From ea568baff71037badb8e5c7e479c33c0082412b9 Mon Sep 17 00:00:00 2001
From: Jimmy
 <1fe240cd1a8cf775f6f3060f115e5a303181f3abf28ad4cb0c2515f4a02b36a8@meshllm.communities.buzz.xyz>
Date: Sat, 12 Sep 2026 15:52:52 +1000
Subject: [PATCH 1/2] fix: restore external model endpoints on Mesh 0.76

Pin the protocol-3 SDK, retain the install ID, and document the provider workflow and its limits.

Signed-off-by: Michael Neale <14976+michaelneale@users.noreply.github.com>
---
 Cargo.lock             | 114 ++++++++++++-------
 Cargo.toml             |   3 +-
 README.md              | 253 +++++++++++++++++++++++++++++++++++++----
 src/lib.rs             |  13 ++-
 tests/host_protocol.rs |  76 +++++++++++++
 5 files changed, 394 insertions(+), 65 deletions(-)
 create mode 100644 tests/host_protocol.rs

diff --git a/Cargo.lock b/Cargo.lock
index baf71ae..c87f3c1 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -34,7 +34,7 @@ checksum = "9035ad2d096bed7955a320ee7e2230574d28fd3c3a0f186cbea1ff3c7eed5dbb"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -55,6 +55,12 @@ version = "0.22.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "72b3254f16251a8381aa12e40e3c4d2f0199f8c6508fbecb9d91f575e0fbb8c6"
 
+[[package]]
+name = "base64"
+version = "0.23.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ac07cdecf99051d9a5238b80f35af32cdeba5b336e55d957b318b50137e18da5"
+
 [[package]]
 name = "bitflags"
 version = "2.11.1"
@@ -129,9 +135,9 @@ checksum = "773648b94d0e5d620f64f280777445740e61fe701025087ec8b57f45c791888b"
 
 [[package]]
 name = "darling"
-version = "0.23.0"
+version = "0.24.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "25ae13da2f202d56bd7f91c25fba009e7717a1e4a1cc98a76d844b65ae912e9d"
+checksum = "ed17f5901b6630b993ca003def43f2f8ef4014fc13b047b57aad617ff32bc2ec"
 dependencies = [
  "darling_core",
  "darling_macro",
@@ -139,26 +145,26 @@ dependencies = [
 
 [[package]]
 name = "darling_core"
-version = "0.23.0"
+version = "0.24.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9865a50f7c335f53564bb694ef660825eb8610e0a53d3e11bf1b0d3df31e03b0"
+checksum = "6837e2cf7485aaae18f86181d2f0e9a7ed297a025e220aeabf63fdebd3a2ddff"
 dependencies = [
  "ident_case",
  "proc-macro2",
  "quote",
  "strsim",
- "syn",
+ "syn 3.0.5",
 ]
 
 [[package]]
 name = "darling_macro"
-version = "0.23.0"
+version = "0.24.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ac3984ec7bd6cfa798e62b4a642426a5be0e68f9401cfc2a01e3fa9ea2fcdb8d"
+checksum = "2ac7135c3ef02b2f7833bbeb1be5ba7f966dcde8a87c6b87f65a778d71a02785"
 dependencies = [
  "darling_core",
  "quote",
- "syn",
+ "syn 3.0.5",
 ]
 
 [[package]]
@@ -169,7 +175,7 @@ checksum = "1ac70aa55017e108007fbaf5aa0f54b021c98f92ff8af59d42eda9da96e3dd4f"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -319,7 +325,7 @@ checksum = "e835b70203e41293343137df5c0664546da5745f82ec9b84d40be8336958447b"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -512,7 +518,7 @@ version = "0.1.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "96547c2556ec9d12fb1578c4eaf448b04993e7fb79cbaad930a656880a6bdfa0"
 dependencies = [
- "base64",
+ "base64 0.22.1",
  "bytes",
  "futures-channel",
  "futures-util",
@@ -762,8 +768,8 @@ checksum = "6b947ae49db0d222b1dbc6b113ce7248a3fc3a6ca21b696717bfc000ba4484d8"
 
