getAuthToken captured expiresAt before refreshing, then validated the new
token against that stale value. When the old token was already past expiry
(not just inside the 5-minute buffer), a successful refresh was thrown
away and null returned, so the first call after long idle failed despite
valid credentials. Re-read the expiry from the refreshed auth info.
* fix(hub): don't forward recoverable agent errors to dashboard peers
Recoverable error events are in-run notices, not turn outcomes: the
MistakeTracker emits one for every recorded mistake (e.g. a plan-mode
guard-blocked run_commands call) while the run continues. The hub
dashboard forwarded every error event to peers, so the webview dropped
out of the sending state and appended an error row mid-turn — the same
host bug fixed for VS Code and the CLI in #12953.
Gate the forward on recoverable, matching those hosts: the tool failure
is already shown inline via the failed tool_event, and the turn's
outcome stays decided by how it actually ends (turn_done or a
non-recoverable error). Recoverable errors are logged server-side.
* fix(hub): forward recoverable flag to peers instead of filtering server-side
Per review: the server is a translation layer between agent events and
the webview protocol, so it should not embed display policy or console
logging. Forward every agent error with its recoverable flag on the
peer message and let each peer decide — the webview keeps recoverable
errors out of the transcript and keeps the turn state, matching how the
CLI gates display on the same flag while the information stays
available to any peer that wants it.
* add custom model selection to the vertex provider
* fix race conditions from PR review
* fix linter warnings
* fix test failures
* refactor(vscode): drop Vertex global-endpoint picker filtering
The SDK catalog is live (models.dev), so a static host allowlist of
global-endpoint-capable models lags every model launch and silently hides
new models from users on vertexRegion=global. Remove the allowlist, the
host override that injected supportsGlobalEndpoint, and the picker filter;
show the full catalog for every region.
An unsupported pick now fails loudly at request time: map Vertex's
'model not available in region: global' (and Google's Publisher Model
locations/global not-found body) to recovery guidance in the error row.
Also drop Anthropic's universal pricing from the Vertex Fable 5 overlay —
Vertex bills region-dependently, so the copied price understated recorded
cost; the record now carries no pricing instead of a wrong one.
---------
Co-authored-by: Saoud Rizwan <7799382+saoudrizwan@users.noreply.github.com>
* fix: remove stale Double-Check Completion feature tip
The rotating feature tips still told users to enable "Double-Check
Completion" in settings, but that toggle was removed in the new UI —
the Features section now offers Auto Compact, Feature Tips, Background
Edit, Checkpoints, Worktrees and Hooks. Following the tip sent users
searching the settings panel for something that isn't there.
Drop the tip. The remaining ten were checked against the current UI and
all still hold, including the "Settings → Features → Feature Tips" path.
* chore: remove dead CLI settings e2e page object and orphaned test
`page-objects/settings.ts` asserted the CLI settings Features tab shows
"Double-check completion" — the same removed setting behind the stale
feature tip. Nothing in the live tui-test suite (apps/cli/src/tests)
imported it; only chat.ts and auth.ts page objects are in use.
Its one importer, apps/vscode/tests/e2e/cli/interactive.test.ts, is a
leftover from the pre-2026-06-02 SDK migration squash: all three of its
imports resolve to files that don't exist, there's no tui-test config in
that tree, and no npm script runs it. It cannot execute.
* fix: respect user max output tokens in compaction summarizer requests
The compaction summarizer hardcoded max_tokens to 1024 and the VSCode host
never mirrored the user's Max Output Tokens onto providerConfig, so summary
requests were always capped at 1024 tokens. Reasoning models can spend that
entire budget thinking; the reasoning stream is discarded, so no summary
text arrives and compaction is skipped on every attempt.
- Mirror maxTokensPerTurn onto providerConfig.maxOutputTokens in the VSCode
session factory so consumers that build handlers straight from it (the
compaction summarizer) honor the user's setting, matching the CLI.
- Resolve the summarizer output budget from explicit config, then model
info, then knownModels, before the default; raise the default to 4096.
- Log a diagnostic warning (reasoning chars, incompleteReason, likely
cause) when the summarizer returns no summary text instead of silently
skipping.
