Field-Fixing the Hermes Harness on a DGX Spark — When the NIM Won't Stream Tool Calls, and Other Rough Edges

The Harnesses series already has, by its keystone, a Hermes agent that installs clean, serves from a local lane, survives a hostile-tool-call battery, and drives a curated slice of fieldkit over MCP. What it does not have yet is an honest account of the rough edges between “installed” and “actually pleasant to drive” — the small, infuriating mismatches you only hit when you sit down and use the thing against a real NVIDIA NIM on the box.

This article collects them. The lead fix is the one that bit hardest: the DGX-Spark build of the Nemotron-Nano-9B-v2 NIM ships a deliberately non-streaming tool parser, so vLLM raises NotImplementedError("Tool calling is not supported in streaming mode!") the instant an agent sends stream=true alongside tools — which Hermes does on every first tool-bearing turn. Three blind alleys precede the real fix (the misleading display.streaming config, the top-level streaming block, a Hermes source path that hardcodes the streaming probe), and the real fix turns out to be a one-line bind-mount: NVIDIA publishes a fully-implemented streaming parser in the same model repo, registered under the same module name, so overlaying it onto the path the NIM already loads makes streaming tool calls just work — no edit to the NIM image, no patch to Hermes.

More fixes will land here as the harness gets desk time: the reasoning-model output-token trap that silently truncates <think> before the answer, the config keys whose names lie about what they gate, and whatever the next session surfaces. Think of it as the field-repair log for running an agent harness on a single Spark — the entries that are too sharp-edged to belong in any of the first four articles, but too useful to lose.