Four worker-terminal-manager tests SIGKILL a terminal worker whose
terminal program runs in the test's temporary home. On Windows that
program ends only when the dead worker's console closes, a moment
later. afterEach then removed the home with a 0.5 s retry budget, and
Windows refuses to remove a directory that is a live process's working
directory, so cleanup failed with EBUSY.
The tests now kill a worker through crashWorker, which records the
worker and its terminal programs. afterEach waits up to 10 s for them
to end before it removes the temporary dirs.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(server): keep Changes available when Git metadata watching fails
A valid checkout could lose its first comparison when metadata observation exceeded the directory limit. Keep working-tree observation active and share a backing-off metadata sample per Git root so quiet comparisons stay idle.
* fix(server): probe only revisions used by degraded Changes streams
Reuse each comparison's initial Git resolution as the baseline. Batched ref lookups run after subscription setup, stay bounded when the repository has a large ref table, and refresh only comparisons with changed inputs.
* fix(server): report Changes metadata fallback in daemon diagnostics
Log the watcher failure once per root and count comparisons currently using the fallback. Exercise the transition from an active watcher to polling and the release of the last comparison.
* fix(server): keep Changes probes working across Git versions and worktrees
Use line-delimited batch input for older Git versions. Resolve shared refs in the common Git directory so a removed worktree cannot block the remaining comparisons.
* fix(server): isolate invalid revisions in Changes polling
Git aborts a batch when an upstream or reflog expression cannot resolve. Resolve expressions independently so a failed lookup leaves other comparisons current.
* fix(timeline): keep replies visible across idle notifications
Resolve notification membership at the manager's live event boundary. An idle notice gets a display turn without starting a run, and in-flight notices join the open run.
* test(timeline): keep notice paging assertions compatible with async history
* fix(timeline): reuse managed turn assignment for notices
Resolve notices before waiter matching with the same rule used by live persistence. Verify untagged foreground runtime notices through runAgent and idle Pi notices through timeline paging.
* fix(browser): let agents choose which desktop opens a tab
Registration order sent new tabs to the wrong desktop window. Let the agent discover and select the intended window by name, and require a choice when several windows can open tabs.
* test(desktop): run browser host selection in a fresh app session
Closing the second window can leave the existing browser journey in a different renderer state. Give both journeys their own windows and daemon so the routing test does not change the old test setup.
* fix(browser): keep window names visible and stable across reconnects
Frameless desktop windows need an in-app place to see and copy their name. Give each window one browser host identity so a reconnect replaces a lingering connection without creating a duplicate choice.
* test(desktop): cover host selection in the browser journey
Open the second window after the existing browser checks finish so their setup remains isolated within one CI run.
When Paseo used an OpenCode v2 service it did not start (a discovered
`opencode serve --service`, or a configured server URL), nothing
reported that service going away. The connection's `exited` signal
resolved only when Paseo itself stopped using it. A turn that was
running when the service died was never failed, the event stream
retried the dead port every 100 ms, and Stop, send and reload were
refused with "active run cancellation was not acknowledged".
The shared connection now treats a refused connection as the service
exiting, the same signal an owned server's process exit gives. The
session fails the turn, stops its stream, and reconnects on the next
operation.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(server): let agents watch pull requests
* fix(server): isolate PR watch test homes
* fix(server): preserve watch delivery during subscription changes
* fix(server): batch co-due PR observations and drain admitted updates
* fix(server): settle combined poll requests with await
* fix(server): deliver overdue PR updates together
* fix(server): include failing check links in PR watch updates
* test(server): verify PR watch check links on live GitHub
* fix(server): deliver PR digests through the shared agent queue
* test(server): verify PR and child updates share the next turn
* fix(server): withdraw queued updates before provider admission
* feat(server): show PR watches as agent background work
Render active watches beside heartbeats so users can see and stop them while the agent is idle. Bind Stop to the original subscription so a delayed request leaves a replacement intact.
* fix(codex): offer the Flex speed and thread it through
Speed choices came from the model list's `serviceTiers`, and Codex never
advertises Flex there: `ModelInfo::supports_service_tier` treats it as an API
request option that is valid whatever the model advertises. So the Speed menu
could never list it, choosing `flex` threw "Codex speed 'flex' is not
available", and a stored `service_tier: "flex"` was rewritten to Normal before
the first turn.