* fix: clamp summarizer default output budget by model metadata instead of adopting it
Model maxTokens is reported capability, not a product default: without an
explicit configuration the summarizer now requests the 4096 default, lowered
by model metadata when the model reports less, never raised by it. Explicit
values still win as-is.
* feat(vscode): remove YOLO mode setting, migrate old users to auto-approve all
The SDK extension's YOLO toggle was cosmetic: nothing in the approval
path read it, so runs were silently governed by the per-action
auto-approval settings underneath (cline/cline#13114). Instead of
keeping a parallel override system, remove the setting entirely and
make the auto-approve menu the single source of truth:
- drop yoloModeToggled (and the equally dead autoApproveAllToggled)
from state keys, settings handlers, state posts, telemetry, the
remote-config yoloModeAllowed transform, and the settings protos
(field numbers reserved)
- remove the Yolo Mode toggle from Settings -> Features (the whole
Experimental section, it was the only entry) and the
"Auto-approve: YOLO" AutoApproveBar takeover
- add a v3 storage migration that folds a previously-enabled YOLO /
auto-approve-all toggle into autoApprovalSettings by enabling every
action, so previously-unattended setups keep running unattended;
the dead keys are cleared from the file store
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* refactor(vscode): keep dead yolo keys in place instead of clearing them
Current builds never read the removed keys (the state loader only visits
known keys), so deleting them buys nothing - and the file store is shared
with older builds that still know them, so clearing would flip YOLO off
for a user who downgrades. Same downgrade-safety rule as the v1 export.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* chore(vscode): rename wasUnattended to shouldEnableAllActions in yolo migration
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* chore(vscode): drop dead toggleActModeForYoloMode and stale yoloModeAllowed comment
The method was a legacy-controller carryover nothing called, and it set
the mode without rebuilding the session, which is wrong for the SDK
architecture. The comment cited yoloModeAllowed as a live remote-config
example; it no longer maps to anything.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* chore(vscode): refresh checked-in proto descriptor_set.pb
The tracked descriptor set had not been regenerated since the repo
move and still advertised long-changed schemas (including the removed
yolo_mode_toggled fields) to gRPC reflection clients. Sync it with the
output of bun run protos.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* desktop: paste clipboard images into the composer as attachments (CLIENTS-78)
Pasting a screenshot into the composer did nothing: only drag-and-drop
and the paperclip file picker fed the attachment pipeline. Add an
onPaste handler on the composer textarea that extracts image files from
the clipboard, renames them to timestamped pasted-image-*.png files, and
routes them through the existing onAttachFiles flow. Text pastes are
untouched.
* desktop: only extract clipboard images in formats message serialization supports
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* desktop: context-aware welcome suggestions for non-code folders (CLIENTS-98)
* desktop: treat pending branch discovery as its own state for welcome cards
The welcome-card classifier read the "no-git" sentinel as a confirmed
non-repo, but page.tsx also used that value for the initial state and
while a workspace switch was awaiting branch discovery, so a git repo
could briefly show the plain-folder cards. Branch state is now null
while discovery is pending: the welcome screen shows no cards until the
folder is classified, and chat-mode cards (which never depend on git
state) still show immediately. Other branch consumers keep the string
contract via a "no-git" fallback.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* desktop: carry nullable branch state to all consumers
Propagate the pending-discovery null through ChatInputBar,
WorkspaceSelector, and the welcome workspace controls instead of
coercing to "no-git" at the page boundary, so only display leaves
fall back and the welcome classifier is the single consumer that
distinguishes pending from confirmed non-repo.
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* desktop: never let 'Add project…' fail silently; add manual folder path entry (CLIENTS-73)
- sidecar picker tries zenity then kdialog on Linux and throws a descriptive
error when neither exists, instead of returning null (indistinguishable
from user cancel); picked paths are trimmed of trailing separators
- picker failures now surface as visible error messages in both workspace
selectors, with a manual path-entry fallback (typed absolute or ~ paths
in the search box offer an 'Open folder' action)
- failed workspace switches (invalid/nonexistent paths) show an inline
error instead of silently doing nothing
- validate_workspace_directory expands ~ and returns the resolved path
* desktop: keep workspace menu search/error state through catalog refreshes
The welcome-screen workspace picker reset its search text and error
message whenever onRefreshWorkspaces changed identity, which happens on
every session-history poll. Typing a path or reading an inline error
raced against the timer: the menu would silently wipe mid-interaction.
