# Shared runtime source candidate (SDK >=0.7.4)

This source declares shared-runtime-v1 and stateless-v1. It is not certified or a claim of live vendor acceptance.
The exact archive requires independent review, signer approval, and two-Workspace invocation acceptance.

- Configuration and credentials are read per invocation. No SDK process-global
  settings are modified.
- Host-managed state and the SDK connection binding supply authority. User config
  strings and adapter business parameters are not tenant identity.
- Explicit HTTP(S) self-hosted endpoints remain supported, including private service
  addresses. Endpoint validation is syntax validation, NOT an SSRF/network sandbox.
  Shared deployments MUST enforce their own egress policy against metadata/control
  endpoints and foreign tenant services. Do not certify a deployment without that
  network-policy evidence. Environment HTTP proxies are no longer implicitly inherited.
- Redirects are not followed by HTTP clients. Configure the final endpoint/attachment
  URL; redirect-dependent workflows must update their URL rather than forwarding
  credentials to an unreviewed destination.
- Tests use installed SDK types and actual local transports/processes, with explicitly
  simulated vendor replies. They are not paid/live vendor-account tests.

## Optional asset gateway

The listener belongs to the SDK-created Runner component (one per artifact), not
a process-global SDK singleton. Concurrent runs with identical host/port/timeout
share the component listener, with distinct revocable invocation capabilities and expiring tokens. Last release
closes the listener and drains accepted tasks. Conflicting fixed-port settings fail
instead of silently reusing or mutating an active listener. Separate installations require an explicit ingress route to their individual run tokens.
Limits: 16 accepted connections, 16 KiB headers, 1 MiB request/response, 32-message
MCP batches, finite request timeout <=120s and token TTL <=3600s. Tool RPC cancellation
does not promise rollback of an already-dispatched Host operation.

## Synchronous vendor SDK isolation

Every SDK operation executes in a new killable Python subprocess, using only the
request's credentials and a minimal environment. Linux/Unix `resource` support is
required; native Windows is not supported by this candidate. Worker address-space
budget is 512 MiB; frame budget 1 MiB; stream budget 16 MiB. OS pipes apply bounded
backpressure. Timeout, cancellation and early close kill/reap the child before
returning. Host run concurrency/cgroup limits remain necessary. No daemon threads
or executor uploads survive cancellation. Parent-owned private temporary directories
are removed after reaping, including killed Tbox uploads. This is not proof of the
production nsjail subprocess/rlimit policy; exercise that policy before signing.

## Identity and per-invocation scope

This plugin derives no upstream user identity. Unlike `coze`/`dify`/`n8n`/`tbox`/
`weknora` there is no `pkg/scoped_identity.py` here and no user id is sent upstream.
(It never had a `_get_user_tag` helper.)

Read per invocation from `ctx.config` (never cached on the shared component):
`app-type`, `api-key`, `app-id`, `remove-think`, `timeout` and the asset-gateway
settings. Each run builds its own `DashScopeClient`, and every synchronous
DashScope SDK call is executed in a fresh, killable subprocess with a private
temporary cwd and a minimal environment (never the parent's cwd, secrets or
import paths). `external.conversation_id` comes from Host-managed run state.

Nothing is binding-keyed: the runner keeps no process cache, no detached
tenant-bearing tasks and no module-level mutable tenant state, so
`on_installation_revoked` has nothing to release.