Flex now joins Normal as a tier Paseo knows, on every Codex model. The Speed
control and the tier checks read one list, so what the control offers is what
the session runs: `turn/start` and rewind's `thread/fork` carry
`serviceTier: "flex"`, a saved Flex survives open and model switches, and a
catalog that does advertise Flex keeps its own entry instead of duplicating it.
Ref: #9773
* Verify Codex Flex in the browser Speed menu
* Gate unadvertised Flex on Codex 0.156
Codex 0.155 silently strips Flex from outgoing requests when the catalog omits it. Use the connected app-server version to gate the locally added option and normalize unsupported saved selections. Keep the Speed icon highlight for any explicitly selected tier and verify Flex follows that behavior.
---------
Co-authored-by: Mohamed Boudra <boudra.moha@gmail.com>
Keep Ultracode independent of effort in the Claude composer and retain support for saved Ultra Code choices.
Co-authored-by: fivetaku <zipsanyang.ai@gmail.com>
An expired browser_wait condition reached the agent as "The browser did
not respond before the timeout. Try again or check the browser host."
The host did respond. summarizeBrowserError replaced every
browser_timeout message with that fixed text, so the host's own message
("Timed out waiting for browser text: ...") never reached the agent.
The tool text is now the producer's message for browser_timeout. The
broker's round trip timeout, the one case where the host did not
answer, now says so in its own message.
Closes#10128
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
When a plugin provider closed a session with session.closed, the next
prompt went stale and Paseo tried to reopen the agent. The reopen first
closes the old session, and that close sent session.close for a session
the plugin runtime had already forgotten. The send failed with "Unknown
provider session", the reopen stopped, and the prompt was rejected with
"Provider session <id> is stale after its plugin reloaded". Every later
prompt failed the same way.
A session that the provider closed now skips the native close and only
drops its routing, so the reopen continues and the prompt runs.
Closes#10151
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(omp): map tool approvals that offer session or grant choices
OMP 18.x adds "Approve <tool> for this session" and directory/file grant
choices to its rpc-ui approval select. The mapper only recognised the exact
["Approve", "Deny"] options, so these prompts fell through to a generic
"OMP select" question; allowing it from the CLI or a notification answered
with no value and OMP treated the call as denied.
Recognise approval selects whose one-call choice is "Approve" or
"Approve Once", whose last choice is "Deny", and whose other choices are
approval labels. Surface the extra choices as allow actions and answer with
the selected label.
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
* fix(omp): allow one call when a grant is labeled Approve
OMP names the one-call choice "Approve Once" whenever it also offers
grants. A grant for a target that is not a file is then labeled
"Approve", so the select arrives as
["Approve", "Always Approve", "Approve Once", "Deny"]. Taking the
first "Approve" as the one-call choice made a plain allow pick the
grant, which remembers more than the call. Prefer "Approve Once" when
it is offered.
Co-authored-by: Dhruv Singal <dhruvsingalabc@gmail.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* refactor(omp): read the picked approval choice from its action
Each extra approval choice's action label is the select value OMP
offered, so the response reads it from request.actions instead of a
second copy in metadata.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Dhruv Singal <dhruvsingalabc@gmail.com>
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Use current Hub permissions when reconnecting a daemon
* Back off when the Hub repeatedly rejects session permissions
* Show the permission recovery commands in Hub status
* Keep a finished reply visible when Claude starts a turn after a background command
When a background command finishes while Claude is idle, Claude Code sends the
task notification first and opens its own turn after it. The notification row
was stored with no turn id. A row with no turn id joins the turn before it to
the turn after it, so under Collapse turns the app showed both turns as one and
folded the finished reply behind the summary row.
The Claude provider now keeps a task notification that arrives between turns
and emits it inside the turn Claude starts for it, so the row carries that
turn's id. A notification no turn follows is emitted as before: when Claude
reports idle (a command the user stopped), before the user's next message, and
on close.
Rows already stored without a turn id are not changed.
Refs #10030.
* Show a stopped background command's notification at once
A stopped background command starts no Claude turn. While a background
subagent is alive Claude also sends no idle, so the held row stayed hidden
until the subagent finished and then took the subagent's wake-up turn id.
Only task notifications Claude reacts to (completed or failed) wait for a
turn now. A stopped one is emitted right away, as before.
Refs #10030.
