Repository navigation
Conversation
agent-shell-prompt-ranges walked every font-lock-face change in the buffer, and chat mode calls it on every shell event. On long sessions this dominated streaming: a 1.6 MB buffer (10 turns, mostly collapsed tool output) took ~0.2s per relabel. Cache the prompt runs before the next-to-last one and rescan only from there, since a busy shell streams between the submitted prompt and the live one. An after-change-functions hook drops the cache on any change at or before that run (clear, session reload). Narrowed calls scan without touching the cache.
77d4fe3 to
c125764
Compare
moyotar
changed the title
agent-shell-prompt-ranges stopped at every font-lock-face change, which grows with the session and slowed chat mode relabeling on long buffers. Prompts end where comint's output field gives way to input, as comint-next-prompt relies on, so walk field instead for faster results.
|
Thanks for the profiling and the PR! Before caching, I'd like to see if we can make the scan itself cheap: a cache depends on every earlier change running modification hooks, which nothing enforces. I've pushed a change that walks |
|
Tried 78a2c94: same ranges as the old scan, and the stuttering is gone. Per relabel:
Agreed the cache isn't worth its fragility, closing this. FWIW the cost still grows with the buffer, ~14 ms per MB on my DeepSeek sessions: |
|
Thank you for retesting and profiling.
Increasing |
agent-shell-prompt-rangeswalked everyfont-lock-facechange in the buffer, and chat mode calls it on every shell event, so relabel cost grew with the whole session.I noticed Emacs stuttering while a fast agent was streaming into a long session. Over 40s of streaming, the profiler put 71% of CPU in
agent-shell-chat--relabel-buffer→ the prompt scan. That buffer was 1.6 MB: 10 turns, ~97% collapsed tool output, ~29k face runs. Each relabel took 0.17–0.27s.Cache the prompt runs before the next-to-last one and rescan only from there, since a busy shell streams between the submitted prompt and the live one. An
after-change-functionshook drops the cache on any change at or before that run (clear, session reload). If the cached position no longer holds a prompt face, it falls back to a full scan. A narrowed buffer is scanned in full and leaves the cache alone.Same buffer (1.7 MB by now):
agent-shell-prompt-rangesagent-shell-chat--relabelOver ~2800 calls in normal use, including turn boundaries, the cached result matched a full scan every time I compared them.
Not addressed:
agent-shell-prompt-rangesstill checks each run for the end-of-prompt marker (~1 ms here).font-lock-facein agent-shell or shell-maker.Checklist
M-x checkdocandM-x byte-compile-file.