
are now offered among the agent's other tools, each under its key, and a selection
of the whole server as one item.
* 🏷️ fix: Check Listed Tools at Creation and Label Them Against the Servers in Use
Readiness for a schedule being created read only the grant's consents, so a tool
listed on an undeclared server was set aside and its server never probed. The label
for such a tool split its key at the last delimiter, which is wrong for a server
name that contains one; the view now names the servers in use and the key is read
against them.
* ✍️ feat: Authorize a Schedule Again in the Same Step That Edits It
Editing a schedule ended its grant, and the edit form had no way to make it again:
the owner saved, then had to open the authorization dialog and confirm. The edit
form now carries the same section as the create form, starting from the grant the
schedule has, and the edit writes the new grant with it.
* 🚧 fix: Fence an Edit's New Grant on the Grant It Was Decided Against
An edit that carries an authorization writes the grant with the schedule, fenced on
the schedule's edit generation. Consent writes do not move that generation, so the
edit could undo a revocation, or drop a pin or mark a run added, made after the new
grant was decided. The write now names the grant it replaces and is refused unless
that is still what is stored, untouched.
* ✅ test: Expect the Lifetime the Edit Form Now Starts At
* 🔂 fix: Check the Grant an Edit Replaces on the Decision's Final Read
The decision reads the stored grant twice and compared the owner's expected revision
with the first read only. A revocation landing between the two was then taken as the
grant to replace, and the edit's fenced write matched it. The final read is now held
to the same revision.
An owner who authorizes a schedule tool by tool had no way to allow its agent's
skills: a skill is not a tool anyone could tick, so the skill tools were refused.
Skills are now one item per agent that uses them. Allowed, the agent runs with its
skills as in chat. Left out, none is resolved and no skill tool is registered.
The lifetime field started at one hour in the create form and the authorization
dialog. A schedule authorized without changing it lost its tools an hour later,
before a daily schedule's next run. It now starts at the operator's limit. An API
key's grant still starts at one hour.
With the whole agent ticked the dialog only named the tools a schedule still asked
about. For a server its agents select as a whole there is no tool box to untick, so
the owner had no way to stop it short of revoking. The note is now a checkbox.
* 🛤️ feat: Session Rail for Subagents and Background Tasks
A column beside the conversation, opened from the chat header, that lists what
the conversation has running: its subagents, then its background tools. It
replaces the header popover, whose control only appeared once there was
something to list. The control is now always offered, and the rail keeps its
own column past the side panel, so a subagent opened from it stays in view of
the list. A running subagent's row says what the child is on now, read from
the thread the page already follows for the card.
interface.sessionRail (default true) turns it off and restores the popover.
* 🪤 fix: Hold Focus in the Session Rail's Small-Screen Sheet
The sheet covers the conversation, so it now traps Tab and answers Escape
through the shared focus trap the other sheets use, and says it is modal.
Focus still starts on the close button rather than on Stop all.
* 🪶 perf: Render Nothing but the Rail When the Session Rail Toggles
Opening or closing the rail changes the chat column's width once. The composer
bar kept its row's measured width in state, so it and everything under it
rendered for every frame of the composer's width change: 6 times on open and
18 on close, and the same on a sidebar toggle. It now reads only whether its
chips fit, through useElementSizeSelect, and renders when that answer changes.
The scan that titles background tool rows from the loaded messages, which runs
on each update of a streaming thread while the rail is open, now passes over
every tool call no task names before building its keys.
Adds rail.perf.spec.ts beside the sidebar's benchmark.
* 🪶 perf: Keep an Open Session Rail Still While the Thread Streams
Measured by replaying a recorded reply into long agent chats with the rail
closed and then open and listing running work.
A once-a-second clock held by the header control rebuilt every row object, so
every row rendered each second while anything ran. The clock now belongs to
the duration text: one shared timer, read only by the running rows' elapsed
times, and a row renders when what it shows changes.
A running row's title no longer sweeps. Each sweep is three compositor layers,
and every frame of the streaming thread paid for them in style, layerize and
commit: six running rows took the page from 9 layers to 33. The row keeps its
pulsing dot and its counting duration.
