// the take
AutoSwarm-Shadow-Adapter is a FastAPI/React sidecar for HKUDS's AutoSwarm multi-agent framework that lets a DAG pause a task for human sign-off or remote GPU execution without blocking the engine, parking state in its own SQLite WAL store and writing results back into AutoSwarm's store.db. It's for teams running AutoSwarm's 'AI company' setup who need a real human (or a remote compute node) to gate specific roles like legal review or architecture approval.
The non-blocking handoff (return AWAITING_HUMAN, free the execution thread immediately) solves a real problem: a human review step that takes minutes or hours would otherwise time out or stall parallel DAG branches. Keeping a separate SQLite WAL store instead of writing straight into AutoSwarm's own database limits blast radius if the adapter misbehaves, and SHA-256 hashing on uploaded deliverables is a sane baseline for artifact integrity. Offering both an embedded (same process) and standalone sidecar/Docker deployment mode covers both a solo dev and a small team without forcing either into the other's setup.
It resumes AutoSwarm by reaching directly into that project's internal, undocumented SQLite schema (`delegation_work_items.phase`) instead of a stable public API — any upstream schema change breaks this silently, and there's no compatibility layer for that coupling. The CLI examples pass plaintext passwords and provider API keys as command-line arguments (`--password "secure_pass_1"`), which end up in shell history and process listings on every worker node — a real problem for anything described as handling legal or enterprise review. The README leads with an 'Overemployment Engine' pitch for secretly running multiple full-time jobs at once, which is a terminable or fraudulent use case at most employers and reads like it's optimizing for attention rather than describing the actual engineering. Star count (4265) against 2 forks is a strange ratio for a tool that specifically invites people to run their own remote worker nodes against it — real adoption of a BYOC tool usually shows up as forks and worker variants, not just stars. It's self-labelled 'Production Release v0.1.0' with no mentioned security review, incident history, or real deployments, so 'production' is aspirational rather than demonstrated.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →