HiDream API Gen Skills

A local HiDream/Vivago image + video generation package for this workspace.

Runtime host: vivago.ai

What this package is for

In this workspace, hidream-api-gen is the default multimodal generation gateway for local API-backed image/video generation.

That includes not only ordinary image/video requests, but also cases where a higher-level workflow has already narrowed down to actual multimodal asset generation, such as:

• one story video / OSV asset-generation steps
• comic / manga / 漫画生成 flows
• local comic-gen-style workflows once they reach the actual image/video generation step

Important boundary:

• higher-level project/workflow routing should still be handled by the appropriate domain/router skill
• but once the task is specifically “generate the asset through the local multimodal API path”, this package is the preferred gateway

Canonical entrypoints

Use these entrypoints in order of role:

• scripts/dispatcher.py → canonical generation entry
• scripts/requery.py → explicit follow-up for an existing task_id
• scripts/healthcheck.py → chain validation / smoke test

Treat raw per-model scripts as implementation backends. Do not build new default workflows around calling them directly.

Current scope

Active

• image generation
• video generation
• re-query / polling discipline
• healthcheck / smoke-test path

Recognized but not fully wrapped

• text family
• audio family

requery.py can poll text / audio kinds, but full generation wrappers + dispatcher integration for those families are not yet landed.

Current support surface

Supported / active

• Kling
• Seedance 1.0 Pro
• Seedance 1.5 Pro
• Seedream
• Nano Banana

Removed from active surface

• Sora 2 Pro
• Hailuo 02

For current alignment details, see:

• references/model-mapping.md
• references/status-board.md
• references/verification-log.md

Status semantics

Current operative contract:

• 0 = queued
• 1 = success
• 2 = in progress / non-terminal
• 3 = generation failed
• 4 = safety review failed

Hard rule:

• if a task is at task_status=2, do not classify it as failure
• continue polling or use scripts/requery.py

Runtime requirements

• HIDREAM_AUTHORIZATION → required
• HIDREAM_ENDPOINT → optional override
• X-source → optional upstream routing/review marker

Quick start

Healthcheck

python3 scripts/healthcheck.py
python3 scripts/healthcheck.py --live

Re-query an existing task

python3 scripts/requery.py --task-id <TASK_ID> --kind video --module hidream-Q2

Example backend CLI usage

python3 scripts/kling.py --version "Q2.5T-std" --prompt "A flying car"

Where the details live

• runtime gateway guidance → SKILL.md
• model alignment → references/model-mapping.md
• canonical entry → scripts/dispatcher.py
• explicit follow-up → scripts/requery.py
• chain validation → scripts/healthcheck.py

Security notes

• reads HIDREAM_AUTHORIZATION from env/config
• reads user-provided local media files for upload/base64 conversion when needed
• sends requests only to the configured Vivago endpoint