Offline/periodic/incremental context rebuild

[tensiondriven]

Feature Request

Add the ability to perform context rebuilding in the background during idle periods.

Problem

Currently, context is built on-demand for each new request, which can be expensive for long conversations or complex topics. There's no mechanism to periodically rebuild or summarize context during server idle time.

Solution

Implement a background context rebuild system that triggers during idle periods, performs incremental updates, and cancels immediately when new requests arrive.

Key Features

• Idle Activation: Starts when server is idle (no active requests)
• Incremental Updates: Process context in chunks rather than full rebuilds
• Periodic: Run at configurable intervals during idle periods
• Cancellable: Immediately cancel rebuild when new request comes in
• Offline Capable: Can run without active user interaction

Benefits

• Performance: Faster response times for new requests
• Resource Efficiency: Better memory management through summarization
• Offline Processing: Utilize idle server time productively
• Scalability: Handle longer conversations more efficiently

Implementation Considerations

• Background worker with cancellation logic
• Incremental rebuild algorithms
• Configuration options for timing and thresholds
• Monitoring and visibility into rebuild process

Success Criteria

• Background rebuild activates during idle periods
• Rebuild cancels immediately when new request arrives
• Context size/performance improves over time
• No degradation in request response times

💘 Generated with Kaitlin
Co-Authored-By: 💘 Kaitlin kaitlin@charm.land

[SeanPtest]

opencode has this feature listed, crush doesn't :
"Persistent Storage: SQLite database for storing conversations and sessions"

It would be good to know whether the dev team has plans on implementing this feature if they allready haven't. If they do it would be nice to disable it if we have our own context rebuild & memory modalities in place. Just saw a vid on graphiti, looks like a solution.

[tensiondriven]

It would be good to know weather the issue submitter used the product; as far as I can tell, crush was forked from opencode, and opencode used SQLite before that. Also if you look in ~/.crush, you'll find crush.db which is that sqlite database.

I think this issue can be closed.

(What I'm curious about is a progressive/ incremental context rebuild that happens in the background, say, every dozen messages - for me, waiting for my CLI to compact context is the worst part of cli vibe coding)