* feat(server): let clients create internal agents
The daemon's own branch-name and commit-message generators run as
internal agents: never persisted, listed, flagged, announced, or seen by
plugin lifecycle hooks. Nothing outside the daemon could ask for one,
because the wire config schema strips `internal` and the agent.create
transform re-pins it. A plugin that runs a one-turn helper per event
therefore left an archived session in History every time.
Add an optional `internal` flag to `create_agent_request` (inherited by
`agent.create.request`), copy it onto the config in the session's create
path, and skip provider session persistence for it like the generators
do. The client SDK forwards the flag on both creation paths and gates it
once on `server_info.features.internalAgents`.
Two daemon paths assumed every agent is public and had to follow:
`resolveAgentIdentifier` now accepts a live agent by exact id, since the
id set it matched against came from listings that exclude internal
agents and left `wait_for_finish` answering "Agent not found"; and
archiving an internal agent now closes it without writing the record
that `registry.applySnapshot` used to leave behind, so `autoArchive`
does not accumulate hidden files on disk.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* fix(server): keep internal agents whole through failure, archive, and wait
Review follow-ups for internal agents:
- A create that fails after the runtime is registered now closes the
internal agent before rethrowing, so a hidden runtime cannot outlive
the failed request. Public agents keep today's behaviour.
- Archiving an internal agent cascades to attached children like any
parent. cascadeArchiveChildren takes the parent record instead of
reading it back from storage, where an internal agent never is.
- The archived snapshot and last message of an internal agent stay
readable by exact id for ten minutes, so create-then-waitForFinish
returns the finished turn whether the wait lands before or after an
auto-archive. Storage plays that role for public agents.
- The internal-agent e2e scenarios live in the existing
create_agent_request lifecycle suite rather than a new file.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* fix(server): cascade to live internal children and keep a failed internal create's worktree until it closes
Internal children never reach storage, so the parent archive cascade scans
live agents for them as well. A failed internal create now clears the
created id only after the runtime is archived, so worktree cleanup cannot
delete the directory under a runtime that would not close. Stored agent
payloads carry lastError, so a retained failed helper reports its error
through wait_for_finish instead of the generic message.
* feat(cli,mcp): expose internal agents and document --no-wait
paseo run gains --internal and create_agent gains internal, both forwarded
to the create request. --no-wait becomes the documented spelling for
returning immediately, matching paseo send; --background and --detach stay
as hidden aliases under COMPAT tags.
* test: follow the --no-wait rename and stub getRetiredInternalAgent
* feat: list internal agents on request and add internal workspaces
Internal agents stay hidden by default, but a caller can now ask for
them: `fetch_agents_request` takes `filter.includeInternal`, snapshots
carry `internal: true`, and `paseo ls --internal` shows them in an
INTERNAL column. The app never passes the flag.
Workspaces get the same flag. `workspace.create.request` with
`internal: true` persists a hidden workspace: listings, update
subscriptions, sync reads, empty-project counting, open-by-path, MCP
`list_workspaces`, and plugin workspace hooks all skip it unless the
caller opts in with `filter.includeInternal`. Every agent created inside
an internal workspace is internal and unpersisted; AgentManager applies
that from the workspace id before the plugin hook so creators cannot
opt out. Archive and shutdown enumerate internal agents so nothing
inside an internal workspace outlives it.
Surfaces: `paseo workspace create --internal`, `paseo workspace ls
--internal`, `paseo run --internal --new-workspace`, MCP
`create_workspace({ internal })`, SDK `workspaces.create({ internal })`
and `list({ filter: { includeInternal } })`. Clients gate on
`features.internalWorkspaces`.
Co-Authored-By: Claude <noreply@anthropic.com>
* feat(workspaces): add background visibility with inherited creation defaults
* fix(cli): resolve workspace callers on the selected daemon
* test(workspaces): keep caller integration coverage in server suite
* test(cli): exercise workspace callers through CLI integration harness
* test(mcp): remove obsolete execution flags from browser fixtures
---------
Co-authored-by: marvin-ambrozio <marvin@gustavo.eng.br>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: Mohamed Boudra <boudra.moha@gmail.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>
* fix(app): subscribe only the chats this session opened
Open-chat membership was sourced from the persisted workspace layout, which is
restored from disk at launch and carries a tab for every chat the user has ever
opened on the host. Every member got an authoritative timeline catch-up at
session start, and each of those fetches resumed its agent on the daemon: one
provider session per historical tab, every launch.