Hold the refresh callback in a ref so the reset only runs when the menu
actually opens.
* desktop: format welcome-workspace-controls test
* desktop: distinguish picker launch failures from user cancellation
A zenity/kdialog rejection with a non-ENOENT spawn error (EACCES, EMFILE,
ENOMEM) or a crash signal was classified as a user cancel, which skipped
the kdialog fallback and suppressed the inline error - recreating the
silent no-op this branch is meant to eliminate. Only a clean exit code 1
from a dialog that actually opened now counts as cancellation; broken
backends fall through to the next candidate and surface a descriptive
error otherwise. Picker logic moved to sidecar/workspace-picker.ts with
an injectable exec so the classification is unit-tested.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Classify picker launch failures separately from user cancellation
zenity/kdialog failures like EACCES, EMFILE, or ENOMEM were treated as
user cancellation, suppressing the kdialog fallback and the inline
manual-entry error. Only a clean exit code 1 now counts as a cancel;
any other failure falls through to the next backend or throws the
picker-unavailable error. Picker logic moved to sidecar/folder-picker.ts
with an injectable exec so the classification is unit tested.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Revert "Classify picker launch failures separately from user cancellation"
This reverts commit 24d27a004d.
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(core): preserve queued prompts across user-initiated aborts
Pressing stop while prompts were queued silently destroyed them:
abort() called clearAborted(), which emptied the pending prompt queue
with no way to recover the typed input. The prompts vanished from the
UI, were never sent, and left no trace in session artifacts.
Aborting now only stops the in-flight turn. Queued prompts stay in the
queue and drain once the abort settles, matching the drain behavior
that already existed for self-aborted turns (loop detector / mistake
limit). The thrown-abort path (completeAbortedInteractiveTurn) now
schedules the same drain that runTurn schedules for turns resolving
with an aborted finish.
* fix(core): full-stop semantics and abort-window queue edits for surviving queues
Follow-up to the queued-prompt survival change: aborting a user turn keeps
the queue and auto-runs it, but two gaps remained.
1. No full stop: aborting a queue-initiated turn also kept draining, so
every Escape consumed one queued prompt and started a fresh provider
call - a session with queued messages could never be brought to rest.
Aborting a drained turn now discards the remaining queue: the first
Escape skips to your queued follow-ups, a second Escape stops the
queued work too.
2. Queue operations were still rejected while an abort settled: a prompt
typed right after Escape was silently dropped, and queued prompts were
briefly uneditable and undeletable even though they were about to
auto-run. enqueue/update/delete now work during the abort window;
scheduleDrain/drain still wait for the abort to settle.
* test(core): cover abort + host restart + seeded recovery durability
Adds an e2e regression guard for the reported "cancel a turn, lose the
conversation" failure: a cancelled turn, a daemon restart, a
client-side recovery seeded from disk, and a second restart before that
replacement ever runs a turn. Reverting the eager seeded-history
persistence makes the final read come back empty.
Materializing a seeded session at start also left its history row with
no prompt and no title, since there is no first prompt to derive one
from. Seed the title from the inherited transcript using the same
inference listSessionHistory hydration applies, so forks and recoveries
stay identifiable in unhydrated surfaces too.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(core): retitle seeded sessions from their first user prompt
Eagerly-materialized seeded sessions kept the interim transcript-
inferred title forever, a behavior change from pre-eager persistence
where a fork's history row was titled by the first post-fork prompt.
The interim title now only covers the window where no turn has run
(previously those rows were simply absent), and the first user prompt
after the seed backfills the row's prompt and retitles it — unless the
user renamed the session in the meantime, in which case only the prompt
column is backfilled. The resident manifest and session metadata are
updated in step so the end-of-turn usage-metadata merge cannot clobber
the title back through a stale in-memory fallback.