* Keep a row that belongs to no turn out of the turns around it
A turn starts with the user's message or the provider's own turn start
(Claude's init). A background command's notification that arrives
between turns starts nothing, so the daemon writes it with no turn ID.
The shared turn rule read a missing ID as legacy history and let the
row continue both neighbours. The reply before it and the turn after
it then read as one turn, and Collapse turns folded the first reply.
Now a row with no turn ID never continues a row that has one, so it
cannot join two turns. History with no IDs at all keeps the trigger
rule. The app's fold and the daemon's whole-turn paging share the rule.
This replaces the earlier Claude provider hold, which had to guess from
the notification's status whether Claude would start a turn for it.
Refs #10030
* Give a background task's notification the turn it starts
A background task's notification that arrives while no turn runs starts
a turn. Claude's next run belongs to it unless Claude announces another
start first: a later notification starts the next turn, and a queued
command (the user's message, a scheduled wake-up) starts its own. When
Claude does not respond, the notification is a turn in itself.
So a reply Claude writes for a finished background command shares the
notification's turn, and the reply before it stays in its own turn and
stays visible under Collapse turns. A stopped command's row shows at
once as its own turn, and the agent stays idle.
Claude sends no frame for a Monitor event, so a run it starts for one
belongs to the last notification before it.
This replaces the shared turn-rule change, which drew the notification
as a turn between the two.
Refs #10030
* Keep a subagent's own job notification out of the main session's turns
A background job a subagent started reports to that subagent, and
Claude starts no main-session run for it. Its notification gave the
next main-session run a new turn, so when it arrived between a main
command's notification and Claude's reply to it, the reply was drawn
in a different turn from the notification it answered.
Refs #10030
When OMP changed the thinking level itself, for example on `/switch @role`
with a role that carries a level, Paseo dropped the `thinking_level_changed`
event because the OMP runtime schema did not list it. The thinking selector
and `runtimeInfo.thinkingOptionId` kept the old level while OMP ran the new
one.
Parse the event and update the session's level, preferring OMP's
`configured` selector (for example "auto") over the effective level, the
same way the Pi provider handles it (#6741).
Closes#8249
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Give worktree teardown commands a shared one-minute deadline. Projects can override it with `worktree.teardownTimeoutMs` in `paseo.json`. Drain output so successful teardowns can exceed the old 1 MiB buffer. Bound the output retained for diagnostics.
At the deadline, try to terminate the command, close its output streams, and keep the worktree on disk. An exit status of zero in response to the stop signal still counts as a timeout.
Closes#9857.
Co-authored-by: Okan YILDIRIM <63511519+ragokan@users.noreply.github.com>
* fix(terminal): delete a stopped worker's copy on Windows
Windows keeps a just-exited worker's files open for a moment, so the
single rm in pruneTerminalWorkerRuntimes failed and left the copy until
the next daemon start. Retry the delete for up to 5.5s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(terminal): bound the Windows retry of a worker copy delete
rm's maxRetries repeats at every directory level, so a file that stays
locked deep in node-pty would hold daemon start and stop far past the
intended window. Retry one rm per attempt for up to 3s instead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(workspaces): preserve commits when archiving worktrees
A removed worktree can be the only reference to its HEAD. Keep that commit before deletion and record the checkout so detached workspaces restore at the same commit.
Co-authored-by: Okan YILDIRIM <63511519+ragokan@users.noreply.github.com>
* fix(workspaces): restore saved checkouts through one git path
Keep existing archive and deletion behavior when Git cannot read a worktree HEAD, and record every readable checkout for a common restore path.
* fix(workspaces): preserve git HEAD identity through restore
Record Git’s full HEAD ref or commit so restore can pass branch and detached checkouts to Git without guessing from a short name.
* fix(workspaces): scope saved HEAD to each archive
An archive with no readable checkout must not retain the HEAD saved by an earlier archive.
* test(workspaces): verify detached checkout after restore
Update the browser journey for detached restore and allow the daemon restart test enough time on Windows runners.
---------
Co-authored-by: Okan YILDIRIM <63511519+ragokan@users.noreply.github.com>
* Let a steer wait out a Claude compaction instead of cancelling it
While Claude compacted, the Claude session answered every steer with
"unavailable". The agent manager reads that as "replace the turn", so
it interrupted the turn and Claude Code printed "Compaction canceled.".