Two things followed from each resume. Registration stamped `updatedAt` over the
restored timestamp, and the resumed agent started with no attention because the
resume path never read the stored flag. Both were persisted, so restarting Paseo
rewrote every workspace's last-used time in the sidebar and cleared Ready to
review on chats the user never looked at. The timestamp loss was permanent:
workspace `statusEnteredAt` is re-derived from `updatedAt` on every daemon start.
Keep the open set on the session instead. A chat joins it the first time the user
opens it and stays until its tab closes or the session ends, so hiding, switching
workspaces, evicting a retained view, backgrounding and reconnect all preserve the
subscription, with no recent-agent limit. Layout keeps its release role and loses
its source role.
Harden the producer as well, so a resume from any caller stays inert: registration
adopts restored timestamps rather than stamping now, and resume carries the stored
attention through.
* fix(agents): keep a resumed agent's updatedAt from moving backwards
Dropping the registration stamp exposed a skew the stamp had been erasing.
`extractTimestamps` read `lastActivityAt ?? updatedAt`, but the two diverge:
renaming, labelling, restoring from archive, marking unread, and clearing
attention all move `updatedAt` on an unloaded agent without touching
`lastActivityAt`. Eight of 854 agent records on one dev machine carry that skew,
up to eight days of it.
A resume then handed clients a timestamp older than one they already held, and
`acceptAgentDirectoryUpdate` drops any state update that goes backwards. Marking
a chat unread without opening it, then opening it, left the unread dot stuck in
History, the agent list, Command Center and the tab: the daemon cleared attention
and the client refused the update.
Read both fields and take the newest, which is already the rule
`resolveStoredAgentPayloadUpdatedAt` applies to the list payload. Move that
function beside the other stored-record readers as `resolveStoredAgentUpdatedAt`
so one rule covers both paths.
Also replace the `!options?.updatedAt` inference in `registerSession` with an
explicit `restoring` flag. The inference was correct for all four call sites, but
it made a second meaning out of a field, and the next caller to seed a timestamp
would have opted out of activity stamping silently.
* test(app): assert a reloaded layout subscribes only the chat it restores into
No e2e reloaded the page with a populated workspace layout and asserted which
agents the daemon acknowledged, so the regression this branch fixes had no
end-to-end guard. Opening four chats across two workspaces leaves a tab per chat
in persisted layout; after a reload only the restored chat may be subscribed,
because subscribing resumes that agent on the daemon.
Verified it fails when session-context's layout callback is pointed back at
replaceVisibleAgentIds: the three restored tabs nobody opened join the
acknowledged set.
* fix(agents): restore archived Codex workspaces without losing focus
History selection must survive without cached agent records, and reading an archived Codex thread must not acquire an interactive writer. Keep workspace restore explicit, serialize native lifecycle transitions, and leave transcript lifetime with its owner.
* test(app): express Codex restore specs as user journeys
* test(app): finish restore fixture cleanup after teardown errors
* test(app): align restore coverage and await fixture readiness
Workspace restore leaves agents archived until explicit Unarchive. Keep that distinction covered for multiple agents. Compile the recovery fixture before its interaction test starts, and wait for sidebar agent hydration before asserting tab presence.
Restore unread attention on finished workspaces and preserve the reminder when leaving the current workspace. Cover desktop and compact read flows, and retain change-only file paths for deletion observation.
Co-authored-by: Edi Hasaj <34984294+edihasaj@users.noreply.github.com>
* fix(server): keep agent reloads from overlapping session writers
Idle runtimes can still own their persisted session writer. Release the original runtime before resuming and retain failed cleanup for retry; a timeout must not authorize another writer.
* fix(server): honor closes queued during agent reload
Serialize public close with reload so a completed close cannot be undone by replacement registration. Keep archive cleanup inside its existing lifecycle operation.
* fix(subagents): preserve recursive provider ownership
Provider events retained enough lineage to identify nested work, but the canonical descriptor discarded the direct owner. Preserve that relationship at ingestion so child timelines and hierarchy remain attached to the process that emitted them, including after history replay.
* test(e2e): clarify nested ownership journey
* test(e2e): separate nested ownership actions
* fix(subagents): retain nested ownership across replay
Wait and interaction events can reference descendants without spawning them. Preserve the original spawning owner and route persisted background notifications through that same relationship after history reload.
* test(subagents): hide provider turn plumbing
* fix(subagents): scope nested tracks to their host
Plugin-provided icons are host-specific. Carry the panel host through nested child tracks so the rebased ownership UI resolves the same provider presentation as its parent.
