Claude Code sends a tool_progress heartbeat every 30 s while a tool is
still running, with that tool's own id as parent_tool_use_id. The Claude
adapter routed any frame with a parent id to the sidechain tracker, so
a main-thread call that ran past 30 s (MCP, WebFetch, WebSearch) became
a "Claude subagent" row with a child action per heartbeat that never
completes, and its card was relabeled Task.
A heartbeat now names no sidechain, so it never reaches subagent
routing.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(claude): load history from a transcript under another project folder
Claude Code resumes a session by id from any working directory and keeps
writing to the transcript in the project folder where the session started.
When the agent's cwd encodes to a different folder (a moved project, a
migrated daemon, a Windows binary under WSL), the agent kept working but
reopened with an empty timeline. Look the transcript up by session id
across the config dir's project folders when it is not where cwd points.
Refs #4540
* fix(claude): report a failed project folder listing instead of an absent transcript
Take the transcript lookup's inputs as one object, and rethrow read errors other than a missing projects folder so the history load logs them as a failure.
* test(claude): skip the unreadable projects folder case when running as root
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(claude): open autonomous turns only when Claude starts a turn
Paseo guessed from frame shape whether the main session had started work
on its own, and only a result could close the turn that guess opened. A
background subagent's task frames, its own background jobs'
task_notification and rate limit events all arrive with no parent tool
use id, so they looked like main-session work and opened a turn no result
would ever close. The agent then reported "running" while its main
session sat idle, and the app's message queue never drained.
Claude brackets every main-session turn itself: system/init opens it and
exactly one result closes it, whether a prompt, a wake-up after background
work, a slash command or an API error started it. Background subagents
never emit either on the main stream. An autonomous turn now opens only on
that init, and only when no foreground turn is open. The frame heuristic
is removed.
- Paseo launches the CLI with CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS=1,
added as an overlay in buildSdkEnv so the SDK and the usage probe share
one env object. An "idle" session state closes an autonomous turn as a
safety net, never a foreground one. "running" opens nothing: Claude Code
2.1.280 stays "running" for as long as any background subagent is
alive. State frames never reach the timeline.
- An init clears a leftover pendingInterruptAbort. A new turn has begun,
so the interrupted one is over, and a stale flag would swallow the new
turn's result.
Tests that opened autonomous turns from bare frames now send the init
Claude sends. New coverage: helper frames never open a turn while the main
session is idle, a wake-up turn opens on init and closes on its result, a
folded steer keeps one turn while a turn Claude starts after the result
gets its own, an idle closes a turn no result closed, session state never
opens a turn, and the env flag reaches the SDK launch.
docs/agent-lifecycle.md records that only system/init and result bracket
a main-session turn, since the code cannot tell a reader why frame shape
fails.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(claude): interrupt Claude only while its main turn runs, and spare helpers
Every Stop, and every send that replaces a running turn (paseo send, MCP
send_agent_prompt, voice, the app's queue drain, Send now), ends in
query.interrupt(). Without the SDK's perTaskStopAffordance declaration
that call kills every running background subagent, even when the main
session is idle and there is nothing of ours to stop.
- Declare perTaskStopAffordance, so an interrupt aborts only the main
turn, like Esc in Claude Code. Claude stops a helper itself with its
TaskStop tool.
- Track whether a main-session turn is in flight from Claude's own
brackets: set by system/init and by a pushed message's
command_lifecycle "started" (/compact works for seconds before its
init), cleared by the result or an idle session state.
- Call query.interrupt() and set pendingInterruptAbort only when that
turn is in flight, or the cancelled turn's own message has already been
started. Otherwise withdraw our not-yet-started messages with
cancel_async_message instead: an idle interrupt reaches only helpers,
and with no result to follow it the flag would linger.
- Track the foreground prompt with the queued steers until Claude starts
it, so Stop withdraws a message still waiting in the queue. A turn
stopped while the Claude process is still starting never sends it.
- Accept command_lifecycle "cancelled" as terminal, so a cancelled steer
no longer waits to be withdrawn or keeps denying permission cards.