Claude Code already holds a `next`-priority message until compaction
finishes, so the session now accepts the steer during compaction.
Co-authored-by: NightMachinery <36224762+NightMachinery@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Steer only while the compaction is still running in the real Claude test
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: NightMachinery <36224762+NightMachinery@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The MCP agent tools serialized live agents without the background work
summary that the WebSocket snapshot carries, so a parent agent saw an
idle worker with a running background shell as plain idle. Both tools now
add the same summary the session adds, omitted when the agent has none.
Co-authored-by: Orcun Simsek <1612171+ortschun@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Changing the thinking level of a running Claude agent marked its query for a
restart. The next turn, mode change or model change then retired the Claude
process, which ended every background shell, workflow and subagent it ran with
"connection lost".
A move between two effort levels now goes to the running query through
applyFlagSettings({ effortLevel }), as Fast mode already does, and marks no
restart. Switching thinking off, to ultracode or back to the default still
restarts, because those change how the query is launched. If the running
process does not take the change, the old restart path runs, so the change is
never lost.
Row: app-built-from-one-set
Paseo: c6210e21-fc13-4dfa-adfc-43a050c7d75a
Co-authored-by: gilad-rubin <6134299+gilad-rubin@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A force-cancel finalizes turn X in the daemon while the provider keeps
running it, so the next run can be bound to the same X. handleStreamEvent
dropped every terminal for a finalized turn id, including the synthetic
turn_canceled that cancelAgentRunNow dispatches for the tracked run and
the provider's own terminal. The run never settled, cancelAgentRunNow
waited on it forever, and every later cancel, reload and send queued
behind it until the daemon restarted.
A terminal for the tracked run's own turn id is no longer dropped.
Fixes#6355
Supersedes #8405
Co-authored-by: Dominic Anaeto <dominic.anaeto@gmail.com>
On web and desktop, Download over a direct connection now fetches the file and saves it from a blob instead of navigating the window to the host URL. A rejected or unreachable download shows an error in the download toast and the app stays on the workspace. The daemon's HTTP CORS check uses the same origin rule as the WebSocket upgrade.
Closes#9895
Co-authored-by: panta82 <4291324+panta82@users.noreply.github.com>
Brings in the shipped 0.16.1 CHANGELOG.md and CHANGELOG.zh-CN.md entries
exactly as tagged, the 0.16.1 version in every package.json and
package-lock.json, and the F-Droid changelog files for version codes
160011-160014.
* fix(server): treat canceled Codex starts as interruptions
* fix(server): report a replaced pending start as undelivered to its sender
waitForAgentRunStart followed the agent rather than the sender's run, so a
send canceled by a replacement before reaching Codex resolved once the
replacement started. Keep the sender's own pending run and reject when it is
cleared before its turn starts.
The canceled-start tests drain streamAgent, because a fresh Codex agent has no
session handle before its first turn (#8393).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(server): keep a sender's start wait across a stale-session retry
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(server): expect the start options passed to a foreground run
Session create and scheduled prompts now pass the sender's start options,
so the two call-argument assertions include them.
Co-authored-by: 3ae3ae <3ae3ae@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(server): keep a Codex message canceled before it starts in history
A Stop or replacing send during Codex preparation settled the run as
canceled without recording the sender's prompt, so the message vanished
from history after a reload and on other clients. Record it as a
rejected attempt, as the start-failure branch already does, so it stays
in history without being delivered.
Co-authored-by: 3ae3ae <3ae3ae@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(server): return each send's start from dispatch
A send waited for its turn by looking up the agent's pending run when the
wait began. When a provider plugin's session was stale, the first start
failed and its run was cleared before the sender started waiting, so the
send rejected with "has no pending run" while the retry answered it.
startAgentRun now returns the start of the prompt it dispatched. Every run
created for that prompt, including the stale-session retry, is bound to
it, so the send resolves when its own turn starts and rejects when its own
run ends unstarted. This replaces the retry hand-off between runs in the
agent manager.