The e2e mock's updateSession now mirrors the real persistence-service
contract (row + manifest file), and the durability e2e covers both the
retitle and the rename guard; removing the retitle call fails the
'now add tests' assertion.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* refactor(core): collapse seeded-session titling to the old mechanism
The interim transcript-derived title, retitle flags, rename comparison,
and resident-manifest syncing existed only to title forks that never
run a turn - a new nicety, not parity. Dropping it collapses the whole
design back to what rows did before eager persistence: the persistence
service derives the title from the prompt when a row gains one, so the
host only needs to backfill the promptless row with the first user
prompt via updateSession. Renames win automatically because the service
preserves an existing title when no explicit title is passed.
Net production change vs main is a single 20-line backfill block in
executeTurn. The e2e mock's updateSession now models the service's
title semantics (explicit title wins, existing title preserved,
untitled rows derive from prompt), and the durability e2e asserts the
raw row stays untitled until first prompt while history hydration
infers a display title from the transcript.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
Pressing stop while prompts were queued silently destroyed them:
abort() called clearAborted(), which emptied the pending prompt queue
with no way to recover the typed input. The prompts vanished from the
UI, were never sent, and left no trace in session artifacts.
Aborting now only stops the in-flight turn. Queued prompts stay in the
queue and drain once the abort settles, matching the drain behavior
that already existed for self-aborted turns (loop detector / mistake
limit). The thrown-abort path (completeAbortedInteractiveTurn) now
schedules the same drain that runTurn schedules for turns resolving
with an aborted finish.
* fix(core): keep a hung MCP server from taking down session creation
A stdio MCP server that never finishes initializing used to hold its
connect open for the full DEFAULT_MCP_CONNECT_TIMEOUT_MS (doubled across
the newline/framed attempts). MCP tool discovery runs on the
session.create critical path, so that wait blew past the 30s hub command
timeout and the CLI tore the whole interactive session down instead of
just skipping the bad server.
- Bound MCP tool loading during session build with a startup budget that
is safely under the hub command timeout. Servers that connect in time
contribute their tools; slower/hung servers are skipped for the session
(their error still surfaces via the MCP manager) instead of failing
session creation. Budget is overridable via CLINE_MCP_STARTUP_BUDGET_MS
for tests.
- Add StdioMcpClient.close() (and optional McpServerClient.close) that
marks the client disposed so an in-flight connect() aborts its retry
loop instead of respawning the framed fallback.
- Dispose the manager by closing clients up front, outside the per-server
operation locks, so a server hung in initialize can no longer stall
teardown for the full connect budget.
Adds regression tests covering both the non-blocking build and prompt
disposal while a client is hung in connect().
* refactor(core): simplify hung-MCP-server fix to a startup budget
Replace the bespoke per-server race/tracking in loadConfiguredMcpTools
with a small withStartupBudget() wrapper around the existing
Promise.allSettled: a server that exceeds the budget becomes a normal
rejection that the existing loop already logs and skips. The connect
budget, MCP settings display (initialize timeout 30s), and the rest of
the loader are left untouched.
The client close()/manager.dispose() cleanup is kept minimal: it is what
lets teardown abort a still-in-flight connect instead of blocking on the
per-server lock (and clears the pending request timer).
* fix(mcp): cap the default initialize budget at 3s to protect session creation
Supersedes the startup-budget approach on this branch with the simple
constant fix.
MCP initialize runs on the session.create critical path, which the hub
caps at 30s, and connect() can spend the budget twice (newline then
Content-Length framing). The 30s default from #13067 meant a server that
never initializes held session.create for up to 60s, so the hub RPC
timed out and the CLI tore the whole session down and exited.
Return to the pre-#13067 shape with a bigger probe: 3s instead of 1.5s.
That still covers the ~2s starters the old probe killed (#13035) and
keeps the worst case at ~6s per server, far under the hub deadline.
Genuinely slow starters (JVM-based servers like Oracle SQLcl) now need
an explicit timeout in cline_mcp_settings.json, which continues to
override the default in either direction.
Tests: update the slow-start regression tests to the new policy (2s
connects by default, 4s connects with a configured timeout), refresh the
displayed initialize-timeout assertions, and add an invariant test that
keeps the doubled default well under HUB_DEFAULT_COMMAND_TIMEOUT_MS so
the budget cannot silently creep past the session deadline again.
* fix(desktop): stop opening a session from replacing the remembered model
The composer's ModelSelector mirrored every provider/model prop change
into the remembered last selection (localStorage), which seeds new
sessions via getInitialChatConfig(). Opening an existing session drives
those props to that session's config, so merely viewing an old session
silently replaced the user's explicitly picked default model.