Interrupt tests now wait for the pump to read the init before stopping,
since that is when Paseo learns a turn is in flight. New coverage: a stop
before Claude starts the message withdraws it without interrupting, a
stop while only helpers run never reaches Claude, a stop mid-turn
interrupts once and drops the aborted result, a new turn clears a flag
no result will consume, the SDK option is passed, and a cancelled steer
is not withdrawn again.
docs/providers.md described the old unconditional interrupt; it now states
when the Claude adapter interrupts and why it declares the affordance.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(claude): keep a helper spawned in the background out of a turn's cancel
When a turn is canceled, Paseo terminalizes the subagents that were
running in its foreground and skips the backgrounded ones, which outlive
it. It only learned a child was backgrounded from task_updated's
patch.is_backgrounded, which Claude sends when a foreground task moves to
the background later. A child spawned in the background carries the flag
on task_started instead, and a resumed subagent always does, so the
track marked those helpers "canceled" while they kept working.
Read is_backgrounded from task_started as well. docs/agent-lifecycle.md
said task_updated was the only signal; it now names both.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(app): drain queued messages as steers
The composer queue drains when an agent goes idle. It sent each message
with no activeTurnBehavior, which the daemon treats as "interrupt": if
the agent had just started a turn of its own (a Claude wake-up after
background work finished), the drained message stopped that turn and
replaced it.
Send drained messages with activeTurnBehavior "steer". A drain that races
such a turn folds into it; with no turn running the daemon starts a
normal one, as before. "steer" is an existing protocol value, so older
daemons need nothing new.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(claude): cover background helpers against real Claude
Every other test in this file tells Claude "Do not use a background
task", so a message sent or a Stop pressed while a backgrounded subagent
runs was never exercised. Four real-Claude scenarios cover it:
- "a background subagent keeps running after a message is sent while it
works": one helper, the main turn ends, the parent is not reported
running, a message in that window is answered, and the helper is never
read as canceled and completes.
- "a helper's own nested background job does not phantom-run the parent
or kill the helper": a helper starts its own background Bash job, then
keeps working in the foreground. The parent stays idle across the
nested job's completion (sampled for 30 s around a 5 s job), a message
sent in that window is answered, and the helper reports CHILD_DONE.
- "the app's default steer survives a backgrounded helper while the main
session works": the app's default send behavior (steer) does not kill
a helper when it interrupts the main session's own turn.
- "Stop cancels the main session's own turn but spares its background
helper": the same for Stop.
The wake-up turn Claude starts after a helper finishes lands a moment
after the helper reads completed (0.1 to 5 s in recorded traces), and the
daemon announces it as a pushed "running" agent_update, not as a
turn_started stream event. The checks poll for it and read the final idle
once the wake-up settles. waitForFinish can return a stale "idle" right
after a send, and the push stream can miss a message's final text when it
resolves, so the tests read fetchAgentTimeline like the rest of the file.
Helper status comes from agent.provider_subagents.list, since agent_stream
carries no provider_subagent events.
The scenarios need the previous commits together. Run against the frame
heuristic this branch removes, the nested-job test fails: the parent
flips to running once the nested job finishes and stays there.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(claude): interrupt a stopped prompt Claude has already taken
Stop before Claude announces a prompt started withdraws it with
cancel_async_message instead of interrupting. Claude can dequeue a
message for its next turn before announcing it, and withdrawing it is
then a no-op that reports cancelled: false, so the stopped prompt kept
running. Interrupt whenever a message could not be withdrawn.
Co-authored-by: Isak Persson <isak.a.persson@gmail.com>
* fix(claude): withdraw a stopped prompt only once
When the withdrawal of a stopped prompt failed, the interrupt path
withdrew a second time, which could cancel a prompt sent while the first
withdrawal was in flight. Withdraw once and interrupt on failure.
The real-Claude steer scenario now waits for the background helper to
exist before steering, so it proves the helper survives the steer.
Co-authored-by: Isak Persson <isak.a.persson@gmail.com>
---------
Co-authored-by: Isak Persson <isak.a.persson@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>
A file name that fits in Latin-1, like café.txt, went out as a raw byte in
Content-Disposition filename= with no filename* parameter, and Chromium saved
the file under a name taken from the URL (download.txt). Build the header so
every non-ASCII name carries filename*=UTF-8''… with an ASCII fallback.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* Keep reporting a Linux daemon as running after its VM resumes from a host sleep
A VM paused while its host sleeps keeps its uptime, but its wall clock
catches up on resume. The boot instant derived from the two then lands
after the running supervisor's lock, so daemon status reported it as
stopped and daemon stop released its lock without stopping it.
