* fix: install marketplace plugins and MCP servers in-process instead of spawning a cline binary
The desktop app sidecar and cline-hub shelled out to 'cline plugin install'
and 'cline mcp install' for marketplace installs. Packaged GUI apps inherit
launchd's minimal PATH on macOS and most desktop users have no cline CLI
installed at all, so installs failed with a red
'Executable not found in $PATH: "cline"' error.
Install via @cline/core's installPlugin/installMcpServer in-process instead,
matching what the VS Code extension already does. Also fix
parseMcpInstallArgs in @cline/core to treat the marketplace catalog's '--'
separator as end-of-options; previously the separator itself became the
stdio command.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* refactor: drop test-injection plumbing from marketplace installers
Call @cline/core's installPlugin directly instead of threading an
installer option through the marketplace entry points; tests stub the
core module instead.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* revert: keep cline-hub marketplace installs CLI-backed
The hub dashboard is launched via 'cline dashboard', so a CLI is always
present and CLINE_WRAPPER_PATH resolves it; the PATH bug only affects
the desktop app, which does not ship a CLI.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
Each session row shows a trash icon on the right while hovered (or
when the button itself is focused), opening the same delete
confirmation dialog the row's context menu uses. The row is a button
and buttons cannot nest, so the trash is an absolutely positioned
sibling inside a group/row wrapper, overlaid where the timestamp sits:
row hover hides the timestamp, shows the trash, and moves the row's
hover background to the wrapper group so it holds while the pointer is
on the trash itself.
Quitting the mac app beach-balled for ~5-7s. The shutdown POST was
built from the ws transport URL (appending /shutdown lands inside the
query string), so the sidecar was never told to exit, and stop() then
polled the child for up to 7s on the main thread - on macOS inside
applicationWillTerminate - before SIGKILLing it.
stop() now sends SIGTERM and returns immediately. The sidecar handles
SIGTERM with the same bounded (5s) graceful shutdown as the /shutdown
endpoint and exits itself, finishing session persistence as an orphan.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* feat(desktop): merge schedule details into one view and open run-now sessions
The schedule details dialog drops its Overview/Runs tabs: one scrollable
column with the meta grid, the configuration JSON (capped at max-h-64
with internal scroll so it cannot crowd out what follows), and a Runs
section beneath it showing the three most recent runs with a ghost
"Show all N runs" expander (collapsed again whenever a different
schedule's details open). The "Full configuration for this schedule"
subtext is gone; the dialog passes aria-describedby={undefined} so
Radix does not warn about the missing description.
Run now hands you into the session it starts. The trigger command
queues the run and returns before the runner attaches a session id, so
after the toast the handler polls the schedule overview once a second
for up to 15 seconds — which doubles as keeping the page's run status
fresh (refreshSchedules now returns the fetched overview to make that
single-stream) — until the triggered execution reports its session id,
then calls onOpenSession. Guarded so it never auto-navigates after the
user left the page.
* feat(desktop): hide runtime steering messages from transcripts
Scheduled/automation runs inject user-role steering messages each
iteration ("[SYSTEM] This run is not complete until you call
submit_and_exit...", plus a team-obligations variant). The chat view
rendered them as user bubbles, as if the person had typed them — in a
scheduled session the transcript was mostly [SYSTEM] noise.
They are machinery talking to the model, not something the person said
or needs to read, so the transcript now hides them entirely:
MessageBubble renders null for any [SYSTEM]-prefixed user message.
Grouping still treats them as working-row machinery via a single
isSystemSteeringMessage predicate — they collapse into the run's work
span, are never a turn boundary, can never be mistaken for a run's
answer, and never advance the run count even when metadata is missing —
so work-block folding and checkpoint/edit run numbering stay correct.
A finished scheduled session now reads as prompt, work summary, answer.
* fix(desktop): poll history while an attached session's event stream is dead
Opening a scheduled session while (or right after) it runs left the
view stuck on the thinking shimmer until the user switched away and
back. Root cause is in core: the hub daemon executes scheduled runs on
a private LocalRuntimeHost inside createLocalHubScheduleRuntimeHandlers,
while the hub server only projects live events from its own session
host — so session.attach succeeds but no assistant/tool/status events
ever flow. And since multiple hub daemons share cron.db, a run claimed
by a different daemon is invisible to this hub regardless. The proper
core rewiring is tracked as ENG-2474.
Client-side heal that covers every case: while an attached history
session reports a busy status and no chat_event chunk has arrived for
five seconds (and no assistant bubble is mid-stream), poll every three
seconds — re-read canonical history, merged through the same dedupe
path hydration uses, and the session record's status — so the
transcript and the thinking indicator settle in place. Locally driven
turns keep chunks flowing, so the quiet-window guard keeps the fallback
inert there.
* chore(desktop): format workspace selector components
Biome formatting drift that landed on main; picked up by a formatter
pass over components/views/chat.
* fix(desktop): keep stale-stream poll inert during locally driven turns
The fallback poll could fire between a local submit and the model's
first chunk (optimistic user bubble added, stream quiet past the
window, no assistant bubble yet). It then replaced the optimistic
bubble — raw prompt text — with its canonical history twin, which is
stored wrapped in a user_input envelope. The rekey handler that runs
when the stream starts looks for a trailing user bubble matching the
raw prompt, finds only the wrapped copy, and appends a second bubble:
duplicated messages in normal interactive chat.
The poll now stays inert while a local turn is in flight
(turnEpoch !== turnSettledEpoch, or outstanding optimistic user
messages), checked both before polling and again after the snapshot
returns. Hydration marks the turn settled — the mount defaults
(epoch 0, settled -1) otherwise read as an open turn and would keep
the fallback inert forever for the scheduled-session case it exists
for. Applying a polled snapshot also rebuilds the live tool routing
keys, same as hydration, so later tool events update canonical rows
in place instead of appending.
* fix(desktop): keep the working indicator alive for narrating scheduled runs
Watching a scheduled run live: the first tool row appeared, then the
thinking indicator vanished with nothing streaming, and the rest of
the run (final answer, submit_and_exit) only showed up seconds later
in one lump.
inferHydratedChatStatus treats a "running" session record whose
transcript ends on an assistant message as a session that died without
a status flip and reports "completed". That heuristic is right for
stale records, but scheduled/automation models narrate between tool
calls, so a polled snapshot can genuinely end on assistant text
mid-run — the completed flip hid the working indicator, folded the
run early, and disarmed the stale-stream poll (status left the busy
set), dead-ending live updates until an in-flight poll happened to
deliver the finished run.
The heuristic now only applies once the transcript has actually gone
quiet (newest message older than two minutes — comfortably past model
latency plus tool runs). A recently active transcript keeps the
record's "running" verdict, so the indicator stays up and polling
stays armed until the record itself settles.
* fix(desktop): stale-stream poll mirrors the session record instead of inferring
Replaces the previous fix for the vanishing working indicator (the
time-window guard added to inferHydratedChatStatus) with a version
that adds no inference at all: the heuristic is restored to exactly
its long-standing form, and the poll now maps the session record's
status verbatim (mapSessionRecordStatus).
The record is the right authority in the poll's context: the sessions
this fallback serves have a live host maintaining their record, and it
flips to a terminal status when the run ends. Transcript-shape
inference belongs only where it has always lived — hydrating sessions
whose records may be orphaned — and would misread a mid-run snapshot
ending on assistant narration as a finished session, hiding the
working indicator and disarming the poll.
* fix(desktop): address review findings on steering detection and run-now matching
Steering detection additionally requires the injected-message marker
(meta.userRunSpan === 0) beside the [SYSTEM] prefix, so a person's
genuine prompt that happens to start with "[SYSTEM]" stays visible
and turn-counted. The failure direction is deliberate: an unstamped
injected reminder would merely show as a user bubble, while the
content-only check could hide a real prompt.
Run-now only follows the execution id the trigger reply itself named;
the newest-execution-for-this-schedule fallback could open a previous
run's session when the trigger failed to enqueue one.
* fix(desktop): report a failed run-now instead of confirming a start
A trigger reply without an execution means no run was enqueued (the
schedule may have been disabled or deleted since the page loaded). The
handler previously toasted "Run started" regardless and then silently
skipped the session-open polling. It now shows a destructive
"Run not started" toast, refreshes the schedule list so the row
reflects reality, and skips the polling entirely.
* feat(desktop): split Customize into Installed and Marketplace pages
The Customize hub previously embedded a Browse section inside every tab
that had a catalog. That inlining made each tab long and buried the
catalog. Customize is now the installed inventory only (skills, MCP,
plugins, rules, hooks, tools tabs pass marketplaceVariant="installed"
to the embedded MarketplaceView; McpServersContent grew the same prop),
with an outline Marketplace button in the header.
Browsing moved to a dedicated Marketplace settings section that renders
the previously dead "directory" variant of MarketplaceView: one list
across all catalog types with type-filter chips, wrapping tag chips,
and light rules separating the filter tiers from each other and from
the results. The Clear control now renders inline at the end of the tag
row only while a tag is active, the Updated date is gone, and the
header hosts an Installed button mirroring the one on the Customize
page. Directory subheader copy: "A curated set of plugins, MCP
servers, and skills from the Cline community."
Tag and type chips wrap to new lines instead of scrolling
horizontally.
* feat(desktop): sidebar time view with sections, sort toggle, and scheduled detection
Restores the time-sorted session list as the default sidebar view, with
collapsible Pinned / Scheduled / Tasks sections (headers appear only
once something is pinned or scheduled) and the page-fill effect that
grows the fetched history window until a Show-more click makes visible
progress. Project grouping stays as the alternate mode behind a
one-click sort toggle whose icon reflects the active mode — the old
dropdown cost an extra click for a two-option choice.
Scheduled sessions are detected two ways: the hub-schedule origin
trigger in session metadata, plus a fallback that asks the hub which
session ids belong to schedule executions (list_routine_schedules,
fetched on mount and every two minutes, merged into a rolling set).
The fallback matters because locally executed scheduled runs do not
reliably stamp the trigger into session metadata — a real scheduled
session created today carried only {mode:"user"} provenance. The
scheduled clock icon now leads the row, left of the title; pin and
timestamp stay on the right.
The initial visible page grows from 10 to 30 rows so a tall sidebar
fills instead of stranding a stub of rows over empty space (history
fetches already start at 50).
The expanded sidebar's Customize row now hosts indented Installed and
Marketplace sub-tabs while a customize section is open; the active
sub-tab carries the full selected background while the parent keeps a
subtler one so the two simultaneous highlights read differently.
Also fixes the hover-card flash on click (logo card and session-row
cards): Radix HoverCardContent sits on a DismissableLayer, so a click
on the trigger registers as a pointer-down outside the card and
dismisses it, and the trigger's focus event immediately reopens it.
onPointerDownOutside preventDefault suppresses the dismissal; cards
still close on pointer leave.
* feat(desktop): schedule page row, dialog, and details UX polish
Schedule cards are now click targets: clicking anywhere on a card
outside its controls opens the details dialog (guarded via
closest("button,...") since every inline control, including the Radix
switch, renders a button element), with Enter/Space keyboard support.
The redundant eye button is gone. The remaining edit / run / pause /
delete buttons grow from the 12px icon-sm size to 28px targets with
16px icons, sized consistently with the adjacent enable toggle — the
icons use explicit size-4 classes so the Button base svg rule cannot
shrink them back.
The new/edit dialog gains breathing room between field labels and their
inputs (space-y-2 per field wrapper).
The details dialog no longer scrolls as a whole when the schedule JSON
is long: the dialog is a flex column capped at 85vh, the JSON pre
shrinks to the remaining space (min-h-0) and scrolls internally, and
the Runs tab list scrolls inside the tab the same way.
* feat(desktop): add Retina DMG background tooling
* feat(desktop): customize the macOS DMG layout
* ci(desktop): validate DMG background assets
* fix(desktop): adjust DMG Applications icon position
* ci(desktop): drop redundant DMG artwork validation from publish workflow
Tauri's beforeBuildCommand already runs dmg:background (with its own
validation) at the start of the build/sign/notarize step, and the
release/beta config overlays do not override the build section, so this
step duplicated work the publish job performs anyway. PR-time coverage
lives in desktop-test.yml.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <7799382+saoudrizwan@users.noreply.github.com>
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(vscode): stop showing cost estimates for subscription-billed providers
Providers whose usage is covered by a flat-rate subscription (ChatGPT
Plus/Pro via openai-codex, ClinePass) are marked with
metadata.usageCostDisplay = "subscription" in the SDK, and the CLI
already suppresses dollar figures for them. The VS Code host collapsed
that value into "show" before it reached the webview, so the task
header and model pricing rows rendered API-rate cost estimates that
users read as real charges on top of their subscription.
Pass all three usageCostDisplay values ("show" | "hide" |
"subscription") through the catalog listing and render cost only when
the value is "show", matching the CLI's shouldShowCliUsageCost
policy.
* feat(llms): mark Claude Code as a subscription-billed provider
Claude Code is typically authenticated with a Claude Pro/Max
subscription, but its models reuse Anthropic API pricing metadata, so
Cline rendered per-token prices and API-rate cost estimates for usage
that is covered by the subscription. Set usageCostDisplay =
"subscription" on the provider (picked up by the CLI and the VS Code
webview) and suppress the price rows in the Claude Code settings card.
The Claude Code CLI can also run on API-key billing, where a real cost
exists; the provider cannot distinguish the two, so we prefer showing
no number over a misleading one.
* fix(vscode): suppress cost display until provider listings load
While the ListProviders request is in flight (or after it fails), the
usage-cost hook had no listing to consult and fell back to "show",
flashing the API-rate estimate at subscription users on every chat-view
mount — the exact display the previous commit removes. Return
"unknown" whenever listings are absent; consumers already render cost
only for "show", so they suppress it during that window with no
changes. Briefly hiding a real cost is harmless, briefly showing a fake
charge is not.
* capture richer workspace information for vs code extension
* fix(shared): redact credentials from workspace remotes
* fix(shared): avoid regex backtracking in remote redaction
---------
Co-authored-by: Max Paulus 🥪 <max@cline.bot>
The model picker renders a radiogroup of styled buttons with role=radio
and aria-checked; biome's useSemanticElements flags the role as an
error, which fails the sdk-test Quality Checks lint for every PR
touching sdk/ or apps/ paths. Suppress with a justification — switching
to input type=radio needs a restyle and belongs to the desktop settings
work.
* feat(desktop): merge customization pages into a Customize hub with inline marketplace
Replaces the Plugins page and the dedicated Marketplace page with a single
Customize hub. Tabs: Skills, MCP, Plugins, Rules, Hooks, Tools, each with
live counts. Tabs backed by a marketplace catalog render the installed
items followed by an inline browsable Browse section (CLI-hub style), so
installing from the catalog immediately reflects in Installed above.
- Installed cards restyled to mirror the browse-card anatomy: bg-card p-4
containers, absolute top-right xs Uninstall matching Install, truncating
semibold titles, primary-tinted icons, real Badge components instead of
ad-hoc bordered spans, un-indented line-clamped descriptions
- Rules/Hooks/Tools rows brought into the same card language; redundant
intro paragraphs (duplicating the page description) removed; Tools group
headers match the Installed header style with counts
- Marketplace section header renamed to Browse; duplicate 'N results' row
removed (the header count is the single source)
- MCP embedded view now shows the full marketplace instead of
installed-only
* feat(desktop): overhaul sidebar sessions and navigation
Sessions list:
- Sort toggle removed; sessions are always grouped by project, with pinned
sessions leading each group (both subsets ordered by recency). The
Pinned/Scheduled/Tasks category sections and their time-mode paging
machinery (page-fill effect included) are deleted
- Scheduled sessions get an inline clock icon next to the pin position;
pin + clock render together when both apply, and the running/unread
status dot now coexists with them
- One font size (text-sm) across the list: titles, timestamps, project
headers, show-more buttons, empty states. sidebarText needed !text-sm
because the default button size's text-base wins the twMerge conflict
- Gradient fade under the Sessions header once the list scrolls, so rows
fade out instead of hard-clipping
- The session-detail hover card is controlled from the sidebar and closes
on scroll (Radix receives no pointer events while scrolling, so it used
to float over moving content)
- Sidebar min resize width raised 224->260px; the per-project show-more
label truncates so its nowrap text can't force rows to overflow and clip
timestamps at narrow widths
Navigation:
- Customize replaces the Plugins/Marketplace/Hooks/Rules/Tools sidebar
entries; Schedules and Customize are hidden from the expanded settings
nav (their top rows cover them) but stay reachable when collapsed
- The settings gear always opens General instead of resuming the last
section; the Account no-op hover special case is gone
- The New row highlights (aria-current) while the fresh not-yet-started
task page is showing and hands off to the session row once the task
starts; hitting New also focuses the prompt input via a window-event
signal (lib/prompt-input-focus.ts) since the sidebar and composer sit in
distant subtrees
- Fixed the xs button size collapsing any icon-bearing button to 12x12
(leftover has-[>svg]:size-3 from when xs was a micro button) — this was
why Uninstall buttons rendered broken next to Install
* feat(desktop): polish settings pages and chat composer
Models page:
- The provider detail panel is always open: no X button, no empty
no-selection state. It defaults to the first connected provider (falling
back to the first in the catalog), which also removes the layout shift
that happened when the page swapped between full-width and panel
variants on selection
- Fixed the list pane becoming unscrollable while the panel was open:
grid items default to min-size auto, so the pane grew past its track
inside the overflow-hidden grid and its ScrollArea had nothing to
scroll; wrapped it in a min-h-0 min-w-0 cell
- Add Provider opens a Dialog instead of swapping the page
(AddProviderContent gained a dialog variant that renders only the form)
- Embedded inputs (provider search, model search, detail fields) share one
EMBEDDED_INPUT_CLASS stripping the Input component's own border/dark bg
tint/shadow/ring, which rendered as a mismatched inner box; the model
search box uses the same h-9/px-3 frame as the provider search
- Model list flows with the page instead of a max-h capped inner scroller
Other pages:
- Account uses the shared PageFrame/PageHeader: left-aligned, text-3xl
title, Sign Out in the header actions slot
- Desktop notifications is one General section: header row plus the
Event/Notify/Sound matrix nested in a card, so its rows no longer read
as top-level peers of Dark mode; 'Available in the desktop app' label
removed
- Schedule page retitled from Schedules with a real description; Customize
description rewritten
Chat composer:
- The voice dictation button only renders once a voice model is
configured (Settings -> Voice); the unconfigured deep-link state is
gone (prop type kept for an easy restore)
* Redesign desktop Model Providers page and split voice input into its own settings page
- Group providers into Connected / Popular / All with auth-kind hints and
connection status instead of per-row enable toggles
- Show browser sign-in (not an API key field) for OAuth providers, with a
collapsed manual-key escape hatch where supported, plus explicit
Connect / Disconnect / Sign out actions
- Move voice input to a dedicated Settings > Voice page that only offers
connected transcription-capable providers, preselects a default model
(streaming preferred), and stays disabled in the sidebar until a
provider is connected
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Show native tooltip on the disabled Voice settings nav item
Disabled buttons drop pointer events, so the 'connect a model provider'
hint moves to a wrapping span for the browser tooltip to render.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Drop letter avatars and gray provider ids from provider rows and voice chips
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Drop model counts from provider list rows
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Rename provider Connected status to Configured and drop the green styling
A settings entry is configuration, not a live connection; neutral gray
text avoids implying an active link, since the user still picks which
configured provider to use per chat.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Resync provider catalog from disk when a settings save fails
Connect/disconnect/credential edits update the list optimistically; a
failed save now reloads the catalog instead of leaving the optimistic
state (and the view's module cache) claiming a configuration that was
never persisted.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Rename oauthProvider test fixture to dodge CodeQL name heuristic
CodeQL's clear-text-storage query flags any identifier matching 'oauth'
as a credential source and traced the fixture's provider id into the
favorite-models localStorage write, which stores only provider/model id
strings. Renaming the fixture removes the false-positive source.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Guard catalog reloads against races and resync detail drafts on failed saves
Optimistic provider mutations now bump a generation that discards any
in-flight catalog response, so a failed-save recovery reload can't
overwrite a newer edit with an older disk snapshot. The recovery also
remounts the provider detail panel via a reset token so its local field
drafts reflect the reloaded on-disk state instead of unpersisted edits
or an optimistically cleared disconnect.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Fix failed-save recovery ordering and retry superseded reloads
Remount the provider detail only after the authoritative catalog reload
lands, so its drafts re-seed from disk state rather than the optimistic
values that failed to persist. When a concurrent edit supersedes the
recovery's in-flight response, retry the reload (bounded) instead of
dropping it, since that edit performs no reload of its own.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Add Pinned/Scheduled/Tasks categories to desktop app sidebar
Replace the Schedules and Favorites filter-menu options with visible
collapsible category sections in the session sidebar, and rename the
Favorite action to Pin across the sidebar and sessions view.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Grow full history window when Tasks show-more outpaces loaded tasks
loadMoreSessions treats its argument as a limit on all sessions, but the
Tasks show-more count only tracks Task rows, so once pinned/scheduled
rows pushed the loaded total past the requested count the call no-oped
and clicks went dead. Grow the whole history window via
loadOlderSessions instead, and only when the loaded tasks cannot fill
the next page. Addresses Greptile review on #13528.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Auto-fill the Tasks page instead of fetching once per show-more click
A single 50-session window growth can consist entirely of pinned or
scheduled sessions, leaving a show-more click with no visible Tasks
progress. Replace the one-shot fetch with a page-fill effect that keeps
growing the history window until the requested Tasks page fills or
history runs out. Addresses the follow-up Greptile review on #13528.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Halt page-fill retries after a failed history fetch
A failed fetch leaves the task count and has-more state unchanged,
which are exactly the conditions the page-fill effect fires on, so one
failing request would retry and re-toast forever. Halt the effect after
a failure and let the next explicit show-more click retry. Addresses
the third Greptile review on #13528.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* desktop: clean up sidebar navigation chrome
- Give New Task its own full-width labeled row below the logo row
instead of an ambiguous icon next to the agenda toggle
- Wire the New Task row to the home action so starting a new task
clearly takes you home (the logo still works as a fallback)
- Swap back/forward chevrons for browser-style arrow icons and
bump their size
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* desktop: sidebar New/Schedule/Customize rows and always-visible search
- Stack New (plus icon), Schedule, and Customize as full-width labeled
rows below the logo; whole row highlights on hover via sidebarItem
- New starts a fresh task (home), Schedule opens Settings > Schedules,
Customize opens the Customizations sections (Plugins first)
- Show the session search bar permanently above the sessions list
instead of hiding it behind a search icon toggle
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* desktop: move session search into a dialog behind a logo-row icon
- Replace the inline sidebar search bar with a search icon in the
logo row that opens a cmdk command dialog listing sessions
- Selecting a result opens that session and closes the dialog
- Remove the agenda/tasks toggle the icon replaces, along with the
now-unreachable sidebar Agenda panel (the welcome screen still
surfaces agenda tasks)
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* desktop: load full session history when the search dialog opens
Addresses Greptile review on #13533: the dialog only searched the
currently loaded history batch, so older unloaded sessions could not
be found. Opening search now kicks off loadAllSessions() (the hook's
purpose-built global-search loader), and the empty state reads
'Searching older sessions...' while more history is streaming in.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* chore(vscode): remove per-tool MCP auto-approve checkboxes from webview
MCP auto-approval is now governed solely by the global 'Use MCP servers'
toggle; the SDK approval path (shared with the CLI and desktop app) has no
per-tool granularity, so the per-tool and 'Auto-approve all tools'
checkboxes were no-ops that implied control that no longer exists. Remove
them from the MCP settings view and chat tool rows. The autoApprove arrays
in cline_mcp_settings.json and the toggleToolAutoApprove RPC are left
intact for the legacy extension.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* chore(vscode): hide per-tool MCP auto-approve checkboxes behind a flag
Keep the checkbox components, handlers, and RPC plumbing intact but gate
rendering behind SHOW_MCP_PER_TOOL_AUTO_APPROVE=false: the SDK approval
path (shared with the CLI and desktop app) is all-or-nothing via the
global 'Use MCP servers' toggle, so the per-tool checkboxes were no-ops.
