@oh-my-ai-sdk/environment

Turn-scoped execution Environment for AI SDK agents: (host, fs, user, cwd) behind one ctx feeding a whole tool suite. bash and every fs tool resolve through the SAME registry, so "read locally, execute remotely" is structurally inexpressible.

import { EnvRegistry, LocalEnv, SSHEnv } from "@oh-my-ai-sdk/environment";
import { makeEnvironmentTools } from "@oh-my-ai-sdk/environment/tools";

const reg = new EnvRegistry();
reg.register(await LocalEnv.open("local"));
reg.register(await SSHEnv.open("build-box", "google-charles"));

const tools = makeEnvironmentTools(reg); // { bash, read, write, edit, grep, glob }
// generateText({ model, tools, ... })

Design (proven in ria lab/env-lab, four generations of live-model probes):

• One persistent shell per environment — local bash child or ssh host bash over ControlMaster — sentinel-framed, so cwd/env/background jobs persist across calls. The tool is stateless; the state lives in the Environment, which is turn-scoped by construction.
• Uniform one-off target on the whole suite. A call names its environment or inherits the registry binding; no tool call mutates the binding (chosen over sticky switching — the binding a model believes it has is the binding it has, regardless of context length).
• binding_note on any result produced under an explicit target: the accident-asymmetry mitigation from the ria consult.
• Visibility invariant: every exec result carries a [user@host:cwd] $ cmd header.
• Unknown target = loud error. Never a silent fallback to local.

Tests: bun run test (local-only, real bash). SSH transport: ENV_SSH_HOST=<alias> bun smoke/ssh-smoke.ts.