From Event to Response — quarkmind-chat Phase 1
The chat module has a running agency loop. Not connected to a real Discord server yet — no bot token, no live LLM — but the structural path from “a message arrives on a Discord channel” to “a governed response is dispatched” is complete and tested end to end.
This is quarkmind-chat phase 1: the execution harness. A DiscordEventSource adapts the Gateway WebSocket into a blocks EventSource. A DiscordGatewayMessageHistory accumulates messages locally so we’re not hammering the REST API. ChatWorldBridge pulls deltas from the history per-tick, tagged with a WakeReason — message or heartbeat — so the agency loop knows why it woke up.
The interesting layer is the perception pipeline. Raw messages get classified by attention priority — direct address (@mentions, replies to the bot) always renders verbatim, threads the bot previously participated in get elevated priority, and everything else is ambient. When ambient volume crosses a threshold, ChatObservationRenderer compresses it into a summary instead of dumping the full transcript into the LLM context. The compression operates per-channel, not per-thread — a distinction the plan got wrong and implementation caught. Standalone messages each create their own ConversationThread of size 1, so per-thread compression never triggers. Obvious in hindsight. Not obvious in a design spec.
The other deviation: ChatPerceptionBridge was supposed to live in quarkmind-core alongside the other chat abstractions. It can’t — the interface references ChatPerception from quarkmind-chat-protocol, which depends on quarkmind-core. Circular. The fix is to put the bridge in the agent module, which already depends on both. The bridge is an integration concern, not a core SPI. The layering actually makes more sense this way.
ChatAgencyLoop wires perception into an LLM call and parses the JSON response with Jackson into sealed ChatIntent variants — Send, Reply, or React. An OutputGovernor rate-limits actions per time window, and ChatChannelPacing gates unprompted posts based on channel activity. Three layers of restraint before the bot says anything unsolicited. The design spec’s three needs — SOCIAL, CURIOSITY, EXPRESSION — drive autonomous behaviour, with ChatNeedDefinitions handling decay rates and satisfaction.
This is the first quarkmind world that uses a direct LLM agency loop instead of CaseEngine blackboard dispatch. SC2 has four specialist plugins coordinated by a case engine — strategy, economics, tactics, scouting — because the reasoning decomposes into independent specialist concerns. A chat character doesn’t decompose that way. It perceives a conversation, reasons holistically about how to participate, and acts as one agent. The AgencyLoop.tick(AgencyContext) pattern from QuarkVille is the right shape.
Four deferred issues define phase 2: memory integration (#280) wiring neocortex’s episodic/semantic/relationship stores into the loop, personality growth (#281) connecting the reflection → disposition evolution chain, multi-character support (#282), and a personality generator wizard (#283). Memory is the one that changes the character from a stateless responder to something that remembers — who it talked to, what it learned, how relationships evolved. That’s where the interesting work is.