The remembered selection is now written only from the explicit picker
handlers (provider select and model select). Passive prop changes, such
as opening a session, no longer touch it.
* fix(desktop): re-seed remembered provider/model on chat reset
reset() kept the previous config's provider/model and only cleared the
session ID, so a chat pane that had hydrated a historical session could
carry that session's model into the next chat. Re-seed provider/model
(and apiKey when the provider changes) from the remembered defaults --
the same source a freshly mounted thread uses -- so reset and remount
behave identically.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): stop treating leftover plugin install dirs as installed
isOfficialPluginInstalled() only checked that the marketplace install
directory existed. A failed or interrupted install can leave that
directory behind with no plugin inside, and the next install attempt
then short-circuited with a fake 'already installed' success: the
marketplace button flipped to Uninstall with no error while nothing
actually worked, and the installed-entries listing kept reporting the
broken entry as installed.
The check now requires a loadable plugin module inside the directory
(via discoverPluginModulePaths) before reporting the entry as
installed, so partial directories fall through to a real install
attempt whose outcome is surfaced to the UI.
* fix(desktop): reclaim leftover partial plugin install dirs with --force
* fix(desktop): canonicalize diff panel paths against the session cwd
Tool calls address the same file inconsistently across a session: one
edit uses a workspace-relative path (journal.txt), a later one the
absolute path (/tmp/ws/journal.txt). mergeToolDiffs keyed entries by the
raw string, so the same file was listed twice in the diff panel with
split +/- counts and inconsistent naming, most visibly after git was
initialized mid-session and the model switched to absolute paths.
Diff paths are now canonicalized against the session cwd before
merging: entries for the same file collapse into one, files inside the
cwd display as workspace-relative paths, and files outside it display
their resolved path. Without a cwd the previous raw-key behavior is
kept.
* style: collapse editorReplaceEvent signature per biome format
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): collapse dot segments and keep root cwd in diff path keys
* fix(desktop): compare Windows diff path keys case-insensitively
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(vscode): hide View Changes on completion rows until there are changes to show
The button previously always rendered on the latest completion row, faded
and disabled when the count check came back 0 - which covers both 'nothing
changed since your last message' and 'no checkpoint to compare against'
(non-git workspace, repo with no commits, comparison failure). A dead
button with a misleading tooltip in the non-git case is worse than no
button: now the row renders nothing until the host confirms there are
actual changes, and the button is always enabled when shown.
* fix(vscode): reset View Changes state when showViewChanges toggles
Greptile review: a stale positive hasChanges from a previous evaluation
could flash the button before the host confirms the new comparison when
showViewChanges flips false and back true on the same row. Reset to
'still checking' whenever the effect re-runs.
* fix(core): never run a foreign compiled plugin-sandbox bootstrap for a source host
When @cline/core runs from source (e.g. the desktop hub daemon in dev)
with CLINE_WRAPPER_PATH set, resolveBootstrap() picked the compiled
plugin-sandbox-bootstrap.js from a separately installed CLI platform
package (such as a published version sitting in the package-manager
cache) before falling back to the source bootstrap. That bootstrap
resolves modules against the other installation's layout, so every
plugin failed to load with "Cannot find module '@cline/core'" - and
the settings pipeline swallowed the failure, leaving Settings > Tools
showing "No plugin tools found" and plugins showing no contributions
even though the same plugins loaded fine in chat sessions.
Bootstrap selection now prefers, in order: a compiled bootstrap next to
this module (always matches the host build), the source bootstrap when
the host runs from source, and only then wrapper/executable-derived
bootstraps - which remain the path for compiled binaries where
import.meta points inside the bunfs bundle.
* chore(core): restore untouched settings-service formatting
* fix(core): keep session context durable across aborts and hub restarts
Users on slow self-hosted endpoints reported sessions losing their entire
conversation after cancelling a long-running request: the TUI still showed
the transcript, but the next turn greeted them like a brand-new session.
Root cause is a stack of two failures:
1. The hub daemon exits on any unhandled rejection that is not an
AgentRuntimeAbortError, so a floating abort-family rejection from a
cancelled provider stream kills every resident session.