 [[package]]
 name = "mesh-llm-plugin"
-version = "0.66.0"
-source = "git+https://github.com/Mesh-LLM/mesh-llm.git?branch=codex%2Fplugin-cli-commands#c0dc2fef75bf1b2919843f196d97797b561d0c59"
+version = "0.76.0"
+source = "git+https://github.com/Mesh-LLM/mesh-llm.git?rev=557630fd0a543aecd2f2b8ef5232ec1829865115#557630fd0a543aecd2f2b8ef5232ec1829865115"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -840,6 +846,7 @@ dependencies = [
  "mesh-llm-plugin",
  "reqwest",
  "serde_json",
+ "tempfile",
  "tokio",
 ]
 
@@ -865,7 +872,7 @@ checksum = "a948666b637a0f465e8564c73e89d4dde00d72d4d473cc972f390fc3dcee7d9c"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -960,7 +967,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "479ca8adacdd7ce8f1fb39ce9ecccbfe93a3f1344b3d0d97f20bc0196208f62b"
 dependencies = [
  "proc-macro2",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -997,7 +1004,7 @@ dependencies = [
  "prost",
  "prost-types",
  "regex",
- "syn",
+ "syn 2.0.117",
  "tempfile",
 ]
 
@@ -1011,7 +1018,7 @@ dependencies = [
  "itertools",
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -1128,7 +1135,7 @@ checksum = "b7186006dcb21920990093f30e3dea63b7d6e977bf1256be20c3563a5db070da"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -1166,7 +1173,7 @@ version = "0.12.28"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "eddd3ca559203180a307f12d114c268abf583f59b03cb906fd0b3ff8646c1147"
 dependencies = [
- "base64",
+ "base64 0.22.1",
  "bytes",
  "encoding_rs",
  "futures-core",
@@ -1216,14 +1223,14 @@ dependencies = [
 
 [[package]]
 name = "rmcp"
-version = "1.7.0"
+version = "3.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0810a9f717d9828f475fe1f629f4c305c8464b7f496c3a854b58d29e65f4058e"
+checksum = "b88db56b8ae316560e9e868b6b978ea940f27cb883323fc90e07435e44158f5c"
 dependencies = [
- "async-trait",
- "base64",
+ "base64 0.23.1",
  "chrono",
  "futures",
+ "indexmap",
  "pastey",
  "pin-project-lite",
  "rmcp-macros",
@@ -1234,19 +1241,20 @@ dependencies = [
  "tokio",
  "tokio-util",
  "tracing",
+ "uuid",
 ]
 
 [[package]]
 name = "rmcp-macros"
-version = "1.7.0"
+version = "3.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6aefac48c364756e97f04c0401ba3231e8607882c7c1d92da0437dc16307904d"
+checksum = "873b730df6f0a9b74b13eb514e0dca4c2db0d8b68b74af98a2e9bf3f9d436585"
 dependencies = [
  "darling",
  "proc-macro2",
  "quote",
  "serde_json",
- "syn",
+ "syn 3.0.5",
 ]
 
 [[package]]
@@ -1339,7 +1347,7 @@ dependencies = [
  "proc-macro2",
  "quote",
  "serde_derive_internals",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -1404,7 +1412,7 @@ checksum = "d540f220d3187173da220f885ab66608367b6574e925011a9353e4badda91d79"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -1415,7 +1423,7 @@ checksum = "18d26a20a969b9e3fdf2fc2d9f21eda6c40e2de84c9408bb5d3b05d499aae711"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -1510,6 +1518,17 @@ dependencies = [
  "unicode-ident",
 ]
 
+[[package]]
+name = "syn"
+version = "3.0.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "12df2e0110f65b775f769bb17ef989067a1d931b2eb822bd4346631eeada89f9"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "unicode-ident",
+]
+
 [[package]]
 name = "sync_wrapper"
 version = "1.0.2"
@@ -1527,7 +1546,7 @@ checksum = "728a70f3dbaf5bab7f0c4b1ac8d7ae5ea60a4b5549c8a5914361c99147a709d2"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -1581,7 +1600,7 @@ checksum = "ebc4ee7f67670e9b64d05fa4253e753e016c6c95ff35b89b7941d6b856dec1d5"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -1619,7 +1638,7 @@ checksum = "385a6cb71ab9ab790c5fe8d67f1645e6c450a7ce006a33de03daa956cf70a496"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -1719,7 +1738,7 @@ checksum = "7490cfa5ec963746568740651ac6781f701c9c5ea257c58e057f3ba8cf69e8da"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -1773,6 +1792,17 @@ version = "1.0.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b6c140620e7ffbb22c2dee59cafe6084a59b5ffc27a8859a5f0d494b5d52b6be"
 
