KaimonGateExt + Point to current repo / branch of gRPCServer 1.0.0

CONTRIBUTING.md

@@ -3,7 +3,7 @@
 Developer notes for working on ReactantServer.jl. For using the server, start with
 [Getting Started](docs/src/manual/getting_started.md).
 
-After cloning, instantiate the workspace (gRPCServer resolves from the `scelles-merge` tag of
+After cloning, instantiate the workspace (gRPCServer resolves from the `main` branch of
 github.com/csvance/gRPCServer.jl via the workspace `[sources]`):
 
 ```
@@ -120,6 +120,36 @@ the two stub files are hand-maintained:
    `control_server_stubs.jl` and the `<Service>_<Rpc>_Client` constructors into
    `control_client_stubs.jl`, preserving the hand-tuned deadlines already in those files.
 
+## Kaimon gate tools (dev only)
+
+`packages/ReactantServer/ext/KaimonGateExt.jl` registers four `rserver_*` GateTools with a running
+[Kaimon](https://github.com/kahliburke/Kaimon.jl) gate, so an agent (or a REPL) can bring a worker
+up in the session where the code is loaded: `rserver_start` (model directory, optional model
+allowlist, `accelerator`, GPU `device`, ports), `rserver_status`, `rserver_models`, and
+`rserver_stop`. KaimonGate is a weak dependency, nothing in `src/` references it, and a session
+that never loads it pays nothing. This is a development surface; deployment is still
+ReactantServerNode, one worker subprocess per GPU.
+
+Two properties are worth knowing before using it:
+
+- `device` is the ordinal among the devices VISIBLE to the process, the same numbering CUDA uses.
+  Which physical card that is stays out of band (`CUDA_VISIBLE_DEVICES` or whatever the process
+  prefers), exactly as it is for a worker under the supervisor.
+- Reactant's XLA client is process-global and is created by the first server, so `mem_fraction` and
+  the preallocated arena are fixed for the whole session. The extension records what the first CUDA
+  start committed to and refuses a later start that disagrees rather than letting it silently
+  inherit the existing arena.
+
+`rserver_start` returns before the server is up (bundle loading and compilation run on a background
+task, and a Kaimon tool call has a hard deadline), so the agent polls `rserver_status`. Argument
+and configuration errors are raised by the call itself.
+
+To drive the tools from a REPL rather than an agent, or after a manual `KaimonGate.stop()`:
+
+```julia
+Base.get_extension(ReactantServer, :KaimonGateExt).reinstall_kaimon_tools()
+```
+
 ## Documentation
 
 The docs are built with Documenter on its default HTML theme, with two plugins:

Project.toml

@@ -1,6 +1,6 @@
 # Workspace root for the ReactantServer monorepo. This environment is not itself a package;
 # it lists the member packages under [workspace] (they share the single root Manifest) and
-# declares the [sources] for gRPCServer (the s-celles-merge branch of the upstream repo) and
+# declares the [sources] for gRPCServer (the main branch of the upstream repo) and
 # the inter-package ReactantServerCore dependency, so every member resolves them consistently.
 
 # packages/ReactantServerExport is intentionally NOT listed here: it is the offline export
@@ -26,9 +26,9 @@ gRPCServer = "608c6337-0d7d-447f-bb69-0f5674ee3959"
 
 [sources]
 ReactantServerCore = {path = "packages/ReactantServerCore"}