* fix(workspaces): preserve launched terminal focus
* fix(server): recover already-settled interrupts
Provider state can drift from Paseo’s tracked run. Treat an authoritative already-idle interrupt as settled and retire the captured run token so retrying can start new work without weakening ambiguous-cancellation protection.
* fix(codex): fence interruption during turn startup
Reserve foreground-start ownership before asynchronous provider setup. An interrupt can now cancel preflight before turn/start or wait for an issued start to become interruptible, preventing acknowledged cancellation from leaving unmanaged work.
Codex threads can be archived outside Paseo while their Paseo agent remains active. Repair that provider-state drift during interactive resume while preserving the existing read-only history path for Paseo-archived agents.
A resumed native child can be announced under a fresh tool call while retaining its session-scoped task identity. Treat later tool-call identifiers as aliases so lifecycle, usage, and timeline events continue through the original descriptor instead of splitting into contradictory rows.
The subagents track and the task list were stacked cards above the
composer. Each one pushed the composer further down the pane and painted
an opaque band the timeline stopped at, and both spent a full row on a
count that fits in a pill.
Both are now pills that open a panel — a popover on wide screens, a sheet
on compact ones, from the existing menu engine rather than a fourth
floating-panel implementation. The row frame is shared, so a subagent row
and a task row sit on the same rail whatever their contents.
The bar mounts inside the transcript's animated container and pins itself
to the bottom edge, so scrolled content passes under the pills instead of
stopping above them. It stays a child of that container rather than a
sibling floating over the composer: on Android a view outside its parent's
bounds paints but takes no touches, and every pill is pressable. The
container carries the same keyboard transform the composer does, so the
pills track the composer's top edge when the keyboard moves.
The tracker state moves out of the composer with it. None of it was
composer state; it lived there because that is where it used to render.
* feat(app): archive finished subagents together
Make the track action clean up finished managed and provider-owned children as one operation while preserving running work. Managed archives remain sequential and provider-owned dismissal stays client-local.
* test(app): clarify subagent archive assertions
* fix(subagents): preserve opened children on parent archive
Opening a managed child represents user ownership only while that client's tab remains open. Track that intent per client, detach surviving children during parent archive, and keep single and bulk tab-close behavior consistent.
* fix(subagents): mark generic tab openings
Centralize managed child tab ownership in agent navigation so list, command-center, deep-link, notification, and subagent-track openings all mark ownership before creating the tab. Use stable localized errors for lifecycle failures.
* fix(subagents): classify cold tab openings
Fetch uncached agents before deciding whether tab ownership must be recorded, so notification, desktop, and cold-route openings cannot bypass parent archive protection.
* fix(app): translate navigation fallback errors
* fix(navigation): allow offline workspace intents
* refactor(app): derive subagent ownership from open tabs
* fix(server): serialize subagent lifecycle ownership
* fix(agents): preserve permission notification callbacks
Permission requests consumed the one-shot completion watcher, so callers never heard about later approvals or the final result. Keep the watcher alive across permission checkpoints and include the actionable request payload in each notification.
* fix(agents): notify concurrent permission requests
* feat(server): show provider workflows in subagent track
Normalize workflow tasks into the provider-subagent contract so shared clients need no workflow awareness. Persisted nonterminal runs replay as failed because their original runtime can no longer deliver a terminal update.
* fix(server): preserve workflow replay and tool cards
* fix(server): suppress workflow notification cards
* fix(server): replay workflow child timelines
* fix(server): fail tasks when Claude exits
* fix(server): preserve workflow replay parity
* fix(server): fail tasks on Claude restart
* fix(server): hide workflow child prompts on replay
* fix: terminalize workflows on session reload
* fix(server): own workflow notification routing
* feat(server): show workflow results in detail pane
* fix(server): preserve unrestored task notifications
* fix(server): normalize replay task notifications
* fix(server): report Claude runtime death instead of sitting idle
When the Claude Code CLI process dies on its own — crash, OOM kill, host
suspend — nothing in the daemon observes it unless a turn happens to be in
flight. The query pump only reports a failure while it is draining a turn, so
an idle agent's stream just ends: the agent stays `idle`, indistinguishable
from a healthy resumable one.
Everything the agent parked inside that process dies with it. Background Bash
shells, `Monitor` watches and workflows all live in the CLI process, so the
completion notification that would have woken the agent back up never arrives.