2. When the CLI recovers the missing session it rebuilds from the persisted
messages file - but aborting a turn never flushed the transcript, and
lazy session persistence (SDK 0.0.70) kept seeded history (mode-switch
restarts, forks, previous recoveries) memory-only until the first
completed turn. Recovery then seeds an empty session: silent context wipe.
Fixes:
- completeAbortedInteractiveTurn now flushes the transcript to disk, so an
aborted exchange survives a hub restart.
- Sessions started with initialMessages persist them (and any compaction
sidecar) immediately; brand-new empty sessions stay lazy, so closing an
unused runtime still leaves no empty history entry.
- The hub daemon ignores abort-family unhandled rejections (DOMException
AbortError, Node ABORT_ERR) the same way it already ignores
AgentRuntimeAbortError, instead of exiting with every session resident.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(core): write seeded history atomically with session materialization
Greptile review flagged a residual crash window in the seeded-session
persistence: ensureSessionPersisted created the session row (with an empty
messages file) and only then called persistSessionMessages, so a crash
between the two left a discoverable session whose seeded history was gone.
Close the window by threading initialMessages/systemPrompt through
createRootSessionWithArtifacts: the messages artifact is now written with
the seeded transcript before the session row is committed, so every crash
point leaves either nothing discoverable or complete data. The follow-up
persistSessionMessages call at session start is gone; the seed travels
inside session materialization.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* Revert "fix(core): write seeded history atomically with session materialization"
This reverts commit 5a7e0b37f1.
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
Turns drained from the pending-prompt queue resolve their errored
AgentResult inside PendingPromptsController.drain(), which discards it,
and the legacy 'error' agent event had no projection in the hub's
session-event projector — so a failed queued turn never produced any
terminal hub event. Interactive clients (e.g. the desktop app) hung on
'Thinking...' with no error shown.
The projector now publishes run.failed (with the error text and a core
session snapshot) for non-recoverable lead-agent error events, but only
when no RPC-driven turn is awaiting sessionHost.runTurn for that
session — the awaiting run.start handler already publishes the
authoritative terminal event, so this avoids double-reporting a turn
that resolves through both paths.
* fix(core/cli): drain queued prompts after self-aborted turns and surface the stop
When a run ends with finishReason "aborted" without a user abort request
(loop detector hard escalation or the consecutive-mistake safety stop),
runTurn skipped the pending-prompt drain, stranding user-queued messages
forever, and the CLI rendered nothing - the task appeared to silently
stop with queued messages never consumed (#13030).
- core: schedule the drain after every completed turn, including
aborted/error finishes. User-initiated aborts are unaffected because
abortSession() already clears the queue, and drain() stops after one
failed send so an erroring provider cannot spin the queue.
- cli: when a turn comes back aborted without the user having requested
an abort, append a "Task stopped before completion." status entry
instead of ending the turn silently.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(core): hold queued prompts on error finishes instead of consuming them
Addresses the Greptile P1 review on #13061: a drained prompt whose turn
resolved with finishReason "error" returned normally, so the
exception-only requeue path treated the send as successful - the failed
prompt was consumed and draining continued firing the remaining queue
into a failing provider.
- drain() now stops the chain when a drained send resolves with an
error finish. The errored entry itself is not requeued (its turn ran:
the prompt is in the conversation and the error is surfaced), but the
rest of the queue is held.
- runTurn() no longer schedules a drain after "error" finishes (the
skip is removed only for "aborted", which is the #13030 fix).
Held prompts still drain via the existing enqueue/update/delete
triggers or the next successful turn.
- Two new unit tests cover both layers.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* revert(cli): drop the 'Task stopped before completion' status line
Keep the change scoped to the queue-drain fix in @cline/core. The CLI
no longer prints a notice for non-user-initiated aborted finishes;
apps/cli is back to parity with main. When messages are queued, the
drain itself makes the stop visible (the queued message runs); richer
stop-reason surfacing can be a follow-up.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
The litellm builtin spec pinned protocol: "openai-responses", so every
request went to POST {baseUrl}/responses. Self-hosted LiteLLM proxies
commonly implement only /chat/completions, so all prompts failed with
404 Not Found on the SDK path (CLI, and now the Next extension bundle).
Drop the override so litellm inherits the openai-compatible family
default (openai-chat -> /chat/completions), matching every sibling
openai-compatible builtin and the Legacy extension behavior.