Flip the flag back on if the SDK gains per-tool approval granularity.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(core): prevent search_codebase from crashing the process on giant single-line files
searchWithRipgrep buffered all of rg's --json stdout into one string. Each
JSON event embeds the full text of the matched line (--max-columns is
ignored in JSON mode), so searching a directory of serialized trace dumps
(single-line multi-hundred-MB JSON files) accumulated gigabytes of stdout
until string concatenation threw RangeError: Out of memory inside the
stream data handler. That throw is outside the tool's try/catch, so it
escalated to an uncaughtException and killed the CLI/hub daemon.
Parse rg's JSON events incrementally line by line, drop events larger
than 256KB, truncate matched/context lines to MAX_LINE_CHARS, and stop
reading once maxResults is reached. The fallback regex scan now skips
files larger than 10MB (reporting the skip count) and truncates its
context lines the same way.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* simplify search_codebase crash fix to a minimal diff
Replace the incremental JSON-event parser with three small guards: stop
buffering rg stdout past 10MB, drop the trailing partial event before
parsing, and slice fallback context lines to MAX_LINE_CHARS. Drops the
fallback file-size skip and skip-count reporting.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
The /shutdown handler queued teardown on a microtask, which runs before
the event loop's write phase, so the daemon could process.exit() before
the accepted 202 was handed to the socket. Unix masked it (uv_try_write
lands small loopback writes synchronously); Windows has no such fast
path and lost the race regularly — the recurring shutdown.e2e.test.ts
'socket hang up' failures on windows-latest. Start teardown from the
response's write callback instead, with an idempotent 1s fallback so a
client that vanishes mid-write cannot strand the daemon, and send
Connection: close so the client gets a FIN rather than an abort.
Since the flakiness this compensated for is fixed at the source, restore
maxWorkers: 2 for the Windows core suite (serializing it cost ~3 min of
CI per run), and raise the e2e daemon discovery hang guard 10s→30s —
it guards against hangs, not runner speed.
ci-node-smoke.ts installs the packed SDK tarballs with a plain npm
install in a fresh temp dir, where the repo root package.json overrides
do not apply. When @sap-cloud-sdk 4.9.0 shipped (2026-08-24) it broke
@sap-ai-sdk/ai-api 2.14.0 (via @jerome-benoit/sap-ai-provider in
@cline/llms) with ERR_PACKAGE_PATH_NOT_EXPORTED, failing the smoke step
on every PR even though the root already pins @sap-cloud-sdk/* to 4.6.0.
Copy the root overrides block into the generated sandbox package.json
so the smoke install resolves the same pinned versions as the repo and
future third-party releases cannot break it independently.
* fix(hub): cap hub-events db size so it can't fill the disk
Row/time retention alone didn't bound disk usage: envelopes carrying
full session snapshots reach hundreds of KB each, so retained rows
could total tens of GB, sweeps only ran hourly, and DELETE never
shrinks a SQLite file. Enforce a 64 MiB size budget in prune() (oldest
rows first, VACUUM to return the space), and also prune after every
16 MiB appended so bursts can't outrun the hourly timer.
Fixes#13505
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(hub): tolerate VACUUM failure on a full disk
VACUUM needs scratch space and can fail in exactly the state a
ballooned event log causes. The byte-budget deletes already bound live
data, so swallow the error and let the next sweep retry the reclaim
instead of aborting startup pruning and disabling the durable log.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(hub): count the size budget in UTF-8 bytes, not characters
envelopeJson.length (UTF-16 units) and SQLite LENGTH() (characters)
undercount multibyte text by up to 3x, which could leave a CJK-heavy
log settled above budget and re-running VACUUM every sweep. Use
Buffer.byteLength and LENGTH(CAST(... AS BLOB)) instead.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(vscode): resolve hook workspace identity from the window, not shared global state
Hook discovery, hook cwd selection, and the workspaceRoots metadata passed
to hook scripts all read the workspaceRoots/primaryRootIndex global state
keys. Global state lives in ~/.cline and is shared by every Cline instance
(all VS Code windows, the CLI, the JetBrains plugin), and nothing writes
these keys anymore, so hooks resolved against whatever project some other
or older instance last recorded. With a second window open on another
project, a workspace's .clinerules/hooks scripts were never discovered.
Resolve workspace roots via a single guarded helper backed by
HostProvider.workspace.getWorkspacePaths() (in-process, window-scoped,
same as refreshHooks): blank paths are filtered, a host-bridge failure
degrades to no workspace roots instead of silently disabling global hooks
or skipping blocking PreToolUse guards, and one resolution is threaded
through hooks-dir discovery, cache misses, cwd selection, and hook input
metadata so they can't disagree (previously up to four host lookups per
hook execution — real gRPC round trips in the standalone host). Roots and
hooks dirs are matched on whole path segments with the longest root
winning, so prefix-sharing or nested workspace roots resolve to the right
project. The adapter creates the runner once per event and skips no-op
runners, making creation the single resolution point; the separate
hasHook/getHookInfo checks are removed. The dead workspaceRoots and
primaryRootIndex state keys are dropped, and the four hand-rolled
HostProvider.workspace test stubs are consolidated into one shared
helper.
* test(vscode): add e2e coverage for workspace-scoped hook discovery
Boots real VS Code with the packaged extension against the workspace
fixture, sends a prompt, and asserts the fixture's UserPromptSubmit hook
was discovered from the open window's workspace, executed with that
workspace root as its cwd, and received the same root in its
workspaceRoots input — the end-to-end contract the hook workspace
identity fix establishes.
* test(vscode): isolate the e2e hook fixture from the shared workspace
The UserPromptSubmit fixture hook lived in the shared e2e workspace, so
every prompt-sending spec executed it (hooksEnabled defaults to true) —
and its cold PowerShell spawn on Windows pushed chat.test.ts past the
5s expect timeout. hooks.test.ts now overrides workspaceDir to a
dedicated workspace-hooks fixture, so only the hooks spec pays the hook
spawn.
The $4.99 first-month promo is ending, so the CLI's first-launch "Try ClinePass" dialog should no longer advertise it. Also drops the leftover CLI_PROMO_CODE plumbing, which has been an empty string since the promo-code flow was removed.
* fix(vscode): honor MCP auto-approve settings for SDK tool calls
The SDK extension required both the global 'Use MCP servers' auto-approve
toggle AND each tool's per-tool autoApprove flag before silently approving
an MCP call, while the legacy extension treated them as either/or. Restore
the legacy OR semantics so toggling MCP auto-approve works again.
Also key toolPolicies by the registered SDK tool name (via
defaultMcpToolNameTransform, now exported from @cline/core) instead of raw
server__tool. Servers whose names contain sanitized characters (e.g.
marketplace names like github.com/user/repo) or exceed 64 chars produced
policy keys that never matched the registered tool, so those MCP tools ran
without any approval gate; the live auto-approve lookup now re-applies the
transform instead of string-splitting the name.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Revert "fix(vscode): honor MCP auto-approve settings for SDK tool calls"
This reverts commit 86c568fbba.
* fix(vscode): auto-approve all MCP tool calls when the MCP toggle is on
The SDK extension only auto-approved an MCP call when the global 'Use MCP
servers' auto-approve toggle AND that tool's per-tool autoApprove flag were
both set, so toggling MCP auto-approve appeared to do nothing and users had
to opt in each tool individually. The toggle alone now governs all MCP
tools; the per-tool flag is no longer consulted.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
The task.completed fallback lived only inside shutdownSession, but
stopSession/dispose route interactive sessions with a terminal reported
status through releaseSessionRuntime, which never emitted. Truthful
session-status reporting (shipped in 4.1.11) re-routed a large share of
interactive stops onto that branch and silently dropped the event.
Route the emission through a single choke point,
emitTaskCompletedOnTeardown, called from both shutdownSession and
releaseSessionRuntime. The completion criterion no longer reads
session.status: interactive sessions use the recorded final-turn
outcome (lastInteractiveTurnFinishReason), non-interactive sessions
keep the existing input.status === "completed" logic. A new
taskCompletedEmitted flag (also set by the submit_and_exit observer)
enforces exactly one task.completed per session. failSession now
records the errored final turn so a stale "completed" from an earlier
turn can never leak into the teardown emission. Telemetry only; no
user-facing behavior changes.
Four consecutive SDK publish runs failed on windows-latest, each on a
different test, all of them plain timeouts: two @cline/shared SQLite
tests at the 5s vitest default, core's bash executor at 10s, and the hub
singleton endpoint test at 10s. The 2-core Windows runner spawns forks
and takes SQLite locks slowly enough to blow those budgets under load.
These timeouts guard against hangs; they are not timing assertions (the
one suite that does assert elapsed time, shutdown.e2e, was fixed by
removing file-level parallelism instead). Raise core to 20s and give
@cline/shared an explicit 15s in place of the inherited 5s default.
singleton.e2e.test.ts (added in #13468) spawns real daemons and runs for
~15s. Vitest's default file parallelism let it run alongside
shutdown.e2e.test.ts, whose assertions are wall-clock bound: discovery
within 10s, exit within 5s, and a 2s shutdown watchdog. On the 2-core
windows-latest runner that contention alone broke those budgets, failing
the shutdown test two different ways across runs — once never observing
discovery, once with the daemon forced to exit before its HTTP 202
flushed (socket hang up). The test passed on Windows before #13468 and
has failed every SDK publish run since.
* fix(core): seed tools capability when custom model capabilities are synthesized from boolean flags
For a models.json entry with no explicit capabilities list, toStoredModelInfo
synthesized a capability array purely from boolean convenience flags (e.g.
supportsReasoning: true -> ["reasoning"]). modelSupportsToolCalling fails open
only for a missing or empty list, so the synthesized non-empty list read as an
authoritative denial and silently stripped every tool definition from requests
to custom OpenAI-compatible models (#13463).
Seed "tools" whenever the list was not explicitly authored and the boolean
projections made it non-empty, preserving the fail-open contract. Explicitly
authored capability lists remain authoritative and can still disable tools.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* test(core): cover stale catalog capability overrides
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix: treat stored capability lists as non-authoritative for tool calling
The hasExplicitCapabilities guard still let two producers of tool-less
lists through:
- The VS Code legacy-override migration (legacyModelInfoToOverrides)
persists explicit partial lists like ["prompt-cache"] into models.json
for custom OpenAI-compatible models, which then read as an authoritative
"cannot call tools" and drop every tool - same symptom as #13463.
- Any hand- or UI-authored partial list on a non-catalog model.
Stored entries and user-authored provider metadata have no way to declare
"cannot call tools" (there is no supportsTools field, and every writer
that authors a full list includes "tools"), so seed "tools" into any
non-empty list for a language model. Only generated catalog capabilities
remain authoritative - a genuine no-tools catalog model stays that way -
and non-language models (e.g. image generation) never gain a tools claim.
Also make legacyModelInfoToOverrides write "tools" into the arrays it
fabricates, matching the providers.json migration, so models.json stops
being poisoned for older readers.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* feat(hub): add drain and upgrade commands with replay support
* handles disconnection
* feat(hub): wire bot profiles, drain, and durable event/run-queue into the live transport
Completes the wiring the previous commits' primitives needed:
HubServerTransport gains isDraining(), hub.drain/hub.status/profile.get
command handling, and replayEventsAfter() (backed by the durable event
log), plus the sequence/sinceSequence wire types they depend on in
shared/hub.ts. run-queue-handlers.ts reads the active bot profile's
plugin roots when executing durable runs.
Also adds hub/profiles/: profile.json (identity/rules/plugins) ->
system prompt composition, --profile / CLINE_HUB_BOT_PROFILE
resolution, and the bundled cline-dad profile with its
cline_hub_support read-only diagnostics tool.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* Revert "feat(hub): wire bot profiles, drain, and durable event/run-queue into the live transport"
This reverts commit 6696d5d202.
* fix(hub): dedupe replayed events by eventId, not just sequence
HubEventLogStore.append() returns a new envelope stamped with a
sequence rather than mutating the input, so a pending approval
re-issued sequence-less by subscribe() (it predates any durable-log
append) and its later sequence-stamped copy from the durable log are
two different objects carrying the same eventId. The replay-then-live
buffer in browser-websocket.ts only deduped by sequence, so the
sequence-less copy's guard never tripped and it was delivered a second
time when the buffer flushed after replay.
Track delivered eventIds alongside the sequence cursor; eventId
survives the append/stamp round-trip unchanged, so this dedupes the
exact-same logical event regardless of which copy arrives first.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* fix(hub): wire drain, durable event log, and run queue into the live transport
CI on this branch failed bun run build:sdk: browser-websocket.ts,
client/index.ts, and hub-websocket-server.ts (already on this branch)
reference sequence/sinceSequence, HubServerTransport.isDraining(), and
the "hub.drain" command — but the commit that reverted bot profiles
out of this branch also reverted this wiring, since it shared a commit
with the profiles work. That wiring is a hub concern, not a
bot-profiles one; split it back out.
- shared/hub.ts: sequence/sinceSequence types, run.enqueue/run.list/
hub.drain/hub.status/stream.replay capability, command, and event
names. profile.get intentionally excluded — stays bot-profiles-only.
- context.ts: isDraining() on HubTransportContext. botProfile field
intentionally excluded.
- hub-server-transport.ts: eventLog/runQueue fields and start/stop
lifecycle, publish() appends to the durable log, handleCommand cases
for run.enqueue/run.list/hub.drain/hub.status, drain-refusal check,
replayEventsAfter()/lastEventSequence(). startBotProfile()/
startHubSupportTool() and the profile.get case intentionally
excluded.
- run-queue-handlers.ts: added without handleProfileGet (needs
ctx.botProfile, which doesn't exist here).
- hub-upgrades.test.ts: added without its two bot-profile-injection
tests (they need a resolved bot profile to assert against).
Verified bun run build:sdk exits 0 (the exact CI command) and
bunx vitest run src/hub passes (311/312; the one failure is the
same pre-existing environment-timing flake already present before
this change).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* fix(hub): export instance-lock, event-log, and run-queue from the hub barrel
These landed as internal modules only; hub-server-transport.ts and
hub-websocket-server.ts import them by direct path, but nothing
re-exported them from the public @cline/core/hub surface the way
sibling discovery/server modules already are.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* fix(hub): wire the instance lock into the daemon entry point
The singleton lock (discovery/instance-lock.ts) and its consumption in
startHubWebSocketServer/ensureHubWebSocketServer were already on this
branch, but the daemon entry point's own half was not: retrying a bind
when a retiring predecessor still holds the lock, and exiting with a
distinct code (3) instead of the generic fatal path when a live Hub
already owns the data directory. Without this, a daemon racing a
retiring predecessor could fail outright instead of waiting the lock
out, and losing the singleton race looked identical to a crash.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* fix(hub): address drain/upgrade review findings (#13478)
- cline hub upgrade: check idleness at least once (--wait 0 works), reject
non-numeric --wait, and un-drain on every abort path so an aborted
upgrade can never leave the hub refusing new work
- add cline hub drain --off and the off query param to requestHubDrain so
POST /drain?off is reachable from shipped code
- HubEventLogStore/HubRunQueue: WAL journal mode + busy_timeout, and stamp
sequences from lastInsertRowid instead of SELECT MAX(sequence)
- HubInstanceLock.acquire: degrade to an unheld lock when SQLite is
unavailable instead of refusing hub startup; only BUSY/LOCKED still
raises HubLockHeldError
- ensureHubWebSocketServer: retire an unusable discovered hub through the
shared retireDiscoveredHub (busy hubs are attached to, drain precedes
shutdown, discovery cleared only when the hub actually retired)
- replay adapter: advance the cursor past eventId-deduped events, cap
replay pages, stop when the cursor stalls, and drop the dedupe set after
the buffered flush so it cannot grow for the socket lifetime
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
* fix(hub): derive the singleton e2e challenger cwd portably
The challenger's working directory was derived by round-tripping the
discovery path through a file: URL and stripping the last pathname
segment. On Windows that yields a POSIX-style '/C:/...' path, which is
not a valid spawn cwd, so the spawn fails ENOENT before the singleton
lock is ever contested and the Windows SDK test job goes red.
The data dir is simply the discovery file's parent: use dirname().
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Saoud Rizwan <7799382+saoudrizwan@users.noreply.github.com>
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
The PublishNightly environment gained required reviewers, so each cron
run parked on approval, held the workflow's concurrency group, and
silently cancelled every scheduled run queued behind it. 20 consecutive
scheduled nightlies died this way between 2026-07-31 and 2026-08-21;
the only nightlies that shipped in that window were manual dispatches.