The lock now records Linux's boot id, and a lock that carries one is
judged by it. Locks without one keep the wall-clock check.
Closes#6019
* Give the OMP idle-deadline test a deadline a slow runner cannot pass before the first wait
With a 1 ms deadline, the gate's first check, made synchronously when the
turn ends, could already be past it on a loaded Windows runner. The turn
then failed before the test's first scheduler wait, and the test timed out.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(server): encode Unicode download filenames in Content-Disposition
* test(server): cover Korean download filenames
* test(server): validate download token and agent ID formats
* fix(server): set the download header through Express instead of a new dependency
The download route already runs on an Express response, and res.attachment()
builds Content-Disposition with the same content-disposition encoder that
Express ships. Use it and drop the direct content-disposition dependency and
its type package. The header values the download tests assert are unchanged.
Co-authored-by: jasonhnd <3293256+jasonhnd@users.noreply.github.com>
---------
Co-authored-by: Dillon <dillonzq@outlook.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: jasonhnd <3293256+jasonhnd@users.noreply.github.com>
* fix(acp): apply configured models the agent does not advertise
A model added through a provider's models or additionalModels was listed
in the picker, but an ACP session only accepted ids the agent advertised
in session/new, so the agent silently ran on its default model. The
registry now passes configured model ids to the session, which applies
them through session/set_model. Unconfigured, unadvertised ids are still
skipped with a warning.
* fix(acp): send configured models when the agent advertises none
Also covers the resume path in the regression test.
* test(acp): keep configured-model tests in the ACP agent suite
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(claude): let an agent leave Auto mode when Claude Code uses Bedrock or Vertex
Changing the mode of a Claude agent whose process is not running starts the
process first. It was started in the agent's current mode, so an agent in Auto
on Bedrock or Vertex failed the Auto check before the new mode applied, and
every mode change and message failed the same way.
Start the process in the requested mode instead.
Closes#5878
* fix(claude): relaunch a query whose mode change Claude Code rejected
A mode change that starts the query now launches it in the requested mode. If
Claude Code then rejects the mode, the agent keeps its previous mode, so mark
that query for relaunch instead of reusing it in the rejected mode.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(server): keep a subagent card labeled when two background agents launch together
The daemon's stream coalescer kept only the last update per tool call
inside its 60 ms window. When Claude launches two background Agent calls
in one message, the first child's task_started arrives a few ms after the
second call starts streaming, so its labeled card is buffered and then
overwritten by the unlabeled completed Agent update. The coalescer now
folds updates with the same merge the timeline projection applies, so a
labeled detail is not replaced by an unknown one.
Closes#6252
* refactor(server): compute the merged tool call before building the coalescer entry
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* Keep an archived agent archived when a send to it fails
Sending a message to an archived agent unarchives it before loading the
session. When loading failed, for example because the agent's working
directory was removed, the agent stayed unarchived. The send now restores
the original archive timestamp when it fails, matching resume_agent.
Closes#6245
* Restore the archive only when the send cannot load the agent
Concurrent sends share one load, so a failed load leaves no session for
the restore to close. A failure after the load no longer re-archives an
agent another send may already be running.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* website: match left and right edge padding on ranked plugin rows
The rank sat right-aligned in a fixed 16px column and the row had 8px of
padding, so a single digit started ~15.6px from the row edge while the
install count ended 8px from it. The rank now sits at the row's padding,
which is 16px on both sides, and the list's negative margin matches so
the rank still lines up with the section heading. Tabular figures keep
the tile column aligned without the fixed width.
* website: give plugins section headings room above their content
Section headings sat 16px above their cards while sections sat 56px
apart, so a heading read as closer to the section above it. A shared
PluginSection now owns the header-to-content gap (32px), used by the
index sections and by the browse results under the sort tabs.
* website: center the Build your own plugin card
The heading, description, and buttons were left-aligned in a full-width
card, leaving most of it empty. Center them and keep the description at
its readable measure.
* website: order What's new by first-listing date
newestFirst and addedAgo read submittedAt. The registry publishes the
first-listing date as publishedAt, so What's new now sorts and dates by
it. The fixture's publishedAt values become first-listing dates.
