The Brain panel now auto-scans your machine for a running local agent and connects in one tap, nothing to type. It probes the usual suspects — Ollama, LM Studio, Jan, vLLM, llama.cpp/LocalAI, Text-Gen-WebUI. If you run something it doesn't know yet, let's teach it to find yours.
What: add your local server to the discovery probe list so opening "Your API" finds it automatically. Good candidates not covered yet: GPT4All (:4891), KoboldCpp (:5001), Msty, LocalAI on a nonstandard port, or whatever you actually run at home.
Start here:
host-python/src/dreamlayer/ai_brain/server/backends.py — _LOCAL_AGENT_PROBES (~line 178). Add one dict: label, provider ("ollama" only for the Ollama /api/tags shape; everything OpenAI-compatible is "custom"), base_url (saved verbatim — use the /v1 form for OpenAI-compatible servers; the no-double-/v1 rule in _build_request handles the rest), and models_url (the endpoint that lists models).- Every probe must be a
localhost address. That's the invariant that keeps one-click connect from ever wiring a remote endpoint — a discovered agent is on-device by construction. Don't add a LAN or public host.
_models_from(data) already parses both the Ollama (models[].name) and OpenAI (data[].id) shapes — reuse it. If your server returns a third shape, extend that one function.
Test (no network needed):
host-python/src/dreamlayer/tests/test_primary_api_brain.py → TestOneClickDiscovery. discover_local_agents(getter=fake) takes an injected getter, so you fake the HTTP response and assert your agent shows up with its model list. Mirror test_discover_returns_only_reachable_with_models.
Done means: your server is discovered when it's running, absent when it isn't, the base URL round-trips through _build_request without doubling /v1, and the new test passes with the rest of the suite (pytest -m "not hardware"). Sign your commits (git commit -s) — CI checks DCO.
No hardware needed — everything runs in the test suite.
The Brain panel now auto-scans your machine for a running local agent and connects in one tap, nothing to type. It probes the usual suspects — Ollama, LM Studio, Jan, vLLM, llama.cpp/LocalAI, Text-Gen-WebUI. If you run something it doesn't know yet, let's teach it to find yours.
What: add your local server to the discovery probe list so opening "Your API" finds it automatically. Good candidates not covered yet: GPT4All (
:4891), KoboldCpp (:5001), Msty, LocalAI on a nonstandard port, or whatever you actually run at home.Start here:
host-python/src/dreamlayer/ai_brain/server/backends.py—_LOCAL_AGENT_PROBES(~line 178). Add one dict:label,provider("ollama"only for the Ollama/api/tagsshape; everything OpenAI-compatible is"custom"),base_url(saved verbatim — use the/v1form for OpenAI-compatible servers; the no-double-/v1rule in_build_requesthandles the rest), andmodels_url(the endpoint that lists models).localhostaddress. That's the invariant that keeps one-click connect from ever wiring a remote endpoint — a discovered agent is on-device by construction. Don't add a LAN or public host._models_from(data)already parses both the Ollama (models[].name) and OpenAI (data[].id) shapes — reuse it. If your server returns a third shape, extend that one function.Test (no network needed):
host-python/src/dreamlayer/tests/test_primary_api_brain.py→TestOneClickDiscovery.discover_local_agents(getter=fake)takes an injected getter, so you fake the HTTP response and assert your agent shows up with its model list. Mirrortest_discover_returns_only_reachable_with_models.Done means: your server is discovered when it's running, absent when it isn't, the base URL round-trips through
_build_requestwithout doubling/v1, and the new test passes with the rest of the suite (pytest -m "not hardware"). Sign your commits (git commit -s) — CI checks DCO.No hardware needed — everything runs in the test suite.