dsh-plugin-adaptive-agent-policy

A recurring pain point when running multi-step tasks in DeepSeek Harness is that the Agent Loop applies the same step budget and tool-result accumulation regardless of task scope. A read-only analysis pays the same loop cost as a cross-module refactor, so large tasks balloon in tokens and wall-clock time while small ones lack guardrails to stay on track.

The plugin applies control proportional to task. It first classifies the work into one of five tiers - read, small, frontend, large, batch - using a text-only router with no model call, then enforces per-tier maximum output tokens, soft-checkpoint steps, and a plain-text hard stop. At each soft checkpoint the risk router picks at most one high-confidence, previously uncovered check from nine candidates and backs off when evidence is thin. Tool results are trimmed through three progressive levels (moderate, tight, critical) that protect recent outputs and skip rewriting history entirely when size and savings thresholds are not met. Policy state is delivered via a non-persistent system section, so input cost stays constant instead of growing with every step.

If you regularly run multi-step workloads - cross-module refactors, batch jobs, frontend debugging - in Harness and want to cut redundant loops without changing your model, provider, or toolchain, this is a lightweight workflow-level adjustment. It does not touch permissions, sandboxing, or model selection, and requires no separate installation of the upstream pruner.