Race condition on auto-saving long (in time) AI responses

[Korolen]

Doesn't auto-save response to latest prompt

Context

The auto-save functionality for keeping the specstory saved .md file up to date always.

Expected / wished-for behavior

File is auto-saved upon prompt submission and thereafter updating the file repeatedly with new response tokens as they stream in live. Saving to disk with every token, or possib. every 0.25-1 second as the response streams in, until it's finished.

Current behavior

Instead, auto-save only ever fires immediately upon prompting itself, saving no reply data as it's not yet available at that moment in time.

No matter how long it's left afterwards, the response is never added to the .md file.

Thanks so much

This would be hugely helpful for me to be able to keep all my systems, throughout, incl. git and supporting tools vastly more tightly in-sync with the current moment during my flow of work.

Super-useful tool btw !! much appreciation and enthusiasm for the idea of it and this change i think would take it to the next level for rubber-meets-the-road utility for that most-fresh in-the-moment "tying-things-together" developer experience ! :DD >.> :P :)

[belucid]

Your description isn't totally accurate @Korolen in terms of how SpecStory works, but the essence of your concern still holds. We don't actually know anything "upon prompting itself", we get no invocation or signal of that happening at all.

Instead what we're doing is watching the sqlite DB Cursor file that uses to store the chat history. We watch that for changes at the OS level, and whenever it changes we look for updated chats that have a new date since the last time we auto-saved that chat.

This does explain the delay that's often seen between the time a human/AI interaction takes place and when the content makes it into an auto-save file, as we are dependent on the chat interaction making it out to the DB on disk, which doesn't happen in real-time (or "streaming" as you suggest). Cursor writes to in memory DB pages, and those get flushed to disk periodically. We only see the changes when they get flushed to disk.

What can happen though is that the AI response takes a while (it maybe thinks for a while on something, or has tool use that's slow, or what have you) a partial and incomplete portion of the AI response may get flushed to disk. Later, when the rest of the AI response makes it to disk, we check the time on the AI interaction, and see that it's one we've already seen, so we don't update the auto-save unless/until there is another prompt from the user (or, you can force a regeneration by deleting the auto-saved markdown file itself). This is problematic, and we'll use this issue #57 to track improving this.

[belucid]

@Korolen I updated the issue title.

[jakelevirne]

Confirming that this happens pretty regularly for me, given that most AI responses now take 10s of seconds.

• Autosave gets triggered (either by poll or by database update)
• The AI response in Cursor chat is still streaming.
• The resulting markdown contains only a portion of the AI response.
• I never enters another prompt in that same chat session.
• The autosave markdown will never contain the complete response from the AI.