Drop the cron rather than leave a trigger that cannot succeed unattended.
* fix telemetry session propagation
* feat telemetry client version metadata
* fix(core): address Langfuse review feedback — hub client identity + delegated agent session grouping (#13475)
* fix(core): rebuild hub session client identity from request headers
Hub-backed sessions do not transport extensionContext (it is local-only),
so the daemon's runtime built traces without the clientName/clientVersion
metadata even though the hub client bakes X-CLIENT-TYPE / X-CLIENT-VERSION
into the session's provider headers. Reconstruct extensionContext.client
from those headers during local runtime bootstrap so hub-backed Langfuse
traces carry the same client identity as local runtimes, and the daemon's
header re-resolution stops clobbering the original X-CLIENT-TYPE.
* fix(core): propagate parent distinctId/sessionId to delegated agents
Delegated agents (spawned sub-agents, configured agents, teammates) were
built without distinctId and sessionId, so their Langfuse traces had no
userId or sessionId and did not group with the parent user or session.
Thread the host-resolved distinctId through RuntimeBuilderInput and the
root sessionId through the delegated-agent config provider, and copy both
onto the delegated AgentConfig.
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
* Treat an empty preserved capability list as unspecified when seeding tools
toSdkModelInfo guarded the tools seeding with a strict
preservedCapabilities === undefined check, but modelHasCapability —
the runtime's own reader — treats undefined AND length === 0 as
"unspecified". A custom OpenAI-Compatible model whose stored
capabilities field is a defined-but-empty array (a config carried over
from before the field existed, or one round-tripped through a boundary
that defaults it to []) skipped the seeding; the first boolean
projection to run afterwards (e.g. supportsReasoning) then populated
the array, the runtime gate read the non-empty, tool-less list as
authoritative, and every tool definition was silently dropped from the
session (#13463).
The guard now covers the empty array too, matching the reader's
unspecified semantics.
* test: satisfy the store's isModelInfo gate so the empty-capabilities case actually reaches knownModels
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: yzxcj797 <yzxcj797@users.noreply.github.com>
Co-authored-by: Saoud Rizwan <7799382+saoudrizwan@users.noreply.github.com>
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(vscode): enforce enterprise MCP controls on the Customize marketplace
The unified Customize marketplace replaced the old MCP marketplace
without carrying over enterprise remote-config enforcement: the catalog
RPC returned every MCP entry and installs were never policy-checked,
so orgs with mcpMarketplaceEnabled=false or an allowedMCPServers
allowlist saw (and could install) all marketplace MCP servers.
- Filter MCP entries out of getMarketplaceCatalog when the marketplace
is disabled, and restrict entries to the allowlist when configured
(matching entry id, display name, installed server name, or source
repo URL, mirroring legacy GitHub-URL allowlist ids)
- Reject installMarketplaceEntry requests that violate the policy
- Map the published catalog's repo/homepage fields onto
sourceUrl/homepageUrl so URL-based allowlists can match
- Update the enterprise MCP server controls docs
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* refactor: simplify MCP marketplace policy enforcement
Fold the policy check into marketplace-helpers, drop the dedicated
test suite, and trim the docs edit to the strictly necessary line.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* chore(deps): update Langfuse packages and bump app versions
Update @langfuse/otel to v5.10.1 and add @langfuse/vercel-ai-sdk v5.9.1 for improved observability with Vercel AI SDK.
Bump versions for @cline/code to 0.0.14 and @cline/ui to 0.2.0-next.6, updated via bun.lock.
Other Changes:
Added optional userId to AgentRuntimeConfig.
Propagated userId, sessionId, conversationId, runId, iteration, provider, and model context into AI SDK telemetry.
Added AI SDK 7 runtimeContext with explicit includeRuntimeContext.
Added stable OTEL_SERVICE_NAME=cline-sdk.
Added runtime metadata assertions in agent tests.
* add taskId
* Revert "add taskId"
This reverts commit f20d31d96d.
* fix(core): keep hub session status truthful across queue-drained turns
Queue-drained turns settle only through the event stream, but the hub
runtime host mistranslated their lifecycle in two ways:
- session.updated events carrying only a snapshot (persistence updates)
defaulted the projected status to "running". When one trailed the
final idle update after a turn, clients that track busy state from
status events (the desktop sidecar's workspace restore gate) stayed
busy forever. Use the snapshot's real status and emit nothing when
neither source reports one.
- the per-run agent.done dedup was only reset by run.started, which the
daemon-side queue drain never publishes, so a drained turn's done was
swallowed as a duplicate of the previous turn's. Reset the dedup on
session.pending_prompt_submitted, and suppress stale run.completed
events that land inside a drained turn's window so they can neither
emit a phantom done nor consume the drained turn's dedup slot.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* test(desktop): cover restore unlock after an event-settled queued turn
Exports the sidecar's core-session event handler so the queued-turn
lifecycle (busy via status events, cleared by the done agent event,
restore allowed afterwards) is testable end-to-end.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(core): start interactive sessions without a prompt as idle
The runtime host reported every new session as "running" until its
first turn ended. Interactive hosts (the desktop app) start sessions
with no prompt and dispatch turns through separate send calls, so a
created-but-never-prompted session stayed "running" forever — wedging
clients that gate workspace operations (checkpoint restore, message
edit) on active turns.
Interactive no-prompt starts now begin idle, start emits the session's
actual status (resumed sessions no longer masquerade as running), and
markTurn* transitions keep tracking in-memory status for lazily
persisted sessions so the first turn still reports running -> idle.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* style: format hub-runtime-host test filter
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* refactor: drop the drained-turn done bookkeeping, keep the minimal fix
The stuck restore is fully explained by the two status defects (fabricated
"running" from snapshot-only session.updated events, and never-prompted
interactive sessions reporting "running"). The done-dedup machinery for
queue-drained turns addressed a separate cosmetic gap (queued turns emit no
chat_done, pre-existing) and required fragile run-window heuristics, so it
is removed to keep this change reviewable. Sidecar test now settles the
queued turn through the status event, matching the shipped mechanism.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* docs(sdk): document the truthful session-status contract
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
The edit preview computed proposed content with an exact old_text match, but
the SDK executor normalizes old/new text to the file's own line endings before
matching (#12305) - reads strip CR, so models emit LF-only text even for CRLF
files. Any multi-line old_text in a CRLF file therefore failed the preview's
match: the diff edit view silently never opened while the executor applied the
edit. Single-line edits (no line break in old_text) were unaffected, which is
why the diff view appeared to trigger inconsistently.
Mirror the executor's EOL normalization (and its literal $-sequence insertion)
in the preview computation.
Fixes#13296
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
Classic Cline truncated long conversations by omitting an index range of
api_conversation_history from every API request (keep the first
user-assistant pair, drop everything through the range end, strip
orphaned tool_results from the first kept message). The range was
persisted on the history item while the full history stayed on disk.
legacyApiHistoryToSdkMessages ignored conversationHistoryDeletedRange
and converted the entire file, so resuming a migrated long task handed
the SDK an untruncated working context that could exceed the model's
context window by millions of tokens - every request failed with
'prompt is too long' and every compaction restarted from the full
history (#12996, confirmed by the reporter: the task was migrated from
an older version and broke after a restart, with each compaction
starting from ~3M tokens).
The migration now replays exactly what the classic extension sent:
slice out the deleted range and drop orphaned tool_results, mirroring
ContextManager.getTruncatedMessages (see origin/main). Malformed ranges
fall back to the full history (previous behavior).
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(core): watch agenda task specs via the resolved long path
fs.watch on a path with 8.3 short components (e.g. C:\Users\RUNNER~1
temp dirs) trips a libuv assertion in fs-event.c on Windows and aborts
the whole process. Since the agenda task manager landed, every hub
server test spins up its spec watcher on such a path on hosted Windows
runners, killing the vitest worker and failing the sdk-test Windows job
on every branch. Resolve the specs dir with realpathSync.native before
watching so libuv only ever sees the long form.
* test(ui): stub ResizeObserver for @pierre/diffs in tool-diff tests
jsdom does not implement ResizeObserver, so every ToolFileDiff render
logged a ReferenceError from @pierre/diffs to stderr. Tests still
passed; this just silences the noise the same way the constructable
stylesheet shim does.
* fix(core): skip the agenda spec watcher when the dir does not resolve
Falling back to the raw path on realpath failure would reintroduce the
Windows short-path abort; log and go without the watcher instead.
* fix(hooks): collect PostToolUse hook output and honor its control
tool_result (PostToolUse) hooks ran fire-and-forget with stdout
ignored, so their entire JSON output — contextModification and cancel —
was discarded. Legacy awaited PostToolUse, injected its
contextModification into the conversation, and honored cancel.
- Run tool_result hook commands blocking (same 120s default timeout as
tool_call) in both the hook-config-file layer and the agent-hook
subprocess layer.
- Map their output: cancel stops the run with the hook's error message
as the reason; otherwise context is injected via afterTool
appendContext.
This restores legacy blocking semantics: tool results now wait for
tool_result hooks, but only in sessions that have one configured.
Ref: https://linear.app/cline-bot/issue/CLINE-2987
* fix(hooks): bound tool_result hook wait and isolate cancel reason
Address review findings:
- The agent-hook subprocess layer forwarded an unset timeoutMs
unchanged, so a tool hook command that never exits would block the
agent indefinitely. Default both tool_call and tool_result to the
120s bound the hook-config-file layer already used.
- A cancelling hook's error message was folded into the same context
field as other hooks' injectable context, so merging controls could
leak unrelated hook context into the cancellation reason. Carry it as
a separate cancelReason, and surface it as the stop reason for
beforeTool cancels too.
* fix(hooks): prefer errorMessage as a cancelling hook's stop reason
When a cancelling hook returns both contextModification and
errorMessage, the context-first parse precedence made the injectable
context the cancel reason and discarded the actual error. Parse the two
fields separately: errorMessage wins as the cancel reason (matching
legacy), and a lone errorMessage still folds into injectable context
for non-cancelling hooks as before.
* fix(vscode): honor PostToolUse hook cancel and contextModification
The adapter awaited PostToolUse hooks but discarded their output
entirely. Map cancel to a stop control (with errorMessage as the
reason) and contextModification into the runtime appendContext channel,
matching the PreToolUse mapping and legacy semantics.
* fix(hooks): whitespace-only errorMessage no longer suppresses the cancel reason
A cancelling hook returning meaningful context alongside a blank
errorMessage lost both: the parsers selected the whitespace as the
reason and the result mappers trimmed it away. Require a non-blank
errorMessage before it wins, so context serves as the fallback reason.
Apply the same fallback in the extension adapter's stop mapping.
* fix(hooks): deliver tool hook contextModification to the model
On the next engine, a tool_call (PreToolUse) hook's contextModification
was parsed into HookControl.context and then silently dropped: the
runtime beforeTool/afterTool result contract had no channel for
injecting conversation context. Legacy consumed it (ToolExecutor /
ToolHookUtils pushed <hook_context> blocks into the next user turn), so
this was a regression of documented behavior.
- Add appendContext to AgentBeforeToolResult/AgentAfterToolResult.
- AgentRuntime collects appendContext across hooks during an
iteration's tool executions and appends one <hook_context> user
message after the tool results, keeping tool-result parts contiguous.
- Map HookControl.context into appendContext in both subprocess hook
layers (skipped when the hook cancels, matching legacy, where the
message doubled as the error).
- Truncate injected context at 50KB per hook output, matching legacy.
- Concatenate appendContext across merged hook layers.
tool_result (PostToolUse) hooks still run detached with stdout ignored;
making them blocking so their context can be collected is a follow-up.
Ref: https://linear.app/cline-bot/issue/CLINE-2987
* fix(hooks): stamp tool identity on injected hook context blocks
Contexts are batched into one message after the tool results, and
parallel tool execution collects them in completion order, so position
alone cannot attribute a block to its tool call. Add tool_name and
tool_call_id attributes to each <hook_context> block.
* fix(hooks): sanitize hook context block markup
Attribute values (tool_name, tool_call_id) are stripped of quote/angle
characters and embedded </hook_context> closers in hook output are
neutralized, so neither provider-supplied ids nor hook text can corrupt
or spoof a block's stamped identity.
* fix(hooks): neutralize forged opening hook_context tags in hook output
The previous sanitization only neutralized closing tags, so hook output
could still open a forged <hook_context> block claiming another tool's
identity. Escape both opening and closing embedded tags with one rule.
* fix(hooks): hide injected hook context from user-facing transcripts
Stamp the injected hook-context user message with displayRole 'system'
(the compaction-summary convention) so it reaches the model but does
not render as a user bubble in live or replayed transcripts. Without
this, resuming a session showed the raw <hook_context> block as if the
user had typed it.
* fix(hooks): neutralize case-variant embedded hook_context tags
The tag-neutralization regex was case-sensitive, so hook output could
still smuggle a forged tag as <HOOK_CONTEXT>. Match case-insensitively.
* fix(vscode): map PreToolUse contextModification into runtime appendContext
The extension's hooks adapter bridged file hooks into the SDK runtime
but forwarded only cancel/errorMessage, so a PreToolUse hook's
contextModification never reached the model. Map it into the runtime's
appendContext channel; HookFactory already truncates it at 50KB.
* fix(vscode): hide hook-injected context from replayed transcripts
Live sessions never rendered the injected <hook_context> user message,
but session reload replayed it as a user bubble (and post-resume turns
kept doing so). Treat these messages as synthetic in the user-message
mapping: honor the displayRole 'system' stamp the runtime sets, with a
text-prefix guard for paths where metadata is unavailable. This also
keeps edit/regenerate ordinal mapping aligned with visible bubbles.
* fix(hooks): run file hooks through exactly one layer per host
The VS Code extension registered two independent hook execution layers:
its own hooks adapter (config.hooks) and the SDK core's file-hook
extension from the runtime bootstrap. When both discover the same hook
files, every hook executes twice per event — and with context injection
wired, each contextModification would be injected twice.
Add a 'hooks' runtime config extension kind (in the default set, so the
CLI keeps core file hooks unchanged) and gate the bootstrap's file-hook
extension on it. The extension excludes 'hooks' at session start, so
its adapter — which also provides the hook status UI and the
hooksEnabled setting — is its single execution path.
* fix(vscode): discover hooks from the session workspace, not only global state
Hook discovery read workspaceRoots from global state shared across
every Cline instance, so another window repointing it made workspace
hooks silently stop being discovered. With the extension's adapter now
the single hook execution layer, that meant no hooks at all.
HookFactory takes an optional sessionWorkspaceRoot and unions that
root's .clinerules/hooks into discovery (and into cwd resolution), fed
from the session config's cwd. Shared-state discovery still works, so
behavior in the single-window case is unchanged.
* fix(hooks): keep sanitized hook attribute values distinguishable
Replacing every markup delimiter with the same underscore could
collapse two tool call ids that differ only by such a character into
identical stamps. Escape each delimiter with a distinct token instead.
* fix(hooks): make hook attribute sanitization injective
Escaping the underscore itself turns the attribute escaping into a
uniquely decodable code, so no two distinct tool call ids can collapse
to the same sanitized stamp (previously an id containing a literal
escape token could collide with an id containing the delimiter).
* fix(vscode): reconstruct hook status rows when replaying transcripts
hook_status messages are emitted live but never persisted, so reloading
a session dropped every hook row. The injected <hook_context> blocks
carry the hook source and tool name, so the replay translator now
rebuilds a completed hook status row from each block. The injection is
also no longer treated as a user turn boundary, so the final turn's
completion retag is unaffected by it.
The job sets working-directory: apps/vscode, but that default applies to run
steps only, not to `uses:` steps. Since #10961 moved the extension under apps/
and added that default, the artifact path has resolved against the repo root,
matched nothing, and every failing run logged "No files were found with the
provided path: test-results/playwright/" instead of uploading recordings.
Widen to test-results/ so Playwright's error-context snapshots ship alongside
the videos.
The ui-publish smoke check pins a set of Tailwind candidates the packed
sources must emit; #13410 grew the SearchCombobox options list from
max-h-56 to max-h-64, so the publish run failed on the stale candidate.
All other pinned candidates verified against the current sources.
* feat(ui): sectioned model picker support in SearchCombobox
Adds option sections with headers, badges (NEW/Free pills), keyboard
navigation (arrows/Home/End/Enter with active-row tracking and
aria-activedescendant), substring match highlighting, a configurable
panel width, a trigger chevron, and a cleaner borderless search row.
All additions are backwards compatible; bumps @cline/ui to
0.2.0-next.6.
* feat(desktop): recommended and free model tiers in the composer picker
The composer's model selector showed raw provider/model ids and listed
the entire catalog alphabetized by id. It now labels providers and
models by display name and, for the cline provider, leads with the
Recommended and Free tiers from the recommended-models feed (NEW/Free
badges, descriptions) ahead of an All models section — matching the
CLI's featured picker and the kanban selector. cline-pass gets
Subscribed/Free tiers. A new list_cline_recommended_models sidecar
command exposes @cline/core's fetchClineRecommendedModels (display-ready
names, bundled offline fallback); feed ids resolve against the catalog
with a unique-slug fallback for Vercel/OpenRouter alias spellings, and
unresolvable entries are dropped rather than rendered unselectable.
* fix(desktop): widen the provider trigger for display names
Provider labels are now display names (e.g. "Cline Usage-Billing"),
which truncated badly at max-w-28.
* chore(desktop): drop unused featured-models test helper
* style(desktop): align workspace/branch picker search rows with the model picker
The composer's workspace/branch popover and the welcome screen's
workspace and branch pickers used a boxed inner search shell that now
clashed with the model picker's borderless search row sitting next to
them. Behavior unchanged.
* feat(ui): center the selected option when SearchCombobox opens
Opening a long list previously scrolled the selection just into view at
the panel edge; it now lands centered, and keyboard/hover navigation
falls back to minimal nearest-edge scrolling.
* style(desktop): picker row contrast, transparent search fields, centered open
The workspace/branch pickers' rows had a nearly invisible
surface-hover-lighter hover; rows now hover with surface-hover and mark
the current entry with the accent background plus check, matching the
model picker. The search inputs drop the Input base class's
dark:bg-input/30 tint that rendered a gray box inside the panel in dark
mode. Opening a picker now centers the current workspace/branch via a
shared scroll helper instead of starting at the top of the list.
* fix(ui): visible option hover/selected states and no scroll-jump on hover
The option row stacked bg-transparent with the conditional state
backgrounds; at equal specificity the later-sorted bg-transparent
utility won, so hover/selected rows rendered with no background at all.
The background classes are now mutually exclusive.
Mouse-driven active-row changes also reused the keyboard scroll-into-
view effect: hovering a row at the panel edge scrolled it into view,
which moved the list under the cursor and re-triggered hover — an
endless jump. Scroll mode is now per-source: center on open, nearest
for keyboard/typing, none for hover.
* fix(desktop): show only subscribed and free tiers in the cline-pass picker
The ClinePass offer is exactly the feed's subscribed + free tiers, but
stale bundled/cached catalog entries (e.g. a nemotron model) leaked
into an "All models" tier. Match the CLI's featured picker: hide
catalog leftovers, and only fall back to the full catalog when the
subscribed bucket is empty so a subscriber is never limited to free
models offline.
* fix(ui/desktop): strengthen the selected-row highlight in light mode
The selected row used the semantic accent surface (violet step 3),
which is nearly white in light mode. SearchCombobox and the desktop
workspace/branch pickers now highlight the selected/current row with
accent step 4 (with a fallback to --accent), which reads clearly in
both themes without touching the shared --accent token that shadcn
hover states depend on.
* fix(desktop): fit full provider display names in the composer trigger
"Cline Usage-Billing" — the default provider — truncated to
"Cline Usage-Bi…" at max-w-36; the trigger now allows up to max-w-56,
which fits the longest built-in provider names.
* style(ui/desktop): animate picker panels open like the shadcn dropdowns
The thinking-effort Select (shadcn/Radix) animates open while the
model/provider/workspace/branch pickers popped in instantly. All picker
panels now share the same open treatment — 150ms fade + slight zoom,
sliding from the trigger side. SearchCombobox uses a self-contained CSS
keyframe (consumers may not ship tw-animate-css); the desktop's custom
panels use the app's tw-animate utilities. Both respect
prefers-reduced-motion.
* chore(desktop): drop stale eslint-disable comments in picker search rows
This repo lints with biome; the jsx-a11y/no-autofocus disables were
inert leftovers. Flagged in review.
