NeMo Helix model selection
Recommend a model for a new agent from NIM or another provider configured on
the running Platform. Explain the capability fit first and benchmark evidence
second. Return the model choice in a form suitable for the
Platform-managed agent.yaml.
Pre-flight
1. Load the benchmark cache
test -f packages/nemo_helix_ext/src/nemo_helix_ext/skills/nemo-model-selection/references/benchmark_cache.json && echo "cache_present" || echo "cache_missing"
If cache_missing, proceed with the static table in this file. Tell the user once that benchmark data is stale and they can refresh it with:
python scripts/refresh-benchmark-cache.py
The cache (schema v6+) carries four things the rest of this skill reads:
models[]— editorial entries for a curated set of NIMs withstrong_at,watch_out_for,intent_hints,derived_fromlineage, and direct/inferred scores.upstream_index.bfcl_v4andupstream_index.arena_elo— full BFCL and per-category Arena Elo tables for ~84 and ~360 models respectively. Use these to look up scores for ANY model name, not just the registered ones.namespace_to_type[]— namespace-prefix → NAT_typemapping used only for legacy NAT workflow output.name_decomposition_rules[]— pattern→hint rules for synthesizingintent_hintswhen an unknown model name lands.
2. Fetch the live model list for a Platform-routed harness
nemo inference models list --output-format json 2>/dev/null | python3 -c "import json,sys; d=json.load(sys.stdin); print('\n'.join(m['id'] for m in d.get('data', []) if m.get('id')))" 2>/dev/null || echo "PLATFORM_UNREACHABLE"
Interpretation:
- Model IDs returned → these are the candidates the user can actually pick from. Each
idis the routable VirtualModel identifier inworkspace/nameform: it is what the inference gateway's OpenAI-compatible route accepts as the requestmodeland what the compatibility probe sends in its request body. Carry it through to Step 1+ unchanged. The probe's URL argument andmodels.default.modelinagent.yamltake only the part after the/(seereferences/harness-compat-check.md): a deploy binds the agent to its own workspace's gateway route, which accepts a bare name or only that workspace's prefix. List models in the workspace the agent will deploy into. Do not substitute an id fromnemo models list, which returnsModelEntityrecords whose ids are Platform entity ids and are not routable. PLATFORM_UNREACHABLE→ platform isn't up. Recommendations may continue from the curated cache, but a model for Platformagent.yamlcannot be finalized until the live Platform model list and harness-specific inference route can be checked.
Use the nemo CLI rather than constructing a Platform URL or calling
/v1/models directly. The CLI resolves nemo --base-url, NHX_BASE_URL, the
active CLI context, authentication, and workspace consistently with subsequent
agent commands. Do not hardcode localhost, 127.0.0.1, or port 8080.
This list is authoritative only for models routed through Platform. For a
native-provider harness such as claude, use the configured provider's native
model catalog and validation tooling instead.
Step 0 — Pick the conversation direction
Before the profile questions, ask which path the user is on:
Quick check before I ask the design questions:
A. **You're choosing a model.** Walk me through what the agent does and I'll
recommend from what's deployed.
B. **You already have a model in mind.** Tell me which one — I'll assess
whether it fits your task and surface what to watch for.
Which are we doing? (If unsure, A is the default.)
If the user picks A → continue to Step 1 with the recommend flow. If the user picks B → continue to Step 1 with the assess flow:
- Skip the "recommend a model" framing entirely in Step 2.
- Still ask the three profile questions in Step 1 (you need them to evaluate fit).
- In Step 2, produce a fit assessment for the user's named model rather than a recommendation — same evidence-quality flagging, but the output frames as "here's what we know about for " not "use this model."
Step 1 — Profile the agent
Ask all three questions in a single message. Skip any that the conversation has already answered.
Before I recommend a model, three quick things about what the agent will do:
1. **Tool density.** How many tools, and how do they interact?
- One tool (search, lookup, or similar)
- 3–5 tools whose outputs chain into each other
- Many tools, often called in parallel
2. **Primary capability.** What does the model spend most of its time doing?
- Calling APIs or tools reliably (MCP, structured function calls)
- Working with code (reading repos, editing, writing patches)
- Reading and reasoning over long documents
- General conversation and instruction following
3. **Deployment.**
- Cloud (NVIDIA Build API, OpenAI, Anthropic, etc.)
- Self-hosted on a GPU — if so, roughly how much VRAM is available?
- Not decided yet
Do not propose a model before all three answers are in. Push back on "you decide" — commit to a default and announce it ("I'll assume cloud and a tool-heavy agent. Tell me if that's wrong.").
Step 1.5 — Build the candidate set and pick a presentation pattern
Identify the selected harness and whether it uses a Platform-routed or native provider path before building candidates. Read it from the source config or conversation; ask if it is still unknown.
