Skip to content

perf(desktop): page speed of WireCat Desktop on a real store #345

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
leemour merged 1 commit into from
Nov 10, 2026
Merged

perf(desktop): page speed of WireCat Desktop on a real store #345

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions performance/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,7 @@
against file watching.
- [agents](agents/README.md): what the CLIs cost the AI agent that reads their answers — output size and meeting workflows.
- [tasks](tasks/README.md): tasks found in messages — duplicate detection.
- [desktop](desktop/README.md): WireCat Desktop — how long each page waits for its data, and how fast clicks paint.
- [command-server](command-server/README.md): `commands --serve` for WireCat Desktop — several requests at once
through one process.
- [startup](startup/README.md): install size and cold start of the five CLIs, and which packages a simple
Expand Down
7 changes: 7 additions & 0 deletions performance/desktop/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
# Desktop

Measurements of WireCat Desktop, the app that runs the CLIs for the owner.

- [page-speed](page-speed/README.md): what each page waits for on a real store. A warm memo server answers in 30 to
330 ms what a fresh process answers in 650 to 960 ms, `people list` takes 3.7 to 4.2 s for any limit, and clicks in
a production build paint in 24 to 96 ms (2026-11-10).
125 changes: 125 additions & 0 deletions performance/desktop/page-speed/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,125 @@
# How fast each WireCat Desktop page gets its data

Measured 2026-11-10 with the scripts in [`scripts/`](scripts/). Totals are in [`results.txt`](results.txt).

## Why we ran it

The owner used WireCat Desktop on his own store and found it slow and buggy: pages waited seconds for memo. He asked
whether Desktop can be made really fast, and what has to be done for that. The questions:

1. Which memo commands does each page run, and how long does each take on a real store: as a fresh process, and
through a warm `memo commands --serve` server?
2. What does a page's set of reads cost when fired together, the way Desktop's pool of 2 servers runs them?
3. Is the interface itself slow once it has the data?

The answers feed the speed plan for Desktop, memo and cli-core (kept in the private working folder).

## Setup

- **Machine.** AMD Ryzen AI 9 HX 470, 30 GB RAM, Linux, Node 24.19.0, Google Chrome 154. Other agents worked at the same
time: load average about 6, swap full, 2 to 4 GB free. Nothing was pinned.
- **Code.** memo 0.131.75 (cli-core 0.624.0) installed globally; WireCat Desktop at `32681529`.
- **Data.** The owner's real store, read only: list and show commands with `--no-record`, answers piped into a counter
of rows and bytes. No content was printed or kept. The store holds about 114,000 people, 500 or more agent findings
batches, 6 deal pipelines and a few hundred deals.
- **Commands.**
- `node scripts/make-commands.mjs /tmp/commands.json` writes the reads each page sends on open (it looks up the
busiest deal pipeline, so the file stays on the machine).
- `node scripts/measure-commands.mjs /tmp/commands.json 3`: each command 3 times as a fresh process, then 3 times
through one server; medians.
- `node scripts/measure-fanout.mjs /tmp/commands.json "Memory:" 3`: the page's reads at once, as fresh processes,
through pools of 1, 2 and N servers (a busy server sends the call to a fresh process, as Desktop does), and through
one server one after another.
- `node scripts/measure-first.mjs '<argv list as JSON>'`: the first and second request on a freshly started server.
- Interface: `pnpm exec vite build && pnpm exec vite preview --port 1487 --strictPort` in Desktop, then
`DESKTOP=<desktop checkout> node scripts/click-speed.mjs --runs 3 [--cpu 4]`. Mock data (`?memory=many`,
`?clients=many`, `?tasks=many`, `?people=many`, `?showcase`). "Frame" is the Event Timing duration of the click
(input to next paint, what INP scores); "long tasks" is main thread work over 50 ms in the 2 s after it.