* website: plugin media with video tiles and a click-to-view viewer
The registry's published screenshots field becomes media and may hold
mp4 and webm videos. Cards and link previews use the first image, the
thumbnail route serves only image entries, and the detail gallery shows
videos as tiles alongside images. Any tile opens one viewer. An image
fits the viewport but never exceeds its natural size. A video plays with
controls and scales to the largest size that fits, small sources included.
The daemon resolver now validates only the ID and artifact it installs.
* server: drop the unreleased pluginRegistryEnabled config key
The key only existed in unreleased betas and nothing reads it. The
config schema is strict, so a config that still sets it now fails
validation and names the key.
* website: Featured section on the plugins index
The registry index gains a required featured list of plugin IDs. The
index shows those plugins above What's new, in the registry's order,
skipping IDs that are not listed, and hides the section when none are.
The index sections now share one container for their spacing.
* Show a Claude subagent's handback report as its final message
Since Claude Code 2.1.271, a subagent in auto mode delivers its report
through a SubagentHandback tool call instead of its final text. Paseo
showed that call as a generic tool card with the report as escaped JSON.
Live sidechain frames and replayed history now map the call to the
subagent's final assistant message and hide its acknowledgement.
Closes#6241
* Keep a handback report's leading whitespace and hide its acknowledgement by call id
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* Show an archived ACP agent's history after its worktree is removed
Opening an archived ACP agent loads its history by resuming the provider
session for reading. The provider process was spawned in the agent's
stored working directory, so once that directory was removed (for
example when its worktree workspace was archived) the spawn failed with
ENOENT and the chat opened empty.
A history load now starts the provider process from the home directory
when the working directory is gone, and still names the original
directory in session/load. An interactive resume keeps requiring the
directory.
Closes#5449
* Only treat a missing working directory as gone for a history load
A stat failure other than a missing path, such as a permission error,
now reaches the caller instead of moving the provider to the home
directory.
* Start the history process from home when the stored path is not a directory
A file left where an archived agent's worktree was made the history load
spawn in it and fail with ENOTDIR.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* feat(plugins): enable registry installs by default
* fix(plugins): explain explicit Git source for missing registry IDs
* fix(plugins): tailor missing registry guidance to installation
* test(app): sample footer geometry in one browser frame
* fix(plugins): make registry source resolution deterministic
Directory existence must not change the meaning of a registry identifier shared by CLI, RPC, and UI installs. Classify explicit host paths in the source parser and explain recovery from registry transport and revision-selection failures.
* fix(plugins): recognize dot and Windows directory sources
* test(cli): wait for readable daemon ownership before cancellation
The cancellation test raced PID-file creation against its first write and parsed empty JSON in CI. Use the public daemon-instance reader to wait for and capture the supervisor PID before asserting cancellation cleanup.
* test(app): expect missing directory install guidance
* fix(omp): reopened chats show a compaction row instead of "Unsupported history record"
OMP stores a compaction as a top-level session entry. History replay sent
it to the unknown-record fallback, so after a daemon restart the chat
showed a "Custom message" row reading "[compaction] Unsupported history
record". Replay now maps the entry to a completed compaction row.
Closes#6206
* test(omp): assert on the replayed custom-message rows directly
* test(omp): match only the unsupported-record row after restart
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(acp): terminate the provider when close requests never return
Archive and delete wait on ACPAgentSession.close(). cancel() and
unstable_closeSession() had no timeout. The ACP SDK does not reject a
pending request when the provider never answers, so close() never
reached terminateProcess. The agent was already marked archived and the
provider process stayed alive.
Both calls now use the same 2s withTimeout as probe close. The existing
catch blocks still continue into terminateProcess.
Example:
- Input: archive an ACP session whose cancel and closeSession never settle
- Before: close() waits forever and the provider process stays alive
- After: both calls time out and terminateProcess runs
* test(acp): archive close terminates a real provider that never answers session/close
Spawn a real ACP provider process over stdio that advertises session
close and never replies to it, then close the session the way archive
does. Without the request bound, close() never returns and the provider
keeps running.
Co-authored-by: Ouyang Xingyuan <galaxyroaming@qq.com>
* test(acp): move the close wait and provider cleanup into helpers
Clear the race timer once close settles, and dispose the provider
process and temp directory in afterEach.
