meetings

Context for the Meetings app's data — where a meeting's notes, diagram, transcript-derived tasks and calendar cache live, what the lifecycle states mean, how the app's agents are driven. Load when ask

SKILL.md

GitHub

Meetings app

The Meetings app transcribes a live meeting and fans each line out to a small crew of background agents. A file-backed agent (widget_type markdown or html) owns exactly ONE output file and rewrites it in full after every batch. An agent whose widget_type is chat has no output file — its output IS its chat replies, it gets no OUTPUT_FILE line, and none of the file rules below apply to it.

Only a markdown agent's output is user-editable in the app (the editable minutes). A user edit never touches the agent's file: it is saved as a separate sidecar under edits/ that the dashboard shows in place of the agent's output, so the agent's next rewrite cannot destroy it. That edits/ tree is a sensitive path — no agent can read or write it through file tools — so the agent's own file is NOT what the user sees once they have edited it. An html output is not user-editable.

Only one meeting may be active at a time; starting a second answers 409. The server enforces the lifecycle transition table, not just the UI: active and paused may only reach ended through reviewing (the action-item review gate) on the status route, reviewing may go back to paused, ended may be reopened to active, and a same-status POST is an idempotent no-op. Stop (POST …/meetings/<id>/stop) is the separate exit that ends a meeting outright.

A live translation panel translates each transcript line as it lands, using one tool-less kirocrew-lite call per line in an ephemeral session — deliberately NOT the agent batch path, because latency is the point. It is sequential per meeting and drops the oldest pending line when it falls behind, so gaps in translations.json are expected, not a bug.

Where the data lives

All paths are under ~/.kiro/crew/apps/meetings/data/:

Path What it is
config.json app config: the agent roster, the task/calendar provider ids, presets
dictionary.toml speech-to-text correction terms ([[term]] blocks)
calendar-cache.json the last calendar sync's events
task-ledger.json tasks filed through the local task provider
meetings/<id>/session.json one meeting's metadata (title, status, attendees, attachments, outputs map)
meetings/<id>/tasks.json that meeting's extracted action items
meetings/<id>/<agent-id>.md a markdown agent's output (e.g. note-taker.md)
meetings/<id>/<agent-id>.html an HTML agent's output (e.g. sketch-artist.html)
meetings/<id>/transcript.jsonl the raw transcript, one finalized speech segment per line — read this instead of asking the user to re-summarize
meetings/<id>/translations.json per-line translations for the live translation panel
edits/<id>/<agent-id>.md the user's edit of a markdown output — not readable by agents (see above)

<id> is the meeting id with : replaced by _. Only [A-Za-z0-9._-] is legal in it — the backend rejects anything else, so do not construct a path from a raw calendar UID.

Lifecycle

  1. idle — the meeting folder exists, nothing is running.
  2. active — transcription is flowing; agents receive batched lines every ~30s.
  3. paused — transcription stopped, the session and its queues are intact.
  4. reviewing — the meeting is over and the user is triaging extracted tasks.
  5. ended — finished; outputs are final.

The agents

Agent Writes Notes
meetings-note-taker note-taker.md structured notes: topics, decisions, action items, open questions
meetings-sketch-artist sketch-artist.html one self-contained HTML/Mermaid diagram, revised in place
meetings-task-extractor tasks.json always runs; the app's core output

A file-backed agent's first message carries an OUTPUT_FILE: line. Write to that exact path, character for character, and rewrite the FULL file after every batch — never accumulate content in memory, because a context limit would lose it. A chat agent gets no such line and writes no file.

Lines prefixed [chat] are typed by the user, not transcribed: treat them as corrections or added context and act on them immediately.

Providers

Two things are pluggable, and which implementation is active comes from config.json:

  • task_provider — where a reviewed action item is filed. The shipped provider is local, which appends to task-ledger.json.
  • calendar.provider — where upcoming meetings come from. none (default, meetings are created by hand) or ics, which reads the iCalendar document at calendar.source (a local .ics file path, or a published https:// URL).

To sync the calendar now, ask the user to press Sync calendar on the Meetings page — do not try to fetch or parse the calendar yourself. The app's /api/apps/meetings/* routes sit behind the dashboard's token auth and no agent tool carries that credential, so a raw HTTP call from an agent is refused. With a provider configured the app also syncs on its own every calendar.poll_interval_secs (default 300) and creates the meeting directory for an event calendar.precreate_lead_minutes (default 15) before it starts, so an imminent meeting usually already exists as idle; calendar.auto_sync: false turns the background poll off, and a lead of 0 keeps the sync but stops pre-creation.

Speech-to-text

Transcription uses KiroCrew's own streaming endpoint (/api/ws/stt), driven from the browser. Cloud transcription is optional (pip install 'boto3>=1.34,<2' 'amazon-transcribe>=0.6.3,<1'); when it is not installed the app says so and the user can still type into the broadcast bar to feed the agents.

Correcting recurring mistranscriptions

When the user complains that a project noun keeps coming through wrong, ask them to add a dictionary term in the Meetings settings (the owner-only route below) rather than correcting it in the notes:

POST /api/apps/meetings/dictionary
{"correct": "DynamoDB", "aliases": ["dynamo db", "dynamo d.b."]}

Matching is case-insensitive with word boundaries and the longest alias wins, so every later meeting gets it right at the point of transcription.