The rail's layout slot is memoized: its parent renders for every stream frame.
The replay benchmark gains FADE_PERF_RAIL, and rail.perf.spec.ts a test that
grows the last reply beside the rail and prints the layer count.
* 🧰 feat: Authorize a Schedule for Its Whole Agent, by Default
A schedule's authorization listed tools one by one, so a scheduled agent ran with
less than the same agent in chat: no skills, no MCP server the operator had not
declared, and nothing added to the agent after the owner confirmed.
A grant can now be made for the whole agent. It lists nothing and leaves nothing
out, including what is added later, and it is where a new authorization starts.
Listing tools one by one stays available for an owner who wants to restrict.
* 🛂 fix: Keep Live Checks, Whole-Server Selections and Credential Preparation Working Under a Whole-Agent Grant
A call on a server the grant holds no consent for now goes through the authority:
the owner's permissions and the deployment's stop are checked as for any granted call.
A selection of every tool on a server is reported under its marker, so a server the
grant holds a consent for stays resolvable when the agents switch to it. Credential
preparation is admitted for what the agents select now, not the list at confirm time.
* 📝 docs: Describe the Standing Check and Whole-Server Selections Under a Whole-Agent Grant
* 🧷 fix: Cover Whole-Server Selections and Later-Added Declared Servers in a Whole-Agent Grant
A grant of the whole agent is built from every declared server its agents use,
including one selected as a whole, so the server has a consent for its credentials.
A declared server taken up after the grant was made is authorized from the operator's
current declaration, under the grant's own lifetime and binding.
* 🧲 fix: Keep Whole-Server Selections in Readiness and in the View's Recipient Check
Readiness dropped a server selected as a whole before connecting, so a whole-agent
schedule could be created or fired without its readiness being checked there. The
view compared a whole-agent grant's recipients with the listed servers only, and so
reported an authorization as active after such a server's declaration had moved.
* 🙋 fix: Keep Per-Use Approvals Through a Whole-Agent Renewal
A stored "ask me first" was dropped when a grant was renewed for the whole agent
after the agents came to select the tool's server as a whole, because no tool is
named there to keep it for. A whole-agent grant now keeps a question for any tool
on a server it holds, a run pauses for it whichever agent calls, and the dialog
says which tools are still asked about.
* 🧭 fix: Report a Whole-Agent Grant as Changed When a Server It Uses Lost Its Declaration
The view treated a server missing from the offer as one the agents had dropped. A
server they still use that is no longer declared, or is declared for another URL,
is also missing there, and its calls are refused. The resolver now reports every
server in use by name, and the view tells the two apart.
* 🧗 fix: Settle a Side Chat's Run From Its Status When Reconnecting Gives Up
* 🔁 fix: Keep a Lost Side-Chat Run Until a Server of Its Protocol Says It Is Over
* 🪪 feat: Authorize a Schedule in the Same Step That Creates It
Where consent is on, a new schedule was created unauthorized and stayed exempt until
its owner edited it or the operator's deadline passed. The create form now asks what the
schedule may use, from the offer for its agent, and the schedule is created with its
grant in one write. A schedule created without a choice is marked as needing
authorization from the start.
The grant dialog's tool list and choice are split out so the form shows the same thing.
* 🔁 fix: Read the Offer Again When a Create's Authorization Is Turned Down
The create form kept the offer it first read. Once the server refused a choice made
from it, every retry sent the same refused digest. The form now reads the offer again,
keeps the ticks that still apply, and holds the create until the new offer is in.
* 🧭 fix: Check Readiness at Creation for the Tools the Owner Allowed
A schedule being created has no stored grant, so its readiness check covered every
MCP tool the agents select, including servers the owner left out. The grant is now
validated first and handed to the check, which covers what it allows.
* 🗓️ feat: Let Any Schedule Be Authorized, With or Without an MCP Server
A grant needed at least one MCP tool, so a schedule whose agents use no MCP server, or
none the operator declared, had nothing to authorize and its other tools were never
gated. A grant can now allow the agents' other tools alone. It has no consent to carry a
deadline, so every grant stores its own lifetime; one made before that is read from its
consents.
