context-pruner

Long-running agent workflows pile up a mountain of tool-call records—stale outputs, duplicated requests, verbose error traces, oversized results—that silently devour the context window, waste token budget, and dilute the signal the model actually needs. context-pruner solves exactly that: it plugs a fully heuristic, zero-LLM-call context-triage engine into the official dsh compaction pipeline, making every decision deterministic, testable offline, and independent of any extra service.

Five independently toggleable screeners each handle a distinct category of bloat: archiving stale tool outputs into compact summaries, collapsing repeated identical calls to the latest one, swapping failed-call arguments for a tight error stub, head-and-tail trimming of oversized tool results, and pruning or trimming expired reasoning blocks. Every action is guarded by a reserve zone (the most recent N user turns are never touched) and an exemption list (stateful tools like task, skill, write, and edit are protected by default). A strict before-and-after size comparison ensures no rewrite is applied unless it is provably smaller, and call-pair-level changes act as atomic groups—no half-finished states. Each triage run produces a structured audit report with per-reason counts, token savings, and item-level details.

Built for developers running multi-turn, multi-tool agent workflows on dsh who want the context window to stay focused on what matters without introducing an extra model dependency. If you need zero-cost, deterministic, and auditable context slimming that slots into the official harness compression seam, context-pruner is your plug-and-play answer.