perf: eliminate idle CPU floor and make replay refresh incremental#395
Merged
Conversation
Render chat tab content and the close affordance within the shared WorkStation tab pill surface. This gives the entire rounded tab a consistent hover state and matches My Station tab composition.\n\nUpdate the tab-bar regression test to verify the close control remains inside the shared surface without nested buttons. Pre-commit hook ran. Total eslint: 0, total circular: 0
Reuse the external-history rescan API from the sidebar refresh action so enabled provider sessions are updated before the list reloads.\n\nAdd configurable polling for the external session currently open in Chat, with localized 3-second, 5-second, 10-second, and 1-minute cadence options. Clarify the scan settings copy across all supported locales. Pre-commit hook ran. Total eslint: 0, total circular: 2
Batch all enabled external-history sources into one cache-only IPC request while retaining per-source date-bucket pagination. Add delayed imported-session turn metadata with compact file rows and hover previews. Normalize monitoring rates over the observation window so request bursts are not reported as extreme polling frequencies. Pre-commit hook ran. Total eslint: 0, total circular: 2
Pre-commit hook ran. Total eslint: 0, total circular: 2
Remove two import cycles reported by the repository-wide Madge scan without changing runtime behavior. - import the session adapter registry from its leaf module instead of the sync barrel - derive completed task report types from the existing local Zod wire schema - preserve the structural compatibility of cloud session comment results Validation: - npx madge --circular --json --extensions ts,tsx --ts-config tsconfig.json src/ - npm run lint - npx tsc --noEmit - 58 focused Vitest tests Pre-commit hook ran. Total eslint: 0, total circular: 2
Run an immediate deduplicated incremental scan for enabled non-manual external sources when the app starts, then continue cadence-based scans from completion time. Batch multi-source updates into one native IPC request with clear disabled and reload the sidebar once after a successful scan. Pre-commit hook ran. Total eslint: 0, total circular: 0
Focused polling relaxes from 5s to 10s — each poll spawns a git status subprocess (~130ms on this repo), a permanent 2-3% single-core floor. A new set_source_control_attention command (reference-counted from the frontend via useSourceControlAttention) restores the fast 5s interval while a Source Control surface is actually on screen. Also corrects the stale '4-6 git processes per poll' note: status was consolidated to a single porcelain=v2 call in January. Pre-commit hook ran. Total eslint: 0, total circular: 0
The cross-agent provenance backfill kept a ~6% idle CPU floor: every file-history read reconciled EVERY imported session in the repo, each file interaction spawned 1-2 git subprocesses for path resolution, and active sessions were re-reconciled on every run because appends change their fingerprint. - reconcile only sessions that touched the requested file, plus a bounded backlog batch per run (pre-filtered so completed sessions never consume the budget) - skip sessions whose transcript changed in the last 10 minutes: live interactions arrive via hook capture, and a quiet session is indexed exactly once because its fingerprint then never changes again - short-circuit repeat backfill requests within 5 minutes of a finished run unless the requested file still has pending sessions (TTL is below the quiescence window, so nothing that would have been indexed is delayed) - resolve workspace root / repository id by walking the filesystem (.git, gitdir pointers, commondir) instead of spawning git rev-parse per interaction, with a per-cwd cache; parser version bumped to v3 so reconciled rows rebuild under one id derivation - new claude_code_history_stat command: cheap (mtime, size) probe so the replay auto-refresh can skip unchanged transcripts Pre-commit hook ran. Total eslint: 0, total circular: 0
The cmd+5 panel only saw axios and tauri invoke, hiding most real traffic: bare fetch() calls to the IDE server and everything the backend pushes. Coverage added: - window.fetch patch while tracking is enabled (HMR noise filtered), which also surfaces tauri ipc://localhost requests with durations - push-traffic recorder with per-name rates: tauri events (via the useTauriListen wrappers), session IPC channel messages, code-editor websocket messages, and SSE stream events - hotspot rates report a measured-window rate instead of extrapolating 2-sample bursts into absurd per-minute figures Pre-commit hook ran. Total eslint: 0, total circular: 0
External-replay and sidebar hot paths issued heavy redundant backend work: - replay auto-refresh (default 5s) re-read and re-parsed the whole transcript every tick, even minimized. Now: skipped while the window is hidden; a cheap stat probe (new statTranscript source hook, wired for claude_code) skips unchanged transcripts entirely; and when new lines do arrive, loadSessionAtom keeps existing object references for already-loaded imported events (append-only transcripts, immutable events; last one excepted for chunk aggregation) so the memoized render pipeline only paints new rows - sidebar subagent hydration fired es_get_child_sessions once per visible session on every list refresh (100+ concurrent calls queuing on SQLite, 2.6s avg observed). Children now re-fetch only when the parent's updated_at changes, in batches of 8 - commit reference cards fetched /commits per card purely to decorate the subtitle; dozens of ~1s requests per replay. Removed: the reference extracted from the message already carries the commit id, subject and repo name Pre-commit hook ran. Total eslint: 0, total circular: 0
Cloud org management moves from an inline panel into a first-class chat-panel tab (new cloud-org tab kind with its own display label and tab-bar treatment), so it participates in tab persistence and navigation like launchpad/workspace tabs. Covered by updated tab display/atom unit tests and the cloud org e2e specs. Pre-commit hook ran. Total eslint: 0, total circular: 0
Pre-commit hook ran. Total eslint: 0, total circular: 0
Collaborator
|
LGTM |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Follow-up to #390: eliminates the remaining idle CPU load (org2 held a ~6% floor with periodic spikes; the renderer burned on external-session replays), plus the external-history session fixes and the cloud-org chat tab that were pending locally.
Idle CPU
git rev-parsesubprocesses per file interaction, and re-reconciled active sessions forever because transcript appends change their fingerprint. Now: priority-file scoped with a bounded backlog batch per run, a 10-minute quiescence gate (live interactions arrive via hook capture; a quiet session is indexed exactly once), a 5-minute TTL short-circuit for repeat requests, and filesystem-based git discovery with a per-cwd cache (zero git subprocesses; parser bumped to v3 to rebuild rows under one id derivation). Verified with process sampling: no orgtrack git processes remain.set_source_control_attentionsignal restores 5s while a Source Control surface is actually visible.External-session replay
(mtime, size)stat probe (newclaude_code_history_statcommand + optionalstatTranscriptsource hook) skips unchanged transcripts; and on real appends,loadSessionAtomkeeps existing object references for already-loaded imported events so the memoized render pipeline paints only new rows.es_get_child_sessionsfired once per visible session on every list refresh (100+ concurrent calls queuing on SQLite, 2.6s avg observed → 0.06s). Children now re-fetch only when the parent'supdated_atchanges, in batches of 8./commitsper card (dozens of ~1s requests per replay) — the extracted reference already carries the id, subject and repo name.Observability
fetch()traffic (includingipc://localhostrequests with durations) and backend push traffic (tauri events, session IPC channels, websocket + SSE messages) with per-name rates, and hotspot rates no longer extrapolate 2-sample bursts into absurd figures. This panel work is what surfaced the findings above.Also included (previously unpushed)
Verification
cargo checkclean; orgtrack (19) and watch tests pass;eslintclean on all touched files;tsc --noEmitintroduces no new errors; vitest suites for session atoms, external history, sidebar menu items and chat tab display pass (15 + 23 + 7). Filesystem git discovery validated byte-for-byte againstgit rev-parseon a real linked worktree.