* refactor(core/desktop): stamp recommended-feed tiers onto ProviderModel in the SDK
Review feedback on the composer picker: tier joining should live where
the SDK serves model lists so each client doesn't fetch and join the
recommended-models feed itself (the CLI and now the desktop each did).
ProviderModel gains description and featured ({tier, rank, tags});
getLocalProviderModels overlays the feed's recommended/free tiers onto
cline models and subscribed/free onto cline-pass via
applyClineFeaturedModels, matching feed ids through the
Vercel/OpenRouter alias rules. The feed access is a new cached wrapper
(getCachedClineRecommendedModels, 5-minute TTL, in-flight dedupe) —
this path runs on every picker open, and the bundled offline fallback
is cached too so offline users don't re-pay the 5s timeout per list.
The desktop webview now reads tiers straight off the models: the
list_cline_recommended_models sidecar command, the webview feed fetch,
and its unique-slug alias matching are all deleted. toProviderModel
also carries ModelInfo.description generally.
* feat(desktop): recommended-feed badges and descriptions in provider settings
Review suggestion on #13410: the provider settings page has room for
more model detail than the composer's picker. The cline/cline-pass
provider cards now refresh their model list through
list_provider_models (the catalog snapshot deliberately skips the
recommended-feed overlay so the startup catalog fetch never blocks on
the feed) and render Recommended/Free tier badges plus feed tags (NEW)
next to the model name, with the model description underneath. The
refreshed list also surfaces the live entries instead of the bundled
snapshot.
* fix(ui): hand focus back to the combobox trigger on selection, close on Tab
Selecting an option (Enter or click) unmounted the focused search input
without a new focus target, dropping keyboard users' focus to <body> —
only Escape restored it. And since the search input is the panel's only
tabbable element, Tab always moved focus outside the component while
leaving the popup open behind the new focus target.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): keep the composer model selection inside the picker's visible offer
The active/remembered model was validated against the provider's full
catalog while the picker can intentionally hide models (the ClinePass
offer is exactly its subscribed/free tiers), so a stale remembered model
could become the selection while being absent from the dropdown.
Remembered and default selections (including on provider switch) now
resolve against the picker's visible options, and an explicitly
configured model that falls outside the offer stays active but is
surfaced under a 'Current model' section so the selection is always
visible and re-selectable.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): scope the settings featured model list to its provider and revision
The fetched featured list was unscoped component state: switching
between cline and cline-pass reused the component instance, so the
previous provider's models stayed visible while the new request was
pending (or forever, when it failed), and the retained copy shadowed
later provider.modelList updates — adding a second custom model
submitted the stale list as the complete configuration and dropped the
first addition.
The fetched list now only applies to the provider and modelList
revision it was fetched for (falling back to the catalog snapshot
otherwise and refetching on membership changes), and add-model submits
the union of the displayed and configured ids so an update can never
silently unconfigure existing entries.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* feat(core): stamp featured tiers onto the provider catalog synchronously
listLocalProviders deliberately skipped the feed overlay so the catalog
never blocks on the network — but that left the composer's very first
picker open after a cold boot rendering an untiered flat list until the
per-provider fetch landed. Blocking was never required: stamp tiers from
a synchronous peek at data already in memory (the cached live feed when
fresh, else the bundled fallback, whose recommended ids resolve against
the bundled cline catalog). The per-provider model-list path still
refreshes with live feed data moments later.
* fix(core): harden featured-tier matching and the feed cache reset
Review findings on the tier overlay:
Vendor-prefix mismatches now match by unambiguous id slug (two-pass, so
a catalog carrying both spellings of a model stamps one row, and a slug
shared by two feed entries stamps nothing) — the bundled fallback feed's
vendor-prefixed ids can otherwise miss cline-free/-prefixed catalog
entries, leaving them untiered in degraded mode.
resetClineRecommendedModelsCacheForTests now bumps a generation so an
in-flight feed request resolving after a reset cannot repopulate the
cache it just cleared.
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* feat(ui): sectioned model picker support in SearchCombobox
Adds option sections with headers, badges (NEW/Free pills), keyboard
navigation (arrows/Home/End/Enter with active-row tracking and
aria-activedescendant), substring match highlighting, a configurable
panel width, a trigger chevron, and a cleaner borderless search row.
All additions are backwards compatible; bumps @cline/ui to
0.2.0-next.6.
* feat(desktop): recommended and free model tiers in the composer picker
The composer's model selector showed raw provider/model ids and listed
the entire catalog alphabetized by id. It now labels providers and
models by display name and, for the cline provider, leads with the
Recommended and Free tiers from the recommended-models feed (NEW/Free
badges, descriptions) ahead of an All models section — matching the
CLI's featured picker and the kanban selector. cline-pass gets
Subscribed/Free tiers. A new list_cline_recommended_models sidecar
command exposes @cline/core's fetchClineRecommendedModels (display-ready
names, bundled offline fallback); feed ids resolve against the catalog
with a unique-slug fallback for Vercel/OpenRouter alias spellings, and
unresolvable entries are dropped rather than rendered unselectable.
* fix(desktop): widen the provider trigger for display names
Provider labels are now display names (e.g. "Cline Usage-Billing"),
which truncated badly at max-w-28.
* chore(desktop): drop unused featured-models test helper
* style(desktop): align workspace/branch picker search rows with the model picker
The composer's workspace/branch popover and the welcome screen's
workspace and branch pickers used a boxed inner search shell that now
clashed with the model picker's borderless search row sitting next to
them. Behavior unchanged.
* feat(ui): center the selected option when SearchCombobox opens
Opening a long list previously scrolled the selection just into view at
the panel edge; it now lands centered, and keyboard/hover navigation
falls back to minimal nearest-edge scrolling.
* style(desktop): picker row contrast, transparent search fields, centered open
The workspace/branch pickers' rows had a nearly invisible
surface-hover-lighter hover; rows now hover with surface-hover and mark
the current entry with the accent background plus check, matching the
model picker. The search inputs drop the Input base class's
dark:bg-input/30 tint that rendered a gray box inside the panel in dark
mode. Opening a picker now centers the current workspace/branch via a
shared scroll helper instead of starting at the top of the list.
* fix(ui): visible option hover/selected states and no scroll-jump on hover
The option row stacked bg-transparent with the conditional state
backgrounds; at equal specificity the later-sorted bg-transparent
utility won, so hover/selected rows rendered with no background at all.
The background classes are now mutually exclusive.
Mouse-driven active-row changes also reused the keyboard scroll-into-
view effect: hovering a row at the panel edge scrolled it into view,
which moved the list under the cursor and re-triggered hover — an
endless jump. Scroll mode is now per-source: center on open, nearest
for keyboard/typing, none for hover.
* fix(desktop): show only subscribed and free tiers in the cline-pass picker
The ClinePass offer is exactly the feed's subscribed + free tiers, but
stale bundled/cached catalog entries (e.g. a nemotron model) leaked
into an "All models" tier. Match the CLI's featured picker: hide
catalog leftovers, and only fall back to the full catalog when the
subscribed bucket is empty so a subscriber is never limited to free
models offline.
* fix(ui/desktop): strengthen the selected-row highlight in light mode
The selected row used the semantic accent surface (violet step 3),
which is nearly white in light mode. SearchCombobox and the desktop
workspace/branch pickers now highlight the selected/current row with
accent step 4 (with a fallback to --accent), which reads clearly in
both themes without touching the shared --accent token that shadcn
hover states depend on.
* fix(desktop): fit full provider display names in the composer trigger
"Cline Usage-Billing" — the default provider — truncated to
"Cline Usage-Bi…" at max-w-36; the trigger now allows up to max-w-56,
which fits the longest built-in provider names.
* style(ui/desktop): animate picker panels open like the shadcn dropdowns
The thinking-effort Select (shadcn/Radix) animates open while the
model/provider/workspace/branch pickers popped in instantly. All picker
panels now share the same open treatment — 150ms fade + slight zoom,
sliding from the trigger side. SearchCombobox uses a self-contained CSS
keyframe (consumers may not ship tw-animate-css); the desktop's custom
panels use the app's tw-animate utilities. Both respect
prefers-reduced-motion.
* chore(desktop): drop stale eslint-disable comments in picker search rows
This repo lints with biome; the jsx-a11y/no-autofocus disables were
inert leftovers. Flagged in review.
* refactor(core/desktop): stamp recommended-feed tiers onto ProviderModel in the SDK
Review feedback on the composer picker: tier joining should live where
the SDK serves model lists so each client doesn't fetch and join the
recommended-models feed itself (the CLI and now the desktop each did).
ProviderModel gains description and featured ({tier, rank, tags});
getLocalProviderModels overlays the feed's recommended/free tiers onto
cline models and subscribed/free onto cline-pass via
applyClineFeaturedModels, matching feed ids through the
Vercel/OpenRouter alias rules. The feed access is a new cached wrapper
(getCachedClineRecommendedModels, 5-minute TTL, in-flight dedupe) —
this path runs on every picker open, and the bundled offline fallback
is cached too so offline users don't re-pay the 5s timeout per list.
The desktop webview now reads tiers straight off the models: the
list_cline_recommended_models sidecar command, the webview feed fetch,
and its unique-slug alias matching are all deleted. toProviderModel
also carries ModelInfo.description generally.
* fix(ui): hand focus back to the combobox trigger on selection, close on Tab
Selecting an option (Enter or click) unmounted the focused search input
without a new focus target, dropping keyboard users' focus to <body> —
only Escape restored it. And since the search input is the panel's only
tabbable element, Tab always moved focus outside the component while
leaving the popup open behind the new focus target.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): keep the composer model selection inside the picker's visible offer
The active/remembered model was validated against the provider's full
catalog while the picker can intentionally hide models (the ClinePass
offer is exactly its subscribed/free tiers), so a stale remembered model
could become the selection while being absent from the dropdown.
Remembered and default selections (including on provider switch) now
resolve against the picker's visible options, and an explicitly
configured model that falls outside the offer stays active but is
surfaced under a 'Current model' section so the selection is always
visible and re-selectable.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* feat(core): stamp featured tiers onto the provider catalog synchronously
listLocalProviders deliberately skipped the feed overlay so the catalog
never blocks on the network — but that left the composer's very first
picker open after a cold boot rendering an untiered flat list until the
per-provider fetch landed. Blocking was never required: stamp tiers from
a synchronous peek at data already in memory (the cached live feed when
fresh, else the bundled fallback, whose recommended ids resolve against
the bundled cline catalog). The per-provider model-list path still
refreshes with live feed data moments later.
* fix(core): harden featured-tier matching and the feed cache reset
Review findings on the tier overlay:
Vendor-prefix mismatches now match by unambiguous id slug (two-pass, so
a catalog carrying both spellings of a model stamps one row, and a slug
shared by two feed entries stamps nothing) — the bundled fallback feed's
vendor-prefixed ids can otherwise miss cline-free/-prefixed catalog
entries, leaving them untiered in degraded mode.
resetClineRecommendedModelsCacheForTests now bumps a generation so an
in-flight feed request resolving after a reset cannot repopulate the
cache it just cleared.
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* feat(ui): sectioned model picker support in SearchCombobox
Adds option sections with headers, badges (NEW/Free pills), keyboard
navigation (arrows/Home/End/Enter with active-row tracking and
aria-activedescendant), substring match highlighting, a configurable
panel width, a trigger chevron, and a cleaner borderless search row.
All additions are backwards compatible; bumps @cline/ui to
0.2.0-next.6.
* feat(desktop): recommended and free model tiers in the composer picker
The composer's model selector showed raw provider/model ids and listed
the entire catalog alphabetized by id. It now labels providers and
models by display name and, for the cline provider, leads with the
Recommended and Free tiers from the recommended-models feed (NEW/Free
badges, descriptions) ahead of an All models section — matching the
CLI's featured picker and the kanban selector. cline-pass gets
Subscribed/Free tiers. A new list_cline_recommended_models sidecar
command exposes @cline/core's fetchClineRecommendedModels (display-ready
names, bundled offline fallback); feed ids resolve against the catalog
with a unique-slug fallback for Vercel/OpenRouter alias spellings, and
unresolvable entries are dropped rather than rendered unselectable.
* fix(desktop): widen the provider trigger for display names
Provider labels are now display names (e.g. "Cline Usage-Billing"),
which truncated badly at max-w-28.
* chore(desktop): drop unused featured-models test helper
* style(desktop): align workspace/branch picker search rows with the model picker
The composer's workspace/branch popover and the welcome screen's
workspace and branch pickers used a boxed inner search shell that now
clashed with the model picker's borderless search row sitting next to
them. Behavior unchanged.
* feat(ui): center the selected option when SearchCombobox opens
Opening a long list previously scrolled the selection just into view at
the panel edge; it now lands centered, and keyboard/hover navigation
falls back to minimal nearest-edge scrolling.
* style(desktop): picker row contrast, transparent search fields, centered open
The workspace/branch pickers' rows had a nearly invisible
surface-hover-lighter hover; rows now hover with surface-hover and mark
the current entry with the accent background plus check, matching the
model picker. The search inputs drop the Input base class's
dark:bg-input/30 tint that rendered a gray box inside the panel in dark
mode. Opening a picker now centers the current workspace/branch via a
shared scroll helper instead of starting at the top of the list.
* fix(ui): visible option hover/selected states and no scroll-jump on hover
The option row stacked bg-transparent with the conditional state
backgrounds; at equal specificity the later-sorted bg-transparent
utility won, so hover/selected rows rendered with no background at all.
The background classes are now mutually exclusive.
Mouse-driven active-row changes also reused the keyboard scroll-into-
view effect: hovering a row at the panel edge scrolled it into view,
which moved the list under the cursor and re-triggered hover — an
endless jump. Scroll mode is now per-source: center on open, nearest
for keyboard/typing, none for hover.
* fix(desktop): show only subscribed and free tiers in the cline-pass picker
The ClinePass offer is exactly the feed's subscribed + free tiers, but
stale bundled/cached catalog entries (e.g. a nemotron model) leaked
into an "All models" tier. Match the CLI's featured picker: hide
catalog leftovers, and only fall back to the full catalog when the
subscribed bucket is empty so a subscriber is never limited to free
models offline.
* fix(ui/desktop): strengthen the selected-row highlight in light mode
The selected row used the semantic accent surface (violet step 3),
which is nearly white in light mode. SearchCombobox and the desktop
workspace/branch pickers now highlight the selected/current row with
accent step 4 (with a fallback to --accent), which reads clearly in
both themes without touching the shared --accent token that shadcn
hover states depend on.
* fix(desktop): fit full provider display names in the composer trigger
"Cline Usage-Billing" — the default provider — truncated to
"Cline Usage-Bi…" at max-w-36; the trigger now allows up to max-w-56,
which fits the longest built-in provider names.
* style(ui/desktop): animate picker panels open like the shadcn dropdowns
The thinking-effort Select (shadcn/Radix) animates open while the
model/provider/workspace/branch pickers popped in instantly. All picker
panels now share the same open treatment — 150ms fade + slight zoom,
sliding from the trigger side. SearchCombobox uses a self-contained CSS
keyframe (consumers may not ship tw-animate-css); the desktop's custom
panels use the app's tw-animate utilities. Both respect
prefers-reduced-motion.
* chore(desktop): drop stale eslint-disable comments in picker search rows
This repo lints with biome; the jsx-a11y/no-autofocus disables were
inert leftovers. Flagged in review.
* fix(ui): hand focus back to the combobox trigger on selection, close on Tab
Selecting an option (Enter or click) unmounted the focused search input
without a new focus target, dropping keyboard users' focus to <body> —
only Escape restored it. And since the search input is the panel's only
tabbable element, Tab always moved focus outside the component while
leaving the popup open behind the new focus target.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): keep the composer model selection inside the picker's visible offer
The active/remembered model was validated against the provider's full
catalog while the picker can intentionally hide models (the ClinePass
offer is exactly its subscribed/free tiers), so a stale remembered model
could become the selection while being absent from the dropdown.
Remembered and default selections (including on provider switch) now
resolve against the picker's visible options, and an explicitly
configured model that falls outside the offer stays active but is
surfaced under a 'Current model' section so the selection is always
visible and re-selectable.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Unify desktop plugins, apps, MCP, and skills into one Plugins hub with a Browse directory mode
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Open the marketplace directory as a modal over the Plugins hub instead of swapping the page
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Rename directory to Marketplace: Browse Marketplace button, Marketplace modal title with icon, search placeholder
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Fix search input focus ring clipped by the Marketplace modal scroll container
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Address Greptile review: keep selected tag chip visible when its count drops to zero, and remount installed tab when a marketplace install completes after the modal closed
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Track marketplace modal mutation flag in a ref so a close click racing a queued render cannot skip the inventory remount
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Make Marketplace its own settings page under Customizations and restore Channels as a standalone page
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Remove icon from Marketplace page header for consistency with other settings pages
* Notify mounted inventory views when the marketplace invalidates the cache so late install completions refresh the Plugins hub
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Fix @ file mentions breaking on paths with spaces
Quote mentions generated by getFileMentionFromPath (Add to Cline /
Fix / Explain / Improve commands) when the relative path contains
spaces, so the mention regex no longer truncates the path at the
first space. Also quote the path part of workspace-prefixed mentions
(workspace:/path with spaces) inserted from the @ context menu, which
previously bypassed quoting because the value does not start with '/'.
Fixes#13338
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Fix import ordering in mentions test (biome organize imports)
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Reduce fix to minimal scope
Revert the webview quoting refactor and extra tests; keep only the
getFileMentionFromPath quoting fix with a single regression test.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Normalize mention paths to posix separators for Windows
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Fix code actions failing with 'command not found' on VS Code 1.134
Code action commands carried arguments (expandedRange, diagnostics),
which routes them through VS Code's CommandsConverter cache. VS Code
1.134 disposes the cached entries before the clicked action executes,
so every lightbulb action failed with 'Actual command not found,
wanted to execute cline.addToChat'.
Drop the arguments so the command id is passed through directly, and
recover the context in the handler instead: getContextForCommand now
expands an empty selection by 3 surrounding lines (matching the old
provider behavior) and gathers document diagnostics intersecting the
range when none are passed explicitly.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Scope gathered diagnostics to the selection/cursor
Match the old CodeActionContext.diagnostics behavior: only include
diagnostics intersecting the range the action was requested for, not
the surrounding lines the text gets expanded to.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Fix @ file mentions breaking on paths with spaces
Quote mentions generated by getFileMentionFromPath (Add to Cline /
Fix / Explain / Improve commands) when the relative path contains
spaces, so the mention regex no longer truncates the path at the
first space. Also quote the path part of workspace-prefixed mentions
(workspace:/path with spaces) inserted from the @ context menu, which
previously bypassed quoting because the value does not start with '/'.
Fixes#13338
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Fix import ordering in mentions test (biome organize imports)
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Reduce fix to minimal scope
Revert the webview quoting refactor and extra tests; keep only the
getFileMentionFromPath quoting fix with a single regression test.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Normalize mention paths to posix separators for Windows
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): don't show 'No sessions found' while session history is still loading
Replace the isLoadingHistory flag with hasLoadedHistory, set only once the
backend has actually answered a list_discovered_sessions request. The sidebar
and Sessions view now keep their loading state until that first definitive
response, so the empty-state copy can no longer appear while history is still
being fetched (or while a failed fetch is being retried).
Also retry a failed initial fetch on the 2s event cadence instead of stranding
the UI until the 12s periodic poll, which is what stretched the misleading
empty state to ~10 seconds after a webview reload when the websocket lost the
race with the page load.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): stop history fast-retry from re-arming after hook unmount
A failed initial fetch that settles after the hook unmounted could schedule a
new retry timer after cleanup had already cleared the refs, leaving the
abandoned hook polling the backend every 2s. Guard scheduleRefresh with a
disposed ref set by the mount effect's cleanup.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): anchor work summary duration on the answer row, not attached pre-tool reasoning
The collapsed 'Worked for Xs' row undercounted wall time whenever a turn's
assistant message contained thinking + tool_use with no narration text: the
canonical projection emitted the reasoning-only row after the tool row (both
stamped before the tool executed), the webview attached that row to the final
answer, and collapseCompletedWork used the answer's earliest attached
reasoning timestamp as the end anchor - excluding the entire tool execution
(e.g. 'Worked for 5s' for a turn with an 8s command).
