AGENT TOOLING / CONTEXT MANAGEMENT
Compaction that preserves what a conversation actually needs
Claude Code compacts long sessions when context becomes scarce. The default summary is necessarily generic and can flatten Polish conversations, omit user messages or remove technical detail important to continued work.
Kompakt replaces that one-size-fits-all step with selectable prompts, presets and output rules. The operator decides what kind of memory the next context window receives.
THE PROBLEM
Different sessions need different memory. A debugging session should preserve evidence and attempted fixes; a planning session should preserve decisions and constraints; a long Polish collaboration should not silently become a generic English abstraction.
THE IMPLEMENTATION
Kompakt connects to the compaction lifecycle through Claude Code hooks. Configurable presets select prompts and output behavior without requiring the user to rebuild instructions every time context becomes full.
THE RESULT
The plugin does not attempt to preserve every token. It preserves the information chosen for the current kind of work and makes that policy inspectable rather than hidden inside a default prompt.
Define what the summary should prioritize and how it should be written.
Switch between policies for coding, planning or detailed archival work.
Keep user intent and important verbatim instructions when required.
Control language, detail and structural expectations for the compacted context.
No. It changes how a session is summarized when compaction is necessary, so the next context window receives more useful working memory.
Because debugging, architecture, content work and long multilingual collaboration require different facts and levels of detail to survive.
AGENT MEMORY
We can make context, compaction and lifecycle rules explicit instead of relying on a growing prompt and hoping the important parts survive.
Discuss agent tooling