+[[package]]
+name = "uuid"
+version = "1.26.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2ef6dac1e96601b4fb3acccccff2139741fcb757cb9a36089bf5be91cfb285ce"
+dependencies = [
+ "getrandom 0.4.2",
+ "js-sys",
+ "wasm-bindgen",
+]
+
 [[package]]
 name = "vcpkg"
 version = "0.2.15"
@@ -1854,7 +1884,7 @@ dependencies = [
  "bumpalo",
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
  "wasm-bindgen-shared",
 ]
 
@@ -1932,7 +1962,7 @@ checksum = "053e2e040ab57b9dc951b72c264860db7eb3b0200ba345b4e4c3b14f67855ddf"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -1943,7 +1973,7 @@ checksum = "3f316c4a2570ba26bbec722032c4099d8c8bc095efccdc15688708623367e358"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
@@ -2099,7 +2129,7 @@ dependencies = [
  "heck",
  "indexmap",
  "prettyplease",
- "syn",
+ "syn 2.0.117",
  "wasm-metadata",
  "wit-bindgen-core",
  "wit-component",
@@ -2115,7 +2145,7 @@ dependencies = [
  "prettyplease",
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
  "wit-bindgen-core",
  "wit-bindgen-rust",
 ]
@@ -2182,7 +2212,7 @@ checksum = "de844c262c8848816172cef550288e7dc6c7b7814b4ee56b3e1553f275f1858e"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
  "synstructure",
 ]
 
@@ -2203,7 +2233,7 @@ checksum = "11532158c46691caf0f2593ea8358fed6bbf68a0315e80aae9bd41fbade684a1"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
  "synstructure",
 ]
 
@@ -2243,7 +2273,7 @@ checksum = "625dc425cab0dca6dc3c3319506e6593dcb08a9f387ea3b284dbd52a92c40555"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.117",
 ]
 
 [[package]]
diff --git a/Cargo.toml b/Cargo.toml
index 2f4b3bb..4e1fec0 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -8,9 +8,10 @@ repository = "https://github.com/Mesh-LLM/openai-endpoint"
 
 [dependencies]
 anyhow = "1"
-mesh-llm-plugin = { git = "https://github.com/Mesh-LLM/mesh-llm.git", branch = "codex/plugin-cli-commands" }
+mesh-llm-plugin = { git = "https://github.com/Mesh-LLM/mesh-llm.git", rev = "557630fd0a543aecd2f2b8ef5232ec1829865115" }
 tokio = { version = "1", features = ["full"] }
 
 [dev-dependencies]
+tempfile = "3"
 reqwest = { version = "0.12", features = ["json"] }
 serde_json = "1"
diff --git a/README.md b/README.md
index f5a2df4..81d2e92 100644
--- a/README.md
+++ b/README.md
@@ -1,48 +1,259 @@
-# openai-endpoint
+# External Model Endpoint
 
-`openai-endpoint` is an external mesh-llm plugin for routing inference to an
-already-running OpenAI-compatible server such as vLLM, TGI, Ollama, or Lemonade
-Server.
+Connect an already-running OpenAI-compatible server such as vLLM, TGI, Ollama,
+or Lemonade Server to Mesh. “OpenAI-compatible” describes the API format; it
+does not mean the service is operated by OpenAI.
 
