When multi-turn coding Agents run long tasks, 80% of the context is "useless lines" from tool returns — test logs, error stacks, diff remnants. SWE-Pruner Pro is counter-intuitive: which lines to delete, the Agent's own middle-layer representations already know — no external classifier is needed. ByteDance Seed's arXiv paper (2607.18213) attaches a small pruning head on top of a frozen Coder LLM, mapping the Agent's hidden state when reading tool output into a "line-level keep/prune" decision, with length-aware embeddings to differentiate tool blocks of different lengths. On two open-source backbones including MiMo-V2-Flash and four multi-turn benchmarks, it saves up to 39% of prompt+completion tokens with almost no added inference latency, lifts SWE-Bench Verified solve rate by +3.8pp, and lifts Oolong long-context accuracy by +2.2pp. The significance is a paradigm shift. Previously, every context-compression method assumed "an independent decision model is needed". Pro uses the Coder's own representations to prove that assumption is unnecessary — the Agent is already the optimal "context-relevance" judge; it's just that nobody has been reading out that signal. The code is open-sourced; the next question: can it be ported to general Agents.