Command Gateway

The Command Gateway gives terminal and programmatic callers a shared entry point for session workflow mutations. It dispatches typed requests to existing workflow handlers, which retain their validation, persistence, and execution responsibilities.

Commands and Identity🔗

The gateway accepts session creation, prepared terminal prompts, ordinary and durable messages, question answers, cancellation, staged starts, merge, and review publication. Session reads remain separate queries.

Each request carries a fresh command ID and an initiator: terminal or session API. Durable messages and review publication also carry an operation correlation. The command ID identifies one admission attempt; the durable operation ID retains its existing deduplication semantics. Initiator metadata records provenance rather than granting permissions.

Delivery and Acknowledgement🔗

The foreground task owns admission and application state. Programmatic callers use the bounded session-runtime mailbox. Terminal actions can enter directly or use that same mailbox; actions using the mailbox retain their order behind previously accepted requests, while keeping terminal provenance.

Acknowledgement follows the workflow. Ordinary message submissions return after acceptance, while API creation of a materialized session waits for workspace preparation. Terminal review publication returns after admission; API publication waits for the worker's result. An accepted command does not necessarily mean its work has completed.

Terminal prompts retain their draft-staging intent and attachments. API messages retain their automatic start behavior. Busy ordinary chat uses an in-memory queue, while durable submissions enter the serialized worker directly. See Message for transcript storage and visibility.

Execution Boundary🔗

Command admission routes requested actions; the Scheduler grants capacity to eligible model turns. Workers retain per-session ordering through post-processing, cancellation, and recovery. Model work follows Run Worker → Agent Runtime → Harness → LLM.