High CPU usage

[utsahi]

Checklist

• I agree to communicate with the author myself (not AI-generated).
• I've read the README's Filing issues section.
• I'm running the latest versions (fill in below).
  • agent-shell: 0.46.1
  • acp.el:
  • ACP package (e.g. claude-code-acp):
  • Agent CLI (e.g. claude, gemini):
• For requesting new agent support, I'm including a link to the ACP-capable agent or related ACP package.
• For issues, I'm including ACP traffic (as per README).

Repro step: enter the following prompt: "emit 10 markdown tables with 10 rows and 10 columns each"

Observed behavior

I narrowed it down to markdown-overlays-put. As each row of the table is streamed, it looks like the overlay is being recalculated from the beginning of the current item.

Suggestion

1. Consider calculating the overlay after the entire item has been streamed.
2. Or provide a built in option to disable overlays and an option to recalculate the overlays

Workaround

To unblock myself, I worked around by evaling (defun markdown-overlays-put ())

[xenodium]

Thanks for filing.

We added some table optimizations recently. Do you have the latest shell-maker (currently 0.89.1)? If so, it looks like we may need more optimizations.

cc: @ewilderj

[utsahi]

Yes, I have this version of shell-maker

;; Package-Version: 0.89.1
;; Package-Revision: 86fbe695085f

[ewilderj]

I spent some time looking at this tonight, purely from the shell-maker side of things. I didn't attempt to change anything about how agent-shell calls shell-maker, ie. repeatedly during streaming. But I did investigate a couple of performance optimizations:

• make regexes constants
• caching tables when rendered so they're not repeatedly redrawn with each call to markdown-overlays-put

You can find that work at https://github.com/ewilderj/shell-maker/tree/feature/perf-optimization (@utsahi if you have the inclination to check it out and try using it instead, I would be interested in your experience)

I wrote a (headless) benchmark which gives the following speedups for this branch:

Streaming benchmark results (shell-maker table rendering)

Simulates the streaming pattern: rows arrive one at a time, markdown-overlays-put called after each chunk.

Pathological case — 10 tables × 10 rows × 10 cols (130 chunks):

Metric	main	optimized	Speedup
Total	3157 ms	582 ms	5.4×
Per-chunk avg	24.3 ms	4.5 ms	5.4×
P50 (median)	25.1 ms	3.0 ms	8.4×
P95	44.2 ms	11.2 ms	3.9×
P99 (worst)	50.9 ms	12.9 ms	3.9×

Typical case — prose response with 1 table (5 rows × 3 cols, 25 chunks):

Metric	main	optimized	Speedup
Total	14.7 ms	4.1 ms	3.6×
P50 (median)	0.34 ms	0.15 ms	2.3×
P99 (worst)	7.73 ms	0.47 ms	16×

For any long-lived session, the sheer number of overlays in the buffer also present a performance problem, so I'm also poking at that as a separate problem.

@xenodium let me know if you want me to propose a PR for this speedup work.

[ewilderj]

OK I have an update on the large amount of overlays: they are definitely contributing to the Emacs CPU usage and sluggishness issue. Over in xenodium/shell-maker#22 I list the results of my investigations there.

If anyone following this issue wants to test and verify speedups, then https://github.com/ewilderj/shell-maker/tree/feature/perf-combined is the branch of shell-maker to test with.

Will await advice on how you'd like to structure PRs for these speedups.

[utsahi]

Hi @ewilderj , thanks for your quick response. These fixes definitely improved the speed. The CPU usage is consistently between 80-100% after around the 4th table. But the higher CPU usage now lasts for shorter period due to the speed improvements.

I tested with this config and relaunching emacs.

(use-package acp
  :ensure nil
  :vc (:url "https://github.com/xenodium/acp.el.git" :rev :newest))

(use-package shell-maker
  :ensure t
  :vc (:url "https://github.com/ewilderj/shell-maker.git" :rev "0e587dbc3e36a390442baa12b11e8876651e0bef"))

(use-package agent-shell
  :ensure nil
  :vc (:url "https://github.com/xenodium/agent-shell.git" :rev :newest))

For comparison, with (defun markdown-overlays-put ()) I'm seeing CPU usage vary between 20-40%.

[xenodium]

Hi @utsahi

I've taken some of @ewilderj's changes and combined with consolidated caching in xenodium/shell-maker#23 can you try that too please?

[utsahi]

That brought down the CPU usage to 60-80%, and ~90% towards the end. Attaching a screen capture.

Edit: file uploaded, but its not playing here. You'll have to download it. Also, I have included the first 30 secs and the last 30 secs. The overall duration was 2 min 40 sec.

cpu2.mp4

[xenodium]

For now, xenodium/shell-maker#23 is merged.