Skip to content

fix(idle-cpu): stop home-screen animation loop after 3s of inactivity - #31

Merged
Blankeos merged 2 commits into
mainfrom
perf/cpu-optimizations
Aug 25, 2026
Merged

fix(idle-cpu): stop home-screen animation loop after 3s of inactivity#31
Blankeos merged 2 commits into
mainfrom
perf/cpu-optimizations

Conversation

@Blankeos

@Blankeos Blankeos commented Aug 25, 2026

Copy link
Copy Markdown
Owner

Fixes #28

The Home cursor blink was pinning a ~60fps render loop indefinitely, keeping the event loop polling at 16ms even when idle. This caused noticeable CPU usage compared to peer CLIs.

Changes: - Track last_user_activity timestamp; only animate Home for 3s after activity - Raise idle poll from 250ms to 30s so the event loop parks when truly idle - Increase autocomplete indexer poll interval from 1s to 30s - Add bench-perf.py script and PERF.md to measure startup / idle-CPU vs peers - Add regression test home_animation_freezes_after_idle

@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Aug 25, 2026

Copy link
Copy Markdown

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

@Blankeos

Blankeos commented Aug 25, 2026 •
edited
Loading

Copy link
Copy Markdown
Owner Author

Performance notes

Startup + idle-CPU vs peer agent CLIs. Update this file after meaningful runs.

just bench-perf
# Prefer the release binary for fair idle numbers:
cargo build --release && PATH="./target/release:$PATH" just bench-perf

Script: scripts/bench-perf.py · recipe: just bench-perf

Section What Tool
A --version startup hyperfine
B TUI first frame PTY + terminal-query replies
C Idle CPU after settle process-tree ps (macOS) / /proc (Linux)

Related: #28 (idle CPU peg).

Latest

2026-08-26 · darwin · Carlos-MacBook-Pro.local · cwd = repo
settle=3s · sample=8s · interval=0.25s · version runs=50

A) --version (lower is better)

Agent mean ± σ min … max
crabcode 8.47 ms ± 1.91 6.46 … 18.29
codex 12.71 ms ± 1.31 11.11 … 18.57
grok 12.86 ms ± 1.49 9.66 … 15.58
opencode 367.74 ms ± 13.62 352.03 … 430.16

crabcode is 1.50× faster than codex, 1.52× than grok, ~43× than opencode.

B) TUI first frame (lower is better)

Agent mean best … worst
codex 70.7 ms 50.4 … 110.0
crabcode 123.1 ms 106.2 … 156.8
opencode 1097.4 ms 973.4 … 1342.9
grok 1537.8 ms 1347.4 … 1650.7

C) Idle CPU after settle (lower is better)

Agent mean p50 p95 max RSS
crabcode 0.1% 0.0% 0.2% 1.0% 51.9 MB
grok 0.9% 0.9% 1.6% 1.7% 97.9 MB
codex 1.0% 0.6% 2.6% 5.1% 192.4 MB
opencode 5.6% 1.5% 27.3% 33.9% 990.1 MB

Verdict: crabcode best (or tied) on idle CPU.

Tip: use a release binary and --settle 5 --sample 10+. Debug builds / short settle can still show Home blink (~60fps) and inflate idle %.

Raw dump 
A) --version startup (hyperfine)
  crabcode     8.47 ms ±  1.91  (min 6.46, max 18.29, n=50)
  codex       12.71 ms ±  1.31  (min 11.11, max 18.57, n=50)
  grok        12.86 ms ±  1.49  (min 9.66, max 15.58, n=50)
  opencode   367.74 ms ± 13.62  (min 352.03, max 430.16, n=50)

B) TUI first frame
  crabcode   first_frame   123.1 ms  (best 106.2, worst 156.8)
  codex      first_frame    70.7 ms  (best 50.4, worst 110.0)
  grok       first_frame  1537.8 ms  (best 1347.4, worst 1650.7)
  opencode   first_frame  1097.4 ms  (best 973.4, worst 1342.9)

C) Idle CPU (settle=3s, sample=8s)
  crabcode   cpu mean=  0.1%  p50=  0.0%  p95=  0.2%  max=  1.0%  rss=  51.9MB  procs=1  n=26
  codex      cpu mean=  1.0%  p50=  0.6%  p95=  2.6%  max=  5.1%  rss= 192.4MB  procs=1  n=26
  grok       cpu mean=  0.9%  p50=  0.9%  p95=  1.6%  max=  1.7%  rss=  97.9MB  procs=1  n=26
  opencode   cpu mean=  5.6%  p50=  1.5%  p95= 27.3%  max= 33.9%  rss= 990.1MB  procs=1  n=26