Building candidates
The candidate set is what the user can actually pick from. It comes from three joins:
- Start with the correct live catalog. Use the pre-flight Platform model
list for Platform-routed models. Use the configured provider's native model
catalog for native-provider harnesses. If the required catalog is
unreachable, the cache may support a conversational recommendation, but do
not finalize a Platform
agent.yamlmodel block. Availability alone does not establish compatibility with an agent harness. - For each available model name, look up evidence in this order:
- Token-match against the editorial
models[]entries → if hit, use the full editorial record (lineage, intent_hints, direct + inferred scores) - If no editorial match, token-match against
upstream_index.bfcl_v4keys → if hit, use that BFCL score withsource: "direct_external" - Same for
upstream_index.arena_elofor per-category Elo - If neither editorial nor upstream matches, synthesize
intent_hintsby walkingname_decomposition_rules[]and collecting every hint whosepatterntoken appears in the decomposed model id. Mark evidence assource: "name_only".
- Token-match against the editorial
- Rank candidates by the user's profile — primary capability axis determines which score field dominates.
When the live Platform model list was unreachable, also tell the user the rest of this flow is operating on the curated NIM set, not their actual deployment.
Verify harness compatibility before config handoff
When the output targets Platform agent.yaml, rank a short candidate list from
the appropriate live catalog and the evidence above, then verify candidates
against the selected harness's actual model contract:
| Harness | Required model contract | Compatibility check |
|---|---|---|
codex |
OpenAI Responses API | Valid v1/responses inference request |
hermes |
OpenAI-compatible chat completions | Valid v1/chat/completions inference request |
deepagents |
Provider-specific; nvidia, openai, and openai-compatible use chat completions |
Valid request for the selected provider path; use v1/chat/completions for an OpenAI-compatible route |
claude |
Native Anthropic provider | Require provider: anthropic and validate the configured credentials/model with native Anthropic tooling; do not route it through Platform IGW |
Before making any inference requests, show the user the candidate names and explain that the checks make real, potentially billable model calls. Ask for explicit confirmation and wait. The original request to select a model or write a config is not confirmation for these calls. For native-provider harnesses, use the provider's native validation path instead of forcing the request through Platform IGW.
For a Platform-routed candidate, use the context-aware CLI (nemo inference gateway model post v1/responses for Codex, v1/chat/completions for Hermes or an OpenAI-compatible DeepAgents setup) with the id returned by nemo inference models list, split into workspace and name as shown in references/harness-compat-check.md (copy-paste commands).
A successful model list lookup, schema validation, Fabric planning, deployment
readiness, empty request, or GET does not establish compatibility for a
Platform-routed model. Only candidates that complete a valid request through
the required Platform model path may be returned to nemo-agent-config. For a
native-provider harness, require its adapter provider contract and successful
native credential/model validation instead. Exclude failed combinations and
try the next ranked candidate. If no candidate passes, stop without emitting a
model block and ask the user to configure a compatible provider or explicitly
choose a different harness. Never switch the harness silently.
Record the selected harness, provider, exact model name, required model contract, and successful check in the handoff. Then apply the normal benchmark ranking only among compatible candidates.
Picking the presentation pattern
The pattern depends on the evidence quality of the candidate that best matches the profile:
Identify the best-matching candidate.
Inspect its evidence for the primary axis from profile question 2:
- "code" → scores.arena_elo.coding AND scores.bfcl_v4
- "tools" → scores.bfcl_v4
- "long documents" → scores.arena_elo.hard_prompts
- "general/instruction"→ scores.arena_elo.overall AND scores.arena_elo.instruction_following
If the relevant scores all have source ∈ {"direct", "direct_external"}:
→ Pattern A (single recommendation, current Step 2 template)
If one or more relevant scores have source == "inferred_from_ancestor",
or the candidate has scores == {} (name-only) but has intent_hints:
→ Pattern B (forced trade-off, withhold model name until user resolves it)
If 3+ candidates have similar profiles and similar evidence-quality:
→ Pattern C (shortlist with explicit "I'd want an eval before deciding")
Pattern B — forced trade-off
Hard rule: when the primary candidate's evidence is anything other than direct, the model name does NOT appear in your response until the user has resolved the trade-off. This is the anchoring guard. Without it, users default to the first model named regardless of caveats.
Present exactly two unnamed candidates, A (matched specialist, with its caveat) and B (measured generalist, strongest direct score on the same axis), recommend a 5–10 prompt eval on both, and name a model only after the user resolves the trade-off. Never fill a slot with an incompatible model: if compatibility filtering leaves one candidate, present it unnamed with its caveat and recommend the eval before committing; if it leaves none, follow the no-candidate rule in Verify harness compatibility before config handoff. Full step list and an example output (code-heavy agent whose natural pick has inferred BFCL data): references/pattern-b-example.md.
Pattern C — shortlist (rare)
Use only when 3+ candidates have similar profiles and similar evidence-quality, and no honest trade-off binary exists. Present them in a numbered list with plain summaries and explicit phrasing that an eval is the only way to actually decide. Do not commit to one. Default to recommending an eval as the next step.
Step 2 — Recommend in plain English
Lead with the capability that matters most for this agent, then name the model as the conclusion. Use this shape (adapt the content, do not copy verbatim):
"For an agent that , the thing that matters most is <one capability in plain words — e.g. 'reliably picking the right tool without hallucinating its arguments'>. That's what is built for — <one sentence on what the model actually does well, not what it scored>. The tradeoff is . Given you're deploying <cloud / self-hosted>,
<model string>is the practical choice."