-## Install
+The install/config identifier remains `openai-endpoint` so existing configs and
+package lookups do not break. Mesh starts the adapter process, **not** your model
+server. Once registered, Mesh forwards requests directly to that server.
+
+## Compatibility
+
+This source targets **Mesh 0.76.0 / plugin protocol 3**, using the SDK pinned to
+that release commit. Published adapter **0.1.2 uses protocol 2** and cannot start
+on Mesh 0.76.0: it fails with `uses protocol 2, host uses 3`. Installing a package
+successfully does not establish host/plugin compatibility. Until a protocol-3
+adapter release is published, use the source-build option below; the ordinary
+install command still selects the incompatible published release.
+
+Protocol-2 hosts need the older adapter, pinned explicitly:
 
 ```bash
-mesh-llm plugins install openai-endpoint
+mesh-llm plugins install Mesh-LLM/openai-endpoint@0.1.2
+```
+
+Do not bypass the host's protocol check. A future protocol-3 latest release will
+not be compatible with protocol-2 hosts.
+
+## Quick start: share your existing server
+
+**Start the server → install the plugin → set its URL → run `mesh-llm serve`.**
+You do not need a second model download, a GGUF, or a separate `share` command.
+These examples assume Mesh and the provider run on the same machine and that
+Mesh's default ports (9337 API, 3131 console) are free. Start with a fresh/default
+private-mesh configuration; existing discovery/publication settings still apply.
+Do not add `--auto` or `--publish` for this private recipe.
+
+### 1. Start your provider
+
+If your server already works, leave it running and use its existing **HTTP** base URL.
+This forwarding path does not support a direct `https://` upstream, even when
+the HTTPS model-health probe succeeds. See the gateway limitation below.
+Otherwise choose one recipe below. Install the provider first using its own
+documentation; model downloads and hardware requirements belong to that provider.
+
+| Provider | Start it | Plugin `url` |
+| --- | --- | --- |
+| Ollama | Start the Ollama app, or run `ollama serve`; in another terminal run `ollama pull llama3.2:1b` (or use a model you already have). | `http://127.0.0.1:11434/v1` |
+| vLLM | On a supported machine: `vllm serve Qwen/Qwen2.5-0.5B-Instruct --host 127.0.0.1 --port 8000` (or substitute your model). | `http://127.0.0.1:8000/v1` |
+| LM Studio | Download/select a chat model, load it, then open **Developer → Start server**. With the CLI installed, `lms server start` starts the API server. | `http://127.0.0.1:1234/v1` |
+| Other OpenAI-compatible servers | Start your existing TGI, SGLang, Lemonade, or other server using its own instructions. Copy its HTTP OpenAI base URL, including any prefix before `/v1`. | The server's actual base URL; do not assume port 8000. |
+
+The small model names above are examples, not required models. For Ollama use
+`/v1`, **not** its native `/api` API. For LM Studio use the server's displayed
+port if it differs from 1234; Just-In-Time loading may list downloaded models
+that are not loaded yet.
+
+Check the provider directly before adding Mesh (Ollama example; change the URL
+for your provider):
+
+```bash
+curl --fail --silent --show-error http://127.0.0.1:11434/v1/models
 ```
 
-You can also install directly from GitHub:
+Expect a JSON `data` array containing at least one model `id`. For this simple
+recipe the endpoint must be reachable without upstream API-key authentication.
+Keep an unauthenticated server loopback-only; do not disable authentication on
+an existing shared service. See the API-key FAQ below.
+
+### 2. Install the compatible adapter
+
+After checking the **Compatibility** section above:
 
 ```bash
-mesh-llm plugins install Mesh-LLM/openai-endpoint
+mesh-llm plugins install openai-endpoint
 ```
 
-## Configure
+If the compatible release is not yet published, use **Build from source** below
+instead. Do not repeatedly reinstall 0.1.2 on Mesh 0.76.0.
 
-Point the plugin at the external server:
+### 3. Set the URL once
+
+Create `~/.mesh-llm/config.toml` if absent, or edit your existing file. For Ollama:
 
 ```toml
+[runtime]
+mode = "on_demand"
+
 [[plugin]]
 name = "openai-endpoint"
-url = "http://localhost:8000/v1"
+url = "http://127.0.0.1:11434/v1"
+```
+
+For vLLM or LM Studio, change **only the URL** to the value in the table.
+Merge into an existing `[runtime]` section and existing `openai-endpoint` entry;
+do not append duplicate tables. No `[[models]]` entry is needed for the provider.
+`on_demand` avoids eagerly loading configured native models; it does not disable
+native-runtime initialization or future native serving. Explicit `--model` or
+`--gguf` arguments still request loading, so omit them here.
+
+### 4. Serve and check
+
+```bash
+mesh-llm serve
+```
+
+Leave it running. In another terminal:
+
+```bash
+curl --fail --silent --show-error http://127.0.0.1:9337/v1/models
 ```
 