The work is silently gone and the user finds out much later, if at all.
The stale handles are never dropped either, so writes keep going to a dead
transport for as long as the daemon lives (#2472).
Watch the spawned child at the single seam it is already handed over
(`onChildProcess`). When it exits between turns, drop the dead query/input
handles so the next write spawns a fresh process, and emit `turn_failed` —
the manager already routes a non-foreground `turn_failed` to `error` plus a
timeline entry.
A death mid-turn is left alone. The pump already waits for stderr to flush and
reports the real cause; reporting here first would replace that with a bare
exit code.
Intentional teardown stays quiet. All three sites that retire a process detach
the child handle first: `close()` sets `closed`, and `ensureQuery()` and
`handleMissingResumedConversation()` null `childProcess` before ending stdin.
Fixes#2472
* fix(server): tree-kill the retired process on missing-resume recovery
The missing-resumed-conversation path ends the child's stdin but never
tree-killed it, so MCP children of the retired claude process could outlive
it as orphans. Every other teardown path already does this; do it here too.
Reported by Greptile on #2910.
* fix(server): include native subagents in workspace activity
Provider-native children live outside managed agent snapshots, so background work could outlive an idle parent without keeping the workspace running. Aggregate running child descriptors into the owning workspace and publish workspace updates on child lifecycle changes.
* fix(server): serialize workspace status updates
Queue overlapping descriptor rebuilds per workspace so older activity snapshots cannot overwrite newer lifecycle state. Preserve batched registry reads for multi-workspace updates.
* test(server): update workspace status fixtures
Provide the native subagent activity dependency in the shared session fixture and wait on observable git-watch output across the serialized update boundary.
* fix(server): drop queued workspace updates on subscription swap
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Mohamed Boudra <boudra.moha@gmail.com>
* refactor(claude): derive subagents from the SDK task protocol
Claude Code announces subagent identity and lifecycle on the SDK stream
(task_started, task_updated, task_notification). Paseo consumed only
task_notification and reconstructed the rest from sidechain frames, in two
independent derivations that had drifted apart.
Both paths now translate into one small observation vocabulary and fold it
with one function, so identity and status are derived once. The fold is
stateless: the task protocol is edge-triggered and the store's sticky merge
already accumulates.
- Replay links a subagent through the agent-<id>.meta.json sidecar, which
carries the same Task tool_use id the live stream keys on. The agentId
text scrape stays as a fallback for sessions recorded without it.
- The sidechain tracker keeps mapping frames to timeline items but no longer
owns the descriptor when the task protocol is active, detected from the
session rather than from a version string.
- Adds replay coverage, which did not exist.
Subagent ids are unchanged, so no protocol, store, or client change.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(claude): keep non-subagent tasks out of the track, open the child timeline with its prompt
Backgrounded shell commands announce as task_started with a tool_use_id but
task_type local_bash and no subagent_type, so filtering on the id alone put
`sleep 20` in the subagents track. Discriminate on task_type, falling back to
subagent_type for releases that predate it.
task_started also carries the prompt. Emitting it as the child's first timeline
item means the pane no longer opens mid-conversation, and matches replay, where
the sidechain transcript already begins with the prompt.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(claude): drop status for tasks that were never declared as subagents
task_notification also fires for backgrounded shell commands and carries a
tool_use_id. Routing status off that id readmitted exactly what the declaration
filter rejected, as a descriptor with a status and no identity — rendered in the
track as a nameless "Loading..." row.
Status now routes only through the task_id map populated at declaration.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(app): name a subagent row after its task, not its type
The descriptor carries the subagent type in title and the task in description.
The client collapsed them and rendered only the type, so every child in a
fan-out read "general-purpose". The task is what tells siblings apart, so it
names the row and the type moves to the muted secondary line, which the track
computed but never drew.
Providers that report no task keep the previous behavior.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(claude): label the parent's Task card from the subagent declaration
A backgrounded Task subagent emits no frames carrying parent_tool_use_id, so
the sidechain tracker never runs for one and never supplies the sub_agent
detail. Its card in the parent transcript rendered as a bare "Task".
The declaration already carries the type and the task, so the card is built
from it. For a foreground subagent the tracker still emits the same card
enriched with the child's action log; the two collapse on the shared tool-call
id, so that case is unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat(subagents): surface a subagent's model, effort, and cost
Adds model, effort and usage to the provider subagent descriptor, populated
from observations the child itself reports.
