Mohamed BoudraandGitHub
8021d2b6c9
* Reapply: Stop completed turns from appearing stuck (#2484) * Reapply: Keep older chat history and image previews stable (#2490) * Lock timeline layout stability with user-driven specs Two reds reproduce the reported jank: - new agent, first prompt: the row paints at top=100 then jumps to 144 through authoritative hydration, because the draft tab and the agent panel each render their own AgentStreamView for the same logical row. - long loaded timeline, submit: scrollHeight collapses 10793 -> 9111 mid-turn, because splitOrderedTail reclassifies measured rows into the virtualized segment where the virtualizer falls back to estimateStreamItemHeight. Also restores centered timeline rows: the pagination anchor wrapper introduced in #2490 was a plain block div, which broke alignSelf: center. * Measure the turn footer blink from when the footer appears Bounding the window at the prompt row conflated two things: the footer arriving a frame after the row, and the footer leaving the layout and coming back. Bounded from the footer's first frame, the blink still reproduces, so the footer-ownership race is real rather than a sampling artifact. * Rename the shadowed prompt row binding in the frame recorder * fix(desktop): allow concurrent worktree dev instances Desktop terminals can inherit another instance's profile path and debugging flags. Scope the profile to the current dev root and select an available debugging port so instances do not share global runtime state. * fix(desktop): expose development providers in dev builds Desktop-managed daemons inherited no Paseo runtime mode, so dev builds constructed the production provider registry. Mark unpackaged internal launches as development while continuing to force packaged launches to production. * fix(timeline): stabilize optimistic turn layout Keep initial history loading dormant until user intent and preserve active-turn layout across delayed transport updates. Add painted-frame coverage for optimistic timeline transitions. * fix(timeline): preserve optimistic prompts through hydration Canonical replacement relied on a transport submission record that can settle before a provider echoes the user row. Keep presentation continuity owned by the unreconciled row while destructive history replacement remains governed by active transactions. * fix(timeline): make submitted prompts canonical Submission settlement previously depended on status updates that could arrive after the optimistic row. Persist every accepted prompt in the canonical timeline and keep the footer active until RPC and canonical acknowledgement converge. * fix(timeline): stop older history pages repeating entries Backward pagination counted source sequence rows, so page boundaries inside a merged projected item could return that item repeatedly. Count projected entries and advance from their display anchors instead, with server and browser regressions covering the contract. * fix(timeline): resume agents from the latest history Use a bounded tail probe on resume so long background runs appear atomically instead of replaying every missed page. Preserve unchanged views, reconcile local/live rows, and keep skipped history reachable through backward pagination. * fix(timeline): reconcile turn state after missed updates Turn completion events can be missed while an agent timeline is unsubscribed, leaving footer activity stale even after the directory reports idle. Reconcile the single-owned turn lifecycle from authoritative resume snapshots while preserving newer live evidence. * test(timeline): cover running state across refresh Protect the hydration boundary where directory state must restore consistent tab, composer, and footer activity. Expose the running indicator semantically so the regression checks user-visible state instead of spinner implementation details. * fix(timeline): keep history loading geometry stable Reserve a fixed history-start gutter regardless of pagination state so showing or hiding the loader cannot shift timeline content. * fix(timeline): keep turn activity coherent Derive footer, tab, and composer activity from one lifecycle and submission projection. Idle directory transitions now request an authoritative resume so missed terminal events converge without treating directory delivery order as turn authority. * fix(tests): align checks with runtime boundaries * fix(desktop): stop dev app surviving its terminal Electron was detached from the desktop runner, allowing it to outlive a stopped workspace terminal with broken output pipes and retain the single-instance lock. Keep Electron owned by the runner and preserve explicit process-group cleanup for detached child trees. * fix(timeline): preserve activity through turn handoff Correlate lifecycle events and snapshots with daemon-owned turn identity, and keep optimistic submissions active until their canonical row arrives. This prevents stale updates and RPC timing from exposing a false idle state. * refactor(timeline): remove submission timing state * fix(timeline): restore prompt identity on resume * fix(timeline): exclude attachments from canonical prompt text Structured provider prompts can include hidden history attachments before the user's text. Persist only plain prompt blocks so authoritative user messages match what the composer displayed while provider input remains unchanged. * refactor(agent): remove unused pending-run message identity * fix(desktop): let Chromium allocate