openral-reasoner

OpenRAL S2 reasoner — the event-driven slow planning loop (CLAUDE.md §3 Layer 4). ReasonerCore consumes a WorldState snapshot plus rolling event buffers (failures, perception events, operator prompts) and emits exactly one typed ReasonerToolCall per tick via the LLM's structured tool-use API — no free-form JSON. The ROS-side reasoner_node (in packages/openral_reasoner_ros/) wraps this core with rclpy subscriptions and dispatch plumbing.

Layer

CLAUDE.md §3 Layer 4 — S2 slow reasoning: event-driven with a ~0.2 Hz heartbeat, sitting between the WorldStateAggregator and the S1 skill executor. Dispatch is direct typed tool calls.

Design

• ROS 2 reasoner supervisor: direct typed tool-call dispatch.
• Symbolic S2 reasoner: authored playbooks, self-maintained memory, success-gated task queue.
• VLM-adjudicated completion, grounding-before-decompose, detection identity.

Public surface

from openral_reasoner import (
    ReasonerCore, ReasonerTickResult,          # the S2 tick loop
    ToolPalette, build_tool_palette,           # registry -> LLM tool palette
    AnthropicToolUseClient,                    # provider clients
    OpenAICompatibleToolUseClient,
    build_tool_use_client_from_env,            # model-first env selection
    ContextRenderer,                           # WorldState/event -> prompt context
    MemoryStore, MissionState,                 # self-maintained memory + mission ladder
    CriticWatchdog, SpatialMemoryQuerier,      # critic gating + spatial recall
    plan_active_search,                        # active-search frontier planning
)

• ReasonerCore — the tick loop: render context → call the LLM with the tool palette → validate into a ReasonerToolCall → bounded replanning ladder on failure (retry → param-tweak → substitute-skill → goal-replan → human-handoff).
• AnthropicToolUseClient / OpenAICompatibleToolUseClient — concrete ToolUseClient implementations; selected at activate-time from the curated openral_core.REASONER_MODELS registry via OPENRAL_REASONER_MODEL. OPENRAL_REASONER_ENDPOINT overrides where the model runs. No hidden default.
• ToolPalette / build_tool_palette — generated from the local skill registry, rebuilt on /openral/skill_registry_changed.
• MemoryStore, MissionState, evaluate_task_verdict — self-maintained memory and success-gated task queue.

See docs/methods/06-reasoning-wam-safety-observability.md for the full symbol inventory and docs/reference/reasoner-design.md for the design walkthrough.