Co-authored-by: 3ae3ae <3ae3ae@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: 3ae3ae <ogyshjsm@naver.com>
Co-authored-by: Bae Seongmin <83870598+3ae3ae@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: bot <bot@localhost>
Co-authored-by: paseo-bot[bot] <paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: 3ae3ae <3ae3ae@users.noreply.github.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
In Bypass, Claude agents again show a permission card for asks from the user's permissions.ask rules (including managed policy), PreToolUse hooks that return ask, compound and piped commands that hit an ask rule, permissions.blockReadsOutsideWorkingDirectories, an organization's ask limit, and sandbox overrides. Claude's own checks are still approved without a card. Upgrades @anthropic-ai/claude-agent-sdk to 0.3.326, which passes the ask's source to canUseTool.
Closes#9916
* fix(server): move updatedAt when archiving an agent that is only in storage
Archiving a loaded agent sets its updatedAt to the archive time
(markRecordArchived). Archiving an agent that is only in storage, which
is also how a parent's archive reaches its children, kept the old
updatedAt. Lists sorted by updated_at, such as History, then left a
just-archived agent where its last turn put it, sometimes days back, and
clients paging such a list by time never reached it.
archiveSnapshotUnlocked now records the archive time as updatedAt too.
Co-authored-by: RoundRedDot <roundreddot@gmail.com>
* fix(server): keep updatedAt from moving back when an archive is restored
Archiving a stored-only agent now sets updatedAt to the archive time.
When the archive time is earlier than the record's updatedAt, as when a
failed send or resume puts an agent's original archive back after
unarchiving it, updatedAt stays where it is. Clients drop agent updates
that are older than the one they hold, so moving it back left them
showing the agent active.
Co-authored-by: luke <luopeihuan@gmail.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(server): keep updatedAt from moving back when a loaded agent's archive is restored
The rule that archiving never moves updatedAt backwards now lives in
buildArchivedAgentRecord, which both archive paths use. A failed send
that resumed the agent before a later loading step failed puts its
original archive back through the loaded path, which had the same
backwards move.
The workspace archive stamp test now uses stamps taken after the
agents' activity, as real archive stamps are.
Co-authored-by: luke <luopeihuan@gmail.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* refactor(server): require the archive time when building an archived agent
buildArchivedAgentRecord no longer accepts an archive that leaves
updatedAt unset, the gap #9920 came through, and its tests now assert
the rule: a later archive time moves updatedAt, an earlier one keeps it.
Co-authored-by: luke <luopeihuan@gmail.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: luke <luopeihuan@gmail.com>
Co-authored-by: RoundRedDot <roundreddot@gmail.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
With the default send setting, a message sent while an OpenCode agent
waits on a question steers the turn, and Paseo rejects the question in
OpenCode. The OpenCode session dropped the question from its own list
without reporting it resolved, so the agent manager and the app kept
showing it. Answering or dismissing it then failed with "No pending
permission request", and the prompt could not be closed.
The session now reports permission_resolved whenever OpenCode accepts a
reply, rejection, or answer, as the other providers do.
Fixes#9879
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
macOS FSEvents can replay the file's pre-subscription write as a create
after the observer starts. The test read the first event for the path,
so it stayed on that create after the delete arrived in a later batch.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A Codex session opened to read history reads the thread and disposes its
app-server right away, but canStopRuntime() vetoed it because it had no
client. Idle cleanup then kept every such chat loaded in the daemon.
Closes#9804
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Spencer Shulem <1703265+Spshulem@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Let Claude model discovery finish a token refresh instead of killing it
Model discovery stopped waiting after 5 s by aborting the SDK query and
tree-killing the CLI. A CLI that is still refreshing an expired OAuth
token at that moment has already rotated the refresh token, so the kill
leaves only the revoked token on disk and logs the user out.
The 5 s limit now only bounds the wait. A CLI that has not answered gets
its input ended and exits on its own; it is tree-killed only after a
60 s grace period or when the daemon shuts down.
Fixes#9314
* fix(server): let Claude model discovery finish a token refresh after it answers
Claude Code answers the model list before its startup token refresh has
saved the rotated refresh token. Discovery killed the CLI right after the
answer, so a refresh still in flight left the revoked token on disk and
every later Claude session failed until /login. Discovery now ends the
CLI's input and lets it exit on its own whether or not it answered, with
the existing grace period and shutdown kill as the bound.
Co-authored-by: daysun-k <8309193+daysun-k@users.noreply.github.com>
---------
Co-authored-by: daysun-k <8309193+daysun-k@users.noreply.github.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
After a daemon restart, an agent's timeline is rebuilt from provider
history, and those rows have no turn id. A reply the agent starts on its
own, such as Claude waking after a background subagent, follows the
previous reply with no user message between them, so the last assistant
message lookup joined the two and wait results and finish notifications
returned the old reply with the new one appended.