History

2026-08-26 · darwin · `Carlos-MacBook-Pro.local` · cwd = repo

2026-08-26 · darwin · Carlos-MacBook-Pro.local · cwd = repo
settle=3s · sample=8s · interval=0.25s · version runs=50

A) --version (lower is better)

Agent mean ± σ min … max
crabcode 7.09 ms ± 0.83 5.84 … 9.76
codex 12.53 ms ± 0.96 11.56 … 15.46
grok 12.60 ms ± 2.40 10.23 … 18.87
opencode 361.40 ms ± 6.30 352.33 … 378.60

crabcode is 1.77× faster than codex, 1.78× than grok, ~51× than opencode.

B) TUI first frame (lower is better)

Agent mean best … worst
codex 50.8 ms 50.1 … 51.4
crabcode 103.4 ms 101.8 … 105.2
opencode 1000.0 ms 907.2 … 1126.8
grok 1684.7 ms 1394.4 … 2103.6

C) Idle CPU after settle (lower is better)

Agent mean p50 p95 max RSS
crabcode 0.1% 0.0% 0.3% 1.7% 51.8 MB
codex 0.6% 0.4% 2.3% 4.3% 195.5 MB
grok 1.0% 1.0% 1.5% 1.8% 103.6 MB
opencode 7.1% 2.6% 18.3% 84.4% 1020.1 MB

Verdict: crabcode best (or tied) on idle CPU.

Tip: use a release binary and --settle 5 --sample 10+. Debug builds / short settle can still show Home blink (~60fps) and inflate idle %.

Raw dump 
A) --version startup (hyperfine)
  crabcode     7.09 ms ±  0.83  (min 5.84, max 9.76, n=50)
  codex       12.53 ms ±  0.96  (min 11.56, max 15.46, n=50)
  grok        12.60 ms ±  2.40  (min 10.23, max 18.87, n=50)
  opencode   361.40 ms ±  6.30  (min 352.33, max 378.60, n=50)

B) TUI first frame
  crabcode   first_frame   103.4 ms  (best 101.8, worst 105.2)
  codex      first_frame    50.8 ms  (best 50.1, worst 51.4)
  grok       first_frame  1684.7 ms  (best 1394.4, worst 2103.6)
  opencode   first_frame  1000.0 ms  (best 907.2, worst 1126.8)

C) Idle CPU (settle=3s, sample=8s)
  crabcode   cpu mean=  0.1%  p50=  0.0%  p95=  0.3%  max=  1.7%  rss=  51.8MB  procs=1  n=25
  codex      cpu mean=  0.6%  p50=  0.4%  p95=  2.3%  max=  4.3%  rss= 195.5MB  procs=1  n=25
  grok       cpu mean=  1.0%  p50=  1.0%  p95=  1.5%  max=  1.8%  rss= 103.6MB  procs=1  n=25
  opencode   cpu mean=  7.1%  p50=  2.6%  p95= 18.3%  max= 84.4%  rss=1020.1MB  procs=1  n=25
2026-08-26 · darwin · `Carlos-MacBook-Pro.local` · cwd = repo

2026-08-26 · darwin · Carlos-MacBook-Pro.local · cwd = repo
settle=3s · sample=8s · interval=0.25s · version runs=50

A) --version (lower is better)

Agent mean ± σ min … max
crabcode 7.66 ms ± 0.67 6.42 … 9.14
grok 12.97 ms ± 2.00 10.31 … 20.68
codex 14.28 ms ± 1.49 11.78 … 19.44
opencode 366.29 ms ± 7.38 356.21 … 388.76

crabcode is 1.69× faster than grok, 1.86× than codex, ~48× than opencode.

