Skip to content

take v0.1.0 — Take your script to the screen

Latest
Latest

Choose a tag to compare

View all tags
@XucroYuri XucroYuri released this 24 Aug 19:31
0.1.0
47c1721

take v0.1.0

Take your script to the screen. Agent-first storyboard & AI film production toolkit — running natively inside Claude Code, Codex, OpenCode and zcode.

script.md ──▶ beats ──▶ shots ──▶ storyboard ──▶ images ──▶ video

The agent thinks; take renders. Text analysis uses the agent's own model (no extra LLM API); images route to GPT-image-2, video to Seedance 2.x with Minimax H3 fallback — through a capability-aware failover router.

Highlights

  • dsh-aligned provider layer: seam triangle (Service Definition / Provider / Consumer), stable error codes (TakeError), effect-based registration
  • Production transport: exponential-backoff retry honoring Retry-After, token-bucket rate limiting, timeout tiers, attribution headers
  • Universal OpenAI-compatible adapter: new vendors = config entries, not code; sync/async-job auto-detection
  • Capability-aware routing: requests a model can't honor are skipped before spending a call; unknown models never rejected
  • Job registry + idempotent resume: .take/jobs.json event log, concurrency pool, input-hash resume so interrupted batches don't re-spend
  • Config v2: ordered failover chains, secrets as apiKeyEnv references, v1 auto-migrated
  • Agent skill with precise USE / DO NOT USE boundaries and controlled shot-language vocabulary

Validated

Full pipeline on a worked example (examples/rain-night): 3-scene script → 5 beats → 10 shots → validate → render → resume → export → audit log.

Quality

107 tests · biome clean · package contract gate · CI on Node 20/22/24.

Quick start

npx @take-ai/cli init my-film
cd my-film
# open in your agent: "分镜这个故事"
take validate && take generate images --mock

MIT licensed. See the CHANGELOG for details.

Assets 2
Loading