The timeline projection already treats a message id change as a new
message. Share that rule with the lookup, so both decide which rows are
chunks of one message the same way.
Closes#9814
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(plugins): append agent instructions on session open
Recompute plugin instructions for each launch so changes apply on reopen without changing saved agent configuration.
* test(providers): verify plugin instructions on reopen
* fix(codex): refresh instructions from restored collaboration mode
A resumed thread can retain its prior collaboration instructions when the mode catalog is unavailable. Rebuild those instructions from the current launch before the next turn.
* fix(plugins): keep launch instructions out of agent records
Provider persistence handles can contain launch configuration. Strip transient instructions before saving and projecting handles.
* fix(codex): preserve built-in collaboration instructions
Pass agent instructions through the thread configuration so Codex can apply its own Plan and Default mode instructions.
* fix(opencode): preserve instructions when reading history
History-only opens do not apply current plugin contributions, so they must not clear the native session instruction entry.
* fix(codex): retain instructions after rewinding
A rewind forks the native thread, which does not inherit its parent launch instructions. Pass the current instructions to the fork.
* Deliver updated instructions to Codex threads
* Cover Codex snapshots after rewind
* Persist provider state when reopening agents
* Limit resume persistence refresh to providers that need it
OMP model labels were built as `${provider}/${name}` and then cut at the
last `/`, so a free-text name such as "... ~60-70 tok/s on 0.2.10, 52/56)"
showed as "56)". Underscores were rewritten to spaces ("IQ3_S" -> "IQ3 S"),
and a colon inside the name cut everything before it.
Pi had the same bug, fixed in #9823. Move that label rule into a shared
model-label module and use it for both providers: free-text names are
kept, path-like names are shortened to their last segment, and a
"Vendor: " prefix is dropped only when it is a prefix.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Rahul Sharma <19687913+rms523@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The private index snapshot behind index and All changes comparisons was a
fresh copy, newer than every entry, so Git skipped its same-timestamp
content check and a same-size save made right after staging read as
unchanged. The copy now keeps the source index's mtime, rounded down to
the second.
Closes#9827
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(server): abort timed-out watcher setups and back off their retries
A watcher subscription that missed WORKSPACE_GIT_WATCHER_SUBSCRIBE_TIMEOUT_MS
was abandoned but its startup scan kept walking the tree to completion, and
recovery retried every five minutes forever, repeating the full scan each
time. On a CPU-starved host that kept dozens of directories rescanning.
- FileObserver.subscribe accepts an abort signal that stops the startup scan
and rejects; WorkspaceGit aborts setups at the deadline and on dispose.
- Consecutive setup timeouts extend the recovery delay past the 5-minute
ladder cap, doubling up to one hour; an accepted subscription resets it.
Degraded polling keeps its 5-60 s cadence meanwhile.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(file-observer): keep an aborted observation's backend from starting
Aborting a subscription closes its observation, which can happen while
start() is still reading the root. The backend then started after its close
had finished, and the native backend opened a root watcher that nothing
would close.
Co-authored-by: Spencer Shulem <1703265+Spshulem@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(file-observer): abort a subscription mid-scan in the abort test
The test aborted before the backend started, so it passed with the abort
listener removed. A tree past the Linux directory limit makes a scan that
keeps walking fail on the limit instead of rejecting with the reason.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Spencer Shulem <spshulem@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Spencer Shulem <1703265+Spshulem@users.noreply.github.com>
* fix(server): queue agent messages until the current turn ends
Steering refusal could replace a running turn and cancel compaction. Hold agent-originated messages and notifications together until the recipient settles, retaining explicit interruption and acknowledging finish subscriptions only after dispatch succeeds.
* fix(server): wake queued delivery at turn and permission checkpoints
A pending follow-up must not hide the permission needed to finish the current turn. Wake ready batches on recipient state changes, let completion waits report permission checkpoints, and preserve admission order and session-mode boundaries when busy dispatches retry.
* fix(server): publish autonomous readiness after run settlement
Idle snapshots emitted while the autonomous run record still existed could not wake queued delivery. Publish terminal state after clearing that run, and keep already debounced busy retries ready for the next turn boundary.