-If you are running the binary yourself instead of installing it through
-`mesh-llm plugins install`, provide the command explicitly:
+Copy an exact `id` from **Mesh's** response into `model` below:
+
+```bash
+curl --fail --silent --show-error http://127.0.0.1:9337/v1/chat/completions \
+  -H 'Content-Type: application/json' \
+  -d '{"model":"REPLACE_WITH_LISTED_ID","messages":[{"role":"user","content":"Say hello."}],"max_tokens":64,"stream":false}'
+```
+
+Success is a completion response with `choices`, not merely a healthy plugin
+process. Your application now uses `http://127.0.0.1:9337/v1` as its OpenAI base
+URL. The provider keeps managing its own models, GPU memory and process.
+
+## Use it from another machine
+
+On a second machine with Mesh installed, join using the invite token printed by
+the serving node:
+
+```bash
+mesh-llm client --join YOUR_INVITE_TOKEN
+```
+
+On that second machine, repeat the two curl checks above against its own port
+9337. It needs neither the endpoint plugin nor a copy of the provider's model.
+Keep the invite private.
+
+Only the serving node needs to reach the upstream URL. The second machine does
+**not** connect to your Ollama/vLLM/LM Studio HTTP port directly. If testing two
+Mesh processes on one machine instead, use separate profiles and distinct API
+and console ports; do not start a second process over a running instance.
+
+## FAQ and troubleshooting
+
+### Does Mesh start or stop Ollama, vLLM or LM Studio?
+
+No. Mesh manages the small adapter process, not your model server. The adapter
+registers the URL and Mesh forwards inference to it directly. Keep your provider
+running; stopping Mesh does not stop that provider.
+
+### Why not `mesh-llm client` on the provider machine?
+
+Use `serve` on the machine sharing the endpoint. `client` is for consuming a
+mesh, not advertising this machine as an inference host. The consumer machine
+can use `client --join` without installing the plugin.
+
+### Is “OpenAI-compatible” an OpenAI account requirement?
+
+No. It describes the HTTP API format. These local recipes need no OpenAI account
+or OpenAI API key. The display name is **External Model Endpoint**; the stable
+install/config name remains `openai-endpoint`.
+
+### What about upstream API keys?
+
+This adapter reads a URL only; it has no provider API-key/header configuration.
+Mesh's endpoint health/model probe makes its own request to `/v1/models` without
+an injected upstream bearer token. Setting `OPENAI_API_KEY` in your app does not
+configure that probe. Do not put credentials in the URL or assume caller
+Authorization headers will authenticate mesh-wide forwarding.
+
+Use the loopback-only unauthenticated recipe for a local server you control.
+An HTTPS or protected provider needs a separately secured gateway exposing
+**loopback HTTP to Mesh** and handling upstream TLS/authentication for both
+model discovery and inference; that setup is outside this quick
+start and must be validated separately. Do not expose an unauthenticated gateway
+to the LAN or Internet to work around this limitation. Do not disable TLS or
+authentication on an existing shared provider.
+
+### Why is the plugin healthy but my model missing?
+
+Plugin-process health and upstream readiness are separate. Check in this order:
+
+1. `mesh-llm plugins info openai-endpoint`: correct install/version and executable?
+2. Provider `BASE_URL/models`: reachable from the serving node, HTTP 200, nonempty
+   `data` with model IDs? A 401/403 indicates authentication; a 404 often means a
+   wrong base path; connection refused usually means wrong port or stopped server.
+3. Mesh `http://127.0.0.1:9337/v1/models`: allow time for health checks/discovery.
+4. Completion: use the exact listed ID and a chat-capable model. A model listing
+   alone does not prove generation; loading or a missing chat template can fail.
+   An `https://` upstream may pass discovery but fail generation: this host
+   forwarding path requires HTTP (see the secured gateway limitation above).
+
+Do not point the plugin at Mesh's own 9337 endpoint: that creates a routing loop.
+`plugins info` describes the installed package, which may not be the executable
+in use when config contains a `command` override. Check that absolute path and
+the startup handshake too; remove the override when switching back to a
+compatible released package.
+
+Restart your own Mesh instance after changing the plugin URL. If the provider is
+in Docker or on another host, `127.0.0.1` means the Mesh process's own network
+namespace, not that other host/container.
+
+### Why does startup say “uses protocol 2, host uses 3”?
+
+The adapter and host releases are incompatible, even if installation succeeded.
+Use a protocol-3 build for Mesh 0.76.0; retain adapter 0.1.2 for protocol-2 hosts.
+Do not bypass the handshake. See **Compatibility** above.
+
+### Does this split my provider's model across Mesh GPUs?
+
+No. It shares an existing inference endpoint. vLLM/Ollama/LM Studio retain their
+own execution and parallelism; attaching their URL does not convert their models
+into Mesh layer packages or distribute their weights.
+
+### Are streaming, tools, vision and every provider certified?
+
+No blanket guarantee. Features depend on the provider/model and Mesh's routing
+path. The compatibility repair was validated on Mesh 0.76.0 with local archive
+installation and non-streaming HTTP-fixture requests through two processes on
+one Mac. The recipes are based on provider documentation, not a claim of live
+certification for every named provider, platform, streaming or authentication mode.
+
+## Provider references
+
+- [Ollama OpenAI compatibility](https://docs.ollama.com/api/openai-compatibility)
+- [vLLM OpenAI-compatible server](https://docs.vllm.ai/en/latest/serving/online_serving/openai_compatible_server/)
+- [LM Studio server](https://lmstudio.ai/docs/developer/core/server) and [OpenAI compatibility](https://lmstudio.ai/docs/developer/openai-compat)
+- [Lemonade API reference](https://github.com/lemonade-sdk/lemonade/tree/main/docs/api)
+
+## Build from source
+
+For Mesh 0.76.0, build this protocol-3 source using Rust and Git:
+
+```bash
+git clone --branch jimmy/endpoint-protocol3-rescue https://github.com/Mesh-LLM/openai-endpoint.git
+cd openai-endpoint
+cargo build --release --locked
+```
+
+Set `command` to the absolute path of that binary in the **same** plugin entry
+from step 3, alongside `name` and `url`:
 
 ```toml
-[[plugin]]
-name = "openai-endpoint"
-command = "openai-endpoint"
-url = "http://localhost:8000/v1"
+command = "/absolute/path/to/openai-endpoint/target/release/openai-endpoint"
 ```
 
