Adding Support for a New LLM Model

This skill guides you through adding a newly released LLM to the model_constructors + llms stack (the endpoint-class router). It replaces the legacy lib/api/llm/clients/* router, which no longer exists.

Mental model

A model reaches production through three stacked layers. Add the new model to each:

1. Model config (front/types/assistant/models/*) — the legacy ModelConfigurationType

describing the model (context, vision, reasoning efforts, pricing tiers). Still the source of truth consumed by the UI, pricing, and the dust layer.

1. modelconstructors (front/lib/modelconstructors/*) — provider-agnostic endpoint

classes, one per (provider, model, region, provider-api). Each class mixes a shared provider base client with a per-model config mixin (input schema, context size, token pricing). This is where the real request/response shape and the narrowed input config live.

1. llms (dust layer) (front/lib/llms/*) — thin Dust-specific wrappers around the

modelconstructors classes that add Dust concerns (display name, byok, endpoint filters, and any caps — e.g. exposing 250k context on a model that natively supports 1M). Registered into DUSTSTREAM_ENDPOINTS.

Endpoints are named and filed as: {provider}{model}{region}{providerapi}.ts e.g. googlegemini36flashglobalagent_platform.ts. The class name is the PascalCase of the same, with numbers spelled out: GoogleGeminiThreeDotSixFlashGlobalAgentPlatformStream.

The **fastest, most reliable way to add a model is to copy the most recent model in the
same family** across all layers and rename. Grep every reference to that model and mirror
each one. This skill lists the reference points; the sibling model is your template.

Before you start: verify against official docs (MANDATORY)

You MUST confirm every value below against the provider's official documentation and leave a URL + date in a code comment next to it. Do not carry values over from memory.

• Specs (context window, max output tokens, vision, structured output):

- OpenAI: https://platform.openai.com/docs/models - Anthropic: https://docs.anthropic.com/en/docs/about-claude/models/overview - Google: https://ai.google.dev/gemini-api/docs/models - Mistral: https://docs.mistral.ai/getting-started/models/models_overview/

• Pricing (input / output / cached input per 1M tokens):

- OpenAI: https://openai.com/api/pricing/ - Anthropic: https://www.anthropic.com/pricing#anthropic-api - Google: https://ai.google.dev/gemini-api/docs/pricing - Mistral: https://mistral.ai/technology/#pricing

• Host / region availability: verify which provider APIs and regions actually serve the

model day-one. Mirror the sibling model's endpoints, but only register an endpoint whose region is actually available. Keep an unavailable-but-anticipated endpoint class defined and unregistered (see Gemini's EU agent-platform example) with a comment saying why.

WebSearch/WebFetch the docs first. If a value can't be confirmed, surface it — don't guess.

Reference points to mirror (grep the sibling model)

Pick the newest sibling (e.g. for "Gemini 3.6 Flash" the sibling is "Gemini 3.5 Flash") and grep -rln its id / const / class-name / model-id string. You will touch, roughly:

A. Model config + central registry

B. Pricing / tiers / reasoning (TYPE-ENFORCED over StaticModelIdType)

Adding the id to STATICMODELIDS makes these fail to compile until updated:

C. model_constructors — the endpoint classes (stream)

D. model_constructors — tests (TDD, see below)

E. llms — the dust layer (stream)

F. SDK + UI + marketing mirror

Batch endpoints (.../batch/...) are a curated subset — only add them if the model
needs batch. They are NOT completeness-enforced. Set supportsBatchProcessing to the real
capability regardless.

The TDD loop (steps to actually run)

The endpoint classes derive their behavior from a shared integration test harness. Let the live API tell you the input contract — never infer it from the sibling model. Sibling expectations are the single biggest source of wrong config: two models in the same family routinely differ on temperature, reasoning efforts, and forced tool use.

The config schema must ALWAYS mirror the API's real behavior as closely as possible. It describes what the provider accepts — not what Dust happens to send today, and not what would be convenient. If the API accepts a value, the schema accepts it; if the API rejects a value, the schema rejects it. Never narrow past the API because an upstream layer already strips the field (the dropTemperature / dropTemperatureWhenReasoning config parsers in lib/llms are a product policy and belong there, not in the endpoint schema), and never widen past it to avoid a union. Concretely: Anthropic reasoning models accept exactly temperature: 1, so the field is z.literal(1).optional().default(1) — not z.undefined(), even though the Dust layer drops it before the endpoint ever sees it.

When a divergence from the API is genuinely wanted (exposing a narrower effort set to control cost, say), it is a policy choice — write it as a comment stating that the API allows more and why Dust doesn't, so the next reader doesn't mistake it for a provider constraint.