A schedule that needs a grant it does not have was refused its MCP calls only. Every
tool call is now refused with the same reason until the owner confirms.
* 🧯 fix: Bound the Held-Run Lookup, Accept a Grant With No MCP Tool at Readiness, and Authorize Other Tools Without MCP Use
The run hook gives its two reads one deadline, so a slow grant lookup is refused
before the SDK discards the hook. Readiness has nothing to connect when the grant
allows none of the selected MCP tools. Authorizing a schedule needs the schedules
permission and the agent; only MCP tools are held to the MCP servers permission.
* 🚦 fix: Ask for the MCP Permission After the Grant Filter at Readiness
A granted run needs the MCP servers permission only for the MCP tools its grant
allows. Readiness asked for it first, so an owner without it who authorized only
other tools had a schedule that could never activate or fire.
* 📈 fix: Keep the Context Gauge Counting While a Reply Streams
* 📊 feat: Feed a Side Chat's Usage to Its Composer's Gauge
* 📉 fix: Drop a Failed Side Chat Reply's Counted Usage
* 🧰 feat: Let a Schedule's Owner Authorize Each of the Agent's Tools, or Everything at Once
A schedule's grant could only be offered for an agent whose tools were all MCP tools. An
agent that also had code execution, web search, file search or an action could not be
authorized at all, and once its owner edited the schedule or the operator's deadline
passed its MCP tools stayed refused with no way out.
Each of an agent's other tools is now its own item in the offer, allowed or left out like
an MCP tool. The grant stores the ones the owner allowed, per agent. Under a grant, one
the owner left out is not loaded for the run and does not stop it; one the owner allowed
loads and runs, with the code environment's tools all allowed by code execution. The
dialog lists them in their own group and gains one control that allows every server and
tool at once. An API key's grant still limits MCP tools only.
* 🧰 fix: Hold a Schedule's Other Tools to the Grant's Deadline and Revocation
An MCP call is checked against the stored grant when it is made. The agents' other tools
have no such call, so a grant that had expired or been revoked still loaded and allowed
them. The run now holds them to the grant it read when it started: none loads from a
revoked grant, and they stop at the grant's deadline.
* 🧰 fix: Leave Out a Schedule's Unallowed Tools Where the Initializer Sets Them Up
The grant narrowed what the loader was asked for, but the initializer also reads the
agent's tool list for the markers that set up code execution and memory. A tool the owner
left out was still defined for the model and only refused when called. The list is now
narrowed once, where the initializer takes the agent, so everything below reads the
owner's choice.
Memory runs as two tools of other names. An allowed memory item now covers both at the
run's hook, as code execution covers the code environment's tools.
* ✨ feat: Support Model Spec Handoff Targets
* fix: Preserve Effective Config and Transcript Identity for Preset Handoffs
* test: Provide Query Client for Search Transcript Handoff Fixture
* fix: Match the Subagent Icon Display Contract
* 🧱 fix: Bound Attachment Reads and Keep Bedrock's Document Limit
`getFileStream` buffered a stored object in full for every provider and every
kind of attachment, and the Bedrock document limit was `configured ??
provider`. The merged file config always carries a size limit (512 MB by
default), so with any `fileConfig` section Bedrock's 4.5 MB limit never
applied: an 8 MiB CSV was downloaded, decoded and serialized before Bedrock
refused it.
The shared reader takes a byte limit and stops at the first chunk past it. A
Bedrock document is refused from its recorded size before storage is opened,
read under that limit so an object larger than its record cannot reach memory
either, and reported as a 413 attachment error. A configured limit can lower
Bedrock's limit and no longer raise it.
* fix: Let a Set Limit Override Bedrock's, Keep Replay Tolerant
A limit the operator set overrides Bedrock's own, up or down, as it does for
other providers. What no longer counts is the 512 MB the merge fills in for
every endpoint, which is how the provider limit was lost whenever a
`fileConfig` section existed.