the debugging port * refactor(agent): remove impossible finalized-turn tracking * fix(timeline): load history after keyboard scrolling Treat observed upward displacement as user intent so keyboard and scrollbar navigation can wake dormant pagination. * test(server): preserve accepted turn stream order The manager-owned accepted turn boundary intentionally precedes coalesced output. Assert that lifecycle event explicitly while retaining the timeline-before-terminal ordering contract. * test(app): scope sidebar fixtures by identity Shared E2E daemon state may contain unrelated projects. Keep stale-route coverage tied to its named missing project and drag only the two rows seeded by the reorder spec. * fix(timeline): gate prompt revisions by client capability * fix(timeline): finalize canonical prompt delivery Treat provider identity as a single stored-row transition so legacy clients receive one deliverable canonical prompt. Release identity-pending rows before later foreground output or the current terminal while preserving stale-turn isolation. * fix(server): preserve legacy prompt pagination * fix(timeline): keep assistant block identity stable Assign canonical block IDs during live promotion so terminal finalization cannot remount image-bearing rows. * fix(timeline): apply covered revisions during resume append Run the exact canonical revision pass before overlapping resume pages append newer rows, while leaving cursor advancement to the incremental path. * fix(server): preserve queued agent transitions Project each queued live snapshot before awaiting stored metadata so later lifecycle state cannot overwrite an earlier transition. * fix(server): snapshot queued agent updates Capture each live agent payload before it enters the per-agent queue so shared nested state cannot change while an earlier storage read is pending. * docs(app): date agent turn compatibility cleanup * refactor(timeline): mint submitted prompt identity at acceptance Canonical prompt rows now keep one immutable wire identity while provider identity remains internal for rewind translation. This removes the unreleased same-sequence revision and legacy projection contract. * refactor(app): remove prompt revision reconciliation Submitted prompt rows are immutable on the wire, so resume and live reducers no longer need to apply already-covered same-sequence updates. * test(app): cover submitted prompt rewind roundtrip * docs(timeline): describe immutable submitted prompt identity Provider echoes now enrich daemon-internal identity only; wire rows keep their acceptance-time identity and rewind performs the provider translation. * test(server): validate mock rewind targets Reject unknown provider message IDs so app rewind coverage exercises wire-to-provider identity translation. * fix(server): validate rewind before interrupting * fix(server): omit rewind identity for daemon prompts * test(app): guard rewind eligibility and rejection * fix(server): preserve prompt identity across hydration * refactor(timeline): remove stale revision remnants * Revert "fix(server): preserve prompt identity across hydration" This reverts commit 6adcf7e5b0bc8601761af59b461a746cb5e395fa. * fix(timeline): settle missed submission acknowledgements
14 KiB
Timeline sync
Agent chat delivery has two paths:
- Live stream —
agent_streamWebSocket messages for immediacy. These may be delta-shaped lifecycle updates. - Authoritative history —
fetch_agent_timeline_requestfor correctness. This always returns full projected timeline items, never lifecycle deltas.
The invariants are:
A continuously subscribed client applies every committed row in order. Opening or resuming an agent establishes the daemon's current tail in one bounded request, with older history reachable through backward pagination.
Tool output is bounded before it enters either delivery path. Canonical shell tool output is sliced to 64 KiB, and the same bounded item is used for durable timeline rows and live stream events. Provider history hydration applies the same rule so reopening an agent cannot restore an oversized tool payload.
Presence is not delivery
Client heartbeat reports presence:
- device type
- app visibility
- focused agent
- last activity time
Heartbeat is used for notification routing. It must not be used as a correctness gate for agent_stream delivery. A stale mobile focus heartbeat may affect whether the user gets notified; it must not make timeline rows disappear from the live stream.
Gap recovery is paged but complete
Large unbounded timeline responses can exceed relay frame limits, so catch-up uses bounded pages. Bounded does not mean partial.
Page limits are projected-item targets. A tool call lifecycle is one projected item even if it spans many source sequence numbers, and assistant/reasoning chunks are merged before counting. The response carries seqStart, seqEnd, sourceSeqRanges, and collapsed so clients can advance sequence cursors without rendering delta rows.
When live delivery detects a sequence gap, the app fetches direction: "after". If the daemon
responds with hasNewer: true, the app immediately fetches the next page from endCursor. Gap
recovery is complete only when hasNewer: false.
Initialization timeouts guard lack of catch-up progress, not the full multi-page sync. A successful page that queues the next after page refreshes the watchdog.
Opening or resuming an agent fetches one bounded latest tail page. Older history remains user-driven by scrolling upward.