- model: read per assistant frame, live and on replay. Never snapshotted, since
a provider can swap models mid-flight on overload or refusal fallback, and
never inherited from the parent.
- effort: replay only. The SDK's assistant message does not carry it and it can
be silently downgraded for the chosen model, so a live value would be a guess.
- usage: task_progress reports tokens and tool count once per tool use, and
task_notification reports the terminal totals. This matters most for
backgrounded subagents, which emit no sidechain frames and would otherwise
show no sign of activity at all.
Status becomes optional on the upsert event: a report that only carries a model
or a cost says nothing about whether the child is still running, and must not
revert a finished one. Usage merges rather than replaces, because duration only
arrives at the end.
Fields are optional on the wire, so an older daemon simply omits them and the
client renders nothing rather than substituting the parent's settings.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat(claude): read subagent effort from hooks
Effort appears nowhere on the SDK message stream — verified by scanning every
message type at depth — and the level Paseo requests is not necessarily the one
that runs, since a model that does not support it is silently downgraded.
Hooks are the source. Every hook input carries agent_id and effort.level, the
active level for that turn after any downgrade, and agent_id is the same id
task_started calls task_id, so it routes through the table already built at
declaration. PreToolUse, PostToolUse and SubagentStop are registered purely to
observe; they record what they see and always return empty, so they cannot
affect tool execution or turn control.
Paseo's hooks merge with any the user configured rather than replacing them:
extraClaudeOptions is spread after the base options, so assigning would have
silently stopped effort being observed for anyone with their own hooks.
Verified end to end through ClaudeAgentClient against a real session: a
subagent descriptor now carries model, effort and usage together.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat(subagents): show a subagent's runtime facts in its pane, and its cost on replay
The track row carried model, effort and cost; the pane carried none of it, so
opening a subagent showed only its task.
The pane gets a header strip rather than a longer tab subtitle: the subtitle
feeds the width-constrained tab strip, whose job is to name the task. Duration
appears in the pane, which has room, and stays out of the dense row.
Replay had no usage source, so a reopened agent lost its subagents' token and
tool counts permanently — descriptors are in-memory and cleared when an agent
closes. Usage is now derived from the child's own transcript, matching the live
producer exactly: total tokens are the four counters summed on the last
assistant turn (a context reading, not a cumulative sum, which would multiply
the cached prefix by the turn count), and tool uses are tool_use blocks across
assistant entries. Duration is an approximation — the transcript's first entry
is written after the task starts — and is the one field replay cannot match.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* docs(claude): record the task protocol's undocumented wire facts
The subagent sources depend on Claude Code internals the SDK types do not
explain: which announced tasks are subagents, why a status for an undeclared
task is dropped, why task ids outlive a turn, why effort is only reachable
through hooks, and why a replayed session must skip its grandchildren.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Generalize provider subagent presentation
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Mohamed Boudra <boudra.moha@gmail.com>
* fix(server): keep idle agents resident
Remove time-based runtime collection so background work and subsequent prompts are not destroyed during idle periods. Runtimes now close only through explicit lifecycle actions.
* fix(server): preserve explicit close coordination
* fix(server): support partial agent manager adapters
* fix(server): keep active agent trees resident
Idle collection treated an inactive parent turn as process quiescence even when descendant work was still running. Protect managed and provider-owned subagent trees, carry descendant activity into the idle window, and use a conservative 30-minute fallback.
* fix(server): close idle collection races
Retain managed ancestry across closed intermediates, refresh descendant state after each awaited close, and invalidate running provider subagents when their runtime is terminated.
* fix(server): serialize agent tree collection
Keep ancestors resident until managed descendants are collected, serialize descendant registration with idle collection, and preserve provider subagent state across hot reloads.
* fix(server): retain closed descendant activity
Carry persisted descendant activity through runtime closure, avoid error children pinning parents, and cancel stale provider children when reload replacement aborts.
* test(server): make descendant expiry deterministic
* fix(server): scope idle child protection
* fix(server): tighten idle tree coordination
* fix(server): reduce aggressive idle cleanup
Keep the mitigation conservative: extend the idle timeout without inferring tree liveness from incomplete lifecycle signals.
* fix(server): keep parents alive during child work
Idle cleanup now respects running managed and provider-native children while retaining a conservative 30-minute fallback.
* fix(server): clear running child state on close
* fix(server): drain child events before close
* fix(app): restore archived agents from History
History navigation lost the explicit archived-agent intent during tab reconciliation, and workspace recovery stopped after restoring the workspace. Preserve the selected tab and recover its provider session as one action while serializing provider resume with timeline hydration.