- webview: end the work span at the answer row's own timestamp, clamped to
the last collapsed row so a fallback answer bubble with a synthetic early
timestamp cannot shrink the duration either
- sidecar: flush pending thinking before a tool_use row so rehydrated
transcripts keep the live-stream order (thinking before its tool call) and
pre-tool reasoning no longer rides on the next answer
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): keep interleaved thinking between the tool calls it separates
Address Greptile review: when one assistant message interleaves thinking
between multiple tool_use blocks, each reasoning segment now projects at its
own position (attached to a text row from its own segment when present,
otherwise as its own row) instead of merging into the first reasoning row,
which displayed later thinking before a tool call it actually followed.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
streamdown renders each Shiki token line as a bare inline span with no
newline text between non-empty lines, and only applies its block line
class when lineNumbers is on. With lineNumbers off (the desktop app's
config) every multi-line fenced block collapsed into one run-on line.
Make the direct line spans under code-block-body display: block in the
shared markdown.css; empty lines keep their height via their lone "\n"
child under white-space: pre.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): treat ClinePass as OAuth-managed in chat credential gate
ClinePass shares the Cline account OAuth credentials (its auth handler
stores under the "cline" provider), so the webview never sees a plain
API key for it. The chat pre-flight check only exempted cline/oca/
openai-codex, so switching to ClinePass while signed in via OAuth
blocked with "Missing API key" even though the sidecar resolves the
stored access token fine (which is why the CLI worked).
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* style: format helpers.test.ts with biome
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(llms): surface provider-executed tool activity as observational events
Provider-executed tool parts (e.g. every tool the Claude Code CLI runs
inside its own session) were dropped by the model-tool guard added for
web search: only declared model tools were re-emitted, everything else
hit continue with nothing yielded. Those sessions modified the workspace
with no tool activity in runtime events, transcripts, or the UI.
Route all providerExecuted parts onto the observational path instead:
emit execution-tagged tool-call-delta and tool-result events, matched by
tool-call ID for providers that omit the flag on the result half. They
stay out of AgentRuntime's execution/approval loop, and the runtime
already persists them as modelToolActivities and projects them for
display.
The AgentModelEvent tool-result variant widens toolName from
ModelToolName to string to carry the provider's own tool names.
* fix(agents): keep turns that are only provider-executed tool activity
A turn consisting solely of observational tool activity has an empty
assistant content array - the activity lives in message metadata, since
projecting it into content would replay tool_use blocks the model never
gets results for. The empty-content guard threw on such turns, erroring
the run and losing the activity from the transcript. Count model-tool
activity as content for the emptiness check (error finishes still
throw); replay stays safe through the codec's empty-content placeholder.
Also drop the trailing text delta from one gateway test so the tool-only
stream shape stays covered end to end.
Publish workflows now always do clean npm installs (no dependency
cache in their test gates), the e2e workflow's cache keys are
exact-match only, and the e2e job drops an id-token permission it
never used.
Slack section blocks reject text longer than 3000 characters. The Slack
action logs that rejection as ##[error] but does not fail the step, so an
over-long changelog drops the release announcement while the run stays
green — cline@3.0.50 (3272 chars) published to npm, tagged, and cut a
GitHub release with no Slack post and nothing red to notice.
Every publish workflow pasted the changelog section verbatim into one
section block, so all six were exposed; the SDK, desktop, and extension
sections were only 150-350 chars under the ceiling.
Add a slack_content output alongside content: unchanged when the section
fits, otherwise trimmed on a line boundary with a link to the full
release notes. Only the Slack payload uses it — GitHub release bodies and
the desktop updater manifest still get the whole section.
* fix(shared): run PowerShell commands with fail-fast error semantics
The run_commands PowerShell wrapper never set $ErrorActionPreference, so
the default 'Continue' applied: a pipeline erroring per item (e.g. a
malformed Where-Object over Get-ChildItem -Recurse) emitted one error
record per enumerated file - tens of thousands of stderr records on
large trees, looking like a hang - and could still resolve as SUCCESS
with exit 0.
Prepend $ErrorActionPreference='Stop'; to the script content executed
by the ScriptBlock so the first error terminates the command with a
non-zero exit and a single error message. Concatenated on the same line
as the user command so error line numbers stay unshifted.
Fixes#13285
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(shared): set the fail-fast preference in the bootstrap scope
Setting $ErrorActionPreference='Stop' by string-prepending it into the
scriptblock source displaced a leading param(...) from its mandatory
first-statement position, so scripts beginning with a param block failed
with CommandNotFoundException. Preference variables are dynamically
scoped, so setting Stop in the -Command bootstrap gives the invoked
scriptblock identical fail-fast semantics while keeping the user script
byte-identical (param works, error positions unshifted) and drops the
doubled-quote escaping.
* docs(shared): document the fail-fast tradeoffs in the PowerShell wrapper
Stop promotes every non-terminating error, not only per-item pipeline
floods: partial-result commands (recursive listings over access-denied
junctions) now stop at their first error, and Windows PowerShell 5.1
turns in-script stderr redirection of succeeding native commands fatal.
State this in the wrapper comment as a deliberate tradeoff, with the
GitHub Actions precedent and the per-command opt-outs.
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
Co-authored-by: Mikołaj Kondratek <19799111+mkondratek@users.noreply.github.com>
Turns that settle through the event stream (queued prompts, including the
first prompt of a fresh session) resolve their send() RPC early, so nothing
cleared the streaming shimmer or reconciled live-streamed content against
the persisted transcript at turn end. A turn whose deltas were incomplete
stayed visually streaming forever and only healed when a later non-queued
send rehydrated history.
chat_done (and chat_session_ended / the queue-drain double check) now clears
the active assistant streaming id and schedules a short-delayed
read_session_messages + applyCanonicalHistory, guarded by turn epoch,
session id, and in-flight send submissions so it never clobbers a newer
turn or duplicates the blocking send path's own finalization.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): show typed slash command instead of expanded skill markdown
The sidecar expands /skill and /workflow tokens into their instructions
before dispatching, so the runtime's persisted transcript only contains the
expanded text. After a turn (and when reopening a session) the webview
re-hydrates from that history and rendered the whole SKILL.md body as the
user's message; queue events echoing the expanded prompt could also add a
second user bubble, and fresh sessions were titled with the markdown's first
line. The CLI never shows this because its TUI keeps the typed text in its
own transcript and only sends the expanded prompt to the model.
Mirror that separation inside the desktop sidecar's display boundaries:
- history projection (readSessionMessages) inverts user text that starts
with a configured command's instructions back to '/name remainder',
which also repairs sessions recorded before this fix
- queue snapshots and chat_queued_prompt_start events echo the typed
prompt recorded at expansion time, so the webview's optimistic-bubble
re-key matches again
- an untitled session sent an expanded prompt gets titled from the typed
command instead of the instructions' first line
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): don't overwrite a mid-turn rename with the typed-command title
The untitled check ran before dispatch, so renaming a fresh slash-command
session while its first turn was running got clobbered by the post-turn
typed-command title. Re-check at write time and only replace a missing title
or the one the runtime auto-derived from the expanded prompt.
Also documents the inherent prefix-inversion ambiguity flagged in review:
text hand-typed with a command's exact instructions persists byte-identically
to that command's expansion, so stored history alone cannot distinguish them.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): stop expanding skill commands; let the skills tool load them
Pasting the skill body into the prompt is why the transcript could ever show
it: the desktop webview re-hydrates from the runtime's persisted history, so
whatever the sidecar splices into the user message renders as if the user
typed it. The runtime already registers the skills tool, whose description
requires the model to invoke it whenever the user references a slash command
— so send the typed /skill text through and let the tool deliver the
instructions as a tool result (previously they arrived twice: pasted and via
the tool). The persisted user message, session title, and queue entries are
then simply the typed command, which deletes the typed-prompt registry, the
queue event/snapshot rewriting, and the title machinery from the previous
approach.
Workflows are not served by the skills tool and keep textual expansion, so
the read-time display inverter stays: it collapses expanded workflow prompts
— and skill prompts persisted before this change — back to the typed
/command in the history projection.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* feat(core): option to keep skill slash commands typed for the skills tool
resolveRuntimeSlashCommandFromWatcher (and the hub snapshot proxy) accept
expandSkillCommands: hosts whose sessions register the skills tool pass
false so the typed /skill goes through and the model loads the instructions
as a tool result, keeping the persisted transcript as what the user typed.
Workflows always expand — the tool does not serve them. isSkillsToolAvailable
exposes the catalog check hosts use to decide (yolo preset and the skills
tool toggle leave textual expansion as the only delivery path).
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(cli): skill slash commands load via the skills tool instead of expanding
The TUI user-command wrap and buildUserInputMessage now keep a typed /skill
as-is when the session's mode/toggles register the skills tool, matching the
desktop app; workflows keep expanding, and yolo (zen) keeps expanding skills
because its preset has no skills tool. This also fixes CLI resume/history
surfaces showing the skill body: the persisted user message is now the typed
command.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(vscode): keep configured skill slash commands typed for the skills tool
expandSlashCommands no longer splices a configured skill's instructions into
the model text; the SDK session's skills tool delivers them as a tool result
(previously they arrived twice). Builtin pseudo-skills like /deep-planning
are not served by that tool and keep expanding, as do workflows.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* refactor(desktop): use the shared skill-expansion option in the sidecar
Replaces the sidecar's workflow-detection dance with core's
expandSkillCommands option and gates on isSkillsToolAvailable, restoring
textual expansion where the tool is missing (yolo mode or the skills tool
toggle) — a gap in the previous desktop-only change.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* refactor(desktop): drop the display inverter for expanded transcripts
Accepted trade-off to keep the change minimal: sessions recorded before
skills switched to the skills tool, workflow sends (deprecated), and
yolo-mode skill sends persist expanded instructions and now render that text
as-is instead of being collapsed back to the typed /command at projection
time.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(cli): make TUI dialog colors follow theme changes live
Dialog content previously read the static palette constant, so open
dialogs (including the theme picker itself) kept the default dark-blue
accents while scrolling through theme previews. Add getDialogPalette /
useDialogPalette, which resolve dialog colors from the active theme's
dialog accents and re-render on every theme change, and migrate all
dialog-rendered components to it.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* feat(cli): derive dialog panel background from the active theme
Dark themes now lift their own background one OKLAB step for the dialog
surface, so panels keep the theme's hue instead of the library's fixed
#262626. DialogThemeSync pushes the surface into the dialog container
for new dialogs and repaints open panels, so the surface also follows
live theme previews. Light themes keep the neutral dark panel to match
the dark accent fallback and the light-on-dark dialog text.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix: run_commands object form without args routes through the shell
The structured { command, args? } form of run_commands was always spawned
directly with shell: false. When a model emitted a full command line in
command with no args (e.g. { command: "echo hello" }), spawn failed with
ENOENT for any command containing a space, breaking command execution for
the whole session.
Direct exec now only applies when a non-empty args list is provided; the
object form without args is routed through getShellInvocation like the
string form. Schema descriptions are tightened so models put arguments in
args instead of embedding them in command.
Fixes#13279
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* chore: trim structured-command schema descriptions
The union schema is only used for lenient validation of input the model
already sent; its descriptions never reach a model prompt. Keep them
short instead of restating executor behavior.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* chore: simplify direct-exec comment in shell executor
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* revert: keep original structured-command schema description
The description never reaches a model prompt and the executor now handles
both shapes, so the wording change was cosmetic noise.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix: gate direct exec on args key presence, not array length
Review feedback: an explicit empty args array is intentionally structured
input and stays direct exec; only an object with no args key is treated
as a full shell command line. Matches the key-presence rule already used
by the VS Code host's formatCommandForTerminal. Also replaces the
empty-args shell test (which was PowerShell-incompatible) with a test
pinning the direct-exec contract.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
@cline/ui's generated-media imports @cline/shared/browser, which resolves to
shared's dist output. The build-shared step sat after typecheck/test/build,
so the first ui-publish dispatch since #13025 failed at Typecheck UI with
TS2307. Move the step to right after install.
* feat(ui): share the markdown pipeline, chat polish CSS, and ThinkingBlock
The desktop app and the cloud dashboard both consume @cline/ui yet rendered
assistant output differently, because Markdown policy and the thinking-trace
row lived app-side. This moves the shareable parts into the package:
- components/markdown (new export): the lazy Shiki code highlighter (GitHub
light/dark, pinned language set) and agentMarkdownControls — the standard
Streamdown configuration. streamdown/shiki/@shikijs/* become optional peer
dependencies, mirroring @pierre/diffs.
- components/markdown.css: the desktop's chat polish moves in — chat-scale
headings, outside list markers, single quiet code blocks with a
hover-revealed copy control, table cards. Kept unlayered so it beats
Streamdown's layered Tailwind utilities without !important.
- ThinkingBlock + formatThoughtLabel in agent-chat: the standard thinking
row (brain icon, Thinking/Thought-for-Ns label, streaming shimmer, rail
presentation, capped scrollable body). The shimmer and the
reasoning-hover-suppression rule move into agent-chat.css; triggers gain
the color transition the desktop applied locally.
Version bumps to 0.2.0-next.5 for the dashboard to pick up.
* refactor(desktop): consume shared markdown and thinking primitives from @cline/ui
The local Shiki highlighter, Streamdown controls, chat markdown polish CSS,
streaming-title shimmer, and reasoning hover-suppression rule are deleted in
favor of the @cline/ui versions (the highlighter test moves to the package's
suite). ReasoningBlock becomes a thin wrapper that hands MemoizedMarkdown to
the shared ThinkingBlock, and formatThoughtLabel re-exports from the package
so grouping code and tests keep their import path.
globals.css now imports @cline/ui/components/markdown.css (unlayered, so the
polish keeps beating Streamdown's layered utilities); the app keeps only what
is genuinely app-specific: link/image policy in markdown.tsx, selectability
rules, accent palettes, and the view-enter transition.
* style(ui/desktop): make thinking-trace prose legible
Thinking body text rendered too faint: plain muted-foreground plus the
desktop's font-thin weight. The shared thinking content now leans 75% of the
way back toward the body text color (still slightly de-emphasized), and the
desktop drops the thin font weight.
* feat(ui): add WorkActivity collapsed-run summary and float message actions as a pill
WorkActivity/WorkActivityTrigger/WorkActivityContent fold a finished agent
run's working rows (tool calls, thinking traces, narration) behind a single
"Worked for 4m 12s · 14 tool calls" disclosure built on the shared animated
disclosure primitives, with formatWorkActivityLabel/formatWorkDuration
exported for consumers.
Message hover actions no longer rely on the transcript reserving blank space
below each message: the action row is now a self-backed pill (border,
blurred background, shadow) that floats over whatever follows, so
conversations can pack rows tightly without hover chrome colliding with the
next message.
* feat(desktop): collapse finished runs into a work summary and tighten chat spacing
collapseCompletedWork post-processes the grouped transcript: once a run ends
on assistant text with no further tool calls, its working rows fold into one
expandable WorkActivity row while the final answer stays visible. Runs are
delimited by user messages; the trailing run only collapses when the session
has stopped running and actually produced an answer, so live streams and
cancelled/failed tails keep their rows. Assistant messages carrying images
or media are treated as deliverables and never collapse.
The conversation list gap drops from gap-8 to gap-4 now that hover actions
are self-backed pills that need no reserved space, and user messages add
their own top margin so turn boundaries stay visually distinct.
* refactor(ui/desktop): work summary label wording, flat expansion, stable in-run rhythm
Feedback round on #13315:
- Label reads "Worked for 4m 12s and made 14 tool calls" instead of joining
with a dot; without a duration it falls back to "Made N tool calls".
- Expanded work rows render at transcript level — no rail or extra indent —
since tool rows and thinking traces already carry their own nesting when
expanded. The work content keeps the tight working-row rhythm.
- Live working rows (thinking traces + tool calls) now group into a 'run'
render item with the same tight 0.25rem rhythm, so there is no oversized
gap under a "Thought for Ns" row and every row keeps its exact position
when the finished run folds into the work summary. A trailing
answer-in-progress stays outside the group at transcript level, and pure
prose spans keep normal spacing.
- The transient "Thinking..." indicator moves inside the transcript column
and mirrors a trigger row's geometry, so the first real row replaces it in
place with no jump.
* style(ui/desktop): hover-pill metrics, right-pointing work chevron, scroll and spacing fixes
Another feedback round on #13315:
- Hover action pill: +2px internal padding, a trailing inset after the
timestamp (it sat flush against the pill border), and more clearance
between the message content and the pill (2px -> 6px; the hover bridge
grows to match).
- The work summary chevron points right while collapsed and continues
counterclockwise to point up when expanded.
- Conversation bottom padding drops pb-20 -> pb-8: the composer sits below
the scroller, so the padding only needs to clear a pinned action pill.
- Sending a message scrolls back to the bottom even if the reader had
scrolled up (new AutoScrollOnSend on the user-message count, which ignores
optimistic-bubble re-keying; @cline/ui now exports useConversation for
this).
- An assistant answer directly under its run's working rows pulls itself
0.5rem closer than the full transcript gap.
* style(desktop): leave a visible gap between a pinned action pill and the composer
pb-8 exactly matched the pill's ~40px footprint, so the last row's hover
actions sat flush against the composer top; pb-12 restores ~8px of daylight.
* style(desktop): widen the gap between the pinned action pill and the composer to ~24px
pb-12 left only ~8px of daylight under the pill; pb-16 reads comfortable
without reverting to pb-20's dead space.
* fix(desktop): keep the thinking indicator at the working-row offset mid-run
The indicator matched a trigger row's geometry but sat a full transcript gap
(1rem) below the last working row, while the tool/thinking row replacing it
joins the tight run group at 0.25rem — a visible upward jump. When the last
transcript item is working rows (or streamed assistant output), the
indicator now pulls up to the same tight offset; only at the start of a run,
under the user message, does it keep the normal gap.
* style(ui): calm the hover actions surface per team feedback
Borderless rectangle instead of the bordered pill: radius drops to
var(--radius), the side padding goes entirely (the icon buttons carry their
own hit areas), and the vertical padding halves. Blurred background and
shadow stay so it remains legible over following content.
* feat(ui/desktop): full-band hover reveal and iOS-style disclosure easing
The hover actions only appeared while the pointer was inside the message
box itself. The invisible bridge under each message now spans the full
height of the band the floating actions occupy (full row width), so
hovering anywhere in that strip reveals them. Sibling row types
(.cline-chat-tool, .cline-chat-work, and the desktop's run/tool groups)
become position: relative so they paint above the bridge — their own
content keeps its hover and clicks, and the bridge only wins in the band's
genuinely empty space.
All expandable rows (work summary, tool panels, thinking) open and close on
a 240ms symmetric ease-in-out cubic-bezier instead of the 60ms snap, with
chevron rotation on the same curve. Reduced-motion still disables both.
* revert(ui/desktop): drop the full-band hover reveal; quicken disclosure easing to 180ms
The full-band hover bridge (and the position: relative changes that made it
safe) is reverted per feedback — back to the narrow bridge that only spans
the gap under the message. The iOS-style ease-in-out on disclosures stays
but speeds up from 240ms to 180ms.
* fix(ui): recover live tool diffs that mount as a blank pierre skeleton
Live-streamed edit rows could show an empty diff for the whole run, with the
diff only appearing after the collapsed work row was expanded (fresh mount).
Root cause, confirmed by driving a live session and inspecting the element:
React StrictMode double-invokes @pierre/diffs' ref callback; the first
instance's async highlight work aborts on its immediate cleanup, and the
second instance adopts the abandoned half-rendered shadow tree as if it were
complete prerendered output — zero height, no code, no theme stylesheet,
permanently. A rendered diff always carries style[data-theme-css] in its
shadow root, so ToolFileDiff now checks for it shortly after mount and
remounts FileDiff (bounded attempts) when missing; the fresh host element
takes the normal render path and recovers within ~400ms. Verified live: the
diff now renders during the run.
* fix(desktop): keep interrupted runs expanded even with partial trailing text
The trailing-run collapse gated on 'ended with assistant text', which
misread a Stop that landed mid-answer as a finished run and folded the tool
calls the user wants to inspect. The gate is now the terminal status itself:
only completed (or restored-idle) sessions collapse the trailing run;
cancelled/failed/error tails keep their rows regardless of partial text.
(Greptile P1 on #13315 — matches the PR's stated rule.)