Reaching the history-start threshold loads one older page and preserves the visible content anchor. Cursor progress does not trigger another page. The user must leave and return to the threshold unless the anchored page still leaves the viewport at history start, as with short or compacted content; in that case pagination continues as one loading operation until the page fills the viewport or history is exhausted.
Durable item anchors
Provider message IDs are not guaranteed for every displayed item. Paseo-generated system errors are one example. Rendered item indices are not durable either because pagination and projection can merge source rows.
Actions that address a point in chat history, such as Fork, use the daemon timeline epoch plus the projected item's seqEnd. The app carries that position on the rendered assistant item for both live and fetched history. When adjacent projected chunks merge, the merged item retains the newer chunk's position.
The daemon validates that the epoch is current and the exact source sequence still exists before slicing rows. It slices before projection so later lifecycle updates cannot leak into the selected context.
Resume behavior
Opening, reconnecting, or revisiting after selective-delivery grace fetches the latest tail page. Focus alone does not mutate timeline state; the tail response is compared with the local authoritative range first.
- The same epoch and
window.maxSeqis an exact display no-op. The app advances synchronization bookkeeping without replacing timeline arrays, preserving an upward-scrolled viewport. - When the page overlaps or is adjacent to the local end cursor, only projected items newer than that cursor are applied. Already-covered rows are not replayed.
- A true middle gap, epoch change, or rewind atomically replaces stale canonical history with the latest tail. The replacement reconciles positioned live rows beyond its coverage and unresolved local submissions; it never retains two discontiguous canonical ranges.
The installed tail carries hasOlder, so history skipped by a replacement remains reachable through
ordinary backward pagination. A backward page is accepted only when it is adjacent to the current
history start; a response requested from a pre-replacement range is stale and is discarded.
Client replica lifetime
The host runtime owns each session replica for as long as the host remains registered. React providers attach message handlers and UI integrations to that replica, but mounting or unmounting a provider must not create or clear it. A provider can remount during Fast Refresh or ordinary UI recomposition while the runtime still owns the same directory snapshot and timeline cursors.
Removing the host from the registry is the destructive boundary: it stops the runtime and clears the session and host-scoped setup state together.
The durable replica cache is a display cache, not a synchronization checkpoint. Its timeline record
contains only the focused agentId and a truncated item tail. It never persists a cursor, epoch,
older-history availability, authority status, or sync generation because those facts would describe
the complete source dataset rather than the truncated display dataset.
Restoring that cache produces a painted timeline: the items may render immediately, but the first
daemon timeline request is still tail. A successful tail response atomically establishes canonical
items, range, and older-history availability. Live rows received between cache paint and that tail
response stay in the separate live head, do not advance a cursor or trigger gap recovery, and are
reconciled with the authoritative tail and subsequent catch-up.
Every daemon-derived live item carries its timeline epoch and sequence position. Bootstrap replacement keeps only positioned rows newer than the page it installs, while unresolved local user presentations identify themselves by having client identity without a timeline position. This prevents a page from duplicating rows it already covers without coupling display continuity to the shorter-lived submission registry. Unreconciled local presentations are not persisted in the durable replica cache.
Selective and legacy delivery
The app chooses one delivery policy from server_info.features.selectiveAgentTimeline:
- Selective daemons receive the union of agents visible in every pane. Additions subscribe and catch up immediately. Every visibility-driven removal, including app backgrounding, stays subscribed for a 30-second grace period so brief tab, pane, route, and app switches do not repeatedly unsubscribe and catch up. Losing window keyboard focus does not make a selected pane invisible. Disconnecting and disposal clear pending grace because the subscription itself no longer exists. After grace has expired, revisiting a retained timeline displays its cached state immediately and authoritative catch-up advances it to the current tail.
- Legacy daemons keep globally streaming agent timelines. Visibility still triggers the existing authoritative catch-up, but the app does not issue selective-subscription RPCs.
This policy is owned by viewed-timeline-sync.ts; downstream reducers do not branch on daemon
version.
Projected pages reconcile with live presentation
A projected page is canonical state, not a sequence of live deltas. One projected item can overlap
rows already received live—for example, a tool call retained at its original display position while
its completion advances seqEnd, followed by a merged assistant message. The app uses
sourceSeqRanges to replace overlapping assistant and reasoning projections before applying the
remaining page through the existing stream reducer. It must not append full projected text to a
live prefix.
