Agent-Skills.md

Agent Skills: Add Replicate Model to Catalog

Add new Replicate models to the Renku catalog. Fetches JSON schema, looks up pricing from replicate.com, matches/implements cost function, adds entry to replicate.yaml.

UncategorizedID: keremk/renku/add-replicate-model

Repository

keremk/renku
keremkLicense: MIT
1

Install this agent skill to your local

pnpm dlx add-skill https://github.com/keremk/renku/tree/HEAD/.claude/skills/add-replicate-model

Skill Files

Browse the full folder contents for add-replicate-model.

Download Skill

Loading file tree…

.claude/skills/add-replicate-model/SKILL.md

Skill Metadata

Name
add-replicate-model
Description
Add new Replicate models to the Renku catalog. Fetches JSON schema, looks up pricing from replicate.com, matches/implements cost function, adds entry to replicate.yaml.

Add Replicate Model to Catalog

Add a new Replicate model to the Renku pricing catalog.

Step 1: Gather Information

Ask the user for:

  • Model name (e.g., openai/sora-2) — always in owner/model-name format
  • Model type: video, image, or audio

Step 2: Fetch JSON Schema

Run the schema fetcher to download the model's input schema:

node scripts/fetch-replicate-schema.mjs <owner/model-name> --type=<type>

This creates a JSON file in catalog/models/replicate/<type>/. If the schema already exists, skip this step.

Requires REPLICATE_API_TOKEN environment variable.

Step 2a: Durable Fix Policy (Required)

Treat generated schema JSON files as refresh artifacts, not manual source of truth.

  • Do not rely on hand-editing catalog/models/replicate/**/*.json for long-term fixes.
  • Persist all manual corrections in:
    • catalog/models/replicate/schema-overrides.yaml
  • The refresh scripts re-apply these patches after every upstream fetch.

When you identify a provider/schema mismatch (for example, field constraint corrections), add a patch entry in schema-overrides.yaml for that model and type.

Example:

version: 1
models:
  - name: bytedance/seedream-4.5
    type: image
    patches:
      - op: add
        path: /input_schema/properties/image_size/default
        value: "1024x1024"

Step 3: Annotate Viewer Metadata (Required)

Immediately annotate and validate viewer metadata after schema fetch:

node scripts/annotate-viewer-schemas.mjs --model=<owner/model-name>
node scripts/validate-viewer-schemas.mjs --model=<owner/model-name>

Notes:

  • x-renku-viewer is the UI source of truth for component initialization.
  • If validation reports placeholder-to-be-annotated pointers, ask the user exactly which component should be used for each pointer and update annotations before continuing.
  • Do not leave unresolved placeholders for newly added models unless user explicitly agrees.

Step 3a: Voice-ID Markers for Audio/TTS (Required when applicable)

For audio narration/TTS schemas, propose voice field markers before moving on:

  • Heuristic candidates to inspect (proposal only): voice, voice_id, voiceId, and nested variants like voice_setting.voice_id
  • Ask the user to confirm exactly which schema pointer(s) should be marked
  • After confirmation, annotate the source schema node (or referenced definition node) with:
    • "x-voice-id": true
    • optionally "x-voices-file": "voices/<file>.json" when shared rich voice metadata exists

Important rules:

  • Never place x-voice-id / x-voices-file under x-renku-viewer; they belong to the original schema node.
  • When x-voices-file is present, it is authoritative for voice options.
  • Re-run annotation + validation after applying these markers:
node scripts/annotate-viewer-schemas.mjs --model=<owner/model-name>
node scripts/validate-viewer-schemas.mjs --model=<owner/model-name>

Step 4: Analyze Schema

Read the generated JSON schema file and its x-renku-viewer annotations. Identify cost-relevant input fields:

  • Video models: Look for duration, seconds, generate_audio, mode, resolution, aspect_ratio
  • Image models: Look for image_size, quality, num_images, resolution, width, height
  • Audio/TTS models: Look for text, duration

Step 5: Look Up Pricing

Browse the model's Replicate pricing section directly using the #pricing anchor:

  1. Get Chrome tab context with tabs_context_mcp
  2. Create a new tab with tabs_create_mcp
  3. Navigate to https://replicate.com/<owner>/<model-name>#pricing
  4. Use find to search for "cost per second", "price", or pricing amounts
  5. Take a screenshot to read the pricing cards
  6. Extract the pricing data (per-second, per-run, per-character, etc.) and criteria (mode, audio, resolution)

Step 6: Select or Implement Cost Function

Read docs/cost-functions-reference.md in this skill directory for the full reference.

