The official channel data-plane plugins for Revlm (plugins v3 architecture):
OpenAI:ChannelGroup.type == "openai_compatible", handlesGET /v1/models,GET /v1/models/:id,POST /v1/chat/completions,POST /v1/responses, andPOST /v1/responses/input_tokens.Anthropic:ChannelGroup.type == "anthropic", handlesPOST /v1/messages.
Each plugin is an LD_PRELOAD module that provides extern "C" revlm_handle_v1
(the single /v1 data-plane hook), plus the interposable upstream-prepare and
model-catalog symbols. Non-matching ChannelGroup.type chains via
dlsym(RTLD_NEXT).
Build the modules against the host core headers (a matching Revlm checkout or
an installed RevlmCore package), then package each plugin with
packaging/build-package.py. The script intentionally requires both artifacts,
because production packages must contain backend/amd and backend/arm.
Revlm's multi-architecture container release also builds one image-local
system package per image. It uses --system-target linux-amd64|linux-arm64
with the native module; that package deliberately contains only the image's
own architecture and is never used as an upload artifact. The release Docker
build expands it with packaging/install-system-package.py into the host's
read-only REVLM_SYSTEM_PLUGIN_DIR layout.
See the package format for the ABI and lifecycle contract.