Every path that sends a message to an agent — composer send, dictation accept-and-send, queued
send-now, and the automatic queue drain in HostRuntime — goes through
dispatchComposerAgentMessage with a submission writer. There is no second transport for the same
product action: calling client.sendAgentMessage directly skips the submitted row and the pending
footer, and permanently drops attachments because the daemon does not echo them back.
A submitted prompt is one UserMessageItem row. That row is the authoritative local presentation:
its stable identity, text, timestamp, images, and attachments do not change when the provider
acknowledges it. Submission lifecycle is a separate record keyed by agent, not another row shape.
The transaction registry records two independent settlement facts: canonical acknowledgement and RPC
settlement. Canonical acknowledgement retires optimistic activity immediately. When both facts are
known, whichever arrives second deletes the record. A canonical acknowledgement that arrives first
prevents a later transport error from rolling back a prompt already observed.
The daemon's accepted response waits for the correlated run start and guarantees that the canonical submitted row has been recorded. It publishes the accepted turn's liveness before that row, so the client applies authoritative activity before canonical acknowledgement retires optimistic activity. Timeline render batching does not delay lifecycle application. Directory status never settles a submission. Overlapping sends settle independently rather than collapsing to one newest pending message.
Daemons advertising server_info.features.canonicalSubmittedPrompts guarantee that every accepted
prompt carrying a client message id is recorded and streamed as a canonical user_message with that
same id. This includes daemon-handled commands that do not allocate a foreground turn; their submitted
row is recorded before handler output. The app tracks submission transactions only for hosts with this
capability. Older hosts keep the shipped untracked optimistic-row behavior and roll that row back on RPC
rejection.
Turn activity has one client-side replica. Lifecycle events can attach turnId, and agent snapshots can
expose activeTurn: { turnId, startedAt } | null. An identified terminal cannot close a different
identified turn; an unnamed legacy terminal can close the current turn. Snapshots, stream events,
cancellation requests, and every visible surface update the same per-agent record. The snapshot covers
both user-started foreground turns and autonomous provider turns;
foreground control ownership remains a separate daemon concern. Cancellation request identity is stored
with that record rather than in a React component, so an old request cannot clear a newer one. Submissions
remain a separate pre-turn registry and retire on canonical acknowledgement.
The compatibility boundary for older daemons is snapshot normalization: running/idle status becomes an anonymous active turn or idle state once, and downstream code consumes the same activity shape. The app does not combine anonymous lifecycle events, timestamps, timeline rows, and resume coverage to infer a second running state. Disconnect and replica removal remain destructive close boundaries. Elapsed time comes only from turn liveness, never from submission records or whichever timeline rows happen to be mounted.
The daemon records one canonical submitted user row at acceptance. Its wire messageId is the
submission's clientMessageId, so the row is born with its final identity and remains immutable on
the wire. A correlated provider echo records the provider's native identity internally without
dispatching another timeline event. Rewind resolves the wire identity to that provider identity at
the provider boundary. Daemon-handled prompts follow the same identity rule.
Content matching is limited to the dated compatibility path for daemon timelines created before that field existed. Canonical ingestion may match only an explicit unreconciled local candidate; the draft-create handoff is the one boundary that also permits the legacy canonical twin to have arrived first. Generic reducers and consumers do not reimplement message identity matching.
Ordinary bootstrap, same-epoch reset, and catch-up replacement preserve unmatched locally submitted rows because a provider may never echo them. A known epoch change or rewind replaces history and drops acknowledged local rows omitted by the new canonical epoch; every transaction not yet acknowledged by the provider, and no other local row, crosses that destructive boundary. A cold reset without an existing epoch is destructive because the client has no continuity anchor.
Tail rows are positioned history, so an unmatched local presentation is appended after the canonical replacement rather than ordered by timestamps from different machines. The head is a live overlay: cursorless items stay there during continuity replacement until canonical positions arrive, while a destructive replacement retains only active submission transactions.
Canonical replacement owns both timeline lanes. A matching local row keeps its presentation ID and payload while taking the canonical row's ordered position. If a live assistant head is the canonical assistant prefix, it stays in the head lane. No row may be returned in both lanes.
Relevant code
- Server live stream forwarding:
packages/server/src/server/session.ts - App sync planning:
packages/app/src/timeline/timeline-sync-plan.ts - App viewed-agent synchronization:
packages/app/src/timeline/viewed-timeline-sync.ts - App stream/timeline reducer:
packages/app/src/timeline/session-stream-reducers.ts - Session wiring:
packages/app/src/contexts/session-context.tsx