Skip to content
Vulpine
Esc
↑↓navigate↵open⌘Jpreview
On this page

Meeting events

Typed meeting events from the Meet adapters into the gateway: meeting_started / meeting_segment / meeting_ended / meeting_forget / meeting_opt_out on the owner-local [meet] socket, the per-line acks and closed refusal reasons, how segments land in the recall-by-clock timeline, retention, tombstones and the spoken forget / stop-recording verbs.

Living doc, checked against the code 2026-11-14

meeting.v1 — typed meeting events into the Gateway’s store

  • Status: gateway half built, off by default (ADR 0039). The events travel on an owner-local unix socket between an adapter and the Gateway, not on the WebSocket protocol, so no client sees them.
  • Producers: meetctl (the local two-leg path, tools/meet/) and meet-cloud (the Media API path, tools/meet/meet_cloud.swift). meet-cloud does not send to a Gateway yet. Consumer: the Gateway’s meeting inlet (crates/voice_gateway/src/meetings.rs).
  • Builds on ADR 0034’s events-in contract (JSON lines, Vulpine’s own names only) and the recall-by-clock timeline ([retrieval] timeline_path).
  • Adapters decode; the Gateway owns the transcript store and the recall verbs. A stored segment answers pi_recall_window and “what was said at 22:30” like any other recall by clock. No raw audio and no vendor name cross this seam: a call is a Vulpine id, a leg is near or far, a speaker is a Vulpine label.

Transport

  • One unix socket, [meet] events_socket (default meet/events.sock in the state folder, ~/.local/state/vulpine). The Gateway makes the directory owner-only (0700) and the socket 0600. A socket held by another live Gateway refuses the bind. Owner-local is the authority: there is no key.
  • The inlet runs only when [meet] enabled = true and [retrieval] timeline_path is set (the segments go to that file). Without either, nothing listens.
  • The adapter writes one JSON object per line. Each line is at most 16 KiB. A longer line closes the connection. At most 4 connections at once.
  • The Gateway answers every line with exactly one ack line, in order:
{"ok":true,"type":"meeting_segment","call_id":"m-20261003-2230","seq":12,"stored":true}
{"ok":false,"type":"meeting_segment","call_id":"m-20261003-2230","seq":13,"reason":"call_opted_out"}

stored is false on an accepted line that wrote nothing (a duplicate seq, a draft). An adapter need not read acks. One that does learns of a voice opt-out on its next segment (below).

  • An optional first line {"type":"hello","adapter":"meetctl","contract":1} is acknowledged. A contract other than 1 is refused (contract_mismatch) and the connection closes. meet-cloud’s hello names "adapter":"meet-cloud". The Gateway reads only contract.

The events

All times are unix milliseconds on the adapter’s clock: the Gateway trace’s clock when the adapter runs on the Gateway’s host.

meeting_started

{"type":"meeting_started","call_id":"m-20261003-2230","platform":"meet","started_at":1791090600000,"legs":["near","far"],"notice_given":true}
field rule
call_id [A-Za-z0-9][A-Za-z0-9._-]{0,63} — a Vulpine id the adapter mints (meetctl’s session id). Never a Meet space name or conference record id.
platform [a-z][a-z0-9_-]{0,23} (meet, zoom, browser, fixture)
started_at unix ms, after 2020-01-01 and no more than a day ahead of the Gateway’s clock
legs a non-empty subset of near, far, no repeats
notice_given must be true (ADR 0039 §3: no capture before the notice). false is refused notice_not_given. A fixture replay sends true.

A second meeting_started for an open call is acknowledged (stored:false). For a forgotten or opted-out call it is refused (call_forgotten / call_opted_out): a tombstone cannot be re-imported.

meeting_segment

{"type":"meeting_segment","call_id":"m-20261003-2230","leg":"far","seq":12,"t0":1791091800000,"t1":1791091804200,"text":"Can we move it to Thursday?","speaker":{"label":"far-1","confidence":0.82}}
{"type":"meeting_segment","call_id":"m-20261003-2230","leg":"far","seq":14,"t0":1791091800000,"t1":1791091809000,"text":"Can we move it to Thursday at ten?","replaces":{"from_seq":12,"to_seq":13}}
field rule
call_id a call this Gateway saw start (call_not_started otherwise)
leg one of the call’s legs (leg_not_in_call)
seq per-call, increasing, 0 ≤ seq ≤ 10 000 000; a seq already stored is a duplicate (acked, stored:false)
t0, t1 unix ms, t0 ≤ t1, t1 − t0 ≤ 120 000, t0 not before the call’s started_at − 60 s
text 1–2 000 characters after trimming, no control characters
speaker optional; label [a-z][a-z0-9-]{0,23} (you, far-1), confidence in [0, 1]
replaces optional; from_seq ≤ to_seq < seq, at most 64 lines — the earlier lines are deleted and this one stands for them (the long-form re-decode’s revision id)
final optional, default true; a draft (false) is acknowledged and not stored
source optional, default local; google_artifact (or any other value) is refused source_not_stored (ADR 0039 §2: Google’s transcript is evidence, never the record). meet-cloud’s own decode of the call audio is local, the default.

