Dust LLM

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.

Adding a model is usually only half the task: the model it supersedes has to be retired in the same PR, and a model the provider has switched off needs its agents repointed. See Deprecating or removing an old model.

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.
2. model_constructors (front/lib/model_constructors/*) — 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.
3. llms (dust layer) (front/lib/llms/*) — thin Dust-specific wrappers around the model_constructors 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 DUST_STREAM_ENDPOINTS.

Endpoints are named and filed as: {provider}_{model}_{region}_{provider_api}.ts e.g. google_gemini_3_6_flash_global_agent_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 STATIC_MODEL_IDS 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

At most two versions of a family in USED_MODEL_CONFIGS. The picker groups by maker, so every version left in the list is another near-identical row a user has to read past ("Gemini 3.5 Flash / 3.6 Flash / 3.7 Flash / 3.8 Flash"). When you add a model, keep only it and its immediate predecessor; drop the rest of the family from USED_MODEL_CONFIGS. Count families by product line, not by provider — Gemini Flash, Gemini Flash Lite and Gemini Pro are three families, each allowed two.

Everything a dropped model needs to stay callable lives elsewhere (SUPPORTED_MODEL_CONFIGS, the endpoint classes, pricing), so the eviction only removes it from the picker and the agent builder. Then finish the deprecation properly, or the model rots into a stale default years later:

• Set isLegacy: true + isLatest: false on each evicted config. isLegacy is also what drops it from the public credits page, so an evicted-but-not-flagged model keeps being advertised while being unpickable.
• Repoint every hardcoded reference to it. grep -rn X_MODEL_CONFIG front front-api and fix the ladders and defaults that name it: ORDERED_FAST_MODEL_CONFIGS / ORDERED_SMALL_MODEL_CONFIGS / ORDERED_LARGE_MODEL_CONFIGS in front/lib/api/assistant/models.ts, getFastModelConfig in front/lib/api/assistant/conversation/title.ts, preferredModelConfiguration on the dust-* global agents, and MODEL_STREAMS candidates in front/types/assistant/models/auto.ts. These are hand-maintained lists that no type checks — nothing goes red when they point at a legacy model.

A legacy model still referenced by one of those lists is the failure mode this rule exists for: conversation titles ran on Gemini 3.5 Flash for three releases after 3.6/3.7/3.8 shipped, purely because getFastModelConfig was never revisited.

No marketing mirror. The public credits page fetches /api/marketing/model-credits, which front/lib/api/marketing/model_credits.ts derives at request time from SUPPORTED_MODEL_CONFIGS + MODEL_PRICING. Nothing to copy into marketing/ — but the model only appears there once it has a MODEL_PRICING entry, is not isLegacy, and is released (no availableIfOneOf.featureFlag).

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 INPUT_CONFIGURATION_ERROR 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 (DUST_MANAGED_ANTHROPIC_API_KEY, DUST_MANAGED_GOOGLE_AI_STUDIO_API_KEY, …). Agent-platform/Vertex endpoints need VERTEX_AI_PROJECT_ID plus GCP credentials — a GOOGLE_APPLICATION_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 INPUT_CONFIGURATION_ERROR 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 INPUT_CONFIGURATION_ERROR, 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 NODE_ENV=test+RUN_LLM_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 \
  types/assistant/models/model_tiers.test.ts

• tsgo clean over the files you touched (the satisfies Record<...> maps and STREAM_ENDPOINT_SETUPS are your completeness guardrails).
• sdk_drift.test.ts is a compile-time guard: its it() body always passes, and the Exclude<StaticModelIdType, KnownModelLLMId> assertion only fails under tsgo. So a green vitest run proves nothing here — tsgo is what enforces front ⊆ SDK. Rebuild the SDK types first, or tsgo reads a stale @dust-tt/client and passes on a drifted id.
• model_tiers.test.ts green ⇒ reasoning-effort maps and tier maps in sync with the configs.

Deprecating or removing an old model

Adding a model is normally paired with retiring the one it supersedes. There are two distinct paths — pick by whether the provider still serves the old model.

Path 1: deprecate (superseded, but still served)

Hide it from new work and leave everything else standing, so agents already pinned to it keep working and historical token accounting stays exact. Worked example: Kimi K2.6 added / K2.5 deprecated (f2824da5c5e, #28834).

Retiring a global agent that way keeps it resolvable so past conversations still render, while getGlobalAgents filters it out of list views and isRetiredGlobalAgent gates it out of new conversations. Do not delete the GLOBAL_AGENTS_SID member — the enum values are the sIds persisted in historical messages.

Keep the id in STATIC_MODEL_IDS, SUPPORTED_MODEL_CONFIGS, CURRENT_MODEL_PRICING, STATIC_MODEL_TIERS, STATIC_MODEL_SUPPORTED_REASONING_EFFORTS, and keep its endpoint classes registered. Nothing to do for marketing: isLegacy is what excludes it from the public credits list (front/lib/api/marketing/model_credits.ts).

A deprecated model needs no agent-config migration — that is the point of the path.