B) TUI first frame (lower is better)

Agent mean best … worst
codex 53.0 ms 51.3 … 55.1
crabcode 121.5 ms 101.9 … 159.9
opencode 1020.7 ms 967.7 … 1126.1
grok 1771.7 ms 1505.9 … 2145.1

C) Idle CPU after settle (lower is better) · #28

Agent mean p50 p95 max RSS
codex 0.0% 0.0% 0.0% 0.1% 48.8 MB
grok 0.9% 0.9% 1.4% 1.5% 97.4 MB
crabcode 3.3% 3.3% 3.9% 4.0% 51.9 MB
opencode 5.2% 2.4% 13.3% 42.9% 1014.8 MB

Verdict: loses idle-CPU to codex + grok on this run. Aim: mean ≤ best peer (and ≪ 100% on Linux for #28).

Tip: use a release binary and --settle 5 --sample 10+. Debug builds / short settle can still show Home blink (~60fps) and inflate idle %.

Raw dump 
A) --version startup (hyperfine)
  crabcode      7.66 ms ±  0.67  (min 6.42, max 9.14, n=50)
  codex        14.28 ms ±  1.49  (min 11.78, max 19.44, n=50)
  grok         12.97 ms ±  2.00  (min 10.31, max 20.68, n=50)
  opencode    366.29 ms ±  7.38  (min 356.21, max 388.76, n=50)

B) TUI first frame
  crabcode   first_frame   121.5 ms  (best 101.9, worst 159.9, n=3)
  codex      first_frame    53.0 ms  (best 51.3, worst 55.1, n=3)
  grok       first_frame  1771.7 ms  (best 1505.9, worst 2145.1, n=3)
  opencode   first_frame  1020.7 ms  (best 967.7, worst 1126.1, n=3)

C) Idle CPU (settle=3.0s, sample=8.0s)
  crabcode   cpu mean=  3.3%  p50=  3.3%  p95=  3.9%  max=  4.0%  rss=  51.9MB  procs=1  n=25
  codex      cpu mean=  0.0%  p50=  0.0%  p95=  0.0%  max=  0.1%  rss=  48.8MB  procs=1  n=25
  grok       cpu mean=  0.9%  p50=  0.9%  p95=  1.4%  max=  1.5%  rss=  97.4MB  procs=1  n=25
  opencode   cpu mean=  5.2%  p50=  2.4%  p95= 13.3%  max= 42.9%  rss=1014.8MB  procs=1  n=25

How to refresh

  1. cargo build --release
  2. PATH="./target/release:$PATH" just bench-perf --settle 5 --sample 10 --runs 50
  3. Answer y to Add this to PERF.md? (or pass --write-perf / --no-write-perf).
  4. Optional JSON: --json-out /tmp/crabcode-bench-perf.json

@Blankeos
Blankeos force-pushed the perf/cpu-optimizations branch from cf03202 to 3cb0a45 Compare August 25, 2026 17:27
@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Aug 25, 2026 •
edited
Loading

Copy link
Copy Markdown

Deploying crabcode with  Cloudflare Pages  Cloudflare Pages

Latest commit: 2ad67c3
Status:⚡️  Build in progress...

View logs

@Blankeos
fix(idle-cpu): stop home-screen animation loop after 3s of inactivity (…
1312287 
…#28)

The Home cursor blink was pinning a ~60fps render loop indefinitely, keeping the event loop polling at 16ms even when idle. This caused noticeable CPU usage compared to peer CLIs.

Changes: - Track `last_user_activity` timestamp; only animate Home for 3s after activity - Raise idle poll from 250ms to 30s so the event loop parks when truly idle - Increase autocomplete indexer poll interval from 1s to 30s - Add `bench-perf.py` script and `PERF.md` to measure startup / idle-CPU vs peers - Add regression test `home_animation_freezes_after_idle`
@Blankeos
Blankeos force-pushed the perf/cpu-optimizations branch from 3cb0a45 to 1312287 Compare August 25, 2026 17:47
@Blankeos

Blankeos commented Aug 25, 2026 •
edited
Loading

Copy link
Copy Markdown
Owner Author

BEFORE

A) --version (lower is better)
crabcode 7.88 ms (1.00× best) ←
grok 10.50 ms (1.33× best)
codex 11.94 ms (1.52× best)
opencode 374.19 ms (47.50× best)

B) TUI first frame (lower is better)
codex 50.9 ms (1.00× best)
crabcode 103.5 ms (2.03× best) ←
opencode 1023.1 ms (20.09× best)
grok 1660.6 ms (32.61× best)