Reasoning efforts must ALWAYS mirror the model's official documentation, not merely whatever the endpoint happens to accept. This is the one place where "what the API tolerates" is the wrong source of truth, because gateways are routinely looser than the models they serve:

• The Fireworks gateway validates reasoning_effort against low/medium/high/xhigh/max/none

for every model it hosts, so a live run "passes" on efforts the model never defined.

• DeepSeek documents disabled/high/max and says low/medium are mapped to high and xhigh to

max — accepting them would silently rewrite the caller's choice.

• Kimi K3 is documented low/high/max by Moonshot; medium works through Fireworks but is not

a K3 effort.

• Kimi K2.6 has binary thinking; the graded values are accepted and do nothing (measured:

low produced more reasoning than medium).

• grok-4.5 silently accepts minimal and xhigh, which xAI documents only for other models.

So: find the model author's doc (not just the host's), expose exactly the efforts it lists, and link it in a comment. Where host and author docs disagree, follow the author unless the host documents a model-specific override — generic host guidance is not a contradiction. Then confirm each documented effort actually works on the live endpoint, and record any effort the endpoint accepts but the docs omit, with a note that undocumented efforts can change without notice.

When the product still offers an effort the model does not have, map it in the llms layer with a configParsers entry (mapReasoningNoneToMinimal, mapNonNoneReasoningToHigh, mapReasoningEffortToLowHighMax, forceHighReasoningEffort) — never with a schema .transform(), and never by widening the endpoint schema to swallow it.

1. Widen

Write the config mixin with configSchema set to the broad inputConfigSchema (front/lib/model_constructors/types/input/configuration.ts), marked // TDD SCAFFOLD. Every case must reach the API instead of being short-circuited by a guessed schema.

2. Write the test with every case null

Copy the sibling's key set (so coverage matches) but not its expected values. null runs the case with its default checkers. Starting from the sibling's INPUTCONFIGURATIONERROR markers hides exactly the differences you are trying to find, and lets stale expectations survive — a suite whose expectations were never run green will happily assert things the schema makes impossible.

3. Red run — the whole suite, no --bail

You want every failure at once in order to characterize the contract:

cd front
NODE_ENV=test RUN_LLM_TEST=true DUST_MANAGED_{PROVIDER}_API_KEY=... \
  npm run test -- --config lib/model_constructors/test/vite.config.js \
  lib/model_constructors/test/endpoints/{...}.test.ts

Env-var names live in the sibling's createInstance (DUSTMANAGEDANTHROPICAPIKEY, DUSTMANAGEDGOOGLEAISTUDIOAPIKEY, …). Agent-platform/Vertex endpoints need VERTEXAIPROJECTID plus GCP credentials — a GOOGLEAPPLICATION_CREDENTIALS service-account key works and needs no gcloud auth application-default login. Add --bail 1 or -t "<substring>" only later, when iterating on a single case.

4. Sort every failure into one of three buckets

The bucket decides the fix:

A passing case is evidence. It disproves any assumption that the model rejects that input — including assumptions already written down. Do not keep an INPUTCONFIGURATIONERROR because a code comment says the model doesn't support something: the run outranks the comment.

5. Narrow — including the defaults

Rewrite configSchema to the real contract, with a doc URL + date in a comment next to each value. Three things to pin deliberately, not by inheritance:

• reasoning default effort — read it off the official doc, every time. The .default(...)

is load-bearing: an absent reasoning sends no thinking config, so the provider's own default applies, and that differs per model (adaptive-on for Fable 5 / Opus 5 / Sonnet 5; thinking-off for Opus 4.8/4.7/4.6 and Sonnet 4.6; no thinking for Haiku 4.5; max for Kimi K3 and GLM-5.2). Never carry over a sibling's default or invent one for cost reasons — Kimi K3 sat at low when Moonshot documents max. Mirror the documented default and cite the page; if the product wants a cheaper default, that belongs in defaultReasoningEffort on the llms model config, not in the endpoint schema.

• temperature handling. Sweep actual values against the API rather than assuming — the

rule is per-model. Anthropic reasoning models accept only 1 while thinking is on and any value while thinking is off; some reject the field outright.

• Effort set and forceTool compatibility. Which efforts are genuinely accepted, and

whether a forced tool_choice may coexist with reasoning.

6. Green run

Mark the genuinely-rejected cases INPUTCONFIGURATIONERROR, re-run the full suite until every case passes, then delete the // TDD SCAFFOLD comment.