---------
Co-authored-by: Ouyang Xingyuan <galaxyroaming@qq.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(speech): publish models only after extraction completes
* fix(speech): reinstall local speech models whose extraction was interrupted
The daemon decided whether a local speech model was installed with its own
check in speech-runtime.ts: every required file exists and is non-empty. A
model whose extraction was stopped part-way passed that check, so the daemon
never called the downloader again and dictation aborted on the truncated
file on every start.
The models module now owns "is this model installed", and the runtime asks
it. A model whose archive is still in .downloads did not finish installing,
since the archive is removed only after the model is in place, so it is
reinstalled from that archive. Extraction goes to a directory beside the
archive and moves into place once complete.
Co-authored-by: 3ae3ae <ogyshjsm@naver.com>
* fix(speech): remove the model archive before moving the model into place
* docs(speech): describe when the model archive is removed
---------
Co-authored-by: 3ae3ae <ogyshjsm@naver.com>
Co-authored-by: Bae Seongmin <83870598+3ae3ae@users.noreply.github.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
After a daemon restart, an OMP chat replays from its session file. Three
kinds of entries that the live timeline never shows were replayed as
collapsed "Custom message" rows:
- the rule-violation reminder OMP adds after a tool call matches one of
its rules, a developer message that opens with
<system-reminder reason="rule_violation" ...>, while only the bare
<system-reminder> tag was hidden
- the ttsr_injection entry OMP records alongside that reminder
- the fileMention message an @-mentioned file adds, and the other OMP
message roles the live timeline drops
Closes#6131
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(opencode): agent goes idle after a v2 tool permission is denied
OpenCode 2.0.20 ends the execution when a tool permission is rejected and
reports the interruption with reason "shutdown", the same reason it uses for
an execution it will resume after a server restart. The turn tracker only
treated "shutdown" as final after a dismissed question, so a denied tool
permission left the agent running with nothing pending, and Stop waited on a
turn that could never complete.
Treat any denied request owned by this session as ending the turn when the
following interruption reports "shutdown". Denials of child-session requests
do not interrupt the parent execution, so they are excluded.
* test(opencode): denying a child session's permission keeps the v2 turn resumable
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(opencode): end the v2 turn when the user dismisses a question
OpenCode ends the execution when a question is dismissed but reports the
interruption with the reason "shutdown", which the turn tracker keeps open
waiting for a resume that never comes. The agent stayed running, and Stop,
reload and send were refused. A question dismissed from Paseo now lets that
interruption end the turn as canceled.
* test(opencode): cover the next message after a dismissed v2 question
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(muse): show usage only for the agent account
Usage discovery ignored session scope and exposed all remembered accounts to unrelated agents. Resolve session accounts from their own launch environment while preserving global discovery.
* test(muse): own the usage scope regression in the plugin suite
* fix(usage): scope agent reports to their own logins
Account grouping also merged fetch inputs and cache entries across agent and host queries, hiding the agent login failure behind another harness. Preserve scoped login sets, collect every host failure with source-owned harness labels, and keep refresh results within their requesting view.
* test(usage): expect harness labels in discovered accounts
* fix(providers): launch plugin CLIs on Windows and tolerate Muse startup
Plugin providers bypassed the core Windows launcher and the plugin host exposed an empty server SDK module. Share the launcher through the SDK and bind it in the host. Give Muse initialization its own deadline because native startup traverses session history before answering MSP.
* fix(providers): terminate Windows launcher process trees
A command-script launcher leaves the CLI behind cmd.exe. Expose forced process-tree termination in the server SDK, use it for Muse and SDK ACP teardown, and bound Muse closing so inherited streams cannot leave callers waiting forever.
* test(muse): keep platform setup outside the teardown assertion
* refactor(plugin): expose process capabilities without quoting helpers
* fix(providers): close process trees on uncorrelated protocol errors
* fix(omp): fail the turn when OMP 18.3 rejects a prompt after acknowledging it
OMP 18.3 acknowledges a prompt, then reports a rejection such as a missing
API key only as a second failed response frame for the same request id. The
JSONL RPC transport dropped any response with no pending request, so the
turn stayed running with no error.
