Module worker_binding 

WorkerBindingMiddleware — a SpawnerLayer that propagates the Blueprint-baked per-agent WorkerBinding through Ctx.meta.runtime.

service::task_launch builds an agent name → WorkerBinding map from Blueprint.agents[].profile.worker_binding at launch time and places this layer on the stack (outermost, next to ProjectNameAliasMiddleware). Just before spawn it looks the map up by ctx.agent and, on a hit, inserts the serialized binding into Ctx.meta.runtime under the worker_binding key.

Agents with no declared binding get no entry; the WS thin-path requires_worker_binding fail-loud stays the safety net.

What reads the key: nothing in this repository

Said plainly, because a reader deciding whether to keep this layer on their stack needs it before the rationale. This layer was introduced for the delegate axis (b5eb000, “wire worker binding through the delegate axis”): OperatorDelegateMiddleware had no per-agent OperatorSpawner to carry a compile-time-baked binding, so it read the binding back out of Ctx.meta.runtime at spawn. That axis is gone, and the AgentSpec axis needs no such round trip — OperatorSpawner holds the baked WorkerBinding and hands it to Operator::execute as an argument.

With that reader deleted, a repo-wide search for WORKER_BINDING_KEY and for ctx.meta.runtime’s "worker_binding" entry finds this module’s own tests and nothing else. In particular AgentContextView::from_ctx does not read it (it reads the project / work-dir / metadata / run-id / alias keys), and mlua-swarm-server’s WS session takes the binding from its execute argument.

Two things the layer still does, and they are the whole of why it stays wired:

• It is a published extension point. This type and the key are pub on a library crate, and “downstream Operator / Spawner code reads the injected keys back via ctx.meta.runtime.get(...)” is this crate’s stated convention for out-of-tree implementations — crate::middleware::task_input’s module doc names worker_binding as one of those keys. Having no in-tree reader is not the same as having no reader.
• The value lands in the Ctx snapshot ReplayStore persists for each passed step, so a stored run records the binding every spawn declared.

Neither is load-bearing, and this doc does not pretend otherwise. If the extension point is not wanted, delete the layer, its conditional wiring in service::task_launch::TaskLaunchService::launch, and this key — rather than leaving a claim of a live reader standing. That is a breaking change (both items are pub), and it also needs the prose references in crate::middleware::task_input, crate::middleware::agent_context and crate::core::agent_context updated, one of which is an intra-doc link that would otherwise fail the doc build.

Same shape as ProjectNameAliasMiddleware / CompiledAgentTable: a compile/launch-time table keyed by agent name, looked up via ctx.agent at spawn time, no engine state touched.

Structs

WorkerBindingMiddleware

SpawnerLayer that drops the per-agent binding into ctx just before spawn.

Constants

WORKER_BINDING_KEY

Key under ctx.meta.runtime this layer writes the serialized WorkerBinding to.