koder_exec_kit: pluggable execution backends for agent tools
Status
accepted — 2026-05-23. Owner ratified option 3 (new SDK module engines/sdk/koder_exec_kit) over options 1 (CLI-local) and 2 (gateway-hosted). Implementation begins under engines/sdk/koder_exec_kit#001 (scaffold + local backend), which in turn unblocks AICORE-106 (gateway adopts the SDK) and AICORE-134.2b (MountRoot semantics anchored on the Backend contract).
Summary
Define a Go SDK module engines/sdk/koder_exec_kit that exposes a single Backend interface for executing untrusted (or simply isolated) commands and code on behalf of an agent. Ship MVP backends — local, docker, ssh — and a Register(name, factory) plug-in point for the post-MVP backends Hermes-agent inspired the ticket with (Modal, Daytona, Vercel Sandbox, Singularity, …).
Three or more components on the current roadmap consume execution under different trust models:
- Koder AI CLI (
services/ai/ai/cli/) — already runsexecute_commandlocally; only consumer today. - Koder AI Gateway (
services/ai/ai/gateway/) — would gain aPOST /v1/execproxy for headless agents and parallel-execution scheduling (AICORE-119 spoke AIGW-051). - *ervices/agents