Fixes#13003, fixes#10781
* fix(cli): render MCP tool result text instead of escaped JSON in TUI
MCP tools return {content: [{type: "text", text}]} which
extractFullOutputText JSON-stringified, escaping newlines into one giant
line that word-wrapped across the whole terminal and never triggered the
line-based collapse. Extract the text parts with real newlines so the
existing collapse works.
Fixes#13038
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(cli): keep placeholders for non-text blocks in mixed MCP results
Addresses Greptile review on #13066: text-only filtering silently
dropped image/resource/audio blocks from mixed MCP content. Render them
as [type] placeholders instead.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(cli): surface non-text MCP block metadata in TUI output
Extract embedded resource text, and include resource/resource_link URIs
and image/audio mime types in placeholders so expanded mixed MCP
results keep identifying metadata.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
The stdio MCP client gave servers without a configured `timeout` only
1.5 seconds to answer initialize before killing the process, so
slow-starting servers (e.g. Oracle SQLcl's JVM-based `sql -mcp`) could
never load and were silently skipped at session start.
Raise the default connect budget to 30s, in line with the startup
budget other MCP clients allow. A configured `timeout` still overrides
it in either direction, dead commands still fail fast through the spawn
error/exit path, and the newline -> Content-Length framing fallback is
unchanged.
Fixes#13035
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* feat(desktop): route /team prompts through core runtime
Rewrite desktop `/team` commands as structured user command blocks before sending them to the core runtime. Validate task input and respect the globally disabled Teams tool setting.
Remove legacy agent spawn and team enablement flags from session configuration, and add coverage for prompt rewriting and disabled-tool behavior.
* fix(desktop): preserve team tool defaults
* fix(desktop): display queued /team prompts as their slash form
Queued prompts are stored in their runtime form, so a queued /team
command showed its raw <user_command> envelope in the prompt queue chip
and edit textarea. Fold queue items through formatDisplayUserInput for
display; saving an edit re-resolves the slash form through the sidecar,
so the round trip is lossless.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* chore(hub): align builtin tool catalog flags with the desktop sidecar
The desktop sidecar pins enableSpawnAgent/enableAgentTeams when listing
the builtin tool catalog; the hub's parallel listing did not, so the two
would drift if the preset defaults ever change. Pin the same flags in
the hub and cross-reference the two call sites.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* chore(desktop): drop inert enableSpawn/enableTeams config leftovers
buildCoreSessionConfig no longer reads these keys, so remove the dead
schema fields, default-config initializers, and chat-test payload
entries. The chat-session regression test still sends them on purpose
to prove legacy flags cannot override the runtime's tool presets.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(desktop): reject /team when the mode's tool preset disables teams
The /team guard only checked the global disabled-tools setting, but the
runtime resolves tool availability from the mode's preset, so a preset
without team tools (yolo) would still send the model a spawn-a-team
instruction it cannot act on. Resolve the teams catalog entry for the
session's mode and reject /team when it is unavailable, mirroring the
runtime's own availability logic.
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(core): surface OAuth authorization for SSE MCP servers on 401
A 401 from an SSE MCP server never persisted authorizationRequired: the
fetch-boundary UnauthorizedError was consumed by EventSource and re-thrown
as a status-less SseError, so the instanceof check routed it to
markConnectionError and hosts never offered the OAuth connect action.
Give the SSE stream request a raw fetch so a 401 fails the connection with
the SDK's typed SseError(401), and recognize 401s across transports with a
single isMcpUnauthorizedError predicate at every detection site.
* style(core): apply biome formatting to MCP oauth changes
Toggling Plan/Act while a turn was streaming or waiting on a tool approval
aborted the turn but left the TurnStateTracker on its last live phase: the
aborted session's done event is fenced off as stale once the rebuild
unsubscribes it, so nothing ever settled the phase. The webview then kept
rendering that phase forever - an eternal Thinking spinner with the input
disabled (aborted while streaming), or dead Approve/Run Command buttons wired
to an approval that clearPending had already denied (aborted while awaiting
approval). Users experienced this as 'switched to act mode and nothing
happened / it never wrote the files'.