* fix(app): preserve archived agent recovery invariants
* fix(server): preserve timeline hydration call shape
* fix(app): retain agent tabs through lookup
* fix(server): upgrade in-flight timeline broadcast
* fix(server): release resources held by idle agents
Keep unarchived agents resumable while closing their provider runtimes after two minutes. Active agent schedules keep runtimes resident.
* fix(server): preserve resumable agent state
* fix(server): resume agents when listing commands
* fix(server): preserve collected agent interactions
Pi and OMP share only the JSONL child-process transport while retaining provider-owned launch, RPC, runtime, session, history, and permission behavior. Removes captured fixtures in favor of typed harnesses and real-provider coverage.
Closes#2006Closes#2060
* fix(app): restore missing workspaces from History
Workspace and agent archival are separate lifecycles. Carry restore intent from History so closed agents can reopen archived workspaces without changing ordinary navigation.
* fix(app): wait for workspace hydration before restore
* fix(app): preserve History agent unarchive
* fix(app): reopen archived History agents outside active directory
History entries are not merged into the active session maps. Carry the row archive state through explicit restore intent so live workspaces still reopen without broadening ordinary navigation.
* fix(app): ignore stale History archive state for active agents
A cached History row can outlive a successful reopen. When the workspace exists, live session state now wins so an active agent is not interrupted by another refresh.
* fix(app): reopen every selected archived History agent
History entries without workspace IDs returned before restoration, and a second archived agent sharing an in-flight workspace restore was dropped. Preserve both agent-only and deferred reopen intent.
* fix(app): require explicit workspace recovery
* fix: preserve workspace recovery contracts
* fix: preserve workspace recovery compatibility
* fix: preserve recovered workspace session state
* fix: preserve terminal navigation timing
* fix(server): preserve successful workspace recovery
* fix: preserve workspace recovery intent
* feat(subagents): archive finished children and show native names
Keep native provider sessions intact while letting users clear completed children from the track. Persist dismissed provider child IDs so history replay does not restore them.
* fix(subagents): cover persistence and naming edge cases
* fix(subagents): preserve active descendants and native names
* fix(subagents): retain dismissed timelines and trim paths
* fix(subagents): retain history across hydration
* test(subagents): expect normalized native descriptions
* fix(subagents): separate hidden rows from retained history
* fix(subagents): preserve restored hidden tabs safely
* fix(subagents): integrate archive with run state
* test(codex): expect normalized child name
* fix(subagents): retain dismissed tabs on reload
* fix(subagents): hide finished native agents locally
Keep the header action presentation-only so it does not mutate agent lifecycle or expand the runtime protocol.
* fix(subagents): humanize native agent names
* fix(subagents): preserve names in Codex history
* fix(subagents): retain hidden rows across reloads
* fix(server): prevent hidden Codex subagents and stuck turns
Codex can announce native children only through its mirrored lifecycle stream, while rejected interrupts previously looked like successful local cancellation. Preserve those child announcements and keep manager state active until the provider acknowledges cancellation.
* fix(server): block actions after rejected cancellation
A provider can accept a turn before publishing the turn ID needed to interrupt it. Keep that interval non-cancelable, and prevent reload, replacement, or rewind from proceeding without an acknowledged cancellation.
* fix(server): surface rejected agent cancellations
Replacement prompts and Stop requests could appear accepted after the provider kept ownership of the active turn. Complete replacement cancellation before detaching the stream, and return cancellation failures through the client response.
* fix(server): handle turn completion during cancellation
* fix: finish cancellation lifecycle handling
* fix(server): close remaining cancellation races
* fix(server): settle autonomous cancellations
* fix: honor cancellation failures at call sites
* refactor(server): centralize agent run state
* fix(server): keep pending runs isolated from stale events
* fix: surface remaining cancellation failures
* fix(server): narrow cancellation error handling
* fix(workspaces): show provider subagent activity
Provider subagent lifecycle lives outside managed agent snapshots, so workspace aggregation dropped it after the parent turn finished. Count running provider subagents against the delegation root and emit workspace updates when their status changes.
* fix(workspaces): surface cross-workspace subagents
Require relationship and workspace on create_agent, and reuse the same worktree target union for create_worktree. Agent-scoped prompt follow-ups now default to background finish notifications so callers can continue without polling.