Replayed history skips the recorded-size refusal so an object that is gone is
still reported missing and left out; the bounded read refuses one that is
present. The client shows the size refusal under its attachment-limit
headline, not as a provider failure.
* 🔆 feat: Add GPT-6.2 Sol Support
* test: Correct Sol Request Helper Types
* fix: Apply Sol Constraints at Final Request Boundaries
---------
Co-authored-by: Lia <lia@librechat.ai>
* 🪟 feat: Open a Conversation Beside the Chat From Its Menu
A conversation's menu in the sidebar offers to open it as a side chat
beside the chat on screen. The item appears only where there is a chat
to sit beside and side chats are on, and not for the chat on screen or
an archived one. The conversation's full record is read first, since a
side chat builds its turns from it.
* 🪟 fix: Give a Side Chat the Panel Slot It Was Asked For, Beside the Chat It Was Asked Beside
A side chat opened while an artifact or a child's activity is showing
now takes their place, as each of them takes its. A record lookup that
finishes after the reader has moved to another chat, or to the very
conversation asked for, opens nothing.
* 🪟 fix: Let the Newest Side Chat Request Win, and Test the Real Menu
An older Open beside chat whose record arrives late no longer replaces
the reader's newer choice, or a side chat opened another way meanwhile.
The item's tests now drive the production menu, hooks and strings, and
replace only the request for the conversation's record.
* 🪟 fix: Withdraw a Pending Side Chat for a Newer Panel, and Open It Clear of the Drawer
A side chat still being fetched no longer arrives over an artifact or a
child's activity opened after it was asked for. On a small screen the
sidebar drawer leaves first, the way it does when a conversation is
selected, so the side chat is not opened behind it. Once its
conversation is in, focus moves to its composer.
* 🪟 fix: Ask Again Whether a Side Chat Is Still Wanted Once the Drawer Has Left
The drawer takes a few frames to leave, and the checks that a side chat
was still the reader's newest ask ran before that. The request now
stays theirs until the side chat is written, and is asked about again
at that point.
* 🧪 test: Answer the Real Config Query in the Presentation's Side Chat Tests
* 🗝️ feat: Store an API Key's MCP Grants on the Key, One per Agent
* 🧬 refactor: Let the Consent Service and Enforcement Act for Any Grant Subject
* 🗝️ feat: Check an API Key Run's MCP Calls Against the Key's Grant for the Agent
* 🗝️ feat: Put Remote API Runs Under Their Key's Grant, Behind a Config Switch
* 🩹 fix: Export the Key Grant Storage the Server Wiring Loads
* 🔌 feat: Let an Owner Grant an API Key Its MCP Access From Settings
* 🗝️ fix: Decide Key Grant Access, Mode and Kind of Key in the TypeScript Host
The agent-access check and the grant-mode lookup lived in the legacy route and service
files. The host now builds the check from an injected agent reader and permission
reader, and the key handlers resolve the mode from an injected configuration reader.
Each authorized call read the owner three times and their configuration twice. The host
now reads both once per call and shares them between the access check, the limits and
the offer; nothing is kept between calls.
A background worker rebuilds its request from the admitted record, which names the key
but not its kind, so a key issued to an integration lost its exemption and its channel
runs were refused. The host asks the key store when the request carries no key itself.
* 🔌 fix: Say When a Key Grant's Agent List Is Loading, Failed or Empty
The key grant dialog turned a failed or pending agent list into an empty picker with an
instruction to choose from it. It now shows a spinner while the list loads, an error
with a retry when it could not be read and nothing is cached, and a plain statement when
the owner has no agent the key could run.
* 🧪 test: Hold the Pin Fences on an API Key's Per-Agent Grants
* 🔌 fix: Offer a Key Grant for the Agents the Key Can Run, and Only for Personal Keys
The grant dialog's picker read the app's agent catalog, which lists agents the owner can
view. Running an agent with a key follows another rule: the agent is shared with the
owner for remote use, or is theirs to share. An agent shared only for remote use could be
run but not granted. A new list, GET /api/api-keys/agents, applies that rule to all
agents, and the picker reads it.