-gRPCServer = {url = "https://github.com/csvance/gRPCServer.jl", rev = "s-celles-merge"}
+gRPCServer = {url = "https://github.com/csvance/gRPCServer.jl", rev = "main"}
 
 [compat]
 gRPCClient = "1.1.0"
-gRPCServer = "0.3"
+gRPCServer = "1.0.0"
 julia = "1.12"

README.md

@@ -116,7 +116,7 @@ pulls in the heavy Reactant/XLA stack:
 Offline model export lives in `packages/ReactantServerExport` (a Reactant tracing frontend plus a
 PythonCall-triggered PyTorch extension); it is deliberately **not** a workspace member, so its
 Lux/PythonCall weakdeps stay out of the server images. `gRPCServer` is sourced from the
-`scelles-merge` tag of `github.com/csvance/gRPCServer.jl` through the workspace `[sources]`;
+`main` branch of `github.com/csvance/gRPCServer.jl` through the workspace `[sources]`;
 there are no git submodules.
 
 ## Acknowledgments

docker/Dockerfile

@@ -19,7 +19,7 @@
 # them with a mismatched version.
 #
 # Build context is the repository root. Reactant, HTTP, Prometheus, and gRPCClient are registry
-# deps; gRPCServer resolves from the `s-celles-merge` branch of github.com/csvance/gRPCServer.jl
+# deps; gRPCServer resolves from the `main` branch of github.com/csvance/gRPCServer.jl
 # via the workspace [sources]. NOTE: Manifest.toml is gitignored, so `COPY Manifest.toml` needs a
 # locally generated Manifest in the build context (run `REACTANT_GPU=cuda REACTANT_GPU_VERSION=13.1
 # julia --project=. -e 'using Pkg; Pkg.instantiate()'` first); a fresh clone has none.
@@ -76,7 +76,7 @@ COPY --from=cudalibs /usr/local/cuda/lib64/stubs/libcuda.so /opt/cuda-stub/libcu
 WORKDIR /opt/reactantserver
 
 # Workspace root environment and the member packages. gRPCServer resolves from its
-# s-celles-merge branch on GitHub via the shared Manifest.toml; the remaining forks resolve from
+# main branch on GitHub via the shared Manifest.toml; the remaining forks resolve from
 # the registries. The worker is built on the workspace root; the entrypoint then runs on the
 # ReactantServer member project. .dockerignore drops the package test trees (e.g. the large e2e
 # weights) from the packages/ copy.

docs/Project.toml

@@ -42,7 +42,7 @@ ReactantServerCore = {path = "../packages/ReactantServerCore"}
 ReactantServerGateway = {path = "../packages/ReactantServerGateway"}
 ReactantServerNode = {path = "../packages/ReactantServerNode"}
 ReactantServerExport = {path = "../packages/ReactantServerExport"}
-gRPCServer = {url = "https://github.com/csvance/gRPCServer.jl", rev = "scelles-merge"}
+gRPCServer = {url = "https://github.com/csvance/gRPCServer.jl", rev = "main"}
 # The landing-page plugin, from its git URL so the docs CI can resolve it;
 # a local build can override this with a path entry (see CONTRIBUTING.md).
 DocumenterLandingPage = {url = "https://github.com/csvance/DocumenterLandingPage.jl"}

docs/src/index.md

@@ -86,7 +86,7 @@ through a gRPC control plane, without a restart.
 
 ReactantServer is not yet registered in the General registry, so installation is from the
 repository. The one build dependency beyond the workspace, the gRPC server, is sourced from the
-`scelles-merge` tag of `github.com/csvance/gRPCServer.jl` through the workspace `[sources]`:
+`main` branch of `github.com/csvance/gRPCServer.jl` through the workspace `[sources]`:
 