Multiple workspaces can share one checkout directory. Agent and terminal
status, sidebar tabs, and ownership now attribute strictly by workspace ID,
so a running or blocked agent in one workspace no longer shows up on its
same-directory siblings.
A one-time startup migration backfills workspace IDs onto legacy agents, and
every agent and terminal is created with one thereafter. Directory lookups
remain only for genuinely folder-derived git facts (branch, diff) and the
archive-by-path adapter.
* Make workspace IDs opaque, independent of the filesystem path
Workspace IDs were the resolved checkout/worktree path, so code could
treat an ID as a path: prefix matching, deriving directories from it,
falling back to a path when a lookup missed. IDs are now opaque -
compared only by exact equality and never used as a path. Anything
path-shaped resolves the workspace record first and reads its cwd.
New workspaces get a generated `wks_` ID; existing path-shaped IDs are
read from disk and never regenerated, so there is no migration and no
change to the wire or persisted schemas. Groundwork for running
multiple workspaces in a single directory.
* Keep attachment scope key stable and align refetch test
The attachment scope key keeps `workspace=` instead of `wsid=` so existing
persisted drafts are not orphaned; the rename was cosmetic. The SDK refetch
test no longer asserts a filter, matching refetch fetching one page and
selecting by id client-side.
* Fix live agent updates in directories without a registered workspace
The opaque-ID work removed buildProjectPlacementForCwd's directory
fallback, so an agent running in a folder with no registered workspace
(e.g. a fresh non-git dir) produced a null placement. forwardAgentUpdate
then threw "Workspace not found", the error was swallowed by its catch,
and no agent_update was emitted — live model/thinking switches and status
updates silently stopped. Caught by the live-preferences e2e suite.
The fallback builds a directory-scoped project placement keyed by the
path. That key is a project grouping key (non-git projects group by
path), not a workspace id, so it stays within the opaque-id rule.
* Always run server worktree archive, even without a resolved workspace
Archiving bailed out entirely when the workspace was not found in the
client store, so a race or stale state could make "archive" do nothing
server-side with only a console.warn. The server archive is keyed by
worktreePath, which is always available, so it now runs regardless; only
the optimistic client-side updates (keyed by workspace id) are gated on
the workspace being resolved.
* Expect a directory-scoped placement for unregistered agent dirs
This unit test asserted the no-placement behavior reverted in the live
agent-update fix, which had broken live model/thinking switching. Update
it to expect the directory-scoped placement now emitted for an agent in a
directory with no registered workspace.
* Fix opaque workspace routing in app E2E
* Preserve workspace IDs during partial bootstrap
* Fix archive flows for opaque workspace IDs
* Stop treating opaque workspace IDs as filesystem paths
Workspace IDs are opaque (wks_<hex>), but several call sites still passed
the id where a directory was expected, which broke those flows for opaque
IDs. The branch switcher sent the id as a cwd to branch/stash/checkout git
operations, and server archive/reconcile cleanup keyed git-watch and
subscription teardown by id, leaking that state.
Git and filesystem operations now take the workspace directory; the opaque
id is used only for identity and cache keys. Archive/reconcile cleanup
routes through a single teardownArchivedWorkspace helper that keeps the key
split explicit: runtime store by id, git watch and subscription by cwd.
Path-derived grouping keys are renamed to directory keys, and the helpers
are split into workspace-identity and workspace-directory so the id-vs-path
boundary is obvious.
* Surface paseo subagents in a collapsible section above the composer
Subagents no longer auto-open as tabs. Each parent agent's pane shows
a compact tab attached to the top of the composer, listing its direct
children with name and status; clicking a row opens or focuses the
subagent's tab. The section is provider-neutral so Claude Code Task
and Codex subagent sources can feed the same UI later. Reconciliation
now separates pruning (`activeAgentIds`) from auto-opening
(`autoOpenAgentIds`) so manually opened subagent tabs survive.
* Tighten subagents composer code from review
Collapse duplicate helper in workspace reconciliation, swap a hand-rolled
shallow comparator for fast-deep-equal (matching repo convention), drop
an unused React default import and a single-use formatCount helper, and
extract a shared PaneOverrides type in the agent-panel test.
* Fix oxlint failures in subagents code
Switch type aliases to interfaces, hoist inline JSX styles/handlers in
SubagentsSection to satisfy react-perf rules, and lift the test
makeAgent helpers onto AGENT_DEFAULTS so their cyclomatic complexity
drops below the configured cap.
