中文 | English

agent-dx-executor

Platform-owned FaaS code package for custom-image Agent instances. It starts the executor HTTP server and user processes from a standard YuanRong FaaS initializer. The wheel is installed into the Agent base image and is not a user-facing SDK.

It has no dependency on the agent-dx SDK or the yr package. The image still must contain a compatible openYuanRong Python Runtime because that runtime loads these entries as an ordinary FaaS code package.

Install the Executor into the Python environment used by the YuanRong Runtime. It is image infrastructure and does not use the function's DELEGATE_DOWNLOAD; that remains available for an optional user process code package. Build the image with, for example:

python -m pip install --no-deps agent_dx_executor-<version>-py3-none-any.whl

The package supports Python 3.9 and later without a minor-version-specific dependency. Current primary runtime targets are Python 3.9 and Python 3.11.

FaaS entries:

initializer: yr.agentexecutor.handler.initialize
handler:     yr.agentexecutor.handler.handle
pre-stop:    yr.agentexecutor.handler.pre_stop

Bootstrap commands are read from YR_RUNTIME_BOOTSTRAP_CMD with the same best-effort behavior as the original Python Runtime implementation: at most 64 argv arrays are considered, malformed entries and individual start failures are logged and skipped.

Process stdin is connected to /dev/null. Stdout and stderr are combined into the append-only ${GLOG_log_dir}/bootstrap_cmd_<index>.log; when GLOG_log_dir is unset it defaults to /home/snuser/log/, and when the log cannot be opened both output streams fall back to /dev/null.

The FaaS PRE_STOP_TIMEOUT bounds shutdown. By default the Executor reserves two seconds of that timeout for forced termination and final cleanup, and uses the remaining time as the child-process SIGTERM grace period.