Match the model's pricing to an existing cost function:

| Pricing Pattern | Cost Function | | -------------------------------- | ---------------------------------- | | Flat per run | costByRun | | Per second (simple) | costByVideoDuration | | Per second + audio toggle | costByVideoDurationAndWithAudio | | Per second + resolution tiers | costByVideoDurationAndResolution | | Per second + mode (standard/pro) | costByVideoDurationAndMode | | Per second + mode + audio toggle | costByVideoDurationModeAndAudio | | Per megapixel (video) | costByVideoMegapixels | | Per million tokens | costByVideoPerMillionTokens | | Per character | costByCharacters | | Per second (audio) | costByAudioSeconds | | Per megapixel (image) | costByImageMegapixels | | Size + quality grid | costByImageSizeAndQuality | | Resolution tiers (image) | costByImageAndResolution | | Dimension-based (image) | costByResolution | | Per token (text) | costByInputTokens |

If no existing function matches, implement a new one following the guide in docs/cost-functions-reference.md under "How to Add a New Cost Function".

Step 7: Build & Insert YAML Entry

Read docs/replicate-yaml-format-reference.md for the full format reference and section map.

  1. Read catalog/models/replicate/replicate.yaml
  2. Find the correct insertion point based on model family (see Section Map in the reference)
  3. Build the YAML entry using the appropriate template
  4. Insert using the Edit tool

Step 8: Verify

Run the dry-run verifier:

node scripts/update-replicate-catalog.mjs catalog/models/replicate/replicate.yaml --dry-run
node scripts/update-replicate-catalog.mjs catalog/models/replicate/replicate.yaml --check-diff
node scripts/validate-viewer-schemas.mjs --model=<owner/model-name>

If a model fails due to override patch drift:

  • update catalog/models/replicate/schema-overrides.yaml for that model, then
  • re-run only the model:
node scripts/update-replicate-catalog.mjs catalog/models/replicate/replicate.yaml --update-diff --model=<owner/model-name>

If you modified cost-functions.ts, also run:

pnpm --filter @gorenku/providers check:all

Step 9: Add to Producer Files

Map each new model to the correct producer YAML under catalog/producers/.

9a: Determine Producer Mapping

Read the relevant producer files and the model schemas, then build a proposed mapping table. Present it to the user in this format before making any changes:

| Model | Producer file | Reasoning |
|---|---|---|
| `owner/model-name` | `category/producer.yaml` | why this producer fits |

Also show the proposed field-level mappings for each model:

  • Which producer inputs map to which API fields
  • Any transformations needed (intToString, resolution mode, expand, etc.)

Ask the user:

"Does this mapping look correct? Let me know if anything needs adjusting before I add it."

Wait for confirmation before proceeding to 9c.

9b: Producer Selection Reference

| Model type / schema inputs | Producer file | | ------------------------------------------------ | ------------------------------- | | text (TTS) | audio/text-to-speech.yaml | | lyrics + prompt (music) | audio/text-to-music.yaml | | reference_images array + prompt + duration | video/ref-image-to-video.yaml | | video + prompt + duration (extend) | video/extend-video.yaml | | image + prompt + duration (single image) | video/image-to-video.yaml | | prompt + duration only (text-to-video) | video/text-to-video.yaml |

9c: Insert Mappings

For each model, add an entry under mappings.replicate: in the correct producer YAML, following existing patterns in that file. Use the Edit tool to insert after the last existing replicate: entry. If no replicate: section exists yet, add one at the end of the file.

Common transformation patterns:

  • Plain integer duration: Duration: duration
  • Duration as string: Duration: { field: duration, intToString: true }
  • Aspect ratio only: Resolution: { field: aspect_ratio, resolution: { mode: aspectRatio } }
  • Aspect ratio + preset (two fields): Resolution: { expand: true, resolution: { mode: aspectRatioAndPresetObject, aspectRatioField: aspect_ratio, presetField: resolution } }

9d: Validate Producer Mapping Contracts (Required)

After adding or updating producer mappings, run the producer mapping contract suite from the repo root:

pnpm validate:producers

Do not finish until this command passes. If it fails, fix the producer mappings and re-run.