v2026.3.3

Added

• Sense routing in compose builder -- Route pattern now exposes routing strategy (Broadcast / Keyword / Sense) as inline controls in the dashboard compose builder. Variable service count (3-10) with semantic specialist names (researcher, responder, escalator, analyst, etc.). Collapsible scoring detail shows tag/name/description weights and per-slot quality indicators. Events tab gains a Routing column showing method and score for each delegation event
• Tool search in Cognition panel -- Tool search is now configurable in the dashboard agent creation wizard via the Cognition panel. Function-name picker (resolved from tool_func_map in builder options), auto-pins common functions on first enable, collapsible tuning section. Always visible in the panel (previously gated behind 10+ tools). Info banner appears for agents with 10+ tools suggesting enablement
• Tool search in agent detail -- tool_search exposed in agent detail API response and Config tab. Agent cards and flow nodes show a cyan search badge when tool search is enabled. tool_search added to the features property on AgentSpec
• Cognition panel tooltips -- Every section in the Cognition panel (Pattern, Autonomy, Think, Todo, Tool Search, Guardrails) now has descriptive tooltips. Individual pattern buttons show what each reasoning strategy does
• README Intelligence section -- New top-level section showcasing reasoning patterns, sense routing, and tool search with concrete YAML examples and dashboard references

Changed

• Route pattern flexible topology -- Route pattern is no longer fixed at 4 services. Supports 3-10 services with semantic specialist names from a curated pool. Defaults to 3 (intake + researcher + responder). routing_strategy parameter added to build_compose() and ComposeSeedRequest
• Pattern card descriptions -- Updated compose builder pattern descriptions: chain -> "Linear A -> B -> C pipeline", fan-out -> "Dispatch to all workers simultaneously", route -> "Route to the best specialist automatically" with lime sense badge
• README hero trimmed -- Hero command block reduced from 7 to 3 examples to avoid duplication with the starter table below

Fixed

• Autonomy arrow UX -- Autonomy section in Cognition panel no longer shows a misleading expand arrow when disabled. When unchecked, displays as a flat label + checkbox (matching Think/Todo pattern). Arrow appears only when enabled and sub-settings are expandable
• CognitionPanel toolFuncMap prop -- toolFuncMap prop was declared in the type signature but not destructured in EditorScreen.svelte, causing the variable to be unbound. Added default value {} in the destructuring