feat: AD4M AI-assistant runtime — /v1 tool-calling + server-side subsystem (verified e2e)

[HexaField]

AD4M AI-assistant runtime (stacked on #854)

Turns AD4M into a local AI-assistant host — verified end-to-end on Apple Silicon with Qwen2.5-7B. /v1 gains real tool calling; a server-side subsystem runs assistant turns inside the executor (persistent, concurrent, surviving client disconnect). The WE front end is coasys/we#95.

The AI is an assistant; AD4M agent stays the human (DID). Local models only (kalosm); no cloud dependency.

Commits

1. /v1 tool/function calling (local, constrained decoding) — request tools/tool_choice/parallel_tool_calls; response tool_calls[] + finish_reason:"tool_calls" (oneshot + streaming). openai_compat/tool_grammar.rs compiles each tool's JSON-Schema at runtime into a kalosm ArcParser that masks the sampler, so a local model's output is always a schema-valid <tool_call> (Hermes/Qwen). auto/none recovered from text. No-tools path byte-identical; Cargo.toml unchanged.
2. Assistant-run subsystem (rust-executor/src/assistant_runtime/) — watches perspectives for WE Message turns; the loop resolves Thread→Assistant→model→personalities/skills/MCP, assembles context, calls prompt_messages_stream (reusing the grammar), executes tools, folds results back, repeats; streams by rewriting Message.content (auto-publishes to WE); resumes from a durable RunState. we:// predicates match WE exactly. Boots beside MCP/REST, gated on --enable-assistants (default on).
3. rmcp MCP client — McpToolProvider connects an assistant's McpServers (stdio → child process; http/streamable/sse → streamable-HTTP; websocket unsupported), list_tools→ToolDef, dispatches call_tool; failed servers logged + skipped.
4. cli: --enable-assistants flag — thread the new Ad4mConfig field through the executor cli crate.
5. fix: resolve model name/id before AIService — the loop passed Assistant.model_id straight through, but the LLM channel is keyed by the model uuid, so a name like "qwen2.5" never matched. Resolve via the /v1 model_selector first. Caught by the live end-to-end test.

Verified end-to-end (live — Apple Silicon, metal, Qwen2.5-7B-Instruct-Q4_K_M)

• Tool calling: /v1/chat/completions + tool_choice:"required" → the real model returned a schema-valid get_weather(location, unit) tool call with content:null and finish_reason:"tool_calls".
• Assistant loop: created an Assistant + Thread + completed user Message as we:// links → the runtime resolved everything, called the model, and wrote a real assistant reply (status:complete, RunState:done) back into the perspective — the whole watcher→loop→model→persist path, live.
• Static: cargo check clean (pinned Rust 1.92); 34 unit tests pass (13 grammar + 21 runtime).
• Not yet live-tested: MCP against a real external server (code compile-verified only); the WE UI visual render (needs the WE app).

cc @lucksus @data-bot-coasys

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 307b40f9-decf-4476-9c05-50c9c1e4313c

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/ai-assistant-runtime

---

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.}