Changelog #

0.1.5 #

Perception nodes now carry identifier and value end-to-end to the brain. SemanticsNode parses and re-emits the stable, locale-independent identifier (from Semantics(identifier:)) and the node's value (text-field contents / secure-field bullets); both join ==/hashCode, so a field filling in shows up in the diff. The host emitted these already — the agent model had been dropping them, so the brain only saw them on the native channel.
The bundled agent guide (kDefaultAgentsMd) documents the split: read label to understand what a node is, use identifier as the stable handle for addressing it across locales/sessions, and always act by integer node_id.

Multi-host attach: MultiHostSession attaches to N VM-service hosts at once, merges each host's perception fragment into one observation (side-by-side, keyed by namespace), and routes each tool call to the owning host by namespace (core.* → the Flutter host, native.* → the native channel). A new SessionSurface interface is implemented by BOTH the unchanged single-host LeonardSession and MultiHostSession, so the loop drives either transparently. The agent context-switches by perception, not by hardcoded mode flags.

HandshakeResult gains a capabilities field: host-level features that are reachable but are NOT namespaced tools (so they never appear under extensions) — notably screenshot. The handshake parse reads the new capabilities array and is tolerant of its absence (older bindings parse to an empty list). Lets a driver list screenshot where agents look instead of concluding "no such capability" from the tool manifest alone.

Adopt dartantic as the model-backend seam: a single DartanticModelProvider drives any backend — swift-infer via lenny's custom ChatModel, Anthropic and OpenAI via stock dartantic models. The hand-rolled per-provider classes are removed; the loop keeps retry ownership and the SchemaRejection contract.
Anthropic backend defaults are now compatible with extended thinking: a non-forcing tool_choice (auto) and no temperature override. Anthropic rejects a forcing tool_choice or any non-1 temperature while thinking is enabled, so the previous defaults returned request-time 400s when driving Claude. Thinking stays on; the driver's retry covers a rare prose-only turn.
Fix (Anthropic): per-turn observation context is no longer dropped. The dartantic Anthropic mapper serializes only the tool_result block of a tool-bearing user message and discards sibling text parts, so from turn 1 on the model never saw the observation — it was driven blind after the first turn (the swift-infer path is unaffected; its mapper keeps the text). The Anthropic backend now folds the observation + diff into the tool_result body so the model sees the live screen every turn.
Observation: expose scroll extent on scrollable nodes.

Fix: bound runaway model output. The swift-infer provider now aborts a response once it streams a large amount of reasoning text with no tool call in sight, surfacing a retryable SchemaRejection instead of letting weaker models ruminate all the way to max_tokens — the "endless stream, no tool call" failure. The loop retries with a fresh sample.

Initial release.