fix: restore external model endpoints on Mesh 0.76

[michaelneale]

Summary

Restore the existing plugin workflow for Mesh 0.76.0: share an already-running local HTTP model server without a native model or a new share command.

• Pin the SDK to the exact Mesh 0.76.0 release, fixing the released adapter's protocol-2/host-protocol-3 initialization failure.
• Rename the display to External Model Endpoint while preserving openai-endpoint as the install/config ID.
• Add an executable-level protocol-3 initialization regression and stable-ID/display tests.
• Document simple Ollama, vLLM and LM Studio recipes, second-machine consumption, and FAQ covering lifecycle, discovery, authentication, HTTP-only forwarding and troubleshooting.

User workflow

Start your provider, install a compatible adapter, add this to ~/.mesh-llm/config.toml, then run mesh-llm serve:

[runtime]
mode = "on_demand"

[[plugin]]
name = "openai-endpoint"
url = "http://127.0.0.1:11434/v1" # Ollama; use :8000/v1 for vLLM or :1234/v1 for LM Studio

List /v1/models on Mesh port 9337, then send a completion with an exact listed model ID. Another machine uses mesh-llm client --join YOUR_INVITE_TOKEN.

Protocol and release order

This build targets protocol 3 / Mesh 0.76.0. Published adapter 0.1.2 remains incompatible with that host; this PR does not publish a release. Protocol-2 hosts must retain a compatible adapter (explicit 0.1.2 pin documented). No handshake bypass or host protocol changes.

Companion website guide: Mesh-LLM/mesh-llm#1823

Merge the adapter before the companion website guide; publish a compatible version to repair ordinary catalog installs. The source-build recipe explicitly selects this branch while unreleased. Before deleting that branch, update it to a durable release/main reference. Do not overwrite the existing 0.1.2 release.

Validation

• cargo fmt --all -- --check
• cargo clippy --all-targets --locked -- -D warnings
• cargo test --locked (full package, including executable handshake)
• cargo build --locked
• README fenced TOML and Bash syntax checked.
• Prior final-code package validation: installed with plugins install --archive; host-local and second-node requests returned HTTP 200 and distinct echoed markers through two official 0.76.0 host processes. Final adapter executable SHA-256: 2e3d6313917f…; source/test/dependency state is unchanged since that run (README expanded here).
• Independent source and artifact review by Thinker; HTTPS and source-override documentation findings incorporated.

Limits

Live forwarding evidence is two processes on one Mac against an HTTP fixture, not real-provider/two-machine/streaming/authentication certification. The opt-in direct-provider unit test skips inference unless OPENAI_ENDPOINT_E2E is set; ordinary cargo-test success is not provider certification. Upstream bearer-token injection and direct HTTPS forwarding are not supported by this recipe. No shared instance was modified.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: cc0dceea-8b7c-41dd-8cc2-23da03c68efc

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}