Skip to content
Jarvis

ADR 0044 — The Backend Request: Words Only, and Where the Last Answer Was Cut

Status: Accepted · Date: 2026-09-25

Context

  • ADR 0027 (2026-09-21) settled the backend request: rules and profile in system, memory.db records replayed one message per row by role, the per-turn state as one block on the last user message. Each history line kept a [ts] source: prefix so the model could tell GPT-Live’s words from the backend’s.
  • The local voice chain answers from the backend session directly; the last jarvis_live row is from 2026-09-15. The voice model and the backend conversation model are one model (gpt-5.6-luna since 72f56bd).
  • The model copies the prefix into its answers: up to 11 of 14 replies in the 2026-09-25 replay of Allen’s morning session, and again 3 times in that day’s interruption replays (“[2026-09-25T11:42:05-07:00] jarvis: 好，我继续。”). Who spoke is already the message role.
  • The history still spans days (verbatim since history_since, compaction not yet triggered), so the model needs to know which day a row belongs to.
  • Allen’s 2026-09-25 voice test: when he cut Jarvis off, the next turn had no trace of it; his asks were to mark where the answer was cut and let the model decide whether to continue or take up his new words. The event log already carries a validated heard prefix per spoken answer, folded into the packet’s conversation history.
  • Replay of the real session on gpt-5.6-luna with the cut stated in the state block: a new question after the cut answered 6/6 without a retelling; “你继续说” after a cut factual answer resumed with the unheard part 3/3.

Decision

The backend request is four parts, in this order: system is the rules file plus the profile block ([关于 Allen], one line per profile row, omitted when empty); the history is the records rows after the summary anchor and on or after session.history_since, one message per row with the role read off source (allen is user, every other source assistant), the text alone with the retired envelope tags and any label an answer copied into its own text removed, the first user row of each calendar day opening with a [9月24日 周四] line, adjacent rows of one role joined, and the current summary ahead of them as a user message; the last user message is this turn’s utterance with one [当前状态｜程序提供，不是用户说的话] block prepended, holding the time line, the interaction mode, where the previous turn’s spoken answer was cut if it was not heard whole (after a quoted heard prefix, before any word, or with no answer at all), a pending confirmation ask if one is live, and a one-line count of actions still running; a history that ends on an unanswered allen row folds that row in ahead of the block; tools is the registry surface for jarvis_llm. No other note enters the request. Records before history_since stay in memory.db for search_records / read_records. The backend is not asked for <voice>/<document>.

Alternatives rejected

  • Keep the [ts] source: prefix — measured above: the model writes it into answers, and the only reader it served, a Live turn in the history, has not happened in ten days.
  • A timestamp on every row without the source — the copied shape is the line prefix itself; one date line per day on a user row gives the day and puts nothing at the head of the model’s own messages.
  • Mark the cut inside the interrupted assistant message — tried on the same replay: the marker after the answer and the answer cut down to the heard part both did no better on “继续讲” after a cut story (the model continues past the story’s end in every variant), and cutting the answer lost facts the model needed to resume a factual answer.
  • The per-turn state as its own user message — every turn would carry two adjacent user messages; a provider that merges or rejects them changes behaviour silently.
  • The history as one text block in the first user message — the model reads its own answers as reported speech; the rows already are the conversation.

Consequences

  • Live’s words, if a Live session runs again, are indistinguishable from the backend’s in the history; the Live startup brief keeps its own labelled lines.
  • “继续讲” after a cut story still continues past the end instead of resuming: the history holds the whole written story, and the spoken form compresses a story to about 60 characters anyway.
  • Tool calls and results are still not in memory.db; the time line and the cut line change every turn, so only system and the history are cache-stable.

Back to decisions ·All decision records on GitHub