* fix(server): apply queued modes only after run admission
Preparing a later mode group could change the running turn before busy admission rejected it. Carry the requested mode into the reserved run and apply it immediately before provider startup, leaving earlier and concurrently started turns untouched.
* fix(server): keep later mode groups queued until turn end
Dispatching every mode group at once prepared later modes against the preceding turn. Keep later groups pending and reject busy sends before changing mode, leaving mode validation outside run startup.
* fix(server): deliver provider commands during queued turns
Offer waiting messages individually through the provider-owned out-of-band path while a recipient runs. Recognized commands retain their existing behavior, while ordinary messages keep waiting without starting their dispatch observers.
* fix(server): offer queued commands once while running
Remember provider declines so timer retries do not repeat side effects or reinterpret a waiting message after provider state changes.
* fix(server): defer commands that request a session mode
An explicit mode belongs to a later turn. Keep those commands on the normal idle dispatch path so an out-of-band offer cannot ignore the requested mode or change the running turn.
The reaper signals a worker only after its socket accepts. The two
"cannot be opened for a moment" tests spawned a listening worker and
reaped without waiting for the listen, so on a loaded runner the reaper
found no socket, left the worker alive, and the test failed after
waitUntilGone's 5s deadline. Spawn the worker through a helper that
resolves once the socket is listening, as the other socket tests in
this file already do.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(pi): keep slashes and underscores in model display names
Pi model labels were built as `${provider}/${name}` and then cut at the
last `/`, so a free-text name such as "... ~60-70 tok/s, 52/56)" was shown
as "56)". Underscores were also rewritten to spaces ("IQ3_S" -> "IQ3 S").
The compact model trigger in the app repeated the last-`/` cut.
Build the Pi label from the model's own `name`/`id`: shorten only
path-like values with no whitespace (OpenRouter reports names like
"google/gemini-2.5-flash-lite"), and keep free-text names intact. The
compact trigger now strips only a leading provider namespace.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(pi): cover model labels through the catalog and drop stale ACP copies
acp-agent.test.ts still imported transformPiModels, which the Pi label
fix removed, so its modelTransformer test and a stale copy of the old
Pi label tests failed. The modelTransformer test now supplies its own
transformer, and the Pi label cases run through fetchCatalog with the
reported model names instead of an exported helper.
Co-authored-by: Rahul Sharma <19687913+rms523@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(pi): keep colons inside free-text model names
The vendor-prefix strip cut a free-text Pi name at its first ": ", so
"Qwen3 IQ3_S (tail skip: on, 52/56)" showed as "on, 52/56)". Strip only
when the text before ": " has no parenthesis or comma, which keeps
OpenRouter's "Vendor: Model" names shortened.
Co-authored-by: Rahul Sharma <19687913+rms523@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Rahul Sharma <19687913+rms523@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
A PASEO_HOME on AFS passes the socket directory checks, then refuses the
Unix socket bind with EPERM, and the terminal worker failed to start. The
worker now binds each eligible candidate in turn and moves on when the
filesystem refuses the socket (EPERM, EACCES, EOPNOTSUPP, ENOTSUP). Other
listen errors still fail the start.
Fixes#9213
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A turn the provider starts on its own, such as Claude waking after a
background subagent, has no user message before its reply. The last
assistant message lookup joined every adjacent assistant row, so it
returned the previous turn's reply with the new one appended, and the
220-character notification preview showed only the old reply.
Stop the backward scan at the turn boundary so the lookup returns only
the newest turn's reply, while chunks of one turn are still joined.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The daemon drops a dictation stream when its STT session errors and answers
every later chunk with dictation_stream_error. The app only listened for
acks, so finishing waited 30s and reported a generic timeout. The sender now
fails finish with the daemon's message, and the daemon logs the failure that
dropped the stream once instead of nothing.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(app): browse, install and manage plugins from a built-in gallery
Settings → Host → Plugins becomes a gallery with two tabs. Installed lists
the host's plugins with on/off switches and pending updates. Browse lists the
plugin directory with featured cards, categories, sorting and search. Each
plugin has its own page with its screenshots, overview, details, install or
on/off controls, and its own settings screens inline.
Every install asks first, in a sheet naming the source and version, because
installing runs the plugin's code on the host. Uninstalling also asks first;
its confirm says what is deleted.