-mesh-llm passes `url` to the plugin as `MESH_LLM_PLUGIN_URL`. If neither config
-nor environment is set, it defaults to `http://localhost:8000/v1`.
+Do not add a second plugin entry. Keep that binary in place while it is configured.
+The config URL is passed as `MESH_LLM_PLUGIN_URL`; if neither config nor environment
+sets it, the adapter defaults to `http://localhost:8000/v1`.
 
-## Build
+## Development checks
 
 ```bash
-cargo build
+cargo build --release --locked
+cargo test --locked
+cargo clippy --all-targets --locked -- -D warnings
 ```
 
 ## Release Archives
diff --git a/src/lib.rs b/src/lib.rs
index 53bdfc9..c18e0fb 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -26,7 +26,7 @@ fn build_plugin(name: String) -> mesh_llm_plugin::SimplePlugin {
             plugin_server_info(
                 "mesh-openai-endpoint",
                 VERSION,
-                "OpenAI-Compatible Endpoint Plugin",
+                "External Model Endpoint",
                 "Routes inference to an external OpenAI-compatible server (vLLM, TGI, Ollama, etc.).",
                 Some(
                     "Set MESH_LLM_PLUGIN_URL to point at any server \
@@ -78,6 +78,17 @@ mod tests {
     use serde_json::{Value, json};
     use std::time::Duration;
 