The transport now publishes a response nothing is waiting for like any
other frame, and the OMP runtime reads a failed prompt response as the
prompt_result error later OMP releases send.
* fix(omp): drop prompt results held for an OMP process that was replaced
Request ids restart with each OMP process, so a rejection the old process
reported after its prompt timed out could fail a prompt on the new process
with the same id. Move the real-OMP rejection case into the existing OMP
real E2E suite.
* test(omp): verify late-response routing at the transport boundary
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Mohamed Boudra <boudra.moha@gmail.com>
* fix(server): keep the daemon running when its stdout reader is gone
`paseo daemon run` in the foreground hands its stdout and stderr to the
supervisor. When the program that started it closed its end, as
`paseo daemon run | head -1` does, the supervisor's next write failed
with EPIPE and nothing handled the error. The supervisor exited 1, the
worker lost its parent and exited too, and the CLI reported "Daemon
failed to start".
The supervisor now drops output nobody reads and keeps supervising the
worker. daemon.log still records everything. Other stdout and stderr
errors surface as before.
Closes#6114
* test(server): fail on unexpected daemon.log read errors and check stderr reaches the log
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(omp): finish the turn when OMP's run contains a developer message
OMP 18.5 injects hidden developer messages into a run, such as the
rule-violation reminder after a tool call matches a built-in rule. The
agent_end frame then failed schema validation and was dropped, so the
turn stayed running until it was cancelled.
Closes#6104
* fix(omp): parse every OMP message role so agent_end is never dropped
OMP's RPC AgentMessage also defines pythonExecution, hookMessage,
branchSummary, compactionSummary and fileMention. A prompt that
@-mentions a file puts a fileMention message in agent_end, the frame
failed validation and was dropped, and the turn stayed running.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
OMP reports a prompt it rejects before running the agent, such as one
with no API key for the selected model, as a prompt_result with
agentInvoked false, status error, and the error message. Paseo read only
agentInvoked and ended the turn as completed, so the agent went idle with
no reply and no error. The turn now fails with OMP's message.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
After a restart the parent's timeline is rebuilt from Claude Code's
transcript, and the Task/Agent call went through the generic tool
mapper, which has no subagent branch, so the card replayed as a bare
"Agent" with detail type "unknown". Live, the card is built from the
task protocol's declaration instead.
Replay already reads each Task call's type and description from the
parent transcript for the subagent track. Label the replayed cards
from those same facts, through the builder the live path uses.
The shared tool-call parser is left alone: a completed sub_agent card
from it would replace the action log a foreground child's card
carries live.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(server): name a linked global install as the reason self-update is refused
npm 7 and later print no link flag in `npm ls --json --long`, so a global
install made with `npm install -g <dir>` or `npm link` was never detected as
linked and self-update reported that the daemon was not running from the npm
global install. Detect the link from the install directory being a symlink,
the same way npm decides it.
Closes#6086
* fix(server): fail the self-update inspection when the install path can't be read
An unreadable install path now surfaces as the update error instead of being
treated as a normal install. Tests build real install directories for both
the linked and the normal case.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
When Claude Code calls its EnterPlanMode tool, it reports the new mode
through a system status message with permissionMode "plan". The Claude
provider only read the compacting status from those messages, so the
session kept the previous mode and never emitted mode_changed: the app,
`paseo inspect`, and the stored agent record all kept showing the old
mode.
The provider now records the mode carried by a status message and emits
mode_changed when it differs, sharing the bookkeeping with the init
message so the plan-approval resume mode stays the mode Claude left.
Closes#6063
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(usage): keep details hoverable and restore compact sidebar sheets
Share hover-card mechanics with workspace details so context usage can stay open while users press Refresh. Honor Retry-After for rate-limited usage requests, and let both sidebar entry points share the compact sheet.
* fix(usage): restore settings cog with a fitted popover and compact sheet
Keep shared Usage preferences in the title controls on every surface. Settings rows use the menu surface directly, avoiding the trailing margin and duplicate card border from the page layout.
* fix(app): show account usage before the first agent turn
Keep the context meter available for existing agents without context readings so account usage can be opened before a turn. Show an empty ring and an explicit no-data state until readings arrive.