C) Idle CPU after settle (lower is better)
codex 0.6% mean p95= 2.1% rss= 207.6MB (1.00× best)
grok 1.0% mean p95= 1.8% rss= 96.2MB (1.60× best)
crabcode 1.5% mean p95= 2.8% rss= 49.5MB (2.53× best) ←
opencode 8.9% mean p95= 28.8% rss=1039.9MB (14.67× best)

verdict: crabcode loses idle-CPU to: codex, grok
aim: mean ≤ best peer (and ≪ 100% on Linux)

Gonna try chasing Codex's first paint. brb

@Blankeos

Blankeos commented Aug 25, 2026

Copy link
Copy Markdown
Owner Author

AFTER CHASING CODEX FIRST PAINT STATS (crazy lol)

Performance notes

Startup + idle-CPU vs peer agent CLIs. Update this file after meaningful runs.

just bench-perf
# Prefer the release binary for fair idle numbers:
cargo build --release && PATH="./target/release:$PATH" just bench-perf

Script: scripts/bench-perf.py · recipe: just bench-perf

Section What Tool
A --version startup hyperfine
B TUI first frame PTY + terminal-query replies
C Idle CPU after settle process-tree ps (macOS) / /proc (Linux)

Related: #28 (idle CPU peg).

Latest

2026-08-26 · darwin · Carlos-MacBook-Pro.local · cwd = repo
settle=3s · sample=8s · interval=0.25s · version runs=50

A) --version (lower is better)

Agent mean ± σ min … max
crabcode 7.71 ms ± 0.88 6.82 … 10.81
grok 10.45 ms ± 0.57 9.55 … 12.21
codex 12.61 ms ± 0.93 10.24 … 14.84
opencode 355.67 ms ± 7.51 348.53 … 381.50

crabcode is 1.35× faster than grok, 1.64× than codex, ~46× than opencode.

B) TUI first frame (lower is better)

Agent mean best … worst
codex 50.9 ms 50.5 … 51.6
crabcode 54.9 ms 54.8 … 55.0
opencode 1023.7 ms 969.5 … 1125.0
grok 1743.3 ms 1554.0 … 1936.3

C) Idle CPU after settle (lower is better)

Agent mean p50 p95 max RSS
crabcode 0.1% 0.0% 0.3% 1.3% 52.4 MB
grok 1.0% 1.0% 1.4% 1.5% 95.9 MB
codex 1.0% 0.7% 2.3% 4.6% 204.3 MB
opencode 8.8% 4.8% 28.5% 49.3% 1027.5 MB

Verdict: crabcode best (or tied) on idle CPU.

Tip: use a release binary and --settle 5 --sample 10+. Debug builds / short settle can still show Home blink (~60fps) and inflate idle %.

Raw dump 
A) --version startup (hyperfine)
  crabcode     7.71 ms ±  0.88  (min 6.82, max 10.81, n=50)
  codex       12.61 ms ±  0.93  (min 10.24, max 14.84, n=50)
  grok        10.45 ms ±  0.57  (min 9.55, max 12.21, n=50)
  opencode   355.67 ms ±  7.51  (min 348.53, max 381.50, n=50)

B) TUI first frame
  crabcode   first_frame    54.9 ms  (best 54.8, worst 55.0)
  codex      first_frame    50.9 ms  (best 50.5, worst 51.6)
  grok       first_frame  1743.3 ms  (best 1554.0, worst 1936.3)
  opencode   first_frame  1023.7 ms  (best 969.5, worst 1125.0)

C) Idle CPU (settle=3s, sample=8s)
  crabcode   cpu mean=  0.1%  p50=  0.0%  p95=  0.3%  max=  1.3%  rss=  52.4MB  procs=1  n=25
  codex      cpu mean=  1.0%  p50=  0.7%  p95=  2.3%  max=  4.6%  rss= 204.3MB  procs=1  n=25
  grok       cpu mean=  1.0%  p50=  1.0%  p95=  1.4%  max=  1.5%  rss=  95.9MB  procs=1  n=25
  opencode   cpu mean=  8.8%  p50=  4.8%  p95= 28.5%  max= 49.3%  rss=1027.5MB  procs=1  n=26

History

2026-08-26 · darwin · `Carlos-MacBook-Pro.local` · cwd = repo

2026-08-26 · darwin · Carlos-MacBook-Pro.local · cwd = repo
settle=3s · sample=8s · interval=0.25s · version runs=50

A) --version (lower is better)

Agent mean ± σ min … max
crabcode 8.98 ms ± 1.33 6.53 … 14.01
grok 12.74 ms ± 1.49 10.35 … 17.83
codex 15.24 ms ± 1.18 12.58 … 18.59
opencode 360.30 ms ± 7.27 349.46 … 385.32

crabcode is 1.42× faster than grok, 1.70× than codex, ~40× than opencode.