Recommendation table
| Agent profile | Recommended model |
|---|---|
| Many tools, parallel or chained calls | qwen/qwen3-235b-a22b |
| Many tools, cost or speed is a constraint | qwen/qwen3-30b-a3b |
| Code-focused: reading repos, writing patches | qwen/qwen3-coder-30b-a3b-instruct |
| Reasoning over long documents or many chunks | nvidia/llama-3.1-nemotron-ultra-253b |
| General-purpose, balanced default (platform-curated) | nvidia/llama-3.3-nemotron-super-49b-v1 |
| General-purpose, widely-deployed alternative | meta/llama-3.3-70b-instruct |
| Fast and cheap, simple tool loops | microsoft/phi-4-mini-instruct |
| Self-hosted, single consumer GPU (<24 GB) | qwen/qwen3-8b |
What each model is strong at
Read each model's strong_at and watch_out_for from benchmark_cache.json when present. Fallback text when the cache is missing: references/model-strengths.md.
Step 3 — Mention the evidence only after the plain-English case
After explaining the recommendation, briefly note how we know — framed as "the test that simulates this" not "the score it got".
Right and wrong examples: references/evidence-framing.md.
If the user asks what benchmark was used or wants the raw number, tell them. Do not lead with it.
Step 4 — Output
Show the blocks that fit the user's stage. Default machine-readable output to
Platform agent.yaml. When the chosen model's primary-axis score has
source: "inferred_from_ancestor" or the model relies on intent_hints only,
include an explicit evidence caveat in the human-readable recommendation. Do
not encode benchmark commentary as unsupported config fields.
If they are authoring Platform agent.yaml, emit a default model block:
models:
default:
provider: <configured-provider>
model: <model-string> # Platform-routed: the VirtualModel name, without the workspace/ prefix
api_key_env: <credential-env-var-if-needed>
base_url: <provider-base-url-if-needed>
Use the provider identity configured on the Platform. Omit api_key_env and
base_url when the selected provider does not require user-supplied values.
Keep base_url directly in the model block, not under settings.
Before emitting this block, complete the harness-specific compatibility check above. Do not infer compatibility from provider or model-list metadata alone.
The default model applies to every harness that does not declare its own model. Add a harness-local override only when that harness intentionally uses a different model or provider:
harnesses:
<harness-name>:
kind: <harness-kind>
model:
provider: <configured-provider>
model: <model-string>
Do not emit raw Fabric SDK model objects. nemo-agent-config owns final YAML
placement and validation.
If they are explicitly maintaining a legacy NAT workflow YAML, emit the NAT compatibility block, pick the NAT _type from namespace_to_type[], and pair the model with a workflow type. All of that is in references/legacy-nat-output.md. For Platform agent.yaml, represent credentials with api_key_env and put base_url directly in the model block.
Step 5 — Staleness notice
End with a staleness notice: say the recommendations reflect the last cache refresh, give python scripts/refresh-benchmark-cache.py to refresh, and link the raw leaderboards. Exact text: references/staleness-notice.md.
Verification
This skill writes nothing. Verification is conversational: summarize the recommendation in 3 lines (capability that mattered most, model chosen, one tradeoff) and ask "Does this match what you need?" Do not hand off until the user confirms.
Use nemo-agent-config when the user wants to apply the selected model to agent.yaml.
If verification fails
Wrong recommendation: re-ask only the misread profile question, do not restart. Model not in the table: say so and describe the capability gap versus the closest recommended model. Compatibility probe fails: exclude that combination and test the next ranked candidate; if none pass, stop without emitting a model block. Full symptom/cause/recovery table: references/recovery.md.
Hard rules
- Never name a model before all three profile questions are answered.
- Never lead with a model name, benchmark name, or score.
- Never recommend a cloud-only model when the user said self-hosted.
- Never return an untested model route to a config-writing skill.
- Never make model inference calls without explicit user confirmation.
- Never silently change the selected harness to accommodate an available model.
- Never emit a model identifier without showing the plain-English reason alongside it.
- When the primary candidate's evidence is anything other than
direct, the model name does not appear in your response until the user has resolved the trade-off in Pattern B. Anchoring is the failure mode this guards against — users default to the first model named regardless of caveats. The withhold is non-negotiable. - For
agent.yaml, present the evidence next to the YAML rather than inventing a config field.
Gotchas
- "You decide" needs a committed default, not a silent fill-in. Pick something, name it, tell the user.
- Do not transform model IDs by punctuation convention. Use the identifier
returned by the selected live provider or Platform model listing and pair it
with the correct
provider. Legacy NAT components and Data Designer may use different provider-specific identifiers; preserve the identifier required by that consumer instead of assuming the build skill converts it. - Watch the deployment column. A 235B cloud-API recommendation aimed at a self-hoster with a 24 GB GPU is the most common mismatch and the easiest to catch by re-reading Step 1.