[Feature]: Re-Compress or Re-Compaction of old history

[tisDDM]

Problem

Sometimes I get the case (especially when using Opus on GHCP), that after having everything implemented I am missing headroom for getting all the tests fixed in the same session. The reason might be that the old summaries where a bit "optimistic" and large.
Tried to let Opus recompact them or just delete from the old history - but unfortunately it doesnt work. Not sure if this was never intended or if the interface does not provide this feature. Opus tried - but in vain.
I would try to implement - but not sure if this is wanted.

Proposed Solution

Extend the Compress-Tool so that already compacted information could be deleted or re-compacted

Alternatives Considered

No response

Additional Context

No response

[Tarquinen]

I think for that you should just use normal opencode /compact, the idea behind compress is that it doesn't dillute away information by compressing it multiple times, so it can't compress a compression. I'm still not totally sold on the current system because I agree context can still build up, but I was thinking it's ok that the session doesn't last forever, it just lasts way longer than normal. What do you think?

[tisDDM]

I fully agree to you, that a session shall never (virtually) last forever. I am a friend of using fresh sessions and handovers and I had built a tiny framework for generating docs, handovers, plans and using a 2 phase blueprint gated delegation pattern to subagents. ( at first only intended for internal company use - but reached more than 20 * ).

DCP is one of our daily drivers here, because it just keeps the session clean while implementing a full feature. After implementing, most of the time - even on GHCP with 128k/160k , there enough room to do tests and fix everything before starting with the next feature in a fresh session.

But if things run bad, it gets tight. So I thought about giving the model the ability to revise the summary or simply delete parts of it. E.g. during a 7-Phase implementation there are a lot of summaries written. But in the end it simply doesn't matter how gating between phase 3 an phase 4 had performed - it is more forensics then history.

An other approach is just to change model or provider and push everything through - but that feels unnecessary. What do you think - is it clear what I mean?

[Xynonners]

I would second this FR. generally, when you have a really long running and deep session, it would be great to be able to compact only the useless old summaries instead of having to use the global /compact which loses a lot of information.