The Observation That Doesn't Know Its World
Wacky Manor’s ObservationBuilder had been doing two things: assembling world perception (what’s in the room, who’s here, what you remember from the kitchen) and assembling cognitive state (your goals, your plans, your memories, your last action result). All twenty methods in one static utility, all taking WorldState and CharacterState directly.
This works until a second world wants observations. An SC2 agent or a Godot character has no Room, no CharacterState, no WorldState. But it has the same cognitive needs — goals, memories, reflections, activity history. Those are platform types (AgentGoal, Memory, PartitionedDrain), not manor types.
The split was a WorldObservationProvider SPI in the affordance package — a @FunctionalInterface with a single method: List<ObservationSection> worldSections(). Each world implements it. The manor provides ManorWorldObservationProvider, which captures WorldState, CharacterState, PartitionedDrain, and observer tags, then produces location, exits, objects, characters, remembered rooms, and perception-filtered sections. A separate ManorExchangeObservationProvider handles the simpler two-character dialogue path.
The cognitive side was the more interesting question. Five methods produce ObservationSection from pure platform types: goalsSection(List<AgentGoal>), recentActivitySection(PartitionedDrain), pastExperienceSection(List<Memory>), insightsSection(List<Memory>), relationshipNotesSection(String, List<Memory>). These are formatters — 5-10 lines each, sorting goals by priority, filtering blank memory texts, prefixing relationship memories with “You recall:”. Nothing deep. But they’re the same for every agent, and they belong where the types they produce live.
Where is that? Three candidates.
blocks already depends on eidos-api and neocortex-memory-api. The cognitive formatters are T → ObservationSection factories — they sit alongside ObservationSection.text() and ObservationSection.items() as higher-level constructors that know about domain types. No new dependencies needed.
quarkmind-core has WorldPerception, WorldBridge, AgencyLoop — the agent cognition framework. Conceptually, “how agents represent cognitive state in prompts” fits here. But quarkmind-core is intentionally lean: engine-api and eidos-api only. Adding blocks and neocortex-memory-api would widen a focused core module for five utility methods. And quarkmind-core doesn’t touch rendering at all today — it defines behavioral SPIs, not observation formatting.
Stay in the manor defeats the purpose. The whole point is that a second world shouldn’t have to rewrite goalsSection.
blocks won. Factory methods belong with the type they produce. CognitiveObservationSections went into the affordance package alongside WorldObservationProvider and ObservationSection.
That left four methods that depend on CharacterState — a manor type blocks can’t see. inventorySection, currentThinkingSection, planSections, lastActionResultSection. These stay in the manor’s ObservationBuilder, which is now just an assembly point: call provider.worldSections(), append character state, append cognitive sections from the blocks utility. Eighty lines where there were three hundred and forty.
The three-way split follows the type boundary exactly: methods using WorldState/Room/GameObject → provider; methods using AgentGoal/Memory/PartitionedDrain → blocks; methods using CharacterState → manor. If CharacterState gets abstracted into a platform type someday, those four methods move too. But that’s a different issue with different evidence.
One thing changed beyond the extraction: section ordering. The old code interleaved world and cognitive sections — recentActivity appeared before remembered, keen observations appeared after both. The new layout groups perception together (location, exits, objects, characters, remembered, keen/directed), then character state (inventory, thinking), then cognitive state (goals, plans, activity, memories, insights, last action). Cleaner for the LLM — “what’s around you” before “what you’re thinking.”