* fix(usage): preserve keyboard access and hover card positioning
Context details must remain reachable by keyboard after replacing the tooltip. Keep focus within the card, restore side flipping, and dismiss detached cards when the viewport moves. Prune expired cooldowns even after credential rotation.
* chore: remove QA screenshots from the source tree
* fix(app): let hover cards claim Escape before agent shortcuts
Register hover surfaces with the shared overlay keyboard owner without moving or trapping focus on pointer hover. Component-local Escape handling ran after the global interrupt shortcut and cancelled the underlying turn.
* fix(app): reopen focused hover cards with button keys
Keep focused triggers registered with overlay arbitration after dismissal so activation keys are handled before composer shortcuts.
* fix(app): release overlay ownership when hover cards close
Closed triggers handle activation locally, so a dismissed card does not block pane Find. Open cards retain shared Escape arbitration to protect running turns.
* test(app): pin Linux keyboard identity for composer Find
* test(app): synchronize context usage hover journeys
Return focus to the composer before asserting pointer dismissal, wait for the card to reach the viewport, and park the pointer before reloading so it cannot consume the next scripted usage response early.
* fix(opencode): wait for the server to listen before opening the event stream
OpenCode can accept TCP connections before its HTTP handler is ready. Starting the event consumer at spawn lets a retry land in that window and hang without receiving server.connected.
Gate the first request on the listening promise. Use a five-second watchdog before connection readiness and abandon stalled attempts independently of fetch abort, ignoring any late records. Keep the existing thirty-second watchdog after readiness.
* fix(opencode): forward the listening gate from the agent client
The first startup fix passed the listening gate from the server manager, but the agent client factory dropped it before building the consumer. Lab logs showed four generations retrying before listening and then hitting the first-record watchdog.
Forward the gate through that factory. Exercise the client-owned manager with a real spawned process in the existing provider dispatch test, and assert that no event request arrives before listening.
* fix(providers): show custom context in expandable tool rows
Preserve custom message type and data instead of attributing extension context to the assistant. Apply the same mapping in live output and history for OMP and Pi, and honor hidden Pi messages without suppressing structured extension updates.
* fix(server): mark truncated plain-text tool content
Keep the existing timeline content budget while making omitted text visible to readers of extension context and other plain-text tools.
* fix(acp): answer open permission requests when a turn fails
When an ACP prompt failed while one of its session/request_permission
requests was still open, Paseo stopped showing the request but never
answered it, so agents that refuse new prompts until the request is
settled stayed stuck until the agent was reloaded.
A turn that ends without completing now answers its open permission
requests with cancelled, as interrupt and close already do.
Closes#6040
* test(acp): cover permission requests left open by a cancelled turn
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Claude Code answers a background Agent call right away, then streams the
child's frames. Each frame re-emitted the parent's card from the sidechain
tracker after the tool-input cache had dropped the call, replacing the
settled, labeled card with a bare running "Task". The tracker now emits the
parent card only while the parent's tool call is still open.
Closes#6059
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* website: plugin page design mockups (four directions, throwaway routes)
* website: final plugin page mockup (A directory, B browse frame, list cards, D detail)
* website: browse pages use a four-column card grid in the final plugin mockup
* feat(plugins): keep registry installs behind a daemon flag, off by default
Off restores the pre-registry behavior: bare owner/repo is GitHub shorthand and the
daemon never resolves or contacts a registry on install, update preview, or update.
Enable with pluginRegistryEnabled: true or PASEO_PLUGIN_REGISTRY_ENABLED=1.
* feat(website): ship the plugin directory, browse pages, and single-column detail
Directory: count, build/submit links, What's new, the nine categories, and Most installed
with a This week / This month / All time switch. Browse pages at /plugins/all and
/plugins/category/<slug> with a category sidebar, tabs, and a four-column grid; old
/plugins?category= links redirect. Install counts gain daily KV buckets kept 60 days.
* fix(website): keep a README's leading GitHub alert on the plugin page
* feat(website): search plugins from the directory and browse pages
Typing replaces the URL with /plugins/all?q=<term> (or the current category
with ?q=), which renders matching plugins on the server. Matching covers
name, description, ID, and author. Old /plugins?q= links redirect.