A call holds at most 50 000 stored segments (call_segment_cap).

meeting_ended

{"type":"meeting_ended","call_id":"m-20261003-2230","ended_at":1791093600000,"reason":"call_left"}

reason is [a-z][a-z_]{0,31}. The known values are call_left, adapter_stopped, leg_lost, cloud_lost, error. Late finals after meeting_ended are still accepted, because decode lags the call.

meeting_forget and meeting_opt_out

{"type":"meeting_forget","call_id":"m-20261003-2230"}
{"type":"meeting_opt_out","call_id":"m-20261003-2230"}

Both delete every stored segment of the call and its timeline session row, and leave one text-free tombstone (call_id, state = forgotten | opted_out, the instant). Every later line for that call is refused with the tombstone’s reason. On either, the adapter deletes its own spool for the call (window WAVs and windows.jsonl); the Gateway never holds that spool. The spoken verbs below do the same from the room.

Refusal reasons (closed)

bad_json, unknown_type, line_too_long, contract_mismatch, bad_call_id, bad_platform, bad_time, bad_legs, notice_not_given, call_not_started, leg_not_in_call, bad_seq, bad_text, bad_speaker, bad_replaces, bad_reason, source_not_stored, call_segment_cap, call_forgotten, call_opted_out, store_unavailable.

In the store

  • The call is a timeline session meet/<call_id>, source meet, spanning started_at–ended_at.
  • Each segment is a timeline event at t0, kind meet <speaker label or leg>, text the segment’s text. A recall window renders it as - 22:30 meet far-1: Can we move it to Thursday?. A side table keeps (call_id, seq, t1) for replacements and retention. The tombstones table holds no text.
  • Retention: segments older than [meet] segment_retention_days (default and maximum 90; a larger value is clamped) are swept when the inlet opens the store, with the emptied call sessions. The adapter deletes window audio after 7 days; the Gateway never sees it. Tombstones are kept.
  • Trace: <trace_dir>/meet.jsonl carries started / ended / forget / opt-out / refusal rows and segment counts, never a segment’s text and never a speaker’s words.

The spoken verbs (closed phrase sets)

Routed to the Gateway only while the inlet runs. Otherwise the words are the brain’s.

verb phrases (“please” and the wake word dropped) effect
forget “forget this call / that call / the call”, “forget this meeting / that meeting / the meeting” meeting_forget on the current call
opt out “stop recording”, “stop recording this call / the call”, “stop transcribing”, “stop transcribing this call / the call”, “don’t (do not) transcribe this call”, “don’t (do not) record this call” meeting_opt_out on the current call

The current call is the latest started call that is open or ended within the last 15 minutes. With none, Vulpine says so. The adapter learns of a voice opt-out from its next segment’s ack (call_opted_out) and stops capture.

Minutes (pi_meeting_minutes)

The brain tool pi_meeting_minutes takes a call id, or last: true for the most recent call. The gateway answers it over the brain’s mutation socket, not the events socket (request meeting_minutes, field call_id = the id or last). Nothing is spoken.

  • Read: the call’s stored segments in time order. Each becomes a line [mm:ss] <speaker label or leg>: text, timed from the call’s start. At most 60,000 characters go to the model. A longer call is cut at a line, and the cut becomes a doubt.
  • Model: one tool-free prompt on the brain lane, with the dictation tidy’s launch and a bound of [brain] minutes_deadline_secs (default 45 s, at most 50). The transcript goes between nonce delimiters as data. The prompt pack is a golden (crates/voice_gateway/src/meeting_minutes_prompt.golden.txt):
    • invent nothing;
    • a proposal is not a decision;
    • list only decisions explicitly made;
    • never attribute a line to a person whose label has no name;
    • a label gets a name only when the transcript states it, otherwise unknown;
    • a mandatory doubt section: misheard words, misunderstood points, unclear speakers, name variants.
  • Answer: exactly one JSON object {summary, decisions[], open_questions[], doubts[], speakers[{label, name}]}. Anything else is bad_output. The gateway also enforces the speaker rules:
    • a speaker label the call does not have is dropped;
    • a label the model left out is listed as unknown;
    • a name the transcript never says becomes unknown, with a doubt line.
    • The doubt list is never empty, and the answer is bounded to 14 KiB.
  • Refusals: no_meetings, call_not_found, call_empty, call_forgotten, call_opted_out. last never reaches past a tombstone newer than the newest call: when the most recent call was forgotten, it refuses and does not summarize an older call.
  • Trace: meeting_minutes rows carry counts and reasons, never text. The model call ends in one meeting_minutes {phase: outcome} row: ms (the call), segments_in, reply_bytes (the model’s answer, 0 when none came back), guard {closed_json, unknown_names, doubts_n, capped} (the answer parsed as the one closed object; names turned unknown; doubt lines returned; items dropped to fit the answer) and outcome (ok, timeout or refused:<reason>, e.g. refused:bad_output). On ok it also carries decisions_n and open_questions_n.
  • On request from the desktop: vulpine-glyphctl meet minutes [call_id|last] and the Minutes button on Vulpine.app’s Dictation tab run the same tool through the typed tool door. The door registers the call under its own door-<uuid> id. It hands the runner the session brain’s mutation socket, token and call id in its environment (VULPINE_TOOL_CALL_ID; never on the command line), and answers the socket itself while the runner runs. The runner may live [brain] tool_door_timeout_secs (default 60 s, at least 10 s past the minutes deadline, checked at startup). A second minutes call while one is open on the session is refused with receipt code busy. vulpine-glyphctl prints the typed answer on one JSON line; it is its one verb that prints model text. The tab shows only the counts and the call id. The client-facing answer is bounded at the tool’s 14 KiB plus a 1 KiB envelope.
  • Admin: the events window (GetEvents) carries last_minutes {call_id, decisions_n, open_questions_n, doubts_n, at_ms} for the last minutes drafted since the gateway started. It is absent before the first.

Was this page helpful?