B) TUI first frame (lower is better)

Agent mean best … worst
codex 50.7 ms 50.3 … 50.9
crabcode 54.6 ms 53.7 … 55.1
opencode 914.3 ms 908.1 … 924.6
grok 1518.7 ms 1303.7 … 1843.2

C) Idle CPU after settle (lower is better)

Agent mean p50 p95 max RSS
crabcode 0.1% 0.0% 0.3% 1.4% 52.3 MB
grok 0.8% 0.9% 1.2% 2.1% 96.8 MB
codex 0.9% 0.7% 2.3% 2.6% 209.5 MB
opencode 5.2% 2.5% 21.7% 27.0% 999.1 MB

Verdict: crabcode best (or tied) on idle CPU.

Tip: use a release binary and --settle 5 --sample 10+. Debug builds / short settle can still show Home blink (~60fps) and inflate idle %.

Raw dump 
A) --version startup (hyperfine)
  crabcode     8.98 ms ±  1.33  (min 6.53, max 14.01, n=50)
  codex       15.24 ms ±  1.18  (min 12.58, max 18.59, n=50)
  grok        12.74 ms ±  1.49  (min 10.35, max 17.83, n=50)
  opencode   360.30 ms ±  7.27  (min 349.46, max 385.32, n=50)

B) TUI first frame
  crabcode   first_frame    54.6 ms  (best 53.7, worst 55.1)
  codex      first_frame    50.7 ms  (best 50.3, worst 50.9)
  grok       first_frame  1518.7 ms  (best 1303.7, worst 1843.2)
  opencode   first_frame   914.3 ms  (best 908.1, worst 924.6)

C) Idle CPU (settle=3s, sample=8s)
  crabcode   cpu mean=  0.1%  p50=  0.0%  p95=  0.3%  max=  1.4%  rss=  52.3MB  procs=1  n=26
  codex      cpu mean=  0.9%  p50=  0.7%  p95=  2.3%  max=  2.6%  rss= 209.5MB  procs=1  n=26
  grok       cpu mean=  0.8%  p50=  0.9%  p95=  1.2%  max=  2.1%  rss=  96.8MB  procs=1  n=26
  opencode   cpu mean=  5.2%  p50=  2.5%  p95= 21.7%  max= 27.0%  rss= 999.1MB  procs=1  n=26
2026-08-26 · darwin · `Carlos-MacBook-Pro.local` · cwd = repo

2026-08-26 · darwin · Carlos-MacBook-Pro.local · cwd = repo
settle=3s · sample=8s · interval=0.25s · version runs=50

A) --version (lower is better)

Agent mean ± σ min … max
crabcode 8.47 ms ± 1.91 6.46 … 18.29
codex 12.71 ms ± 1.31 11.11 … 18.57
grok 12.86 ms ± 1.49 9.66 … 15.58
opencode 367.74 ms ± 13.62 352.03 … 430.16

crabcode is 1.50× faster than codex, 1.52× than grok, ~43× than opencode.

B) TUI first frame (lower is better)

Agent mean best … worst
codex 70.7 ms 50.4 … 110.0
crabcode 123.1 ms 106.2 … 156.8
opencode 1097.4 ms 973.4 … 1342.9
grok 1537.8 ms 1347.4 … 1650.7

C) Idle CPU after settle (lower is better)

Agent mean p50 p95 max RSS
crabcode 0.1% 0.0% 0.2% 1.0% 51.9 MB
grok 0.9% 0.9% 1.6% 1.7% 97.9 MB
codex 1.0% 0.6% 2.6% 5.1% 192.4 MB
opencode 5.6% 1.5% 27.3% 33.9% 990.1 MB

Verdict: crabcode best (or tied) on idle CPU.

Tip: use a release binary and --settle 5 --sample 10+. Debug builds / short settle can still show Home blink (~60fps) and inflate idle %.