* feat(desktop): show provider web-search support under the settings toggle
The global Web search toggle silently does nothing unless the session's
provider offers native web search, which made the setting read as if it
worked with any provider. The desktop General settings row now explains
that only providers with built-in web search honor it, and shows a live
status line: which connected providers are ready to use it (no extra
setup needed), or an amber warning with a link to the Models section
when none of them support it.
Support is resolved in the webview via a new providerOffersModelTool
helper in @cline/llms (browser export), sharing the same builtin-manifest
source of truth as the runtime's supportsModelTool attachment check.
* fix(desktop): address review — refetch web-search status on catalog invalidation, clarify per-model support
Greptile P2: the one-time catalog fetch could race an in-flight provider
save and show stale status; the row now refetches when the provider
catalog cache is invalidated (fired after saves complete).
Greptile P1: the ready line implied every model on the provider works;
Vertex excludes Claude routes, so the copy now scopes the promise to
models that support it.
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
* fix(vscode): clear task-scoped settings overlay when task view is cleared or switched
Toggling an auto-approve setting while a task is open writes
autoApprovalSettings into the StateManager's task-settings overlay
(updateAutoApprovalSettings -> setTaskSettings). The SDK controller never
cleared that overlay on clearTask/showTaskWithId (the legacy controller
did), so after New Task the stale overlay kept shadowing global settings
in getGlobalSettingsKey(): toggle RPCs were accepted into global state,
but every posted state still carried the overlay's old version, which the
webview rejects as not newer - the auto-approve checkboxes froze forever.
Restore legacy parity in SdkTaskControlCoordinator: drop the overlay
(persisting pending writes first) in clearTask() and before installing a
different task's proxy in showTaskWithId().
Fixes#13260
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* changeset
* test(vscode): add end-to-end regression test for auto-approve freeze after New Task
Wires the real StateManager, the real updateAutoApprovalSettings handler,
and the real SdkTaskControlCoordinator.clearTask() together with the
webview's version gate modeled on ExtensionStateContext, pinning the
end-to-end invariant behind #13260: checkbox toggles must keep reaching
the webview after a mid-task toggle followed by New Task. Verified the
test fails when the clearTaskSettings() call is removed from clearTask().
* fix implicit any in regression test
---------
Co-authored-by: Saoud Rizwan <7799382+saoudrizwan@users.noreply.github.com>
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
The branch dispatch input was a free-form string with no validation. Both
jobs checked it out and ran full npm lifecycle scripts from it: the publish
job next to VSCE_PAT/OVSX_PAT (and npm run publish:marketplace executes a
script from that same ref with the PATs in env), and the test job with NO
environment approval at all while inheriting the workflow-level
contents/packages/checks/pull-requests write grants. A dispatch pointing at
e.g. refs/pull/N/head would run outside-contributor code with the
marketplace keys behind one approval, or with a repo-write token behind
none.
Remove the input and hardcode the protected legacy-extension branch, drop
the workflow-level permissions to contents: read, and elevate only the
publish job to contents: write (tag push + GitHub release). The branch
input's default was legacy-extension, so normal publishes are unchanged.
publish-extension skill dispatch command updated to match.
The combined-VSIX workflow took legacy-ref as a free-form dispatch input
with no publish-time validation (next-ref has one: publish requires main).
Any typed ref — a PR merge ref, an unprotected branch — would be built
into the published VSIX by the environment-less build job, and the publish
environment approver only ever sees an opaque prebuilt artifact, so the
approval protected the marketplace PAT but not the shipped bytes.
Remove the input entirely and hardcode the protected legacy-extension
branch, which makes that branch's protection rules load-bearing for
releases. The tested-sha pinning between test-legacy and build is
unchanged. publish-extension skill dispatch command updated to match.
* fix(vscode): point Mistral signup URL at the general API keys console
The Mistral provider's signup link led to the Codestral console, which
issues Codestral-scoped keys that fail with 401 on api.mistral.ai — the
endpoint the provider actually calls. Point it at the general API keys
page instead.
Fixes#13288
* fix(vscode): deep-link DeepSeek and Fireworks signup URLs to their API key pages
Both pointed at marketing homepages; link straight to the key-creation
pages instead, matching the rest of the registry and the desktop app's
provider-key-urls map.
* fix(vscode): stop resubmitting the original task text on bare resume (#12975)
A bare Resume after Stop rebuilt the session from task history and
injected historyItem.task into the resumption prompt as 'New
instructions from the user'. The model treated the already-completed
original request as fresh instructions and re-executed it (e.g. re-ran
all terminal commands after stopping a queued follow-up turn).
The preserved conversation history is the source of truth on resume, so
the fallback prompt now just asks the model to reassess the history and
continue, matching the legacy resume prompt which also never resent the
original task. User-typed text still takes precedence when provided.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(vscode): continue the surviving idle session on follow-ups instead of rebuilding
Stopping a turn keeps the session alive, but every idle follow-up (bare
Resume after Stop, and typed follow-ups after a completed turn) tore
that session down and rebuilt it from persisted task history before
sending. Continue the matching idle session in place instead, the same
way the CLI reuses the live session after an abort. Rebuilding from
history now only happens when no live session matches the displayed
task (task opened from history, extension host reload).
A bare resume still needs a prompt to start a turn, so it sends the
neutral [TASK RESUMPTION] prompt (shared with the rebuild fallback and
hidden from the transcript); user-typed content is echoed and sent
as-is. If the send lands while the abort is still settling, the runtime
auto-queues it and drains once the abort completes.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* refactor(vscode): consolidate follow-up send paths in SdkFollowupCoordinator
Now that idle follow-ups continue the live session in place, the
two-mode sendToActiveSession helper was redundant: its non-queued branch
duplicated continueIdleSession minus the bare-resume prompt. Split it
into a single-purpose queueToActiveSession and fold the idle no-task
send into continueIdleSession, flattening askResponse's decision tree
to: queue onto a running turn, continue a matching live idle session,
rebuild from history, or abandon.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* refactor(vscode): reuse the existing neutral resumption prompt for bare resumes
Drop the newly invented long resumption wording in favor of the phrase
that already existed as the no-history fallback and that the transcript
hiding logic and test fixtures recognize: '[TASK RESUMPTION] Please
continue where you left off.' The net change to resumeSessionFromTask
against main is now just deleting the branch that resubmitted
historyItem.task as new instructions.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(vscode): stop resubmitting the original task text on bare resume (#12975)
A bare Resume after Stop rebuilds the session from task history and
injected historyItem.task into the resumption prompt as 'New
instructions from the user'. The model treated the already-completed
original request as fresh instructions and re-executed it (e.g. re-ran
all terminal commands after stopping a queued follow-up turn).
Bare resumes now always use the neutral prompt that already existed as
the no-history fallback; user-typed text still takes precedence. This
matches the legacy resume prompt (responses.taskResumption), which only
ever included user-supplied text as new instructions.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(vscode): hide synthetic prompts from the queued-prompt echo
A send that races a settling abort is auto-queued by the runtime, so a
bare Resume can reach the pending_prompt_submitted echo carrying the
synthetic [TASK RESUMPTION] prompt. Echoing it leaked model-facing text
as a visible user bubble and shifted the visible-user-message ordinals
that edit/regenerate mapping relies on. Filter synthetic prompts with
isSyntheticUserPrompt, keeping user attachments visible (matching
isSyntheticSdkUserMessage semantics).
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
A restore that reuses the source session id rolled the workspace back but
left the persisted transcript describing the discarded turns, so the chat
kept showing turns whose file changes had just been reverted.
Before #13075 the restore reply carried the trimmed messages and the
webview rendered them directly. Now the webview always re-reads through
read_session_messages, which prefers the persisted file over the live
session, so the trimmed history the sidecar puts on the live session is
never read. Persist it as well.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
Beta builds (prerelease versions from desktop-experimental, shipped as
'Cline Code Beta') now identify themselves everywhere users look: a Beta
pill in the sidebar footer, the product name in the sidebar hover card,
an About row in Settings > General with version + channel, the runtime
window title, and the tray menu/tooltip (via package_info, which carries
the overlay's productName).
Channel detection is a pure version-string check (-beta suffix) in the
new webview/lib/app-channel.ts — the version is baked into package.json
at build time and reported by the sidecar's get_process_context, so it
works in both the Tauri shell and web dev mode with no new plumbing.
Stable builds render no channel UI at all.
* feat(desktop): add beta release channel from desktop-experimental branch
Adds a 'channel' input (stable|beta) to desktop-publish.yml. Beta releases
are tagged desktop-vX.Y.Z-beta.N on the desktop-experimental branch, built
with the tauri.beta.conf.json overlay (Cline Code Beta / bot.cline.app.beta,
side-by-side install with stable), published as prerelease GitHub releases,
and served by a separate rolling desktop-beta update feed. Both channels
dispatch from main so the PublishDesktop signing gates are unchanged.
Guards: stable channel now rejects prerelease tags (previously a beta tag
could clobber desktop-latest and auto-update every stable install onto it),
feed selection is fail-closed and cross-checked in the release job, and the
build asserts the compiled binary embeds exactly its own channel's feed URL.
Changelog extraction is exact-version now that stable and beta sections
interleave across branch merges.
Process doc in apps/examples/desktop-app/EXPERIMENTAL.md; publish-desktop
skill now asks stable-or-beta.
* docs(desktop): warn against renaming the desktop-latest feed
* docs(desktop): document the code-trust model for publish approvals
The beta dispatch-from-main invariant protects the workflow definition, not
the checked-out tag's build scripts, which run with signing secrets in scope
for stable and beta alike. Make explicit that the PublishDesktop reviewer
approval is the trust gate for that code, and that desktop-experimental
therefore needs main-grade merge controls.
* feat: add image generation support
* fix(llms): preserve mixed image model behavior
* fix(llms): validate generated image models
* fix(llms): preserve mixed image response streaming
* fix(llms): preserve runtime tool ownership
* fix(llms): address image generation review feedback
* fix(desktop): relay images for attached hub sessions
* chore(llms): regenerate provider and model catalog
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(vscode): preserve SDK model capabilities across the catalog boundary
The new modelSupportsToolCalling gate treats a populated capability list
without "tools" as authoritative. But the VS Code host round-trips model
metadata through the legacy ModelInfo shape, and toSdkModelInfo
reconstructed capability arrays from the legacy booleans alone — which
have no "tools" projection. Every model with any capability flag set
came back as "cannot call tools", so sessions registered zero tools and
the file-edit e2e failed on all platforms (the editor tool call resolved
to "Unknown tool" and the edit never reached disk).
Fix, following the modalities-passthrough pattern so stacked capability
PRs can reuse it:
- Preserve the SDK capability list verbatim on legacy ModelInfo at the
catalog boundary (adaptSdkModelInfo); union user overrides into it
without ever fabricating a list from overrides alone.
- Seed toSdkModelInfo from the preserved list, and when none survived,
emit an explicit "tools" signal (honoring legacy supportsTools=false)
so reconstructed arrays can never silently disable tool calling.
- Add a shared modelHasCapability(model, capability,
{assumeWhenUnspecified}) helper: missing or empty capability lists
carry no signal and each check declares its own default. Future
capability gates should route through it instead of reading
model.capabilities directly.
Verified: file-edit e2e (Single Root + Multi-Roots) passes locally;
shared/core/llms/model-catalog/session-factory suites and typechecks
pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore(llms): refresh generated model catalog
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The Vertex case asserted that a bare providerId resolves to a model
without web search, which only held because the generated catalog's
Vertex default happened to be a Claude route. models.dev has since moved
that default to gemini-3.7-flash, which does support native web search,
so the assertion failed on the next catalog regen while the behavior it
guarded was unchanged.
Drop the catalog-dependent case and cover the default-model fallback
against a synthetic manifest instead, where the excluded route is stated
by the test rather than inherited from upstream data.
The outdated_hub notice reports a state the user cannot act on: this CLI
is already the newer build, the Hub is behind only because retiring it
would kill the sessions it is serving, and the swap happens on its own at
the next launch. A toast that interrupts to say "no action needed" is
still an interruption, and the desktop surface already concluded the same
thing by rendering nothing for this reason.
It also could not deliver the message it existed for. Toast caps at
maxWidth = Math.min(44, width - 4), and the 61-character string did not
wrap, so what actually rendered was "Update finishes the next time Cline"
- a sentence cut off before the reassuring half. Identical at 120 and 200
columns, so widening the terminal did not help.
The classification stays in core and still earns its keep at this call
site: outdated_hub is what stops the update-and-restart prompt from
firing at someone who has nothing to update. Only the rendering goes.
The build_mismatch direction, where the user does have something to do,
is untouched.
Render assistant markdown with internalBlockMode="top-level" so each
top-level markdown block gets its own renderable. The default coalesced
mode merged the entire message into one block that was rebuilt and
re-highlighted on every streamed chunk, flashing settled headings and
links back to raw uncolored markdown (visible ###, unconcealed syntax)
until the async tree-sitter highlight landed, and re-wrapping rows so
the transcript jumped vertically.
Top-level blocks are reused by token identity, so settled content never
re-renders; only the trailing unstable block updates per chunk. Pass
tableOptions style=grid to keep the bordered table rendering coalesced
mode used by default.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(core): stop concurrent Hub installs from retiring each other
Two Cline installations on different builds would shut each other's Hub
daemon down in a loop, and every session died with an abnormal socket
close (code=1006) as its daemon was killed mid-handshake.
The retire decision was a one-sided predicate: each client independently
asked "may I reuse this Hub?", and two clients on differing builds both
answered no. #13177 added build-epoch ordering to break the tie, but left
every unordered case - missing epoch, missing build id - retiring as
before, so any pair involving a build from before epochs were embedded
still looped.
Derive the decision from a total order instead. compareHubBuilds orders
two builds by embedded epoch, then core release version, then build id,
and is antisymmetric by construction, so at most one side of a pair can
ever decide to retire. A Hub that is newer or cannot be ordered is
attached over the compatible wire protocol and left to the build-mismatch
watcher to prompt about. Genuine protocol incompatibility still replaces.
Identity is now read from the same fields on both sides. Filling in a
coreVersion locally that the wire record omits made a build's identity
depend on which role it was playing, and the two directions of a pair were
then decided by different tiers with both concluding they were newer -
a second, independent way to produce the loop.
Also:
- Scope the development Hub owner by build id, so differing dev builds run
their own daemon side by side instead of contending for one record.
Production keeps its singleton.
- Break the circuit after repeated retirements of the same URL, bounding
any future ordering bug to a stale-build prompt rather than an
unusable Hub.
- Report `cline doctor fix` honestly: separate processes that survived a
kill from ones that appeared while the fix ran, name the live parent
respawning a daemon, and mark a startup lock held by a running process
as held rather than leaked.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(cli): only blame a live parent for processes seen during doctor fix
The advice printed under "started during fix" asserted that every such
process was respawned by a live parent, and told the user to go stop it. A
process can also start on its own mid-repair - someone opening a new
session - and then the instruction points at an unrelated process, or at
none at all.
Derive the wording from whether a live parent actually exists: name it
when every process has one, state the facts when none do, and split the
list when it is mixed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* test(core): order the builds in the stale-discovery hub server case
The case stubbed two build ids and expected the second to replace the
first, but supplied nothing that says which came first: no epochs, and both
servers report the same core version. Ordering therefore fell to the
build-id tiebreak, where "new-build" sorts before "old-build" and the
replacement was judged the older of the two.
Give the case the epochs its name implies, and add the missing sibling for
an unorderable pair, which is attached to rather than retired - the
behavior that keeps two installations from shutting each other down.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(core): defer replacing a Hub that is serving live sessions
Retiring a Hub kills its established WebSockets, so replacing one under a
running session ends that turn with an abnormal close (code=1006). The
replacement is correct - the newer build should own the Hub - but the
timing is not the user's to absorb mid-turn.
Defer instead while the Hub reports live sessions: the newer client
attaches to the older Hub over the compatible wire protocol, and the swap
happens once those sessions end. Attaching rather than spawning matters -
a second daemon would race the busy one for the port.
Deferring silently would be worse than the interruption it avoids, because
a long-lived session pins the Hub to old code indefinitely with nothing to
show for it. The build-mismatch watcher only ever prompted in the
direction where updating the client resolves the mismatch; its own comment
notes that older Hubs "are retired and replaced automatically, so
prompting would only flash a stale dialog", which stops being true once
replacement can be deferred.
Add the missing direction as `outdated_hub`, reported only when a mismatch
survives consecutive checks - an idle older Hub is replaced within moments
of being seen, so a single sighting would flash exactly the stale dialog
the original comment warns about. The CLI and desktop dialogs render it as
information rather than an update prompt: nothing to install, the Hub
swaps itself when the sessions end.
The direction is decided by compareHubBuilds rather than reusability,
because a Hub that is newer and one that carries too little metadata to
order are both "reusable" but need opposite advice.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(core): key the outdated-Hub check by daemon instance, not build
The consecutive-sighting check that keeps a routine replacement from
flashing an informational dialog was keyed by build id. Two daemons from
the same build share one, so an outdated Hub replaced by another daemon of
the same older build satisfied the check and reported exactly the churn the
check exists to hide.
Carry a hubInstanceId on the mismatch event - the Hub's own id, falling
back to pid and start time - and key the pending sighting by it. A
replacement instance now restarts the count instead of confirming it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(core): source Hub instance identity from the discovery record
The instance id added in the previous commit was read from the probe
response, but the watcher probes without an auth token and `/health`
deliberately reports only build and address fields - no hubId, pid, or
startedAt. So the id was always undefined in production and the check it
guards still conflated two daemons of the same build. The test missed it by
injecting a hubId into a mocked probe, a shape `/health` never returns.
Take identity from the discovery record instead, which every daemon version
writes with all three fields and which a replacement daemon rewrites as its
own. The probe is still preferred when it does carry an id, since that is
the process just spoken to.
The tests now use the real `/health` payload shape and vary identity through
the discovery record, including the pid-and-start-time fallback for records
written before Hubs carried an id.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(core): confirm the Hub record still describes the daemon just probed
Instance identity is read from discovery before the probe and build data
comes back after it, so a daemon replaced between those two steps was
described with its predecessor's identity - the replacement then satisfied
the prior daemon's pending sighting and emitted the notification the
consecutive-instance check exists to suppress.
Re-read discovery after the probe and report nothing when the record no
longer describes the same daemon. A Hub mid-swap is churn; the next check
sees whatever it settles into.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* revert(core): drop the watcher instance-identity hardening
Reverts the three follow-up commits that keyed the outdated-hub
consecutive-sighting check by daemon instance (42a83beae, 931431371,
9d634f7b3). They guarded one scenario - a different daemon of the same
outdated build swapping in between two watcher ticks - where the only
consequence is an informational dialog showing one interval early or
late. The unauthenticated probe carries no instance fields in
production, which is why the first attempt needed two more patches; the
original reason+buildId consecutive-sighting suppression from this PR's
base commit already covers the case that matters (not flashing a dialog
for a hub that is mid-replacement).
* fix(core): only count sessions that stopping the hub would actually harm
hasActiveHubSessions treated every non-terminal status as busy. But a
session's hub-side runtime outlives its client: a TUI that is killed or
crashes never stops its session, which then sits in the hub with no
participants and a status that never reaches a terminal state. Under the
defer-while-busy rule that pinned the displaced hub as "serving
sessions" forever - it was never retired, every new CLI kept attaching
to the old build, and the outdated-hub dialog recurred with a promise
("replaced once those sessions end") that could never come true.
Verified empirically: a cleanly detached+disposed client leaves its
session status "running" indefinitely.
Busy now means: someone is attached (participants), or a turn may be
executing hub-side (running/pending, which covers headless and scheduled
runs). An idle session with a confirmed-empty participant list is
resumable persisted state, not live work. Hubs from core < 0.0.75 omit
the participants field entirely, so idle stays conservative (busy)
there - an attached client cannot be ruled out.
updatedAt-freshness was considered and rejected as the discriminator:
the sessions row only updates on status transitions, so a single long
agentic turn looks stale while genuinely executing.
* fix(core): gate hub busyness on attached participants only
Simplifies the busy-check to the one signal that cannot go stale:
participants are live socket subscriptions the hub drops the moment a
client's connection closes, so a crashed client can never leave a ghost
that counts as busy. Session status is deliberately not consulted - a
client killed mid-turn strands its session in a non-terminal status
forever, and QA reproduced that pinning an outdated hub as "serving
sessions" until reboot. This replaces the earlier status+participants
heuristic (and drops the aging bound it was growing) with the rule the
deferred-update design stated from the start: the hub is busy while a
client is connected to a session, and replaceable otherwise.
