/model-integration

name: model-integration
description: Connect HTTP models or a native ComfyUI workflow through Nomi's verified certification path.
metadata:
  nomi:
    selectable-in-workbench: true
    version: 1.0.0
    tools:
      - nomi_model_setup
      - nomi_remove_provider
      - nomi_read
    required-providers:
      - text
      - image
      - video
    library:
      kind: skill
      title:
        zh-CN: 模型接入
        en: Model integration
      summary:
        zh-CN: 按已验证的接入流程连接 HTTP 模型或 ComfyUI 工作流。
        en: Connect and verify HTTP models or ComfyUI workflows.
      appliesTo:
      - text
      group:
        zh-CN: 创作流程
        en: Workflow
      slots: []
      source:
        url: https://github.com/aqm857886159/Nomi/blob/7a072f12d4e35d1bf341403d942e5c0c702e6c7b/skills/model-integration/SKILL.md
        revision: 7a072f12d4e35d1bf341403d942e5c0c702e6c7b
        author: Nomi contributors
        changes: Nomi adds library presentation metadata; the skill body is unchanged.
        evidence:
        - https://github.com/aqm857886159/Nomi/blob/7a072f12d4e35d1bf341403d942e5c0c702e6c7b/skills/model-integration/SKILL.md
      preview:
        path: assets/cover.png
        type: image
        provenance: illustration
license: AGPL-3.0-only

Nomi model integration

Use Nomi's integration tools to turn a vendor endpoint or a native ComfyUI workflow into a capability that remains usable after restart. The tools are the contract; do not edit Nomi source, Catalog files, or MCP configuration by hand.

Order

1. **Official contract first.** Start with `nomi_model_setup` (`action: "connect_provider"`) using only public material, and capture the official docs/OpenAPI URL for every `(vendor, model, mode)`. That one call opens Nomi's secure page with the address pre-filled for the user to confirm; never send a key through MCP, and never try to decide where a key is sent — `baseUrl`, `authType`, `authHeader`, `authQueryParam`, `authScheme` and `proxyUrl` are not arguments of any tool. Read the provider docs with web/Bash, probe its public contract, handle pagination and relay quirks yourself, then hand over one complete declaration card with `action: "submit_declaration"`. The card must declare `assetIngestion` explicitly — `{"strategy": "none"}` when the provider has no upload channel; leaving it out is a rejected card, not a fallback. There is no `expectedRevision` to carry: the session fingerprint is filled in for you. Never infer a field from a model name or a neighboring provider. 2. **Use the existing identity graph.** Reuse the matching `modelArchetype`, `Catalog Mapping`, `HttpOperation`, `integrationCertification`, `GenerationRuntime`, `ProductionRun`, and managed assets. One logical model is one catalog row; use the mapping's generic `modeId` discriminator for same-kind modes instead of vendor-specific exceptions or duplicate rows. 3. **Treat propose as the persistence gate.** Nomi validates the public proposal, candidate-to-selection references, workflow shape, owner and exact revision before it changes session state. If it returns `propose rejected: <field> ...`, fix that field in the Agent, keep the same session, and retry with the returned/new revision; do not invent an old discovery or input-resolution call. A rejected proposal does not mutate the session. 4. **Build a canary matrix before network.** For each exact `(vendor, model, mode)` record the official endpoint, required fields, smallest valid parameters, expected output type, upper-bound cost, idempotency key, and one-attempt limit. A model without a complete row stays `documented` or `blocked`; do not infer a cheaper/default field from a neighboring model. 5. **Static gate before network.** Run `pnpm run check:model-certification-coverage` and the root-cause contract checker. Every archetype you add or change must carry a `sources` field (official doc URL + checked date; there is no gate for it any more, so review it by hand). The static gate must prove mapping identity, generated archetype artifacts, and that no secret entered the ledger. 6. **Generic mode discrimination is mandatory.** If one logical model exposes multiple same-task modes, each mapping must carry a generic `modeId` and the request must include it whenever selection is ambiguous. An omitted discriminator must fail closed; never add a provider-specific branch or duplicate model row to hide ambiguity. 7. **Zero-cost loopback and failure matrix.** Exercise create → status/query → result against a local protocol simulator. Cover at least 401, 402/balance, 429, 5xx, timeout, malformed/truncated JSON, unknown status, missing request id, succeeded-without-output, oversized body, and media MIME/magic mismatch. A queued response without a provider request id must fail closed and never be resubmitted. 8. **Provider-owned assets.** Prefer the provider's signed/ephemeral upload API for local references (for example Runway `POST /v1/uploads` → signed multipart → `runway://` URI). Small images may use an official data URI. Anonymous public image hosts are not a debugging strategy and must never be silently retried when the provider has its own upload path. 9. **PR #221 MCP cost gate.** Run the MCP zero-cost journey through spend confirmation and verify `provider request count = 0` before any live canary. Confirmation is immutable and user-owned; an agent cannot invent a receipt or confirm spend. 10. **Live is last and must use the production path.** Only with the user's provider key/credits and explicit canary scope, run one minimal request through Nomi's `GenerationRuntime`/`ProductionRun`, validate the bounded artifact, commit the managed-asset journal, and perform a fresh-process readback. A direct curl/SDK call, provider-only output URL, or loopback pass is not live certification. If managed localization, auth, credits, callback deployment, or network policy blocks the run, keep `status=blocked` with the exact evidence; never retry blindly. 11. Poll `no