tracking(perf): bound Desktop and Runtime Host work as history grows

[Astro-Han]

Problem

Desktop responsiveness and Runtime Host work should depend on the visible window and the amount of new data, rather than the full Session history. Several paths have violated that constraint: long-session hydration, oversized-Turn rendering, Usage tables, WorkHub linkage reconstruction, and artifact persistence/recovery.

This tracker defines the intended architecture, performance budgets, and remaining delivery work. Individual issues retain their reproductions and detailed reviews. #2913 contains the broader incremental-data-flow design; this issue tracks the measured Desktop and Host outcomes across that work.

Desired outcome

A user can open a long Session, read and scroll while output streams, switch Sessions, and open Usage or WorkHub without full-history work blocking the interaction. Reading position, focus, selection, accessibility, and durable execution semantics remain correct.

Target architecture

flowchart TD
    L[Durable execution and domain records] --> H[Runtime Host: lifecycle and query authority]
    H -->|bounded pages, watermarks and changes| M[Desktop Main: bounded sparse replica]
    M -->|incremental IPC| R[Renderer: bounded visible window]
    R --> S[One transcript scroll authority]
    R --> B[Render boundaries within oversized Turns]
    H --> Q[On-demand WorkHub linkage queries]
    H --> P[Delta persistence and bounded diagnostics]

Loading

• Runtime Host owns execution. Session/Turn lifecycle and terminal facts remain authoritative there. RuntimeEvent and the appropriate domain ledgers feed read projections; Desktop does not reconstruct execution state from its rendered history.
• Desktop is an incremental client. Tail bootstrap, bounded pages and generation-checked updates replace full-history loading. Resident data and DOM stay bounded as history grows. One scroll authority owns viewport changes; sub-Turn containment limits rendering work without becoming a second paging or scrolling mechanism.
• WorkHub reads linkage at its owner. The target Session and Coordination ledger establish current delegation facts on demand. Replace/stop revalidate them under the existing admission boundary. No renderer replay or additional durable active-delegation table is required. The current fix(workhub): resolve WorkHub delegation linkage on demand #4699 lookup still scales with target-local assignment history; its cost must be measured rather than called constant-time.
• Persistence writes the change. Artifact metadata uses delta writes, and per-call diagnostics store bounded summaries rather than another copy of the conversation. Remaining recovery scans and repeated full metadata reads need a defined consistency/invalidation boundary before optimization.
• Other views follow the same rule. Usage rows are paged; catalog updates preserve unchanged items; streaming processing consumes new chunks; search and inactive views avoid remote full-history work.

Performance baseline and results

These are measurements or estimates recorded in the linked issues/PRs, not a new benchmark of current main. Each row is a separate workload; the results must not be combined into an overall speedup.

Scenario and source	Starting baseline	Reported result / remaining gap
Tool-rich 1,000-Turn hydration, 30ms RTT — #2913	About 5,000 sequential requests; 150s theoretical transport lower bound, not a wall-clock measurement	Tail bootstrap and byte-bounded paging landed in #2922/#2937. End-to-end latency must be checked against the budgets below.
Hover over an oversized Turn of roughly 15k elements — #4259	Hover p95 59ms; about 20 LoAF >50ms; style recalculation about 1,270ms per measured sweep	PR reports 9ms, 0 LoAF >50ms, and about 100ms respectively. This is a hover result, not proof of faster scrolling.
Cold upward scroll, eight 240px wheel steps — #4259 measurement	Same-build control with inner containment disabled: 0px worst unexpected anchor movement	Old head bfee6e1c: 239.44px, light/dark, three runs each at 1400×900. Later head bd8e0299 changes folded estimates to 32px and uses a 120px test bound; independent remeasurement is outstanding.
Usage Activity log, 409 records — #4531 / #4539	409 rows mounted on each visit	Proposed page size: 50 rows. Before/after tab-entry median and p95 are still needed.
Per-call durable data, 600-message conversation and 60 tools — #4722	405,729 B across diagnostic records and capture file	3,944 B after removing the capture and folding diagnostics; same-machine comparison against 780dc4b4.
Cumulative model-call diagnostics over 500 Turns — #4722	193.44 MiB, preparation CPU 607ms	1.88 MiB, 139ms, under the PR's synthetic workload.
Artifact store, 6,000 records / 400 Sessions — #4722	listPage 13.45ms; create 38.32ms	11.54ms / 37.42ms; means of 50 / 20 operations in one run. Full reads and filesystem durability remain significant.
Host cold start, roughly 11.7k artifacts — #4027 / #4037 / #4038	Historical profile: about 130s total, roughly 90s in the artifact metadata/cleanup phase	Several persistence changes have landed. A fresh startup profile is required; the historical timing is not the current baseline.
WorkHub linkage lookup — #4699	Opening/reconnecting replayed the Coordination history	Current PR reports about 3.8 / 33.8 / 205.3ms for 100 / 1,000 / 6,000 terminal assignments. No comparable before/after wall-clock result is recorded yet.

Target performance

The latency budgets below carry forward the proposed budgets in #2913; they are acceptance targets, not claims that current main meets them.