Raw dump 
A) --version startup (hyperfine)
  crabcode     8.47 ms ±  1.91  (min 6.46, max 18.29, n=50)
  codex       12.71 ms ±  1.31  (min 11.11, max 18.57, n=50)
  grok        12.86 ms ±  1.49  (min 9.66, max 15.58, n=50)
  opencode   367.74 ms ± 13.62  (min 352.03, max 430.16, n=50)

B) TUI first frame
  crabcode   first_frame   123.1 ms  (best 106.2, worst 156.8)
  codex      first_frame    70.7 ms  (best 50.4, worst 110.0)
  grok       first_frame  1537.8 ms  (best 1347.4, worst 1650.7)
  opencode   first_frame  1097.4 ms  (best 973.4, worst 1342.9)

C) Idle CPU (settle=3s, sample=8s)
  crabcode   cpu mean=  0.1%  p50=  0.0%  p95=  0.2%  max=  1.0%  rss=  51.9MB  procs=1  n=26
  codex      cpu mean=  1.0%  p50=  0.6%  p95=  2.6%  max=  5.1%  rss= 192.4MB  procs=1  n=26
  grok       cpu mean=  0.9%  p50=  0.9%  p95=  1.6%  max=  1.7%  rss=  97.9MB  procs=1  n=26
  opencode   cpu mean=  5.6%  p50=  1.5%  p95= 27.3%  max= 33.9%  rss= 990.1MB  procs=1  n=26
2026-08-26 · darwin · `Carlos-MacBook-Pro.local` · cwd = repo

2026-08-26 · darwin · Carlos-MacBook-Pro.local · cwd = repo
settle=3s · sample=8s · interval=0.25s · version runs=50

A) --version (lower is better)

Agent mean ± σ min … max
crabcode 7.09 ms ± 0.83 5.84 … 9.76
codex 12.53 ms ± 0.96 11.56 … 15.46
grok 12.60 ms ± 2.40 10.23 … 18.87
opencode 361.40 ms ± 6.30 352.33 … 378.60

crabcode is 1.77× faster than codex, 1.78× than grok, ~51× than opencode.

B) TUI first frame (lower is better)

Agent mean best … worst
codex 50.8 ms 50.1 … 51.4
crabcode 103.4 ms 101.8 … 105.2
opencode 1000.0 ms 907.2 … 1126.8
grok 1684.7 ms 1394.4 … 2103.6

C) Idle CPU after settle (lower is better)

Agent mean p50 p95 max RSS
crabcode 0.1% 0.0% 0.3% 1.7% 51.8 MB
codex 0.6% 0.4% 2.3% 4.3% 195.5 MB
grok 1.0% 1.0% 1.5% 1.8% 103.6 MB
opencode 7.1% 2.6% 18.3% 84.4% 1020.1 MB

Verdict: crabcode best (or tied) on idle CPU.

Tip: use a release binary and --settle 5 --sample 10+. Debug builds / short settle can still show Home blink (~60fps) and inflate idle %.

Raw dump 
A) --version startup (hyperfine)
  crabcode     7.09 ms ±  0.83  (min 5.84, max 9.76, n=50)
  codex       12.53 ms ±  0.96  (min 11.56, max 15.46, n=50)
  grok        12.60 ms ±  2.40  (min 10.23, max 18.87, n=50)
  opencode   361.40 ms ±  6.30  (min 352.33, max 378.60, n=50)

B) TUI first frame
  crabcode   first_frame   103.4 ms  (best 101.8, worst 105.2)
  codex      first_frame    50.8 ms  (best 50.1, worst 51.4)
  grok       first_frame  1684.7 ms  (best 1394.4, worst 2103.6)
  opencode   first_frame  1000.0 ms  (best 907.2, worst 1126.8)

C) Idle CPU (settle=3s, sample=8s)
  crabcode   cpu mean=  0.1%  p50=  0.0%  p95=  0.3%  max=  1.7%  rss=  51.8MB  procs=1  n=25
  codex      cpu mean=  0.6%  p50=  0.4%  p95=  2.3%  max=  4.3%  rss= 195.5MB  procs=1  n=25
  grok       cpu mean=  1.0%  p50=  1.0%  p95=  1.5%  max=  1.8%  rss= 103.6MB  procs=1  n=25
  opencode   cpu mean=  7.1%  p50=  2.6%  p95= 18.3%  max= 84.4%  rss=1020.1MB  procs=1  n=25
2026-08-26 · darwin · `Carlos-MacBook-Pro.local` · cwd = repo

2026-08-26 · darwin · Carlos-MacBook-Pro.local · cwd = repo
settle=3s · sample=8s · interval=0.25s · version runs=50

A) --version (lower is better)

Agent mean ± σ min … max
crabcode 7.66 ms ± 0.67 6.42 … 9.14
grok 12.97 ms ± 2.00 10.31 … 20.68
codex 14.28 ms ± 1.49 11.78 … 19.44
opencode 366.29 ms ± 7.38 356.21 … 388.76

crabcode is 1.69× faster than grok, 1.86× than codex, ~48× than opencode.

