fast-jev-compaction：用 Jev 判断哪些工具调用该留

Rexei @iamrexei原推发布于 2026.09.19

在 X 上打开原推

The costliest issue with Claude Code isn't the usage limits It’s that moment when the agent has made 50 tool calls, read through a heap of files, and found a bug—only for the compaction process to turn the entire workflow into a three-paragraph summary And along with that, the following details vanish: ▸ the exact file path ▸ the terminal error ▸ the "don't touch this module" constraint ▸ the test result that prompted the agent to dig into the project in the first place That’s where I found fast-jev-compaction It’s a plugin that doesn't ask the LLM to retell the old context Instead, it runs past tool calls and their results through Jev: "Is this action still relevant?" "Is the full result needed?" It discards what isn't needed It reduces some items to brief notes And it preserves—verbatim—what might actually be needed later. The core idea here is more powerful than the repository itself: Standard compaction tries to craft a nice narrative about what happened But an agent doesn't need a story to do its job It needs facts: the command, the path, the error, the diff, the test, and the specific solution The plugin handles both manual /compact commands and automatic compaction. If Jev fails to respond or doesn't compress the context sufficiently, it falls back to Claude Code's built-in summarization I’d use this for long-running tasks: refactoring, migrations, debugging, large repositories, and agent chains involving dozens of tool calls But don't mistake it for "perfect memory." Jev still makes probabilistic decisions about what to discard. Plus, the task state is sent to the external TypeSafe API So, test the results on a low-stakes project first before using it on client code. The context doesn't become infinite It simply stops wasting space on junk and forgetting what matters

这条是引用着另一条写的：x.com/iamrexei/status/2100873212850503745