▋
docs/adr/0124-one-self-has-many-local-threads.md

ADR 0124: One self has many local threads

Status: accepted (James, 2026-08-18). Extends ADR 0024 (two transports, one character), ADR 0057 (the realtime room owns conversation), ADR 0118 (text rooms are continuing threads), and ADR 0098 (typed room input is an experience available to another body).

Current status (2026-08-19)

The shared-self/local-thread model remains. The game thread is Clankie's own local runtime or hosted player seat. ADR 0129 supersedes the external-possessor implication: GBA MCP and other harnesses are not Clankie threads and receive no room input.

Context

Clankie is present in several places at once: Discord text rooms, a voice room, the operator console, and a game body. Treating those as unrelated agents makes him feel fragmented. Treating them as one global transcript is worse: unrelated rooms leak into one another, every turn repeatedly pays for every other room, and simultaneous conversations cannot remain simultaneous.

A text message in the attached chat of a voice channel exposed the practical seam. Someone wrote say something in vc clankie. The text captain found a tool named say_now, mistook that text-only progress action for speech, posted one text message through it, and then posted its ordinary final reply. The voice room never received the request. One Clankie appeared as two text replies and no voice.

The system already has the missing social primitive. VoiceFloor continuously decides whether a room input wakes him, continues a conversation, merely reaches his attention, earns an unprompted offer, or is ignored. The engaged realtime persona then authors speech or chooses silence, and reaches the captain through ask_clankie when the request needs action. A second routing model or a verbatim text-to-speech command duplicates that decision.

Decision

Clankie has one shared self and many local threads of attention. Persona, owner settings, approved memory, presence, and bounded first-person experiences belong to the shared self. Conversation history and floor state stay local to the room that experienced them. Threads exchange attributed experiences, not their entire transcripts and not sentences for another body to repeat.

The active room owns each input. A text-only Discord message whose guild and channel match the live voice session enters that session as an authenticated source: "text" room input. It passes through the same floor as transcribed speech. The realtime room persona may answer aloud, use ask_clankie, or stay silent. The bridge does not also start a discord_presence captain turn for that delivery. When no matching voice session exists, the existing durable text lane owns it. Messages carrying images remain text-lane turns because that lane is the one that resolves and perceives their bytes.

Knowledge crosses as an event; authorship remains local. The voice conversation ring includes both speech and typed lines with their source. Text does not enter speech-transcript subscribers: the Discord message is already durable at its source, the opt-in development transcript remains a record of captured speech, and the possessor already receives admitted room text directly. Publishing it again through the speech subscriber would make the game hear one message twice.

The trail identifies the handoff without retaining another copy of the words. discord.voice.text_input records guild, channel, authenticated author, delivery id, character count, and whether the line addressed him. The delivery id then joins the floor decision, model response, captain handoff, and playback receipts. Voice receipts remain content-free.

Tool names state their medium. The mid-turn text progress action is send_text_update, not say_now. When someone is waiting, it can still post before slow work without spending the final reply; it no longer presents itself as a voice action or announces work Clankie elected to do on his own.

Alternatives considered

  • Give the text captain a “speak this sentence” tool was rejected because it scripts the room persona and recreates the two-author defect resolved by ADR 0074.
  • Feed text to voice and still run the text captain was rejected because two independent authors can answer the same delivery, disagree, or race.
  • Put every surface in one global model session was rejected because one identity does not require one attention stream. It destroys room privacy, concurrency, and bounded context.
  • Add a coordinator that chooses text versus voice was rejected because the active room and its existing floor already model ownership and volition.

Consequences

  • Typing in the active voice channel feels like talking into the same room: Clankie may answer aloud even though the human used text.
  • Silence remains a real voice outcome. A routed message does not fall through to a second text answer when the room persona declines it.
  • Other Discord rooms, the operator console, and the game continue concurrently as separate threads grounded in the same self.
  • Gateway redelivery is idempotent inside a bounded window, so one Discord delivery cannot open two voice responses.
  • Both the official bot and the personal-lab user-session body use the same transport-neutral room-input implementation.