+    #[test]
+    fn targets_mesh_076_protocol_and_preserves_install_identity() {
+        // The published 0.1.2 adapter used protocol 2 and could install but not
+        // initialize on Mesh 0.76. Updating the SDK must preserve this contract.
+        assert_eq!(mesh_llm_plugin::PROTOCOL_VERSION, 3);
+        let plugin = build_plugin(PLUGIN_ID.to_string());
+        assert_eq!(plugin.plugin_id(), "openai-endpoint");
+        let info = serde_json::to_value(plugin.server_info()).expect("server info");
+        assert_eq!(info["serverInfo"]["title"], "External Model Endpoint");
+    }
+
     #[test]
     fn manifest_declares_external_openai_endpoint() {
         let plugin = build_plugin(PLUGIN_ID.to_string());
diff --git a/tests/host_protocol.rs b/tests/host_protocol.rs
new file mode 100644
index 0000000..2f67c02
--- /dev/null
+++ b/tests/host_protocol.rs
@@ -0,0 +1,76 @@
+//! Exercise the executable's control connection, not just its declared URL.
+#![cfg(unix)]
+
+use anyhow::{Context, Result};
+use mesh_llm_plugin::{LocalStream, proto, read_envelope, write_envelope};
+use std::time::Duration;
+use tokio::{net::UnixListener, process::Command, time::timeout};
+
+#[tokio::test]
+async fn executable_initializes_with_mesh_076_protocol() -> Result<()> {
+    // Keep the socket path short enough for macOS, regardless of checkout path.
+    let directory = tempfile::Builder::new().prefix("ep-").tempdir_in("/tmp")?;
+    let socket = directory.path().join("host.sock");
+    let listener = UnixListener::bind(&socket)?;
+    let mut child = Command::new(env!("CARGO_BIN_EXE_openai-endpoint"))
+        .env("MESH_LLM_PLUGIN_ENDPOINT", &socket)
+        .env("MESH_LLM_PLUGIN_TRANSPORT", "unix")
+        .env("MESH_LLM_PLUGIN_URL", "http://127.0.0.1:8000/v1")
+        .kill_on_drop(true)
+        .spawn()?;
+
+    let outcome = timeout(Duration::from_secs(10), async {
+        let (socket, _) = listener.accept().await?;
+        let mut stream = LocalStream::Unix(socket);
+        write_envelope(
+            &mut stream,
+            &proto::Envelope {
+                // Pin the consumer contract independently of the SDK constant.
+                protocol_version: 3,
+                plugin_id: "openai-endpoint".into(),
+                request_id: 1,
+                payload: Some(proto::envelope::Payload::InitializeRequest(
+                    proto::InitializeRequest {
+                        host_protocol_version: 3,
+                        host_version: "0.76.0".into(),
+                        host_info_json: "{}".into(),
+                        mesh_visibility: proto::MeshVisibility::Private as i32,
+                    },
+                )),
+            },
+        )
+        .await?;
+        let response = read_envelope(&mut stream).await?;
+        anyhow::ensure!(response.protocol_version == 3, "wrong envelope protocol");
+        anyhow::ensure!(response.request_id == 1, "wrong response request ID");
+        let Some(proto::envelope::Payload::InitializeResponse(init)) = response.payload else {
+            anyhow::bail!("expected successful initialize response");
+        };
+        anyhow::ensure!(
+            init.plugin_protocol_version == 3,
+            "incompatible plugin protocol"
+        );
+        anyhow::ensure!(
+            init.plugin_id == "openai-endpoint",
+            "changed install identity"
+        );
+        let manifest = init.manifest.context("missing manifest")?;
+        anyhow::ensure!(
+            manifest.endpoints.len() == 1,
+            "expected one external endpoint"
+        );
+        let endpoint = &manifest.endpoints[0];
+        anyhow::ensure!(
+            endpoint.address.as_deref() == Some("http://127.0.0.1:8000/v1"),
+            "configured endpoint URL was not propagated"
+        );
+        anyhow::ensure!(!endpoint.managed_by_plugin, "must not manage the upstream");
+        Ok::<_, anyhow::Error>(())
+    })
+    .await;
+
+    // Reap our own child even when the protocol assertion fails.
+    child.kill().await?;
+    child.wait().await?;
+    outcome.context("plugin handshake timed out")?
+}

