Actor Creator Agent

Wrapping an external API as a marketplace actor is repetitive and demands knowledge of two platforms at once: the upstream API's parameters and auth, and Apify's actor layout, input schema rules, build system, and store conventions. Doing it by hand for dozens of APIs does not scale, and a naive single-prompt generator produces code that fails to build, leaks provider details into user-facing fields, or ships an input schema that disagrees with what the code actually reads. The hard part is not generating code once; it is getting to a verified, published actor with bounded retries and a clear record of what happened.

Work is split into two stages. Alpha ingests the manifest with deterministic parsing (no model calls), structures endpoints into an OpenAPI spec, catalogs input fields, pauses at a human review gate where the operator edits labels and descriptions (including by Excel export and re-import), and generates test cases. Beta runs 21 skills from requirements through code generation, hardening, live testing, schema, README and manifest generation, extended tests, code review, UX review, push and build, log analysis, a run gate, README enrichment, store metadata, and a final push.

Each skill is a row in the database: a versioned Jinja2 prompt template, a model assignment, and retry and timeout settings. Prompts and models are editable from the UI without a deploy, and every save creates a version that can be restored. Skills that need custom logic are Python plugins discovered by a registry; everything else runs through a generic executor.

The orchestrator loop checkpoints after every completed skill, compresses the context snapshot, and supports resume, restart from a step, and skill-requested rewinds (for example a UX review sending the pipeline back to the schema step) with a per-target cycle cap so loops always terminate. Deterministic checks run before any model review: AST parsing, async usage outside the SDK, hardcoded key patterns, banned provider strings, and a cross-check that schema keys match the keys the generated code reads.

Runs execute in Celery workers and stream progress to the browser over a WebSocket. Secrets are Fernet-encrypted in the database and injected into actors as environment variables, never written into generated code.