B) TUI first frame (lower is better)

Agent mean best … worst
codex 53.0 ms 51.3 … 55.1
crabcode 121.5 ms 101.9 … 159.9
opencode 1020.7 ms 967.7 … 1126.1
grok 1771.7 ms 1505.9 … 2145.1

C) Idle CPU after settle (lower is better) · #28

Agent mean p50 p95 max RSS
codex 0.0% 0.0% 0.0% 0.1% 48.8 MB
grok 0.9% 0.9% 1.4% 1.5% 97.4 MB
crabcode 3.3% 3.3% 3.9% 4.0% 51.9 MB
opencode 5.2% 2.4% 13.3% 42.9% 1014.8 MB

Verdict: loses idle-CPU to codex + grok on this run. Aim: mean ≤ best peer (and ≪ 100% on Linux for #28).

Tip: use a release binary and --settle 5 --sample 10+. Debug builds / short settle can still show Home blink (~60fps) and inflate idle %.

Raw dump 
A) --version startup (hyperfine)
  crabcode      7.66 ms ±  0.67  (min 6.42, max 9.14, n=50)
  codex        14.28 ms ±  1.49  (min 11.78, max 19.44, n=50)
  grok         12.97 ms ±  2.00  (min 10.31, max 20.68, n=50)
  opencode    366.29 ms ±  7.38  (min 356.21, max 388.76, n=50)

B) TUI first frame
  crabcode   first_frame   121.5 ms  (best 101.9, worst 159.9, n=3)
  codex      first_frame    53.0 ms  (best 51.3, worst 55.1, n=3)
  grok       first_frame  1771.7 ms  (best 1505.9, worst 2145.1, n=3)
  opencode   first_frame  1020.7 ms  (best 967.7, worst 1126.1, n=3)

C) Idle CPU (settle=3.0s, sample=8.0s)
  crabcode   cpu mean=  3.3%  p50=  3.3%  p95=  3.9%  max=  4.0%  rss=  51.9MB  procs=1  n=25
  codex      cpu mean=  0.0%  p50=  0.0%  p95=  0.0%  max=  0.1%  rss=  48.8MB  procs=1  n=25
  grok       cpu mean=  0.9%  p50=  0.9%  p95=  1.4%  max=  1.5%  rss=  97.4MB  procs=1  n=25
  opencode   cpu mean=  5.2%  p50=  2.4%  p95= 13.3%  max= 42.9%  rss=1014.8MB  procs=1  n=25

How to refresh

  1. cargo build --release
  2. PATH="./target/release:$PATH" just bench-perf --settle 5 --sample 10 --runs 50
  3. Answer y to Add this to PERF.md? (or pass --write-perf / --no-write-perf).
  4. Optional JSON: --json-out /tmp/crabcode-bench-perf.json

@Blankeos
perf: improve tui-paint-speed and reduce idle CPU
2ad67c3 
- Split `App::new` into minimal `new_shell` + deferred `ensure_startup_hydrated` for faster first paint (~50ms saved) - Skip blocking `supports_keyboard_enhancement()` CSI probe (always push flags; opt out via `CRABCODE_DISABLE_KEYBOARD_ENHANCEMENT`) - Defer `SessionManager` SQLite history load until after first draw - Reduce idle poll from 250ms to 30s to fix idle CPU peg (#28) - Resolve theme before first paint to avoid builtin flash - Increase autocomplete indexer poll interval from 1s to 30s - Add `bench-perf.py` script and `just bench-perf` recipe for startup/idle-CPU benchmarking vs peer CLIs - Add `PERF.md` with comparative benchmarks (hyperfine + PTY) against codex, grok, opencode
@Blankeos
Blankeos merged commit a8a1bd6 into main Aug 25, 2026
1 of 2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

High idle CPU: one worker thread pegs a full core per instance

1 participant

@Blankeos