[vchaindz]

Why

Two problems from running OpenDots daily with a voice Dot and a chat Dot against Ollama Cloud:

1. One model for every Dot. OPENAI_MODEL served both, so a voice Dot that needs a fast non-thinking answer and a research Dot that benefits from reasoning had to share one model and one thinking level.
2. Memory depended on the model choosing to look. With a memory server connected over MCP (Second Brain), a long voice thread stopped calling recall. The same question was answered "nothing stored" six times while the server held the answer. A fresh thread answered correctly, so the tool and data were fine.

What

Per-Dot model (commit 1)

• Optional model and reasoningEffort on a Dot. Empty inherits OPENAI_MODEL and the model default.
• dot-agent builds its adapter from dot.model and sends reasoning_effort through modelOptions, which the OpenAI-compatible adapter already spreads into the request body.
• Two inputs in the Edit specialist dialog. updateDot keeps both values when a patch omits them and clears them on an empty string, so callers sending the old shape change nothing.

Recall before every turn (commit 2, off by default)

• A per-Dot autoRecall toggle. When on, the server calls the Dot's connected recall tool with the latest user message before the model runs.
• It uses only a tool the server marks read-only and the owner has not put behind approval. It waits at most 4 seconds, and any failure leaves the turn running without memory.
• The result is attached to the latest user message, not the system prompt. In the system prompt it lost to recent history: the stuck thread kept saying "nothing stored" with the memory in hand. Attached to the question, it answered correctly twice in a row in that same thread. Stored history is not changed.

Verification

• Format, lint, typecheck, 359 tests, and build pass. Knip reports the same findings as main.
• New tests: tests/auto-recall.test.ts (tool selection, approval and read-only gating, failure paths, message attachment) and a WorkspaceStore test for both new columns across omitted and clearing patches.
• Measured against Ollama Cloud with reasoning_effort:

model, effort	latency	reasoning chars
deepseek-v4.1-flash, none	0.5 s	0
kimi-k3.1-pro, low	1.9 s	122
gpt-oss:120b, medium	0.7 s	405

Not covered: no UI test for the new dialog inputs, and narrow-screen layout was not checked in a browser.

🤖 Generated with Claude Code