The accepted cost: a participant-less background run executing at the
exact moment of a hub swap dies with the old hub. Rare, and its next
scheduled tick runs normally on the replacement.
* fix(cli): tell the truth about when the outdated Hub is replaced
The outdated-hub dialog and toast said the Hub is replaced "once those
sessions end". It is not: nothing retires a hub except a fresh launch
running the ensure path, so a user who quits the busy session and
watches sees the old hub stay put and concludes something is stuck
(observed in hands-on QA). Say what actually happens - the newer build
takes over the next time Cline starts after those sessions end.
* fix(cli): speak to users, not architecture, in the pending-update notice
"Cline Hub is running an older build" assumes the reader knows what the
Hub is and why builds differ. The user-relevant facts are only: your
update is not fully active yet, your work is safe, and it finishes by
itself. Say exactly that, in both the TUI and desktop dialogs and the
toast, with the version tucked in parentheses for bug reports.
* fix(cli): drop the outdated-hub dialog for a single quiet toast
The dialog interrupted the user to say that nothing is wrong and no
action is needed - the ideal number of modals for that message is zero.
The TUI now shows one info toast ("Update finishes the next time Cline
starts. No action needed.") and the desktop app shows nothing for the
outdated_hub reason; both dialog components return to their shipped
update-and-restart form, which still appears for the build_mismatch
direction where the user genuinely has something to do. The watcher
keeps reporting outdated_hub - surfaces decide, core informs.
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: Saoud Rizwan <7799382+saoudrizwan@users.noreply.github.com>
* fix(core): bridge protections for updates landing under pre-3.0.55 clients
Three pieces, each proven against real released artifacts:
- postinstall shield: CLI versions <= 3.0.54 restart the hub daemon after a
background auto-update even while it serves live sessions, and their
fingerprint check then rejects every replacement hub, bricking the running
TUI. That code is on users' machines and cannot be patched — but it runs
only after the install completes, and it bails out harmlessly when no hub
discovery record exists. The newly installed package's postinstall sets
the record aside so the old updater never fires.
- superseded-record fallback: the set-aside record is also the only source
of the auth token and pid the next new-build launch needs to retire the
displaced hub (a port probe carries neither); ensure reads it back.
- bind retry: a hub retired on the fixed port can hold it ~2s after acking
shutdown (watchdog force-exit); the replacement daemon retries EADDRINUSE
for up to 5s instead of dying and leaving no hub at all.
* fix(cli): defer auto-update install until no CLI is attached to the hub
Installing while cline processes run swaps the npm package under them:
their respawn paths break on the new build fingerprint, and the updater
then restarted the hub daemon out from under live sessions (the 'Hub
connection closed (code=1006)' incident). Guarding the restart treats the
symptom; the fix is to never install under a running process.
The startup check now only records that an update is available. The
install runs at process exit, and only when the hub confirms no other
cli* client is attached — desktop sidecars and connectors ship their own
binaries, so only cli* clients make the swap unsafe. With nothing old
running at install time, no hub restart is needed at all: the next launch
retires the stale hub through the existing ensure path. Deletes
restartHubServerIfRunning, ensureCliHubServerAfterUpdate, and their
support code; manual 'cline update' still installs immediately and now
just notes that the update applies on next start.
* fix(cli): apply deferred update from the entrypoint exit sequence
The CLI entrypoint always terminates with an explicit process.exit(),
which never emits beforeExit — the hook the deferred installer waited on,
so it would never have run (caught by review). Invoke applyDeferredUpdate
directly from the entrypoint's exit sequence after disposeAll(), where
every normal termination passes; crash paths deliberately skip it. Also
clear the pending update once an install spawns so the apply is
idempotent.
* test(cli): isolate unit tests from the real ~/.cline
A full vitest run could leave a real hub daemon running against the
developer's actual ~/.cline discovery record (observed while validating
this PR: a daemon spawned from the globally installed cline binary,
attached to the real data dir). Point CLINE_DIR, CLINE_DATA_DIR, and
CLINE_HUB_DISCOVERY_PATH at a per-worker temp dir and disable auto-update
before any test file loads; subprocesses inherit the isolation via env.
* fix(core): discard the superseded discovery record once consumed
The set-aside record is one-shot recovery metadata, but nothing deleted
it, and it feeds a pid into retireDiscoveredHub's SIGTERM. Weeks later a
launch that finds no live record (routine after any retirement) could
read the stale file and signal whatever process the OS recycled that pid
onto (review finding by @abeatrix). Unlink it at every ensure resolution
that ends with a live, verified hub; failure paths keep it for the next
attempt.
* fix(cli): harden the exit-time update gate
Three review findings on the deferred-apply path:
- A wedged hub could stall an otherwise-finished CLI for tens of seconds
via the hub client's default timeouts; the whole exit-time query is now
bounded to 3s, with timeout counting as attached (never install unless
the hub positively confirms).
- Sub-second commands exited before the startup version check resolved
and silently dropped the update every time for one-shot-only usage;
exit now grants the in-flight check a 250ms grace.
- client.list can lose a TUI's registration during transport churn while
its session connection survives, so an empty client list is not proof
of safety; cross-check sessions with participants. Participants rather
than session status: finished sessions linger idle forever and must
not pin updates, and participant-less scheduled runs live in the hub
process, which the binary swap does not touch. Verified live: a
session-holding client invisible to client.list defers the install,
and the gate opens once it disconnects.
* docs(cli): fix stale beforeExit reference in the exit-gate comment
* style(cli): apply biome formatting to update deferral code
* fix(cli): let doctor see a hub whose record the update shield set aside
During the shielded update window the discovery record is renamed to
.superseded so pre-3.0.55 updaters cannot restart a busy hub. Doctor
read only the primary record, so in that window it reported the live
daemon - the one serving the user's still-open old session - as a stale
hub daemon and advised 'cline doctor fix', which kills it and reproduces
the exact 1006 incident the shield exists to prevent (found by QA).
Doctor now falls back to the set-aside record the same way the ensure
path does, and doctor fix clears the set-aside file along with the
primary record so a deliberate reset does not leave stale retirement
metadata pointing at a recyclable pid.
* fix(core): keep shielded sessions on one Hub authority (#13244)
* fix(core): recover shielded busy hub discovery
* chore(core): instrument shielded hub recovery
* fix(core): recover shielded hubs with attached clients
* fix(cli): recognize shielded hubs in doctor
* refactor(core): keep shield recovery minimal
* fix(core): retain shared Hub idle helper semantics
* chore(core): align busyness helper with the #13231 wording
The participants-only hasActiveHubSessions here duplicates the change on
bee/hub-lifecycle (this branch needs its semantics for the participant
gate). Matching that version byte for byte lets the two merges resolve
cleanly instead of conflicting. Also restores the module-registry reset
comment this branch dropped - it documents a real local-vs-CI gotcha.
* fix(core): stop concurrent Hub installs from retiring each other
Two Cline installations on different builds would shut each other's Hub
daemon down in a loop, and every session died with an abnormal socket
close (code=1006) as its daemon was killed mid-handshake.
The retire decision was a one-sided predicate: each client independently
asked "may I reuse this Hub?", and two clients on differing builds both
answered no. #13177 added build-epoch ordering to break the tie, but left
every unordered case - missing epoch, missing build id - retiring as
before, so any pair involving a build from before epochs were embedded
still looped.
Derive the decision from a total order instead. compareHubBuilds orders
two builds by embedded epoch, then core release version, then build id,
and is antisymmetric by construction, so at most one side of a pair can
ever decide to retire. A Hub that is newer or cannot be ordered is
attached over the compatible wire protocol and left to the build-mismatch
watcher to prompt about. Genuine protocol incompatibility still replaces.
Identity is now read from the same fields on both sides. Filling in a
coreVersion locally that the wire record omits made a build's identity
depend on which role it was playing, and the two directions of a pair were
then decided by different tiers with both concluding they were newer -
a second, independent way to produce the loop.
Also:
- Scope the development Hub owner by build id, so differing dev builds run
their own daemon side by side instead of contending for one record.
Production keeps its singleton.
- Break the circuit after repeated retirements of the same URL, bounding
any future ordering bug to a stale-build prompt rather than an
unusable Hub.
- Report `cline doctor fix` honestly: separate processes that survived a
kill from ones that appeared while the fix ran, name the live parent
respawning a daemon, and mark a startup lock held by a running process
as held rather than leaked.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(cli): only blame a live parent for processes seen during doctor fix
The advice printed under "started during fix" asserted that every such
process was respawned by a live parent, and told the user to go stop it. A
process can also start on its own mid-repair - someone opening a new
session - and then the instruction points at an unrelated process, or at
none at all.
Derive the wording from whether a live parent actually exists: name it
when every process has one, state the facts when none do, and split the
list when it is mixed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* test(core): order the builds in the stale-discovery hub server case
The case stubbed two build ids and expected the second to replace the
first, but supplied nothing that says which came first: no epochs, and both
servers report the same core version. Ordering therefore fell to the
build-id tiebreak, where "new-build" sorts before "old-build" and the
replacement was judged the older of the two.
Give the case the epochs its name implies, and add the missing sibling for
an unorderable pair, which is attached to rather than retired - the
behavior that keeps two installations from shutting each other down.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* feat(core, llms): Cline custom provider & web search
* fix(llms): preserve reasoning model token parameter
* fix(llms): keep ClinePass provider options on the wire in the shared Cline provider
The shared Cline provider hardcoded the AI SDK provider name to "cline",
but the openai-compatible model reads request-body passthrough options from
providerOptions[<name>]. Option routing emits ClinePass options under the
"cline-pass"/"clinePass" buckets, so gateway reasoning (extended thinking
budgets) silently stopped reaching the wire for cline-pass after it moved
off the generic openai-compatible module.
Thread the gateway provider id through as the provider name, and restore
strictJsonSchema: false for the new "cline" provider-options target so the
wire format matches the previous openai-compatible behavior. Add cline-pass
coverage at both the option-routing and request-body levels.
* feat(sdk): persist provider-executed tool activity (#13077)
* feat(core, llms): Cline custom provider & web search
* fix(llms): preserve reasoning model token parameter
* feat(sdk): persist provider-executed tool activity
* fix(vscode): restore state proto and settings section reverted by merge
The merge of origin/bee/websearch into this branch resolved conflicts by
keeping this branch's pre-#13126 copies of apps/vscode files, which
deleted the auto_approve_all_toggled = 174 proto field (without reserving
the number) and dropped a formatting line in FeatureSettingsSection.tsx.
Neither file is in scope for this PR. Restore both to main's content so
the proto source matches the checked-in generated code again.
* chore(vscode): match main byte-for-byte in FeatureSettingsSection.tsx
The pre-commit biome hook strips a blank line that exists on main, which
kept this out-of-scope file in the PR diff. Commit the exact main content
with --no-verify so the PR no longer touches apps/vscode at all.
---------
Co-authored-by: Saoud Rizwan <7799382+saoudrizwan@users.noreply.github.com>
* refactor(llms): key ClinePass provider options to the shared cline bucket
Both Cline gateway ids (cline and cline-pass) are served by the same
shared "cline" AI SDK provider and hit the same Cline API, so threading
the gateway provider id through as the AI SDK provider name (78dc6f3e7)
was unnecessary indirection. Revert the name threading and instead
normalize option-routing bucket keys: buildProviderAndAliasPatch now
keys both Cline gateway ids to the shared "cline" providerOptions
bucket, which is the only bucket the openai-compatible model reads for
request-body passthrough.
Also tighten the regression coverage that motivated the original fix:
the previous effort-based test rows were vacuously satisfied through the
portable-reasoning early return (effort reasoning never reaches provider
option buckets by design). The rows now use explicit reasoning budgets,
which do flow through the gateway bucket path, and the wire-level test
composes real provider options end to end instead of hand-feeding
buckets.
* revert(llms): drop the cline strictJsonSchema special case in generic-compatible
Restores buildCompatibleProviderOptions to its pre-78dc6f3e7 state. The
strictJsonSchema passthrough is verified inert for the gateway (nothing
in @cline/llms sets a response format), so keeping a hardcoded provider
target in the generic helper bought nothing. If structured outputs are
ever added, strictness for the cline target can be decided deliberately
then.
* fix(llms): claim native web search for openai-native, not the openai alias
supportsModelTool listed "openai", but that id aliases to
openai-compatible (PROVIDER_ID_ALIASES), whose module has no native web
search. The actual native OpenAI builtin id is "openai-native", which is
served by the OpenAI Responses module that does implement
buildModelTools with provider.tools.webSearch(). Without this, the
web_search tool was never offered to native OpenAI users, and was
wrongly offered for the compatible alias.
* refactor(llms): declare model tools in provider manifests
* feat(sdk): project provider tool activity in session history
---------
Co-authored-by: Saoud Rizwan <7799382+saoudrizwan@users.noreply.github.com>
Fix auto-approve checkboxes freezing after "New Task": clear the task-scoped settings overlay when the task view is cleared or switched, so stale task settings no longer shadow global settings
description:Use when preparing, tagging, and publishing a Cline Code desktop app (apps/examples/desktop-app) release. Guides changelog drafting, version bumps in package.json + tauri.conf.json, desktop-vX.Y.Z tags, and the desktop-publish GitHub workflow that builds, signs, notarizes, and updates the auto-update feed.
description:Use when preparing, tagging, and publishing a Cline desktop app (apps/examples/desktop-app) release — stable (desktop-vX.Y.Z from main) or beta (desktop-vX.Y.Z-beta.N from desktop-experimental, shipped as the side-by-side "Cline Beta" app). Guides changelog drafting, version bumps in package.json + tauri.conf.json, tagging, and the desktop-publish GitHub workflow that builds, signs, notarizes, and updates the per-channel auto-update feed.
---
# Desktop App Release
Use this skill when the user asks to release the desktop app, publish Cline Code, bump the desktop version, create a `desktop-vX.Y.Z` tag, or trigger the desktop publish workflow.
Use this skill when the user asks to release the desktop app, publish the Cline desktop app, cut a desktop beta, bump the desktop version, create a `desktop-vX.Y.Z` (or `desktop-vX.Y.Z-beta.N`) tag, or trigger the desktop publish workflow.
> Working directory: run every command below from the repository root.
Desktop releases are macOS-only today (a single signed + notarized universal DMG that runs natively on both Apple Silicon and Intel) and are built entirely in GitHub Actions — there is no local publish path. Installed apps discover new releases automatically through the Tauri updater, so publishing a release is what ships the update to every existing user.
Desktop releases are macOS-only today (a single signed + notarized universal DMG that runs natively on both Apple Silicon and Intel) and are built entirely in GitHub Actions — there is no local publish path. Installed apps discover new releases automatically through the Tauri updater, so publishing a release is what ships the update to every existing user**on that channel**.
## Release contract
- Two channels, one workflow (`channel` input on `desktop-publish.yml`):
- **stable** — tag `desktop-vX.Y.Z` (no suffix; the workflow rejects prerelease suffixes on this channel), cut from `main`, feeds the rolling `desktop-latest` release, ships as "Cline".
- **beta** — tag `desktop-vX.Y.Z-beta.N`, cut from `desktop-experimental`, feeds the rolling `desktop-beta` release, ships as "Cline Beta" (separate bundle identifier `bot.cline.app.beta`; installs side by side with stable). Built with the extra `src-tauri/tauri.beta.conf.json` overlay. Process background: `apps/examples/desktop-app/EXPERIMENTAL.md`.
- Version sources (must match each other and the tag): `apps/examples/desktop-app/package.json` and `apps/examples/desktop-app/src-tauri/tauri.conf.json`. (`src-tauri/Cargo.toml` has its own version but `tauri.conf.json` overrides it; no need to touch it.)
-Release tag: `desktop-vX.Y.Z`, where `X.Y.Z` matches both version files.
- Release prep includes approved release notes, the version bumps, and an `apps/examples/desktop-app/CHANGELOG.md` update.
- Publish path: `.github/workflows/desktop-publish.yml` (workflow_dispatch, requires the tag to exist, point at the checked-out commit, and be reachable from `origin/main`).
-The workflow creates the `desktop-vX.Y.Z` GitHub release (universal DMG + updater artifact + `latest.json`) and refreshes the rolling `desktop-latest` release, which is the static auto-update feed every installed app polls. Never delete the `desktop-latest` release or tag.
- The changelog's top `## X.Y.Z` section is extracted verbatim into the GitHub release body, the Slack announcement, and the updater manifest notes.
-Beta versions are prereleases of the **next** stable: stable `0.0.13` → betas `0.0.14-beta.1`, `-beta.2`, … Once a stable ≥ the beta base ships, the next beta bumps its base (`0.0.15-beta.1`).
- Release prep includes approved release notes, the version bumps, and an `apps/examples/desktop-app/CHANGELOG.md` update — committed on `main` for stable, on `desktop-experimental` for beta.
- Publish path: `.github/workflows/desktop-publish.yml` (workflow_dispatch, requires the tag to exist, point at the checked-out commit, and be reachable from the channel's branch — `origin/main` for stable, `origin/desktop-experimental` for beta).
-**Both channels dispatch from `main`.** This is a security invariant, not a convenience: the run executes `main`'s workflow copy and only the checkout points at the tag, so the signing-secret gates (the `github.ref == main` check and the PublishDesktop environment's main-only deployment-branch policy) hold for beta too. Never add `desktop-experimental` to the PublishDesktop deployment-branch policy.
- The workflow creates the tag's GitHub release (universal DMG + updater artifact + `latest.json`; marked prerelease for beta) and refreshes the channel's rolling feed release, which is the static auto-update feed every installed app on that channel polls. Never delete the `desktop-latest` or `desktop-beta` release or tag.
- The changelog's `## <version>` section (exact-match, not "topmost") is extracted verbatim into the GitHub release body, the Slack announcement, and the updater manifest notes.
- Always ask before pushing commits or tags.
## Workflow
0. Ask which channel this release is for — **stable or beta** — if the user has not said. Everything below branches on it; never guess.
If there is no `desktop-v*` tag yet, this is the first release; use the desktop app's first commit as the baseline and say the baseline is inferred.
For a **beta** release, work on `desktop-experimental` (check out `origin/desktop-experimental`; merge `origin/main` into it first if it is behind — see EXPERIMENTAL.md for the conflict policy) and read the version files from that branch. The last-tag baseline is the newest `desktop-v*` tag of either channel that is an ancestor of the branch.
The sidecar bundles `@cline/core` and friends from the monorepo, so SDK changes ship inside the desktop app too. Fold user-visible SDK changes (providers, models, behavior fixes) into the notes; skip purely internal ones.
@@ -49,13 +60,15 @@ Flat bullet list, user-facing language. Present the draft and wait for approval
4. Decide the version bump.
Ask whether this is patch, minor, major, or an explicit version. Do not guess if the user has not made it clear.
Stable: ask whether this is patch, minor, major, or an explicit version. Do not guess if the user has not made it clear.
5. Update release files.
Beta: apply the versioning rule — base = next stable version, increment `N` (`0.0.14-beta.1` → `0.0.14-beta.2`; after stable `0.0.14` ships, next is `0.0.15-beta.1`). Confirm the computed version with the user.
5. Update release files (on `main` for stable, on `desktop-experimental` for beta).
-`apps/examples/desktop-app/package.json` → new version
-`apps/examples/desktop-app/src-tauri/tauri.conf.json` → same version
- Prepend `## X.Y.Z` (no date) to `apps/examples/desktop-app/CHANGELOG.md` with the approved notes.
- Prepend `## X.Y.Z` (no date; `## X.Y.Z-beta.N` for beta) to `apps/examples/desktop-app/CHANGELOG.md` with the approved notes.
6. Verify before committing.
@@ -77,16 +90,20 @@ Ask before pushing the release commit, then before creating and pushing the tag:
```sh
git push origin HEAD
git tag -a desktop-vX.Y.Z -m "Desktop vX.Y.Z"
git tag -a desktop-vX.Y.Z -m "Desktop vX.Y.Z"# beta: desktop-vX.Y.Z-beta.N / "Desktop vX.Y.Z-beta.N"
git push origin refs/tags/desktop-vX.Y.Z
```
8. Publish.
The release commit must be on `main` and the tag pushed first.
