lossless-compaction cuts Claude Code compaction from 51.8秒 to 0.26秒

The plugin's design is a direct reaction to how the built-in compaction works. Built-in compaction hands the whole conversation to a model and asks it to produce a summary, so time and tokens go into that rewrite, and any detail the summary omits is no longer in the next context. The developer had previously worked around this by manually running /compact at a convenient break before auto compaction hit, but that only shifts the timing and leaves the summarize-and-restart mechanism untouched.

The alternative approach explored, and then abandoned, was to have a model predict which information would matter later. The developer tried using Jev to judge whether a piece of information would be needed, but it did not perform as hoped, which the post attributes to the difficulty of predicting what will be needed 30 turns later from only the current context. Instead, lossless-compaction removes large items from context without deleting them, leaves a ticket with an ID, and lets the model recall the original content when needed. The content is identified by SHA-256 and read back from disk to confirm the write before removal.

The cost story is where the comparison gets less clean. The built-in approach pays at compaction time and then keeps a small context, but if dropped information becomes necessary again, those tokens re-enter the context from the session record. The plugin pays almost nothing at compaction time but carries a larger context into later requests, so when the prompt cache expires it has to rewrite that larger amount. The developer's own framing is that the question is not how to eliminate compaction cost but where and in what form to pay it. How that trade-off lands for a given team is likely to depend on how often its sessions pause long enough to break the cache, and how much of the earlier context later work actually reuses.