From 8059d2c8876f39db67c7794c68bef2d231ea1673 Mon Sep 17 00:00:00 2001
From: Michael Neale <14976+michaelneale@users.noreply.github.com>
Date: Sat, 12 Sep 2026 21:24:31 +1000
Subject: [PATCH 2/2] chore: prepare protocol-3 endpoint release 0.2.0

Co-authored-by: Jimmy <1fe240cd1a8cf775f6f3060f115e5a303181f3abf28ad4cb0c2515f4a02b36a8@meshllm.communities.buzz.xyz>
Signed-off-by: Michael Neale <14976+michaelneale@users.noreply.github.com>
---
 Cargo.lock  |  2 +-
 Cargo.toml  |  2 +-
 README.md   | 27 ++++++++++++++++++---------
 plugin.toml |  2 +-
 4 files changed, 21 insertions(+), 12 deletions(-)

diff --git a/Cargo.lock b/Cargo.lock
index c87f3c1..b244556 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -840,7 +840,7 @@ checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"
 
 [[package]]
 name = "openai-endpoint"
-version = "0.1.0"
+version = "0.2.0"
 dependencies = [
  "anyhow",
  "mesh-llm-plugin",
diff --git a/Cargo.toml b/Cargo.toml
index 4e1fec0..76a7514 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "openai-endpoint"
-version = "0.1.0"
+version = "0.2.0"
 edition = "2024"
 license = "Apache-2.0"
 description = "OpenAI-compatible endpoint plugin for mesh-llm"
diff --git a/README.md b/README.md
index 81d2e92..3e98e71 100644
--- a/README.md
+++ b/README.md
@@ -10,12 +10,21 @@ server. Once registered, Mesh forwards requests directly to that server.
 
 ## Compatibility
 
-This source targets **Mesh 0.76.0 / plugin protocol 3**, using the SDK pinned to
-that release commit. Published adapter **0.1.2 uses protocol 2** and cannot start
-on Mesh 0.76.0: it fails with `uses protocol 2, host uses 3`. Installing a package
-successfully does not establish host/plugin compatibility. Until a protocol-3
-adapter release is published, use the source-build option below; the ordinary
-install command still selects the incompatible published release.
+Adapter **0.2.0 targets Mesh 0.76.0 / plugin protocol 3**, using the SDK pinned
+to that Mesh release commit. Adapter **0.1.2 uses protocol 2** and cannot start
+on Mesh 0.76.0: it fails with `uses protocol 2, host uses 3`.
+
+The unversioned install command selects the latest published GitHub release,
+not a host-compatible version automatically. To select this release explicitly:
+
+```bash
+mesh-llm plugins install Mesh-LLM/openai-endpoint@0.2.0
+```
+
+Existing installs can use `mesh-llm plugins update openai-endpoint` once 0.2.0
+is published. If you used a source `command` override, remove it to use the
+installed package, then restart your Mesh instance. A source build remains an
+option before the release assets are available.
 
 Protocol-2 hosts need the older adapter, pinned explicitly:
 
@@ -23,8 +32,8 @@ Protocol-2 hosts need the older adapter, pinned explicitly:
 mesh-llm plugins install Mesh-LLM/openai-endpoint@0.1.2
 ```
 
-Do not bypass the host's protocol check. A future protocol-3 latest release will
-not be compatible with protocol-2 hosts.
+Do not bypass the host's protocol check. Protocol-3 releases are
+not compatible with protocol-2 hosts.
 
 ## Quick start: share your existing server
 
@@ -232,7 +241,7 @@ certification for every named provider, platform, streaming or authentication mo
 For Mesh 0.76.0, build this protocol-3 source using Rust and Git:
 
 ```bash
-git clone --branch jimmy/endpoint-protocol3-rescue https://github.com/Mesh-LLM/openai-endpoint.git
+git clone --branch v0.2.0 https://github.com/Mesh-LLM/openai-endpoint.git
 cd openai-endpoint
 cargo build --release --locked
 ```
diff --git a/plugin.toml b/plugin.toml
index b4ae936..55efff8 100644
--- a/plugin.toml
+++ b/plugin.toml
@@ -1,4 +1,4 @@
 name = "openai-endpoint"
-version = "0.1.0"
+version = "0.2.0"
 description = "Route mesh-llm inference to an external OpenAI-compatible server."
 command = "openai-endpoint"