Scenario	Target
Local Session open to visible tail	p95 ≤ 150ms
Long Session over 30ms RTT	p95 ≤ 800ms, with initial useful tail in one subscription-open round trip
64 MiB Session	Tail opens and older history remains pageable; no full-history prerequisite
Normal input, scroll and hover	No task/LoAF over 50ms in the fixed interaction trace
Transcript and Sidebar rendering	DOM bounded by the resident window/viewport, independent of total history; #2913's initial transcript ceiling is normally 100 Turns
One catalog update with 1,000 Sessions	Renderer work ≤ 16ms
256 KiB output arriving in 20-character deltas	Cumulative sanitizer CPU ≤ 250ms, preserving cross-delta redaction
Steady-state synchronization and per-call diagnostics	Work proportional to changed data; no full conversation copy per call
Cold-scroll geometry	No stalled/reversed step or snap to tail. Calibrate the numerical anchor-displacement bound against the control; 120px is not yet an accepted product budget.
Host startup, WorkHub and Usage latency	Establish comparable current-main baselines and explicit p95 budgets before accepting these work items. Missing numbers remain open acceptance work.

Implementation plan

• Replace full transcript snapshots with bounded pages and a Desktop sparse replica — perf(runtime-host): replace transcript snapshots with bounded pages #2922, perf(desktop): keep transcript state incremental #2937.
• Reduce repeated rendering work in streamed Markdown, prompt rail and Session rail — perf(ui): incrementally prepare streamed Markdown math #4207, perf(ui): scope prompt rail observation to mounted turns #4203, fix(desktop): keep session workspace action identities fixed #4110.
• Fix native peer snapshot-watch busy loops — fix(peer): stop snapshot watch busy loops #4687.
• Land artifact delta writes, bounded per-call diagnostics and old-capture reclamation — perf(storage): persist artifact metadata deltas #4716, perf: bound what one model call costs to store #4722, feat(storage): reclaim the prepared-request captures already on disk #4738.
• Establish transcript regression fixtures and land the tail-pin correction — test(desktop): make the transcript scroll regression line real #4796, fix(ui): preserve tail pin and correct reader scroll assertion #4831.
• Accept oversized-Turn rendering with both hover gains and cold-scroll/AX validation — perf(ui): oversized single Turn defeats transcript render bounds #4256 / perf(desktop): bound rendering within oversized turns #4259.
• Accept older → newer → latest navigation, retained focus and reading position — feat(desktop): preserve transcript continuity across bounded ranges #4123 / feat(desktop): preserve transcript continuity across bounded ranges #4560.
• Remove WorkHub open/reconnect full replay and validate current linkage under recovery/concurrency — WorkHub Coordination transcript is replayed in full on every open and every reconnect #4647 / fix(workhub): resolve WorkHub delegation linkage on demand #4699.
• Bound Usage Activity rendering and record tab-entry latency — perf(desktop): Usage activity log is slow with large histories #4531 / perf(desktop): paginate usage activity #4539.
• Re-profile Host startup and resolve remaining recovery/batch-retirement costs — fix(runtime-host): cold-start artifact recovery sweeps O(all files) realpath/lstat, blocking Host readiness for minutes #4027 / perf(storage): batch multi-Session artifact purge during session retirement (O(M×N) guard scans) #4038; reassess the residual full-read cost documented in perf: bound what one model call costs to store #4722.
• Reconcile remaining catalog, startup, streaming and search work with perf(desktop): make session data flow incremental and bounded #2913, then enforce the budgets above.

Implementation merged and behavior accepted are separate milestones: the checked items above record landed changes, while the open items retain their outstanding acceptance work.

Validation and completion criteria

Run before/after measurements on the same machine, fixture, viewport, theme and motion settings. Record both commits, OS/runtime versions, fixture size, and whether a result is measured, theoretical, or author-reported. Separate cold from warm runs and hover from scrolling. For latency comparisons, use at least 10 repetitions and report median/p95; keep raw traces with the implementation PR.

Use small and oversized Turns, local and 30ms-RTT Sessions, growing WorkHub histories, the 409-record Usage fixture, and the artifact-startup fixture. Exercise actual Electron/CDP interaction where compositor scrolling, focus or startup is relevant. Storybook covers deterministic layout and accessibility; storage microbenchmarks do not establish end-to-end responsiveness.

This tracker is complete when the open work above is resolved, the performance targets are measured on the delivered implementation, and there is no regression in reading position, focus, selection, accessibility, redaction, or durable execution/recovery. Update the result table and work-item checkboxes as changes land; keep review conversations in the linked PRs.

Related work

• Existing common-path defects remain tracked individually: fix(desktop): archived sessions can remain selected and trigger Skills and Plan queries #4430, Permission mode switch (Auto→Bypass) is not honored by the next turn #3349, Default-model picker in Settings spins on every switch #3827, The Goal chip reads "running" while an armed Goal waits for its first Turn #3362, fix(desktop): connection base URL gate is looser than the Runtime Host contract, so a rejected endpoint reports a misleading error #3672, fix(desktop): agent graph appears frozen while 0/7 operators are running (no heartbeat) #3878, Failed turns hide the provider's own response; show it collapsed, expandable, for every failure class #4502.
• Historical renderer/data-flow work: perf(desktop): stop forwarding unobserved session events #1957, perf(desktop): stop idle-session event-health polling from re-rendering the whole shell #1979, perf(desktop): AppShell subscribes to all session UI state at one granularity, so every stream token re-renders the whole shell #1985, refactor(ui): stream the transcript incrementally instead of re-deriving it per token #2030, perf(desktop): switching back to a long session mounts the whole transcript in one commit, freezing the UI ~0.35s #2052, perf(desktop): remove models.dev metadata from the renderer startup path #2063, perf(ui): seed transcript geometry before progressive fill #2224. Consult these before duplicating a previous optimization.
• Test-suite runtime is tracked separately in tracking(desktop): move every test out of Electron E2E that does not need it #4761.

Prepared with Codex assistance under maintainer direction.

[Astro-Han]

take