A key issued to an integration holds no grants but showed the MCP access button. The key
list now marks such a key, and the button is not shown for it.
* 🙋 feat: Let a Schedule's Owner Approve a Tool Before Each Use
* 🧪 test: Drive a Real Run Through a Schedule's Approval Pause and Resume
* 🙋 fix: Keep a Stored Ask-First Choice When a Run Can No Longer Pause
Where pausing became unavailable, the dialog hid the ask-first boxes and renewed the
grant without them, so a tool the owner approved each time started running unattended.
The dialog now starts from the stored choice either way and always sends it. Where a run
cannot pause it shows only the stored questions, with a note that those tools are refused
for now, and the owner can keep or drop them but add none. The server keeps the stored
questions for tools still allowed when a confirmation says nothing about asking, and
refuses only a new question where a run cannot pause.
* 🙋 fix: Say Why a Schedule Asks for Approval in the Reader's Language
The approval card showed an English sentence sent by the server as the request's
description. The server now sends a code, and the card maps it to a localized string.
Any other description is shown as sent.
* 🙋 fix: Never Offer Always Allow for a Tool a Schedule Asks About Each Use
Where the endpoint enables Always allow, a schedule's approval card offered it for a tool
the owner approves on each use. Choosing it stored a conversation allow the schedule
still overrode at the next use. The offer and the recheck at resume now take a predicate
for anything beyond the policy that asks on every use, and the schedule's grant supplies
it, so such a tool is neither offered nor stored.
* 🙋 fix: Let the Owner Ask First About as Many Tools as They Can Allow
The confirmation accepted up to 25,600 allowed tools per server but only 256 to ask
about, so a large grant could not be confirmed with every tool set to ask first.
* 🙋 fix: Keep a Tool Asking First When the Owner Allows Its Changed Definition
The dialog derived the ask-first boxes from the ticked tools. A tool whose definition
changed starts unticked, so its stored question was dropped, and ticking the tool again
confirmed it as unattended. The ask-first selection is now kept apart from the ticks,
starts from the stored grant, and is narrowed to the ticked tools only when the
confirmation is sent. The owner's edits to it hold only for the grant they were made
against, so a question added to a grant confirmed elsewhere is not dropped unseen.
* 🪟 feat: Continue a Subagent's Thread as a Chat Beside the Conversation
* 🪟 fix: Take a Side Chat's Streamed Attachments Where Main Chat's Go
What a tool hands back while a side chat's generation runs (search
sources, a generated file) reached no handler, so it was missing from
the rows and from the app's file, memory and preview caches. The side
chat's stream now uses the app's attachment handler.
* 📜 fix: Read a Child Thread's Messages a Page at a Time
The child messages route read a thread's whole stored history, and a
card previewing a running child triggers that read without anyone
asking. Every read is now one page (50 rows or 2 MiB, newest first)
with a cursor for the page before it, the followed pane keeps the pages
it has read and offers the earlier ones, and a client or server that
predates paging still gets a list.
* 📜 fix: Start a New Follower From Its Own Newest Page
A follower recreated over rows another one left in the cache kept them
without knowing where they began, so the earlier-messages control was
gone. It now places its first page itself and takes that page's
cursor.
* 📌 feat: Pin an Allowed Schedule Tool to the Definition It First Ran With
* 📌 fix: Bind a Confirmation to the Changed Tools the Owner Was Shown
* 📌 fix: Untick a Tool Marked as Changed While the Owner Is Choosing
The dialog reads the offer again while it is open. A tool marked as changed after the
owner's last tick kept its tick, so confirming would have allowed the new definition
without the warning having been in view when the box was ticked. The edit now records
which tools were marked when it was made, and a tool marked since then is unticked
until the owner ticks it again.
* 📌 fix: Check Pins and Marks in the Writes That Admit, Confirm and Revoke
A use read the grant, then was admitted by a write that looked at neither the marks nor
the pins. A mark or a first pin stored in between did not stop it. Admission now carries
the definitions in use and refuses, in the same write, a tool that is marked or pinned to
another digest. A use refused this way marks the tool when the stored pin disagrees with
what it saw.