Path 2: remove (provider decommissioned it)

Nothing can run on the model any more, so it comes out of the codebase entirely and every agent still pinned to it must be repointed. Worked example: DeepSeek R1 removal (3ca8d834527, #26958). Do Path 1's picker/global-agent repointing first, then:

• Model config + registry: delete the X_MODEL_ID const and X_MODEL_CONFIG from front/types/assistant/models/{provider}.ts, and their entries in front/types/assistant/models/models.ts (STATIC_MODEL_IDS, SUPPORTED_MODEL_CONFIGS).
• The compile-forced trio (§B) then goes red — remove the entries from static_model_reasoning_efforts.ts and model_tiers.ts. For pricing, move the entry from CURRENT_MODEL_PRICING into LEGACY_MODEL_PRICING in the same token_pricing/global.ts (a Record<string, PricingEntry>, so it survives the id leaving StaticModelIdType); that block exists precisely so historical runs still cost out. Deleting the pricing outright silently zeroes past usage.
• model_constructors + llms: delete the endpoint classes, config mixins, and test/endpoints/*.test.ts; unregister from stream/index.ts, setups.ts and llms/stream/index.ts; drop the id from the MODELS array in front/lib/model_constructors/types/models.ts.
• Its global agent, if it had one: retire it via RETIRED_GLOBAL_AGENTS_SID as in Path 1 — that is the normal answer even here, and it keeps historical conversations rendering. Only tear the agent out completely when it must stop resolving at all: delete the factory (global_agents/configurations/{provider}.ts) and the GLOBAL_AGENTS_SID member in front/types/assistant/assistant.ts, which turns every exhaustive reference red — the getGlobalAgent switch and the flag filters in global_agents.ts, global_agent_metadata.ts, prompt_context.ts (Record<GLOBAL_AGENTS_SID, …>) and getGlobalAgentAuthorName. Deleting the member abandons the sIds stored in past messages, so justify it explicitly. (DeepSeek R1 did this in 3ca8d834527, before RETIRED_GLOBAL_AGENTS_SID existed — prefer retirement now.)
• Feature flag: drop the model's flag from front/types/shared/feature_flags.ts once nothing else references it.
• SDK: removing an id from KnownModelLLMId in sdks/js/src/types.ts narrows a public API type — a breaking change. Get explicit sign-off first (see the dust-breaking-changes skill), then rebuild (cd sdks/js && npm run build:types).

Migrating agent configurations

A removal orphans every AgentConfiguration row still pinned to the dead model, so ship a repoint script in the same PR: front/migrations/YYYYMMDD_migrate_<model>_models.ts, built on makeScript. Template: front/migrations/20260608_migrate_deepseek_r1_models.ts.

• Hardcode both source and target model ids as string literals when the source config is being deleted in this PR. The script must stay a frozen snapshot: importing the consts breaks the moment they are gone, and importing a "latest model" pointer would silently retarget the migration when the next model lands. (A pure repoint that leaves the model in the codebase — e.g. 20260810_migrate_sonnet46_medium_to_auto.ts — can import the consts.)
• Scan with AgentConfigurationModel.findAll({ where: { modelId, status: "active" } }) through a ModelStaticWorkspaceAware alias, with dangerouslyBypassWorkspaceIsolationSecurity: true plus the WORKSPACE_ISOLATION_BYPASS comment and biome-ignore lint/plugin/noUnverifiedWorkspaceBypass the linter requires — migrations run across all workspaces.
• Log every matched agent (sId, version, workspaceId, from → to) on the dry run, and gate all writes on execute.
• Write once, batched: a single update({ providerId, modelId }, { where: { id: agents.map((a) => a.id) } }) over the ids already gathered — not agent.update() per row. Scoped to those ids, the update needs no isolation bypass of its own (the cross-workspace scan already happened in the findAll).
• Set providerId alongside modelId — the replacement often sits on a different provider. Reset reasoningEffort too if the target does not support the effort the agent was on.

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
• STATIC_MODEL_IDS + SUPPORTED_MODEL_CONFIGS + 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
• UI model_configs.ts; SDK union updated and types rebuilt before tsgo
• USED_MODEL_CONFIGS holds at most two versions of the family; every model evicted by that rule is isLegacy: true + isLatest: false and no longer named by any hardcoded ladder (ORDERED_*_MODEL_CONFIGS, getFastModelConfig, dust-* global agents, MODEL_STREAMS)
• tsgo clean; types / model_tiers tests green
• Live endpoint test passes (or limitation flagged for follow-up)

Retiring the superseded model (same PR):

• Superseded model isLegacy: true + isLatest: false, dropped from USED_MODEL_CONFIGS
• MODEL_STREAMS candidates and dust-* global agents repointed to the new model, global-agent descriptions updated
• If decommissioned: config + registry + trio + endpoints + global agent + feature flag removed; pricing moved to LEGACY_MODEL_PRICING, not deleted
• SDK KnownModelLLMId narrowing signed off as a breaking change before removal
• Agent-config repoint migration written with hardcoded ids, batched update, dry-run output reviewed before --execute

Troubleshooting

• tsgo fails on sdk_drift.test.ts naming your id → add it to KnownModelLLMId in sdks/js/src/types.ts, then cd sdks/js && npm run build:types (it type-imports the built @dust-tt/client declarations, so the rebuild must come first).
• tsgo on setups.ts / index files → you added an endpoint to STREAM_ENDPOINTS without a matching setup, or vice-versa. Register both.
• model_tiers.test.ts fails → static_model_reasoning_efforts.ts disagrees with the config's supportedReasoningEfforts, or STATIC_MODEL_TIERS is missing an effort the config supports.
• Model not in UI → missing from USED_MODEL_CONFIGS.
• Live test rejects a config → check the bucket first (§4). A provider invalid_request_error means narrow configSchema and mark the case INPUT_CONFIGURATION_ERROR; an input_configuration_error 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 DUST_MANAGED_*_API_KEY twice; the last export wins interactively, so grepping for the first match can hand you a stale key.