Summary

• compress oversized toolResult messages during OpenClaw context-engine assemble()
• keep session transcripts unchanged while shrinking the prompt actually sent to the model
• add plugin config and UI/schema support for tool-context compression tuning

What changed

• add tool-result compression helpers in examples/openclaw-plugin/context-engine.ts
• wire compression into assemble() before returning context messages
• add config fields in config.ts
• add matching uiHints and configSchema fields in openclaw.plugin.json

Default behavior

• enabled for oversized results from exec, process, web_fetch, and browser
• keeps key error/warn lines and essential metadata
• leaves read disabled by default to avoid hiding useful file content
• does not rewrite transcript history; only affects assembled context

Notes

• this PR intentionally does not include our separate local-startup improvements, to keep scope focused on context compression
• our local validation showed large prompt-size savings after compression during assemble()