## Results

### Each command (ms, median of 3)

| Page | Command | fresh process | warm server | rows | KB |
| --- | --- | ---: | ---: | ---: | ---: |
| Memory | memories, facts, agents, roles lists | 631 to 672 | 66 to 88 | 2 to 16 | 1 to 6 |
| Memory | actions list --limit 500 --brief | 718 | 177 | 500 | 359 |
| Memory | relationships list confirmed:false | 686 | 128 | 100 | 77 |
| Memory | relationships list include:ended | 905 | 278 | 100 | 71 |
| Clients | pipelines list kind:deal | 663 | 91 | 6 | 12 |
| Clients | pipelines show | 799 | 212 | 1 | 2 |
| Clients | deals list status:open --limit 500 | 964 | 324 | 500 | 1018 |
| Clients | deals list status:won / lost --limit 20 | 650 to 706 | 76 to 97 | 1 to 20 | 2 to 31 |
| People | people list kind:K --limit 1 (4 kinds) | 4509 to 5074 | 3661 to 3726 | 1 | 1 |
| People | people review --limit 500 | 771 | 327 | 500 | 189 |
| People | people list kind:person --sort interactions --limit 5000 | 5060 | 4248 | 5000 | 1398 |
| Tasks | tasks list, tasks counts, projects list | 660 to 670 | 65 to 101 | 0 to 96 | 0 to 78 |
| Tasks | people list --limit 5000 | 5033 | 3931 | 5000 | 1398 |
| Digests | digest list | 708 | 33 | 0 | 0 |
| Logs | runs list --limit 500 | 732 | 108 | 500 | 453 |
| Settings | groups list | 675 | 63 | 1 | 4 |
| Search | search notes --limit 50 | 1391 to 2342 | 822 to 836 | 50 | 25 |

- `people list` costs 3.7 to 5.9 s through a server whatever the limit (1, 50 or 5000), the kind and the sort.
`--json=ref,name` left the answer at 1398 KB.
- Smaller limits save bytes, not much memo time: `actions list` 177 ms for 500 rows, 104 ms for 50; `deals list`
324 ms for 500, 282 ms for 100.
- A server is ready 62 to 155 ms after it starts. Its first request takes 184 to 323 ms, the next 30 to 55 ms. A bare
`memo --version` takes 20 ms.
- A CPU profile of one fresh `people list --limit 1` (5.3 s sampled): 1.8 s in SQLite, about 1.5 s in memo's
name checks (which names are automated, which are companies) and the query of each person's ties, 0.4 s garbage
collection. The work runs over every person before the limit applies.

### A page's reads at once (ms, median of 3)

| Page | calls | all fresh | pool of 2, rest fresh (Desktop today) | 1 server, queued | a warm server for each |
| --- | ---: | ---: | ---: | ---: | ---: |
| Memory | 7 | 1107 | 1081 | 1039 | 404 |
| Clients | 5 | 1418 | 1439 | 1022 | 561 |

The Clients row fires all 5 at once; Desktop runs `pipelines list` first and the other 4 after it.

### The interface (ms, median of 3)

| Click | frame | long tasks | frame, CPU x4 | long tasks, CPU x4 | DOM nodes |
| --- | ---: | ---: | ---: | ---: | ---: |
| Tools to Memory | 64 | 132 | 120 | 387 | 2374 |
| Tools to Tasks | 40 | 0 | 120 | 105 | 8656 |
| Tools to People | 72 | 0 | 64 | 55 | 1102 |
| Tools to Analytics, Logs, Agents, Search, Digests, Storage, Settings | 48 to 64 | 0 to 61 | | | 400 to 1958 |
| Memory: open a row, approve, filter | 24 to 48 | 0 | 64 | 0 | 2371 |
| Clients: sort, open a deal, board view | 32 to 40 | 0 | 128 to 136 | 114 to 119 | 3326 |
| Tasks: sort | 48 | 0 | | | 8656 |
| Tasks: open a task | 72 | 54 | 288 | 252 | 8969 |
| Tasks: board view | 96 | 89 | 312 | 295 | 8666 |
| People: sort, open a person, another kind | 24 to 32 | 0 | | | 1080 to 1244 |

## What it showed

1. **The time goes to memo, not to the interface.** In a production build, most clicks paint within 24 to 72 ms. The
pages wait 1 to 11 s for their data.
2. **`people list` is the worst single cost:** 3.7 to 4.2 s on a store of about 114,000 people, for any limit. People
runs it 6 times on open (a list and 5 kind counts, 2 at a time), Tasks once to name assignees. The synthetic
[people kinds](../../people/kinds/README.md) run found 0.5 to 0.6 s; a real store is 6 to 8 times slower.
3. **A warm server answers in 30 to 330 ms what a fresh process answers in 650 to 960 ms.** With a pool of 2 and the
rest sent to fresh processes, the Memory page gains nothing over all fresh (1081 against 1107 ms): the page waits
for the slowest call, and that is a fresh process.
4. **One warm server per call would bring Memory to about 0.4 s**, but that is 7 processes of about 165 MB each.
One server that runs several requests at once ([concurrent](../../command-server/concurrent/README.md)) is a single
JavaScript thread; that page found 6 to 54 % off the queued time, so about 0.5 to 1.0 s for Memory (inferred).
5. **A page under 150 ms on its first load is out of reach from memo alone:** the slowest warm reads are 278 ms
(Memory), 324 ms (Clients), 830 ms (Search), 3.7 s (People). Under 150 ms needs the page to draw its last answer
while it reads again, plus faster queries in cli-core.

