Context compression

[AlexKovynev]

Cause MCP good thing but without that context eat in 5 minutes :(

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Add MCP support #19: Both relate to MCP support, though this one focuses on general MCP support while yours is specifically about context compression (confidence: medium)
• Token consumption increases by 300% after adding Capacitorjs because .gitignore files in sub folders are not considered when excluding files #1149: Similar concern about excessive token consumption, though specific to Capacitor rather than MCP (confidence: medium)
• 🐞Critical Exclusion Pattern Failure for Multiple Directories (node_modules, .git) Causes Excessive AI Token Consumption #1307: Both address excessive AI token consumption issues, though this one focuses on exclusion pattern failures (confidence: medium)
• [bug] Token Context Too Large - Unusable with Dyad #693: Both deal with context size making the app unusable, though this was about large app imports rather than MCP (confidence: low-medium)
• Optimize Summarize to Chat #435: Both relate to context optimization, though this was specifically about chat summarization (confidence: low)

Feel free to ignore if none of these address your specific case.

[wwwillchen]

any suggestions on how we do context compression? MCP is inherently quite token hungry.

[AlexKovynev]

https://github.com/Kilo-Org/kilocode/blob/main/src/core/condense/index.ts

@wwwillchen

As i understood it prompt. So you ask some LLM. get results and fit as context then compressed one as base further. Not very great. Can loose some but that all that we have look like on now.

But why not. Can be created settings like in kilocode for border in percent for compress.