 ```
 git clone https://github.com/EnzymeAD/ReactantServer.jl

docs/src/manual/getting_started.md

@@ -19,7 +19,7 @@ ReactantServer is a Julia workspace of five packages under `packages/` (`Reactan
 `ReactantServer`, `ReactantServerGateway`, `ReactantServerClient`, `ReactantServerNode`), plus the
 non-member `ReactantServerExport` for offline bundle export (see
 [Architecture](../design/architecture.md)). It sources its gRPCServer dependency from the
-`scelles-merge` tag of `github.com/csvance/gRPCServer.jl` through the workspace `[sources]`;
+`main` branch of `github.com/csvance/gRPCServer.jl` through the workspace `[sources]`;
 the remaining dependencies resolve from the package registries. After cloning, instantiate the
 workspace:
 

docs/src/tutorial.md

@@ -40,7 +40,7 @@ to one another only through files and the gRPC wire protocol.
    ready: `julia --project=packages/ReactantServerExport/test`.
 2. **Server environment.** The workspace under `packages/` (`ReactantServerCore`,
    `ReactantServer`, `ReactantServerNode`, `ReactantServerGateway`), with gRPCServer sourced from
-   its `scelles-merge` tag on GitHub through the workspace `[sources]`. After cloning,
+   its `main` branch on GitHub through the workspace `[sources]`. After cloning,
    instantiate once:
 
 ```text

examples/object_detection/server/Project.toml

@@ -14,7 +14,7 @@ gRPCServer = "608c6337-0d7d-447f-bb69-0f5674ee3959"
 HTTP = {path = "../../../lib/HTTP.jl"}
 ReactantServer = {path = "../../../packages/ReactantServer"}
 ReactantServerCore = {path = "../../../packages/ReactantServerCore"}
-gRPCServer = {url = "https://github.com/csvance/gRPCServer.jl", rev = "s-celles-merge"}
+gRPCServer = {url = "https://github.com/csvance/gRPCServer.jl", rev = "main"}
 
 [compat]
 gRPCClient = "1.1.0"

examples/transformers/server/Project.toml

@@ -13,7 +13,7 @@ gRPCServer = "608c6337-0d7d-447f-bb69-0f5674ee3959"
 HTTP = {path = "../../../lib/HTTP.jl"}
 ReactantServer = {path = "../../../packages/ReactantServer"}
 ReactantServerCore = {path = "../../../packages/ReactantServerCore"}
-gRPCServer = {url = "https://github.com/csvance/gRPCServer.jl", rev = "s-celles-merge"}
+gRPCServer = {url = "https://github.com/csvance/gRPCServer.jl", rev = "main"}
 
 [compat]
 gRPCClient = "1.1.0"

packages/ReactantServer/Project.toml

@@ -18,12 +18,29 @@ SafeTensors = "eeda0dda-7046-4914-a807-2495fc7abb89"
 YAML = "ddb6d928-2868-570f-bddf-ab3f9cf99eb6"
 gRPCServer = "608c6337-0d7d-447f-bb69-0f5674ee3959"
 
+# KaimonGate is the Kaimon MCP gate's session library, a weak dependency for the extension seam
+# only: when it is loaded alongside ReactantServer, KaimonGateExt registers the `rserver_*`
+# GateTools so an agent can start, inspect, and stop a worker in the session where the code is
+# loaded. Nothing in `src/` references either, a user who never loads KaimonGate pays nothing, and
+# Kaimon appends KaimonGate's OWN environment to JULIA_LOAD_PATH when it spawns a session, so the
+# declaration does not oblige this project's manifest to carry it. It is in test extras so the
+# suite can exercise the extension.
+[weakdeps]
+KaimonGate = "5ee84a8c-75bd-412f-a7b8-4e6463aa635f"
+
+[extensions]
+KaimonGateExt = "KaimonGate"
+
 [compat]
 BFloat16s = "0.6.1"
 DLFP8Types = "0.1.0"
 HTTP = "2"
 InterProcessCommunication = "0.1.6"
 JSON3 = "1.14.3"
+# 1.3 is a floor, not a preference: it is the first release where `serve(tools = ...)` MERGES
+# tools by name on a running gate. Under the older wholesale replacement, loading a second
+# extension (or a notebook worker registering its own tools) silently dropped ours.
+KaimonGate = "1.3"
 NNlib = "0.9"
 Prometheus = "1.5.1"
 ProtoBuf = "1.3.0"
@@ -32,18 +49,19 @@ ReactantServerCore = "1.0.0"
 SafeTensors = "1.2.1"
 YAML = "0.4.16"
 gRPCClient = "1.1.0"
-gRPCServer = "0.3"
+gRPCServer = "1.0.0"
 julia = "1.12"
 
 [extras]
+KaimonGate = "5ee84a8c-75bd-412f-a7b8-4e6463aa635f"
 Logging = "56ddb016-857b-54e1-b83d-db4d58db5568"
 Sockets = "6462fe0b-24de-5631-8697-dd941f90decc"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 gRPCClient = "aaca4a50-36af-4a1d-b878-4c443f2061ad"
 
 [targets]
-test = ["Test", "Logging", "Sockets", "gRPCClient"]
+test = ["Test", "KaimonGate", "Logging", "Sockets", "gRPCClient"]
 
 [sources]
 ReactantServerCore = {path = "../ReactantServerCore"}
-gRPCServer = {url = "https://github.com/csvance/gRPCServer.jl", rev = "s-celles-merge"}
+gRPCServer = {url = "https://github.com/csvance/gRPCServer.jl", rev = "main"}