7. Re-run every endpoint sharing the mixin

A config mixin is shared across regions and provider APIs (e.g. global/anthropic + eu/agent-platform), so narrowing it changes all of them. Run each one.

8. Push the new behavior up into the family's shared config

Shared configs are per family — Opus, Sonnet, Haiku each have their own; a family with a single member (Fable 5) just keeps a standalone config. A family's shared config should track the latest member of that family, because the next model in it is far likelier to repeat the newest behavior than the oldest. So when characterizing a model reveals that its family's shared config was wrong, fix the shared config and put the override on the older models — never special-case the newest one.

The reflex to resist is the opposite: leaving the shared config alone and giving the new model a bespoke schema. That makes every future model in the family inherit stale behavior, and it is how a restriction that only ever applied to one old model ends up applied to all of them. (Worked example: forceTool: z.undefined() sat in the shared Opus config because extended thinking forbids a forced tool_choice. Opus 4.7, 4.8 and 5 all use adaptive thinking and all accept it — verified live — so the fix was to drop it from the shared config, not to override it on Opus 5.)

Do not merge families that happen to agree today. Fable 5 and Opus 5 share every value except one (Fable 5 cannot disable thinking), but they are different families, so they keep separate configs and the coincidence is allowed to drift.

Then re-run the suites of every model in the family (§7), since they all moved.

If you cannot run the live suite (no key / non-interactive), narrow the config from the sibling model in the same family and say so explicitly — flag every expectation as unverified; the live run must still happen before merge.

Without NODEENV=test+RUNLLM_TEST, the test file loads but its cases are skipped; that still validates it compiles and is registered.

Verify (non-live checks that must pass)

cd front
npx tsgo --noEmit                        # whole-project type check
NODE_ENV=test npm run test -- \
  types/assistant/models/sdk_drift.test.ts \
  types/assistant/models/types.test.ts \
  lib/api/assistant/token_pricing/tiers.test.ts

• tsgo clean over the files you touched (the satisfies Record<...> maps and STREAMENDPOINTSETUPS are your completeness guardrails).
• sdk_drift.test.ts green ⇒ front ⊆ SDK (rebuild sdks/js types if it names your id).
• tiers.test.ts green ⇒ reasoning-effort maps in sync with the configs.

Model config properties (quick ref)

Checklist

• Specs + pricing confirmed against official docs, URLs in comments
• Host/region availability confirmed; only available endpoints registered
• Model config added; previous family model isLatest: false
• STATICMODELIDS + SUPPORTEDMODELCONFIGS + model_constructors/types/models.ts
• Pricing/tiers/reasoning trio updated (compile-forced)
• model_constructors: config mixin + endpoint class(es) + stream/index.ts
• Tests: .test.ts per endpoint + setups.ts
• TDD loop run live: widened schema → all cases null → full red run → narrowed schema

with reasoning-default and temperature confirmed against docs → green run → scaffold removed

• Config schema mirrors the API: every value the API accepts is accepted, every value it

rejects is rejected; deliberate divergences commented as policy, not as provider limits

• New behavior pushed up into the family's shared config, with overrides on the older

models rather than a bespoke schema on the new one

• Every endpoint sharing the config mixin re-run green (all regions / provider APIs)
• llms dust layer: dust mixin + endpoint(s) + llms/stream/index.ts
• SDK union updated and rebuilt; UI model_configs.ts; marketing mirror
• tsgo clean; sdk_drift / types / tiers tests green
• Live endpoint test passes (or limitation flagged for follow-up)

Troubleshooting

• sdk_drift.test.ts names your id → add it to KnownModelLLMId in sdks/js/src/types.ts, then cd sdks/js && npm run build:types (the test reads the built @dust-tt/client).
• tsgo on setups.ts / index files → you added an endpoint to STREAM_ENDPOINTS without a matching setup, or vice-versa. Register both.
• tiers.test.ts fails → staticmodelreasoning_efforts.ts disagrees with the config's supportedReasoningEfforts.
• Model not in UI → missing from USEDMODELCONFIGS.
• Live test rejects a config → check the bucket first (§4). A provider invalidrequesterror means narrow configSchema and mark the case INPUTCONFIGURATIONERROR; an inputconfigurationerror under the widened scaffold means a converter or base client is rejecting it, not the API.
• A case you expected to fail passes → the model accepts that input. Fix the expectation (and any comment claiming otherwise) rather than keeping the marker.
• Live suite 401s → check the key you actually exported. A shell profile can define the same DUSTMANAGED*APIKEY twice; the last export wins interactively, so grepping for the first match can hand you a stale key.