Mirror cancelTask: after aborting the turn for the mode change, append a
resume_task ask row and set the phase to resumable, so the footer offers
Resume Task with the input enabled in the new mode.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(llms): retry mid-stream network interruptions before any model output
* fix(llms): scale network retry backoff by network retry count, not shared attempt number
* desktop: native-feel polish and render-path performance fixes
- Suppress the WebView browser context menu on app chrome (keep it for
editable fields and active text selections)
- Make UI chrome unselectable app-wide; opt chat messages, markdown,
code, diffs, and error banners back into text selection
- Contain overscroll so inner scrollers don't rubber-band the window
- Lazy-load Settings/Sessions/Onboarding/Diff views out of the entry chunk
- Memoize ChatInputBar and AgentHeader; stabilize their props in the chat
pane so stream flushes only re-render the affected message bubble
- Stop refocusing the composer textarea on every keystroke (caret flicker)
- Cache slash commands across menu opens (stale-while-revalidate)
- Avoid rebuilding reversed message arrays and ask-question JSX per render
- Drop core info/debug console logging on the streaming hot path behind a
cline:debug-logs opt-in; remove leftover [webview:delete] debug logs
- SearchCombobox (provider/model picker): Escape closes and restores focus
- Remove unused @vercel/analytics, recharts, embla-carousel deps and the
unused chart/carousel UI components
* desktop: surface failed-turn errors instead of leaving the chat blank
On a failed run the runtime reports its error string in result.text.
The webview rendered that as an assistant bubble, which the canonical
history rehydration then wiped (the failed turn is never persisted),
so provider errors like a retired model id left the user staring at a
silently empty chat. Route failed-turn text to a persistent error-role
message added after rehydration instead.
* desktop: fade the welcome/conversation swap instead of hard-cutting
Sending the first message replaced the hero layout with the message
grid in a single commit, which read as a white flash. A 180ms enter
animation now plays when either side becomes visible; disabled under
prefers-reduced-motion.
* desktop: render new-chat panes instantly from the last catalog load
Clicking + remounts ChatThreadPane, which refused to render until the
provider catalog (a large fetch) and workspace list resolved again —
about a second of blank pane plus boot spinner on every new chat.
Seed remounts from a module-level snapshot of the last successful
load; the mount effect still refreshes both in the background.
* desktop: invalidate the provider-catalog snapshot with the cache
Seeding remounted chat panes from the last catalog load left a window
where a pane created right after a credential change could act on the
old keys. The snapshot now lives in the catalog module and is dropped
by invalidateProviderCatalogCache(), so credential edits force the
next remount to wait for fresh data.
* fix(cli): harden tool input/output formatters against malformed payloads
Tool inputs cross the model/tool boundary and may not match their
TypeScript annotations (e.g. run_commands with { command: null }).
truncate() called str.replace() on such values, crashing the TUI with
'.replace is not a function' and making persisted sessions containing
the payload non-resumable, since hydration replays the same input
through formatToolInput().
Normalize untrusted values at the formatting boundary: truncate() now
accepts unknown and safely stringifies null/undefined/objects (including
circular structures and throwing toJSON), formatStructuredCommand no
longer returns non-string commands verbatim, and fetch_web_content
request summaries tolerate malformed entries.
Fixes#13036
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(cli): keep valid empty-string args in structured command summaries
Greptile review: filtering normalized args by truthiness also dropped
genuine empty-string argv entries, so summaries could show a different
argument list than the one executed. Filter only nullish entries before
normalization instead.
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* fix(vscode): fall back to session cwd/Desktop for @-mention search in empty windows
* fix(vscode): use the shared chat workspace as the no-folder fallback root
ensureGitRepository cached a negative probe for the lifetime of the hook
instance, so a session started in a non-git folder never got checkpoints
even after the user ran git init. Cache only the positive answer and
re-probe otherwise; the probe runs at most once per user turn.
* fix(desktop): treat signed-out state as a typed result instead of a command error
* fix(desktop): sign out when the organization balance fetch reports the typed signed-out result
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
* chore: retrigger checks after runner outage
---------
Co-authored-by: Saoud Rizwan <saoudrizwan@users.noreply.github.com>
Saoud Rizwan