The release commit must be on the channel's branch (`main` for stable, `desktop-experimental` for beta) and the tag pushed first. Dispatch from `main` for **both** channels (see the release contract for why).
```sh
gh workflow run desktop-publish.yml -f git_tag=desktop-vX.Y.Z -f confirm_publish=publish
# stable:
gh workflow run desktop-publish.yml --ref main -f git_tag=desktop-vX.Y.Z -f channel=stable -f confirm_publish=publish
# beta:
gh workflow run desktop-publish.yml --ref main -f git_tag=desktop-vX.Y.Z-beta.N -f channel=beta -f confirm_publish=publish
gh run list --workflow=desktop-publish.yml --limit=1 --json url,status,conclusion,createdAt --jq '.[0]'
```
@@ -103,21 +120,24 @@ gh api repos/cline/cline/actions/runs/<run-id>/pending_deployments \
Nothing after `validate` runs — and no signing key is readable — until then.
The workflow builds one universal macOS bundle (`tauri build --target universal-apple-darwin` lipos the aarch64 + x86_64 Rust binaries; the Bun sidecar is lipo'd by `build-sidecar-bin.ts`), verifies every Mach-O in the bundle carries both slices, signs with the Developer ID certificate, notarizes with the App Store Connect API key, signs the updater artifact with the Tauri updater key, creates the GitHub release, refreshes `desktop-latest/latest.json`, and posts to Slack. Notarization typically adds 2–10 minutes.
The workflow builds one universal macOS bundle (`tauri build --target universal-apple-darwin` lipos the aarch64 + x86_64 Rust binaries; the Bun sidecar is lipo'd by `build-sidecar-bin.ts`; beta adds the `tauri.beta.conf.json` overlay), verifies every Mach-O in the bundle carries both slices and that the compiled binary embeds exactly its own channel's feed URL, signs with the Developer ID certificate, notarizes with the App Store Connect API key, signs the updater artifact with the Tauri updater key, creates the GitHub release (prerelease for beta), refreshes the channel's feed (`desktop-latest/latest.json` or `desktop-beta/latest.json`), and posts to Slack. Notarization typically adds 2–10 minutes.
If the workflow fails on missing credentials, see "Publish secrets (one-time setup)" below.
9. Verify the update feed after the run succeeds.
```sh
curl -sL https://github.com/cline/cline/releases/download/desktop-latest/latest.json | head -30
curl -sL https://github.com/cline/cline/releases/download/desktop-latest/latest.json | head -30# stable
curl -sL https://github.com/cline/cline/releases/download/desktop-beta/latest.json | head -30 # beta
```
The `version` field must be the new release and both `darwin-aarch64` and `darwin-x86_64` entries must point at the same new`desktop-vX.Y.Z` universal `.app.tar.gz` asset (each slice of the fat binary requests its own arch key at runtime, so both keys serve the one artifact). Installed apps — including older per-arch installs — pick the update up on next launch or within 2 hours.
The `version` field must be the new release and both `darwin-aarch64` and `darwin-x86_64` entries must point at the same new universal `.app.tar.gz` asset under the release tag (each slice of the fat binary requests its own arch key at runtime, so both keys serve the one artifact). Installed apps on that channel — including older per-arch installs — pick the update up on next launch or within 2 hours.
After a **beta** publish, also confirm the stable feed was not touched: `desktop-latest/latest.json` must still serve the previous stable version. (The workflow guards this fail-closed, but it is cheap to verify and catastrophic to miss — the updater comparator is a plain semver "newer than", so a beta manifest on `desktop-latest` would auto-update every stable install onto the beta.)
10. Final response.
Report: version, tag, changelog updated, commit hash, what was pushed, workflow URL, and the feed verification result.
Report: channel, version, tag, changelog updated, commit hash, what was pushed, workflow URL, and the feed verification result.
# (the branch is hardcoded to legacy-extension in the workflow; it is
# deliberately not an input)
```
npm test suite runs ungated; the publish job waits on the `Publish` environment. This workflow derives + pushes the `v<version>` tag itself and creates the GitHub release — no manual tagging. Publishes to Marketplace **and** Open VSX. The branch is the npm codebase: use `npm`, never `bun`, and expect the old monolith layout (`apps/vscode/src/core/...`).
Paste the "About" diagnostics for your Cline surface. This captures the IDE build, runtime, and host details we need.
Paste the diagnostics for your Cline surface. This captures the build, runtime, and host details we need.
- VSCode Extension: open `Help → About` (Windows/Linux) or `Code → About Visual Studio Code` (macOS), then copy the info.
- JetBrains Plugin: open `Help → About` (Windows/Linux) or `<IDE name> → About` (macOS), then click `Copy` to grab build, runtime, OS, memory, and cores.
- CLI: there is no About dialog. Run `cline --version` and paste the output.
placeholder:Paste the copied About info or `cline --version` output here.
- Desktop App: paste the app version from the Settings view.
- Cloud Platform: paste your browser name and version, plus the page URL where the issue occurred.
placeholder:Paste the copied About info, `cline --version` output, or browser/app details here.
if ! gh release view "$FEED" >/dev/null 2>&1; then
if [ "$CHANNEL" = "beta" ]; then
gh release create "$FEED" \
--title "Cline desktop beta (auto-update feed)" \
--notes "Rolling release backing the beta desktop app auto-updater. The latest.json asset points at the newest desktop-vX.Y.Z-beta.N release. Only beta installs poll this feed; stable installs use desktop-latest. Do not delete." \
--latest=false \
--prerelease \
--target "$(git rev-parse HEAD)"
else
gh release create "$FEED" \
--title "Cline desktop (auto-update feed)" \
--notes "Rolling release backing the desktop app auto-updater. The latest.json asset points at the newest desktop-vX.Y.Z release. Do not delete." \
text: "<https://github.com/${{ github.repository }}/releases/tag/${{ needs.validate.outputs.tag }}|Download DMG> — installed apps auto-update on next launch${{ steps.prev_tag.outputs.prev_tag != '' && format(' | Full Changelog: https://github.com/{0}/compare/{1}...{2}', github.repository, steps.prev_tag.outputs.prev_tag, needs.validate.outputs.tag) || '' }}"
text: "<https://github.com/${{ github.repository }}/releases/tag/${{ needs.validate.outputs.tag }}|Download DMG> — ${{ needs.validate.outputs.channel == 'beta' && 'beta channel: installs side by side with the stable app and only beta installs auto-update; stable users are unaffected' || 'installed apps auto-update on next launch' }}${{ steps.prev_tag.outputs.prev_tag != '' && format(' | Full Changelog: https://github.com/{0}/compare/{1}...{2}', github.repository, steps.prev_tag.outputs.prev_tag, needs.validate.outputs.tag) || '' }}"
Everything here lands through the SDK bundle, so it applies to windows running that bundle.
### Fixed
- Auto-approve every MCP tool call while the "Use MCP servers" toggle is on. The toggle only took effect on tools that had also been opted in individually, so turning it on appeared to do nothing; it now governs all MCP tools on its own.
## [4.1.14]
Everything here lands through the SDK bundle, so it applies to windows running that bundle.
### Added
- Refresh the built-in model catalog. New entries include Claude Fable 5, Grok 4.6 on Vertex, several DeepSeek V4 Flash variants (including the vision preview), MiMo v2.5, Qwen3.8 27B, Gemma 4 26B, LongCat 2.0, Nemotron 3.5 Lightning, and Thinking Machines' Inkling models.
### Fixed
- Restore task completion telemetry for interactive sessions. A share of interactive stops routed through a teardown path that never reported completion after 4.1.11 changed how session status is tracked; every session now reports it exactly once.
## [4.1.13]
Everything here lands through the SDK bundle, so it applies to windows running that bundle.
### Fixed
- Restore tool calling for custom OpenAI-Compatible models whose capability list was inferred from convenience flags like `supportsReasoning`. The inferred list read as an authoritative denial and stripped every tool from the request; an explicitly authored capability list still decides.
- Keep Hub-backed sessions intact across a Hub restart or upgrade. Clients replay the events they missed while disconnected, and the same event is no longer delivered twice when the replay and live streams overlap.
- Carry session and client identity into Langfuse traces for Hub-backed and delegated-agent runs, which previously arrived without their session grouping or client version.
## [4.1.12]
Everything here lands through the SDK bundle, so it applies to windows running that bundle.
### Fixed
- Enforce enterprise MCP controls on the Customize marketplace. MCP entries are now hidden when remote config disables the marketplace, and limited to `allowedMCPServers` when an allowlist is configured.
- Restore tool calling for custom OpenAI-Compatible models whose stored capability list was empty.
## [4.1.11]
Everything here lands through the SDK bundle, so it applies to windows running that bundle — except the last section, which is a legacy-bundle fix.
### Added
- Let models that support it generate images during a task. Generated images render inline in the conversation.
### Fixed
- Fix code actions failing with "command not found" on VS Code 1.134.
- Fix `@` file mentions breaking on paths that contain spaces.
- Show the diff edit view for multi-line edits in files with CRLF line endings.
- Continue the surviving session when resuming a task, instead of rebuilding it from the original task text.
- Clear the task-scoped settings overlay when the task view is cleared or switched, so one task's overrides no longer leak into the next.
- Honor the classic truncation range when migrating legacy tasks.
- Preserve LiteLLM input token limits instead of overwriting them with catalog values.
- Restore custom base URLs for Gemini, and normalize legacy host-root values so they keep working.
- Point provider signup links at each provider's API key page instead of a generic landing page.
- Load skill slash commands through the skills tool instead of pasting their instructions into your message, which previously delivered them twice.
- Stop offering image, voice, and other non-chat models in chat model pickers.
- Deliver a `PreToolUse` hook's `contextModification` to the model again, and wait for `PostToolUse` hooks so their output and `cancel` control are honored.
- Show tool activity a provider runs itself — every tool the Claude Code provider executes inside its own session — instead of dropping it from the conversation.
- Fix `run_commands` failing with ENOENT when a structured command carried a full command line with no arguments.
- Run PowerShell commands fail-fast, so a pipeline erroring per item stops at the first error instead of flooding output and still reporting success.
- Keep remote configuration in step with the SDK: coordinated refreshes, session gating, and a fail-closed opt-out.
### Changed
- Show the billed cost for Cline gateway usage.
- Refresh the model catalog, which adds AMD, Arcee, Echo, Jalapeno, Kosmik, LLM Gateway, RunInfra, and SCNet as providers and updates model lists, pricing, and per-provider default models across the board.
### Fixed (legacy bundle)
- Only treat an Anthropic `invalid_request_error` as a context-overflow when its message says so. An unrelated invalid request (bad tool schema, oversized image, unknown model id) no longer triggers context-overflow recovery.
## [4.1.10]
Everything in this release lands through the SDK bundle, so it applies to windows running that bundle and not the legacy one. The legacy bundle is unchanged from 4.1.9.
### Added
- Let models that support it search the web during a task, with a toggle in Feature Settings to turn it on. Search calls and their results appear in the conversation and persist across reloads.
### Fixed
- Stop two Cline installations on different builds from shutting each other's Hub daemon down in a loop, which killed live sessions with an abnormal socket close. Build identity is now compared through a total order, so at most one side of a pair can decide to retire the other.
- Leave a Hub that is still serving sessions in place instead of replacing it mid-handshake; the swap happens once it goes idle.
- Reclaim idle plugin sandbox processes instead of leaving them running for the life of the session.
### Changed
- Refresh the model catalog, which adds Crusoe as a provider and updates model lists and per-provider default models across the board.
## [4.1.9]
### Changed
- Use the editor's foreground color for diff block text, so diffs stay legible in themes where the previous hardcoded color washed them out.
- Switch the interface to Inter and Geist Mono.
### Fixed
- Don't discard a successfully refreshed Cline token when the old one was already past expiry, which made the first request after a long idle period fail despite valid credentials.
- Stop the legacy-task migration backlog from spamming telemetry, and record a migration outcome only once the seeded session actually persists, so a failed migration is no longer reported as a success.
- Report involuntary Cline logouts (a rejected refresh token) instead of clearing credentials silently.
### Fixed (SDK bundle only)
These land through SDK v0.0.74 and therefore apply to windows running the SDK bundle, not the legacy one.
- Fix the Claude Code provider being unusable for agentic work: it now runs its own native tools instead of receiving tool definitions it cannot bridge, anchors the session on your workspace directory, and loads `~/.claude` plus project settings so your permission rules apply.
- Reject truncated tool-call JSON instead of silently "repairing" it into wrong arguments.
- Fix strict providers rejecting a turn with "user message must have content" when a message's content held only empty text parts.
- Fix a mid-turn crash on streamed tool calls with non-zero or non-contiguous indexes, hit through LiteLLM's Anthropic passthrough.
- Report disjoint per-request token buckets instead of re-counting the whole cached conversation on every request, which inflated per-task totals roughly 5x on cache-heavy sessions.
- The first-launch "Try ClinePass" dialog no longer advertises the $4.99 first-month promo, which is ending
- The hub's event log is now capped at 64 MiB on disk. Events carrying full session snapshots could previously grow the log to tens of gigabytes on a long-running hub, since deleting rows never shrinks the file. Oldest events are dropped first and the space is returned, and pruning runs on volume as well as on a timer
- Refreshed the model catalog. Adds two providers (AgentRouter and Opper) and updates model lists and pricing across providers. The resolved default model changes for Aki.io and NanoGPT, so if you use one of those without pinning a model you will get a different default
## 3.0.57
- Added `cline hub drain`, which stops a hub from accepting new mutating work while it finishes what it is already running, and `cline hub drain --off` to lift it
- Added `cline hub upgrade`, which drains the hub, waits for it to go idle, stops it, and starts a fresh one on the current build. An aborted upgrade lifts the drain again, so the hub is never left refusing work
- Sessions now survive a hub restart. A reconnecting client replays the events it missed while disconnected, deduped by event id so nothing is delivered twice
- Fixed tool calling being silently disabled for custom OpenAI-Compatible models whose capability list was inferred from convenience flags like `supportsReasoning`. The inferred list read as an authoritative denial and stripped every tool from the request
- Langfuse traces now carry session and client identity for hub-backed and delegated-agent runs, instead of arriving without their session grouping or client version
- Refreshed the model catalog, which updates model lists and pricing across providers and changes the resolved default model for several of them (DeepSeek, Crof, CrossModel, Eden AI, Kilo, and NanoGPT)
## 3.0.56
- Models that support image generation can now produce media during a turn. The TUI saves each generated file to a temporary path and prints it so you can open it with your usual tools, HTML session exports embed images inline, and ACP clients receive generated images as image content
- Skill slash commands now load through the skills tool instead of expanding into your message. History and resume show the `/command` you typed instead of the whole skill body, and the instructions reach the model once instead of twice. Workflows still expand, as does zen mode, whose preset has no skills tool
- Image, voice, and other non-chat models are no longer offered in the onboarding and model pickers or ACP model listings, and are rejected for `--model`
- Fixed TUI dialog colors not following theme changes live
- Fixed the account dialog's selection chevron so it matches the other dialogs
- Fixed provider-executed tool activity — every tool the Claude Code provider runs inside its own session — being dropped instead of shown as a tool card
- Fixed `PreToolUse` hook `contextModification` never reaching the model, and `PostToolUse` hooks running fire-and-forget with their output and `cancel` control discarded
- Fixed `run_commands` failing with ENOENT when a structured command carried a full command line with no `args`
- PowerShell commands now fail fast on the first error instead of emitting an error record per enumerated item and still reporting success
- Fixed Gemini custom base URLs configured as a host root
- Fixed `cline schedule` commands against a remote hub, which now register a workspace client so they are authorized under the new workspace-scoped schedule rules
- Usage now displays the billed gateway cost
- Refreshed the model catalog, which adds AMD, Arcee, Echo, Jalapeno, Kosmik, LLM Gateway, RunInfra, and SCNet as providers and updates model lists, pricing, and per-provider default models across the board
## 3.0.55
- Auto-updates no longer install while a CLI is attached to the Hub. The update is recorded at startup and installed on exit, once the Hub confirms nothing else is attached, so a background update can no longer swap the package out from under a live session and kill it with `Hub connection closed (code=1006)`. `cline update` still installs immediately and now tells you the update applies on next start
- Added protections for an update landing under CLI 3.0.54 and earlier, whose updater restarts the Hub mid-session and then rejects every replacement, bricking a running session. The newly installed package defuses that path during install instead of leaving it to fire
- Fixed two Cline installations on different builds shutting each other's Hub daemon down in a loop, which killed every live session with an abnormal socket close. Build identity is now compared through a total order, so at most one side of a pair can ever decide to retire the other (from SDK v0.0.75)
- A newer build no longer replaces a Hub that is still serving sessions — it attaches to it and the swap happens on a later launch, instead of the sessions dying mid-handshake (from SDK v0.0.75)
- Removed the "outdated Hub" notice. It reported a state you cannot act on, and the toast was capped narrower than the message, so it rendered cut off before the reassuring half of the sentence at every terminal width. The prompt for a genuine build mismatch, where there is something to do, is unchanged
- Streaming assistant markdown no longer flashes back to raw text. Settled headings, links, and code stay rendered as new chunks arrive instead of the whole message being rebuilt and re-highlighted on every chunk, which also stops the transcript from jumping vertically mid-stream
- Web search calls and their results from models that run search natively now render in the transcript (from SDK v0.0.75)
- Idle plugin sandbox processes are now reclaimed instead of lingering for the life of the session (from SDK v0.0.75)
-`cline doctor fix` now reports honestly: processes that survived a kill are separated from ones that appeared while the fix ran, a live parent respawning a daemon is named, and a startup lock held by a running process is reported as held rather than leaked (from SDK v0.0.75)
- Refreshed the model catalog, which adds Crusoe as a provider and updates model lists and per-provider default models across the board (from SDK v0.0.75)
## 3.0.54
- Fixed the Claude Code provider being unusable for agentic work: the provider now runs its own native tools instead of receiving tool definitions it cannot bridge, the session is anchored on your workspace directory instead of inheriting the host's cwd, and `~/.claude` plus project settings are loaded so your permission rules apply. File edits under the workspace are auto-approved; command execution stays gated by your own Claude settings (from SDK v0.0.74)
- Fixed truncated tool-call JSON being silently "repaired" into wrong arguments — a payload with an unterminated string is now rejected rather than getting an invented terminator (from SDK v0.0.74)
- Fixed strict providers rejecting a turn with "user message must have content" when a message's content held only empty text parts (from SDK v0.0.74)
- Fixed a mid-turn crash on streamed tool calls with non-zero or non-contiguous indexes, hit through LiteLLM's Anthropic passthrough (from SDK v0.0.74)
- Managed Hub daemons now upgrade directionally: when another Cline install ships a newer Hub build, the CLI attaches to the newer daemon and prompts you to update and restart instead of the two installs repeatedly retiring each other's daemons. Yolo and sandbox sessions, which never attach to the shared Hub, are not interrupted by that prompt (from SDK v0.0.74)
- Fixed the Hub daemon logging an unhandled `hub server close failed` error and exiting non-zero whenever a client was still connected at shutdown (from SDK v0.0.74)
- Fixed per-task token totals being inflated roughly 5x on cache-heavy sessions — token telemetry now reports disjoint uncached-input, cache-read, and cache-write buckets instead of re-counting the whole cached conversation on every request (from SDK v0.0.74)
- Upgrading the CLI now retires an already-running Hub daemon and respawns it on the new code, instead of the upgraded CLI continuing to talk to a daemon executing the previous release
## 3.0.53
- Fixed the CLI reconnecting to a stale Hub daemon after an upgrade. Hub daemons now carry a runtime build fingerprint, so an upgraded CLI retires and respawns a daemon still running older code instead of attaching to it (from SDK v0.0.73)
"\nSome of these were respawned by a live parent (100); stop the parent process listed above, then re-run. The rest started after the fix began and were not targeted.",
return"\nThese processes started after the fix began, so they were not targeted. Re-run to see whether they persist.";
}
if(respawned.length===pids.length){
return"\nThese processes were respawned by a live parent. Stop the parent process listed above, then re-run.";
}
return`\nSome of these were respawned by a live parent (${respawned.join(", ")}); stop the parent process listed above, then re-run. The rest started after the fix began and were not targeted.`;
// biome-ignore lint/correctness/useExhaustiveDependencies: dialogCount re-runs the sync when a dialog opens, covering dialogs opened before the container-level option applied (see docblock).
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.