[Feature] Support NeMo Gym agents as routing targets

[xinyu-dev]

Problem

Using a NeMo Gym agent as a Switchyard target requires writing all the glue
yourself. For one agent we had to build:

• an adapter translating Chat Completions to Gym's Responses API and back
• a buffered-to-streaming shim, because Gym answers in one shot
• a credential proxy, so the provider key stays out of the agent process
• a usage object shaped exactly as Gym's response model demands, or the whole run
  is rejected after it completes

That is ~600 lines sitting between two NVIDIA projects that already speak
compatible protocols.

Why it matters

NeMo Gym is where NVIDIA agents are built, and Switchyard is where traffic is routed. In biology, the harness matters as much as the LLM itself. Many complex workflows require complex agent harness (such as biomni) rather than just a fine-tuned LLM.

Proposed behavior

A Gym agent as a target directly:

targets:
  biology-search:
    kind: gym_agent
    agent: simple_scientific_literature_search_agent
    base_url: http://127.0.0.1:9101

Switchyard would own the Responses API translation, the buffering, and the usage
contract — the parts that are identical for every Gym agent.

Example (very raw) implementation

Here switchyard routes to:

• Opus (LLM endpoint): default
• simple lit search agent: a simple arxiv/pubmed search agent implemented inside NeMo Gym. Nemotron Ultra 3 powers it underneath.
• biomni: a complex, deep research and biology AI scientist agent , impelmented inside Nemo Gym. Claude Opus powers it underneath

Orchestrator/Classifier model:
Nemotron 3 Ultra NVFP4 - chosen due to large context length, very fast performance, great reasoning capability.

Additional fixes

• Increase context length from default 16K to 200K, so classifier model can read into more context & make better routing decision
• Added the [running] visual cues. Biomni agent runs for longer time, user could use a cue that it is invoked. In/Out token stats only returns after it finishes.
• Increase various timeouts
• Other bugs, still WIP

Setup

1. Create the separate Gym and BioMni UV environments once:

 source .venv/bin/activate
 bash examples/gym_biology_agent/setup_gym.sh

2. In terminal A, start the agent stack and leave it running. This launches the NeMo Gym with 2 agents running.

 source .venv/bin/activate
 bash examples/gym_biology_agent/run_agent_stack.sh

3. In terminal B, launch switchyard

 switchyard launch claude \
   --model demo/biomni-agent \
   --config examples/gym_biology_agent/routes.yaml

Q1 : Write a Python merge-sort function

Routes to Opus LLM

Q2 : Explain the mechanism of mRNA splicing

Routes to simple lit search agent in Gym

Gym trace:

Switchyard + Claude Code interface (note the [running] visual cue):

Q3: Compute BSA protein's basic protparam parameters: MW, pI, aa composition, extinction coeffcient (cys all reduced vs cys all forming pairs), half life, instability index

Routes to biomni agent in Gym

Gym trace:

Switchyard + Claude Code interface (note the [running] visual cue):

Results in the UI:

Preserving original hooks

Some users might have hooks with CC or Codex. These hooks might invoke memory bank search, such as below for the hindsight memory plugin. I had to make a few fixes so that the classifier passes the memory blocks (as shown below) to the agents.

Below shows a correct example of memory generated from hindsight OnPromptSubmit hook, being carried over to the agent (as displayed inside Gym trace):

[xinyu-dev]

Related: #240 (classifier selecting among arbitrary named targets), #244 (agent targets), #245 (in-flight target visibility), #246 (classifier truncation), #247 (NeMo Gym agents).

[xinyu-dev]

Working demo of this, running Gym agents as routing targets end to end:
https://github.com/xinyu-dev/switchyard-biomni-demo

It covers three routes — Opus vs Nemotron Ultra, a Gym literature-search agent,
and a purpose router across Opus, literature search, and BioMni — with setup
scripts for the Gym and BioMni environments and a container that sandboxes
BioMni's code execution.

Building it surfaced two gaps filed separately: #252 (the classifier cannot tell
the user's question from the context Claude Code injects around it) and #258
(background requests reach the agent route, so it runs twice per question).