Skip to content
TomoriBot

06: Per-Turn Loop

This folder documents the body of the loop in tomoriChat.ts — the four stages that execute once per responding persona inside runWithChannelLock.

Loop semantics

Section titled “Loop semantics”
for (const turn of turnPlan.turns) {
  const context = await buildChatTurnContext(turn);          // 01
  const responseSink = createChatResponseSink(context);      // 02
  const result = await runGenerationTurn(context, responseSink); // 03
  await runPostTurnEffects(context, result);                 // 04
}
  • Iteration scope: one iteration = one persona’s reply attempt. A 3-persona reply runs the body 3 times under a single channel lock.
  • Carried state: stages 01–03 share the same ChatTurnContext closure. Stage 01 builds it; stages 02–03 may mutate it (responseTarget, tomoriState, contextItems). Stage 04 reads it after generation finishes.
  • No persona-to-persona dependency: each iteration is structurally independent. The next iteration sees the Discord-visible effects of the previous one (the previous persona’s reply now appears in the channel history), but not its in-memory state.
  • Multi-persona reality: when turnPlan.turns.length > 1, only the first persona runs in this loop. Stage 05 (planChatTurns) queues the rest at the front of the channel queue so they replay immediately after lock release. So in practice each loop instance runs body once or twice, not for all matched personas at once.

Stage index

Section titled “Stage index”
#StageFileMission
01buildChatTurnContext01-build-context.mdAssemble the LLM-visible prompt for this turn.
02createChatResponseSink02-create-response-sink.mdResolve Discord delivery target + emit/finalize callbacks.
03runGenerationTurn03-run-generation-turn.mdDrive provider call with fallback chain + key rotation.
04runPostTurnEffects04-post-turn-effects.mdRun the post-generation side-effect sequence.

Cross-references

Section titled “Cross-references”