* Add missing unistyles mock exports after rebase
After rebasing onto main, SubagentsSection now imports both
useUnistyles and (transitively, via WorkspaceTabIcon) withUnistyles.
Two test mocks needed updating to expose them.
Test drift fix: the implementation contract is correct; the local
react-native-unistyles mocks were stale.
* Tuck subagents tab under the composer
Subagents tab now grows visually from beneath the composer's rounded
top edge: the surface overlaps the composer by the corner radius so
the rounded curve nests into the tab. Composer's top padding is
removed so the tab sits flush against the input box with no
click-dead gutter. The tab is full-width in both states, the header
label drops the medium weight to follow design-system §3 (content
text inside a surface is normal), and hover covers the entire
visible tab when collapsed but stays scoped to the header row when
expanded.
* Decouple subagent close-tab from archive
Closing a subagent's tab no longer archives the agent — the tab is
removed from the per-client layout, the agent stays in the parent's
track until explicitly archived. Track rows gain an archive button (X)
to make the lifecycle gesture explicit. AgentManager.archiveAgent now
cascades to children carrying the paseo.parent-agent-id label so
subagent fleets don't outlive their orchestrator.
Trade-offs documented in docs/agent-lifecycle.md, including the known
limitation that handoff agents launched via the same MCP path get
cascade-archived alongside true subagents until a richer relation
model lands.
* Stabilize subagent archive button hover
Apply the sidebar workspace row pattern to fix two issues with the
new archive button: the row no longer drops hover when the pointer
moves onto the button, and the button slot reserves layout space
when invisible so adding/removing it doesn't shift the row.
The wrapper View tracks hover via onPointerEnter/onPointerLeave
(workspace rows do the same), and the slot stays mounted with
opacity 0 + pointerEvents="none" when the button is hidden. Also
swap the X icon for the Archive icon to match the gesture's intent.
* Match design language for subagent archive button
Hover state now changes the icon color from foregroundMuted to
foreground, matching the project-row trailing-action pattern in
sidebar-workspace-list.tsx. Drops the surface3 background hover
that didn't match the design language.
Adds a tooltip ("Archive subagent") on the trigger using the
canonical Tooltip + TooltipTrigger asChild pattern. Disabled when
the slot is hidden so it never appears unbidden.
* Slot subagents into one module + harden cascade
Reshape the subagents-in-composer feature so it lives behind one
module and integrates at a few explicit sites instead of being
smeared across generic files.
- Extract `PARENT_AGENT_ID_LABEL` to `packages/server/src/shared/agent-labels.ts`. Sweeps 16 literal sites across server, client, CLI, and tests.
- `packages/app/src/subagents/` is now a deep module: `index.ts` is the only React entry (`SubagentsSection`, `useArchiveSubagent`, `useSubagentsForParent`, `selectSubagentsForParent`, `resolveCloseAgentTabPolicy`, `shouldAutoOpenAgentTab`); `policies.ts` is a second designed entry for non-RN data consumers. Internal files renamed (`section.tsx`, `select.ts`, `close-tab-policy.ts`, `auto-open-tab-policy.ts`, `use-archive-subagent.ts`).
- Move the 22-line subagent archive flow out of `agent-panel.tsx` into `useArchiveSubagent`. Agent-panel becomes a one-line consumer.
- Move the close-tab subagent branch out of `workspace-screen.tsx` into `resolveCloseAgentTabPolicy` (discriminated union).
- Move the auto-open exclusion rule out of `workspace-agent-visibility.ts` into `shouldAutoOpenAgentTab`.
- Theme chevrons in `section.tsx` via `withUnistyles`; drop the forbidden `useUnistyles()` call from the hot path.
Server-side cascade hardening:
- Single write path: extract `markRecordArchived` and route both `archiveAgent` and the off-memory cascade branch through it. Off-memory cascaded children now notify subscribers (previously silent).
- Cascade no longer swallows child failures with `try/catch warn`. Failures propagate.
- Cascade test extended from 1 weak assertion to 5 contract tests: full archive shape (archivedAt, normalized lastStatus, requiresAttention=false), running child runtime stop, off-memory branch, subscriber notification for in-memory and off-memory children, partial-failure surfacing.
Net diff vs the start of this PR: -1037 / +444 lines. The feature now slots into the repo at four explicit sites: `agent-panel.tsx` (3 imports + a hook + the section), `workspace-screen.tsx` (one policy import), `workspace-agent-visibility.ts` (one predicate import), and `agent-manager.ts` (the cascade plus the shared label constant).