A confirmation and a revocation replace the grant whole, built from an earlier read. A
pin or mark a run added in between was dropped, so the next use could pin a different
definition unnoticed. Both writes now refuse when either list has grown since the read:
the confirmation reports the offer as changed, and the revocation reads again.
* 📌 fix: Tie a Tick on a Changed Tool to the Grant That Marked It
The dialog remembered which tools were marked when the owner last ticked a box, by name
only. If the grant was replaced elsewhere and a run marked the same tool again, the old
tick stood and confirming accepted a change the owner had not looked at. The edit now
records the grant revision, and a tool marked under another revision is unticked until
the owner ticks it again.
* 🪟 feat: Show a Running Subagent's Current Row on Its Card
* 🪟 fix: Draw the Card's Row Live and Skip Label Reservations That Draw Nothing
* 🪟 fix: Show a Card's Row Only While Its Thread Is Followed, and Throttle a Detached Row Too
* 🪟 chore: Sort the Card Spec's Imports
* 🪟 fix: Draw the Card's Row as a Picture That Registers, Claims and Reveals Nothing
* 🪟 test: Give the Artifact Preview Test the Props Its Type Requires
* 🪟 refactor: Keep the Shared Follower Registry in Provider-Scoped Jotai Atoms
* 👀 fix: Look Again for a Run the Caller Knows Of Before Taking It to Have No Stream
* 🗳️ feat: Let a Schedule's Owner Choose Which MCP Tools It May Use
An enrolled scheduled chat could call an MCP tool only when an operator had
listed it in config as read-only, pinned to a hash of its definition. Whether
a tool is safe to run with nobody watching is the owner's call, and read-only
is a poor proxy for it.
The consent offer now lists every MCP tool the agent graph selects on servers
whose recipient the operator declared. The owner ticks servers and tools, and
the grant stores only those. A tool is allowed at run time when the grant
lists it for the calling agent and that agent still selects it; anything else
is refused. Selecting another tool later no longer invalidates the grant.
The operator `mcpConsent.readOnlyPolicy` list is retired: the loader warns
when it is still configured. The stored grant keeps its shape.
* 🗳️ fix: Check the Server Box Only When Every Tool Is Allowed
The shared Checkbox has no partly-selected appearance, so a server with some
tools ticked showed its "allow all" box as checked.
* 🧪 test: Check the Live Catalog at the Manager Boundary Without the Retired Pin
* 🧭 fix: Check Readiness Against What the Owner Allowed, Not What the Agents Select
* 🧭 fix: Turn Away a Tool the Owner Left Out Without Failing the Run
* 🧭 fix: Load Only the MCP Tools the Owner Allowed in a Scheduled Run
* 🧵 feat: Open a Background Subagent's Thread as a Conversation in the Panel
* 🧵 test: Keep the Subagents Host Spec and Add the Follow Cases Beside It
* 🧵 fix: Give the Followed Thread Its Panel Gutter and Assert It in the Browser Specs
* 🧵 test: Do Not Wait on a History Fold That Closed Before It Was Clicked
* 🧵 test: Draw a Steer the Child Took Before the Panel Opened Where It Was Taken
* 🧵 fix: Fall Back to the Durable View Only on a Confirmed Missing Stream or History
* ↩️ feat: Show Guidance a Followed Subagent Has Not Taken Yet, With Its Withdraw
* 🗓️ feat: Align Balance Resets to Fixed Calendar Boundaries
Add balance.refillAnchor, an ISO timestamp with a fixed offset that periods align to: a refill or reset is due at the first anchor-plus-whole-intervals boundary after the last one, so every user resets at the same moment (e.g. each Monday at 00:00 UTC, or the 1st of each month). Omitted, periods keep running from each user's last refill. The gauge names the reset date for anchored periods.
* 🧭 fix: Clear a Removed Reset Anchor During Admission and Shape Balance Responses in TS
A policy built without a record now always carries refillAnchor, null when unset, so removing the setting clears a stored anchor instead of looping admission until its retry bound throws. Balance response shaping moves to toBalanceResponse in packages/api; the CJS controller delegates to it.