Surgical runtime injection for AI agents that switch thinking modes

What it does

dsh-routing-suite is an umbrella installer for two submodules: a runtime injector that manages DeepSeek Harness (DSH) plugins without restarts, and a router-standard preset that classifies incoming tasks into reasoning modes—spec, react, mixed, or weak—then swaps the AI persona accordingly. The README cites P1-P23 benchmarks showing 96% routing accuracy, 100% convergence, and a jump from 0% to 100% open-task completion once persona anchors are locked.

The interesting bit

Rather than triggering a separate API call each turn to nudge the model’s behavior, the preset appends its steering prompt via agent/pre-step, piggybacking on the existing user message request. The README says this eliminated the previous 2× API call overhead while keeping prompt-cache hit rates between 92% and 94%.

Key highlights

• Runtime injector supports hot-loading, sidecar promotion, and self-healing for DSH components
• Model-specific persona tuning: Pro models get speculative planning prompts with few-shot examples, while Flash models receive neutral classification cues
• Three static anchors per persona—review, convergence, and anti-digression—to keep long tasks on track
• Introspection tools exposed as dev_router_status, dev_router_mode, and dev_mode_subagent
• v0.3.0 fixed first-message classification by intercepting agent/inbox/claimed before assembly, so sessions no longer default to weak on the first turn

Caveats

• router-pro is marked as planned but not included in the current v0.3.0 release
• Preset directories must be copied flat into .agent-presets; nesting them under a parent folder causes DSH to ignore them entirely

Verdict Grab it if you are already inside the DSH ecosystem and want to stop paying double for routing guidance. Everyone else is looking at fairly specific PowerShell glue.