## What was not measured

- The commands of the other CLIs (tg, max, wa, b24, zm, drive) that Logs, Agents, Analytics, Tools and Storage run.
- The Rust side of Desktop: JSON parsing, Tauri's bridge and the interface's schema checks on answers of 1 MB.
- The real app. Tauri on Linux draws with WebKitGTK, not Chrome, so the interface numbers are a lower bound; CPU x4
stands in for a slower engine. The mocks are smaller than the real store in places (Memory 300 rows against about
620; People holds pages of 50 against 5,000 rows).
- Desktop's dev build, which the owner was running.
- People's reads fired together, and Desktop's own `page_loads` replay (`cargo test --lib -- --ignored page_loads`):
too heavy for the free memory that day.
- Navigation to Clients in the browser build: the `clients=many` mock opens its stage map dialog, which covers the
sidebar.
87 changes: 87 additions & 0 deletions performance/desktop/page-speed/results.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,87 @@
# memo commands, real store, median of 3 (command, cold ms, served ms, rows, total, KB, status)
Memory: memories list --limit 500 672 79 2 6 ok
Memory: facts list --limit 500 652 88 4 3 ok
Memory: actions list --limit 500 --brief 718 177 500 359 ok
Memory: agents list --limit 30 631 86 10 2 ok
Memory: relationships list confirmed:false 686 128 100 77 ok
Memory: relationships list include:ended 905 278 100 71 ok
Memory: relationships roles list 653 66 16 1 ok
Clients: pipelines list kind:deal 663 91 6 12 ok
Clients: pipelines show <busiest> 799 212 2 ok
Clients: deals list open --limit 500 964 324 500 1018 ok
Clients: deals list won --limit 20 650 76 1 2 ok
Clients: deals list lost --limit 20 706 97 20 31 ok
People: people list kind:person --limit 1 5074 3726 1 113705 1 ok
People: people list kind:organization --limit 1 4563 3699 1 1853 1 ok
People: people list kind:automated --limit 1 4509 3661 1 3854 1 ok
People: people list kind:stranger --limit 1 692 35 0 failed
People: people list kind:unsorted --limit 1 4976 3700 1 30240 1 ok
People: people review --limit 500 771 327 500 189 ok
People: people list kind:person --sort interactions --limit 5000 5060 4248 5000 113705 1398 ok
Tasks: tasks list --sort priority --limit 500 678 62 0 failed
Tasks: people list --limit 5000 5033 3931 5000 113705 1398 ok
Tasks: projects list kind:project 664 65 3 1 ok
Digests: digest list 708 33 0 0 ok
Logs: runs list --limit 500 732 108 500 453 ok
Groups: groups list 675 63 1 4 ok
Tasks: tasks list --sort priority --limit 500 project:* 670 78 0 0 0 ok
Tasks: tasks counts project:* 665 73 1 ok
Tasks: tasks list --limit 500 (chats) 660 101 96 78 ok
Probe: people list --limit 50 4750 5945 50 113705 16 ok
Probe: people list kind:person --limit 50 4794 4006 50 113705 16 ok
Probe: people list kind:person --sort interactions --limit 50 4599 3812 50 113705 16 ok
Probe: people list --limit 5000 --json=ref,name 4872 4216 5000 113705 1398 ok
Probe: actions list --limit 50 --brief 755 104 50 53 ok
Probe: deals list open --limit 100 956 282 100 194 ok
Search: memo search meeting 570 32 0 failed

# fan out (Memory, Clients)
{"page":"Memory:","calls":7,"all_cold_ms":1107,"pool_1_ms":1078,"pool_2_ms":1081,"pool_7_ms":404,"one_server_queued_ms":1039}
{"page":"Clients:","calls":5,"all_cold_ms":1418,"pool_1_ms":1358,"pool_2_ms":1439,"pool_5_ms":561,"one_server_queued_ms":1022}