The daemon fetches the directory from its configured registry
(plugin.registry.list/get, gated by server_info.features.pluginRegistry). The
app never fetches from the registry itself, so a host with a private registry
shows that registry.
README overviews render with the registry's overview contract (the same one
the website uses): only https links and images work, and nothing is read as a
workspace file. Before this, a word like "Node.js" became a broken relative
link.
* fix(app): refresh plugin failures and leave Settings cleanly from a plugin page
The Installed row read the app-side evaluation error without subscribing to the
registry, so a plugin fixed by Reload kept its "failed" badge. Rows and the
plugin page now subscribe through usePluginEvaluationError.
Leaving Settings with no remembered workspace replaced only the top stack entry;
with a plugin page pushed on top, the Plugins list stayed mounted underneath and
reopening Settings stacked a second copy. Dismiss the Settings entries first.
* fix(app): keep Browse working as the registry grows, and test it end to end
Registry listings now reach apps through an open wire schema
(plugin.registry.list/get carry PluginListing*), and the daemon drops a listing
it cannot read instead of failing the whole index. A registry that adds an
artifact kind or bumps schemaVersion no longer breaks Browse on older apps.
The gallery picks its opening tab once, so installing the first plugin keeps
the user on Browse. README links open only https: the parser still autolinks
bare http, www and email text, so a press checks the URL again.
Adds a Browse e2e test against a local registry and Git repository: install
from a row through the confirm sheet, stay on Browse, open the listing page,
uninstall from it.
* test(server): give the oversized-sibling diff refresh the same time budget as its siblings
expect.poll defaults to one second, which a Windows git diff refresh can exceed.
* test(server): wait as long as the other observer tests for native watcher events
expect.poll defaults to one second; macOS FSEvents can deliver later than that on CI.
* fix(app): keep a listed plugin's page usable after uninstalling it, and fetch the directory less
An installed plugin's page finds its listing through the plugin, so uninstalling
a directory plugin from the Installed tab left "not installed or listed" with no
way back. The page now becomes the listing's page, offering Install again.
Updates are re-checked when the host's plugins change and after a failed update,
whose proposal named a revision that no longer exists. The directory, listing
and update queries refetch on mount only once stale, so opening a plugin's page
no longer refetches the whole directory and every update check.
* test(app): install the plugin before opening Plugins in the busy-queue reload test
Plugins opens on Browse when the host has no plugins and stays there after the first install, so a plugin installed through the daemon after the page opened never showed on Installed.
* test(website): wait for hydration before scrolling to the hero in a short window
The router resets the scroll position when the page hydrates. Scrolling to the mockup before that left it below the window, so its start line never came on screen and the opening scene never played. It failed 7 of 10 runs on webkit.
* fix(server): show the registry setup error on Browse instead of asking to retry
The registry's request headers were built inside the try around fetch, so a setup error such as credentials over plain HTTP came back as "Could not reach plugin registry. Retry later." Install already built them first.
* fix(app): show plugin errors without the request details meant for logs
A daemon RPC error's message ends with requestType=… code=…; the daemon's own text is in userMessage. Gallery alerts and toasts now show that text.
* fix(app): show why an update failed on the plugin's page
The update operation's failure was recorded under the plugin ID, but the page only read it to show the spinner, so a failed update looked as if nothing happened. The page now shows the error in its own alert, without the plugin's Failed badge, because the installed version keeps running. While an update is still available the button reads Retry.
* fix(app): show a plugin page's update error only while the update is offered
The failed update stayed in the operations store after the plugin was updated elsewhere or uninstalled, so the page kept saying the update failed. The page now shows it with the Update button it belongs to, as the Installed tab's Updates row does, and uninstalling forgets it.
Approve Claude execution permission callbacks before creating or emitting permission events, preventing approval cards and transient waiting states in Bypass. Keep AskUserQuestion and ExitPlanMode interactive. Add regression coverage for silent approval and live mode changes.
timeline.append accepted any string as the row kind, but the app registers
renderer kinds with /^[a-z][a-z0-9-]*$/. A plugin that appended a kind such as
repro.card got a successful append, then its client setup threw in
addTimelineRenderer and the whole plugin client failed to load.
The daemon now rejects an append whose kind breaks that rule and names the rule
in the error. The wire schema stays a plain string. Document the rule in the
plugin docs.
Fixes#7326
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>