* fix(website): keep the directory's ranking window when searching
* fix(acp): show context-window usage for generic ACP agents
Generic ACP providers never populated Paseo's context meter: the standard ACP
`usage_update` notification was discarded by `handleUsageUpdate` (`void
update`), while the meter only reads `lastUsage.contextWindowMaxTokens` /
`contextWindowUsedTokens`, which `usage_updated` stream events write.
Forward `size` and `used` on a `usage_updated` event for the active turn,
guarded by the same validity rules the meter applies so a provider cannot push
a window the meter rejects, and preserve the turn usage already recorded.
Verified with vitest acp-agent.test.ts (106 passed), the server typecheck, and
an isolated daemon run where a dsh agent reported
`lastUsage={contextWindowMaxTokens:1000000,contextWindowUsedTokens:13623}`.
* fix(acp): emit context-window usage only with both meter fields
Require a valid size and used together so a partial usage_update cannot
wipe the composer meter. Drive the tests through sessionUpdate instead
of the private translator.
* fix(pi,omp): open sessions whose model was removed
Importing or reopening a Pi or OMP session failed with
`Model "<id>" not found` when the session's model was no longer
configured, because the resume launch always passed `--model`.
Resume now launches on the session's own model, which Pi and OMP
replace with their default when it is gone, and switches to the
requested model afterwards only when it differs and is available.
Closes#6039
* fix(pi,omp): fall back only when the requested model is unavailable
Check the provider's model list before switching on resume, so a
failed switch to an available model reaches the caller instead of
silently keeping the session's model.
* fix(pi,omp): match a requested provider:id model on resume
Compare the parsed provider and model id against the session and
available models, so a reference written as provider:id switches
instead of being treated as unavailable.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(server): resume with last live mode, not creation-time mode
Provider-initiated mode switches (ACP current_mode_update, in-session
commands like /bypass, permission-driven plan->acceptEdits transitions)
update currentModeId/lastModeId but never reach config.modeId. On
resume, buildConfigOverrides restored the stale creation-time mode and
applyConfiguredOverrides pushed it back via set_mode — silently
downgrading the session (e.g. bypass -> acceptEdits after restart).
- buildConfigOverrides + buildStoredAgentConfig prefer lastModeId
(the last observed live mode) over config.modeId
- applyObservedMode syncs config.modeId on every provider-reported
mode change (mode_changed, model_changed runtimeInfo, and the
refreshSessionState poll) so in-memory reloadAgentSession carries
the live mode as well
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(server): assert config.modeId sync on mode_changed/model_changed events
Extends the config-drift stream test with a registry so both the
in-memory agent.config.modeId and the persisted record.config.modeId
are verified — covering the event paths Greptile flagged as untested.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(server): report an unknown agent as not found to CLI commands
fetch_agent_request answered an unknown agent ID with an error string, so
DaemonClient.fetchAgent threw instead of returning null. Every CLI command
that looks up one agent checks for null to report AGENT_NOT_FOUND, so
delete, stop, inspect, logs and the rest reported their generic failure
code and skipped the "paseo ls" hint.
The daemon now answers an unknown agent with a null agent and no error.
An empty or ambiguous identifier is still an error.
* test(cli): expect the not-found message agent logs now prints
* test(cli): move the unknown agent ID cases into the delete and stop suites
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Directory search shares two filesystem slots across all searches. Searches
over one tree walk it in the same order, so two Add Project searches that
reach the same folder whose read never returns held both slots, and every
later directory search queued behind them: command center file search,
@-file autocomplete, and file-link lookups in unrelated folders all hung
until the read returned.
Searches waiting on the same path now share its pending filesystem call, so
one stuck folder holds one slot and the other stays free.
Refs #4879
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(server): keep workspaces available after setup config errors
* refactor(server): drop the archive dependency from worktree setup
Setup no longer archives a workspace when it fails, so the setup
workflows no longer take archiveWorkspaceRecord.
Co-authored-by: Bae Seongmin <83870598+3ae3ae@users.noreply.github.com>
* test(server): stop passing the removed archive dependency to the worktree workflow
* test(server): check paseo.json survives without comparing line endings
Git checks the fixture out with CRLF on Windows, so comparing its exact
contents failed there.
---------
Co-authored-by: 3ae3ae <ogyshjsm@naver.com>
Co-authored-by: Bae Seongmin <83870598+3ae3ae@users.noreply.github.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Mohamed Boudra