# clicks, CPU x1
{"scenario":"nav Tools→Memory (many)","frame_ms":64,"busy_ms":132,"longest_ms":132,"nodes":2374}
{"scenario":"nav Tools→Tasks (many)","frame_ms":40,"busy_ms":0,"longest_ms":0,"nodes":8656}
{"scenario":"nav Tools→People (many)","frame_ms":72,"busy_ms":0,"longest_ms":0,"nodes":1102}
{"scenario":"nav Tools→Analytics (showcase)","frame_ms":48,"busy_ms":0,"longest_ms":0,"nodes":965}
{"scenario":"nav Tools→Logs (showcase)","frame_ms":48,"busy_ms":0,"longest_ms":0,"nodes":1958}
{"scenario":"nav Tools→Agents (showcase)","frame_ms":48,"busy_ms":61,"longest_ms":61,"nodes":603}
{"scenario":"nav Tools→Search (showcase)","frame_ms":64,"busy_ms":0,"longest_ms":0,"nodes":409}
{"scenario":"nav Tools→Digests (showcase)","frame_ms":48,"busy_ms":0,"longest_ms":0,"nodes":400}
{"scenario":"nav Tools→Storage (showcase)","frame_ms":56,"busy_ms":0,"longest_ms":0,"nodes":1111}
{"scenario":"nav Tools→Settings (showcase)","frame_ms":56,"busy_ms":55,"longest_ms":55,"nodes":1231}
{"scenario":"Memory: open a row","frame_ms":40,"busy_ms":0,"longest_ms":0,"nodes":2371}
{"scenario":"Memory: approve one","frame_ms":48,"busy_ms":0,"longest_ms":0,"nodes":2379}
{"scenario":"Memory: Kind filter","frame_ms":40,"busy_ms":0,"longest_ms":0,"nodes":2371}
{"scenario":"Tasks: board view","frame_ms":96,"busy_ms":89,"longest_ms":89,"nodes":8666}
{"scenario":"People: open a person","frame_ms":32,"busy_ms":0,"longest_ms":0,"nodes":1244}
{"scenario":"People: another kind","frame_ms":32,"busy_ms":0,"longest_ms":0,"nodes":1080}
{"scenario":"Memory: approve one","frame_ms":24,"busy_ms":0,"longest_ms":0,"nodes":2379}
{"scenario":"Memory: Kind filter","frame_ms":24,"busy_ms":0,"longest_ms":0,"nodes":2371}
{"scenario":"Clients: sort by amount","frame_ms":40,"busy_ms":0,"longest_ms":0,"nodes":3326}
{"scenario":"Clients: open a deal","frame_ms":32,"busy_ms":0,"longest_ms":0,"nodes":3375}
{"scenario":"Tasks: sort by priority","frame_ms":48,"busy_ms":0,"longest_ms":0,"nodes":8656}
{"scenario":"Tasks: open a task","frame_ms":72,"busy_ms":54,"longest_ms":54,"nodes":8969}
{"scenario":"People: sort by last contact","frame_ms":24,"busy_ms":0,"longest_ms":0,"nodes":1100}

# clicks, CPU x4
{"scenario":"nav Tools→Memory (many)","frame_ms":120,"busy_ms":387,"longest_ms":318,"nodes":2374}
{"scenario":"nav Tools→Tasks (many)","frame_ms":120,"busy_ms":105,"longest_ms":105,"nodes":8656}
{"scenario":"nav Tools→People (many)","frame_ms":64,"busy_ms":55,"longest_ms":55,"nodes":1102}
{"scenario":"Memory: open a row","frame_ms":64,"busy_ms":0,"longest_ms":0,"nodes":2371}
{"scenario":"Clients: sort by amount","frame_ms":128,"busy_ms":114,"longest_ms":114,"nodes":3326}
{"scenario":"Clients: board view","frame_ms":136,"busy_ms":119,"longest_ms":118,"nodes":3326}
{"scenario":"Tasks: open a task","frame_ms":288,"busy_ms":252,"longest_ms":252,"nodes":8969}
{"scenario":"Tasks: board view","frame_ms":312,"busy_ms":295,"longest_ms":295,"nodes":8666}

# first request on a fresh server (ready ms, first ms, second ms), 3 fresh servers each
agents list --limit 30: 97/319/33 75/190/36 74/189/31
memories list --limit 500: 105/190/31 115/224/34 67/191/30
deals list --limit 20 status:won: 62/323/54 82/199/45 97/202/45
tasks counts project:*: 85/198/31 96/184/32 61/188/55
search notes --limit 50 meeting: 84/1309/822 69/1108/836 75/1208/831

# other
bare memo --version: 20 ms (3 runs)
search notes --limit 50 meeting, cold process: 2342, 1423, 1391 ms; 50 rows, 25 KB
cpu profile, cold people list --limit 1: 5324 ms sampled; 1797 in SQLite, about 1500 in name checks and tie reasons, 411 GC
Loading
Loading