Data

The bot keeps nineteen tables in one Neon Postgres database, the one DATABASE_URL points at (at run time the bot connects through DATABASE_URL_POOLED). Voice audio is never stored, only text. Each section lists a table’s columns, the allowed values of its enum-like fields and the rules the database enforces. Retention says what is cleared. For how the data is written, see Messages and backfill, The reminder engine, Summaries, Ticket prompts, Meeting transcripts, Voice notes and Asking questions.

What is stored, and for how long

What	Where	How long
Message text	messages	Until you remove it by hand; a Discord delete empties the text and keeps the row
Search vectors of messages	message_embeddings	Removed with the message text; otherwise until removed by hand
Summary text	summaries	30 days, then emptied
Ticket draft text	ticket_prompts	Cleared sooner in some cases, otherwise at the first start after 30 days
Meeting transcript text	meeting_transcripts	90 days by default, then the row is deleted
Meeting summary and action items	meetings, meeting_action_items	Until you remove them by hand
Questions asked of the bot and their tool trace	ask_questions	30 days by default, then the row is deleted
Reminder history	reminder_runs	Until you remove it by hand
Members, channels, settings	the matching sections below	Until you remove them by hand
Interests a member follows	points_of_interest	Until the member removes them

Retention has the exact rules and settings.

Terms

• Linear is the issue tracker the bot files tickets in.
• Backfill is the bot reading a channel’s older history, newest to oldest, to store it.
• Claim means the bot writes a row (or flips its status to claimed) before it does the work, so a retry cannot send it twice. A crash after the claim can lose that one send. Where the bot does retry stuck rows, the table’s section says so.
• DM is a Discord direct message.
• ON DELETE cascade means deleting the parent row also deletes this row. set null clears the link instead. No cascade means the database refuses to delete the parent while this row exists.
• Index types. A GIN index speeds up full-text search; an HNSW index speeds up nearest-vector search. A unique index also forbids duplicates.

Conventions

• Discord ids are stored as text. Linear ids and enum-like fields (fields with a short fixed list of allowed values) are varchar, checked in code. There are no database enums.
• Every table has the three shared columns below. timestamp(3) keeps milliseconds.
• Shared timestamps have no time zone, so the database TimeZone must stay UTC (the Neon default).
• updated_at is bumped by the app, so raw SQL updates do not change it. A member sync that changes nothing does not bump it either.
• The vector extension (pgvector, vector storage and search for Postgres) must exist. Migration 0006 creates it.
• A search column is a generated tsvector (a word index for full-text search that the database keeps up to date), built with the english full-text configuration. Changing the configuration rewrites the table.

Shared column	Type	Meaning
id	uuid, primary key	Version 7 uuid made by the app, with no database default
created_at	timestamp(3), not null	Default now()
updated_at	timestamp(3), not null	Default now(); set again when the app updates the row

Migrations

Run bun run db:migrate before you start a new version, see Scripts. The migration files are in the bot repository, src/db/migrations/0000 to 0019.

Migration	Change
0000	members
0001	channels and messages, with their indexes
0002	Replaces the channel index on messages and adds messages_sent_idx
0003	member_reminder_preferences, reminder_rules and reminder_runs; channels.linear_team_id and members.setup_nudged_at; the three seeded rules
0004	members.left_at; splits the reminder_runs unique rule into the recurring and follow-up indexes
0005	reminder_runs.shift_date and reminder_runs_shift_unique
0006	The vector extension and message_embeddings
0007	summary_configs, summary_recipients and summaries
0008	summaries_status_created_idx
0009	ticket_prompts
0010	members.linear_link_source
0011	Friendlier text of the standup rule, see reminder_rules
0012	Friendlier text of the follow-up and art-eod rules
0013	The six meeting tables: meetings, meeting_transcripts, meeting_participants, meeting_action_items, meeting_excluded_channels, meeting_settings
0014	meetings.linear_project_id
0015	points_of_interest
0016	ask_questions
0017	The search columns and GIN indexes on meetings and meeting_transcripts
0018	meeting_action_items.cycle_choice
0019	ticket_prompts.draft_team_id and meetings.suggested_team_id

If you skip a migration, the bot breaks in these ways:

• A missing column (0010, 0014, 0017, 0018, 0019) makes every query on that table fail, because the bot reads the whole row. Without 0010 that is members, so nearly everything fails.
• A missing table behind a clean-up (0007, 0009, 0013, 0016) logs [summaries] startup prune failed: ..., [tickets] startup prune failed: ..., [meetings] prune failed: ... or [ask] prune failed: ... at every start, even with the feature off. The rest of the bot runs.
• Without 0015, /interests fails and the rest of the bot runs.
• Without 0011 or 0012 the bot runs with the old reminder wording.

members

One row per human in the server. Bots get no row. The bot syncs the list at every start and on every join, update and leave. Rows are never deleted: leaving the server only sets left_at, and joining again clears it.

Column	Type	Meaning
name	varchar(100), not null	Latest server display name, updated when it changes
discord_user_id	text, not null, unique	Discord user id
linear_user_id	varchar(64), unique	The member’s Linear user id, set by name matching or by /linear-admin link; null when not linked. The unique rule means one Linear user belongs to one member
linear_link_source	varchar(8)	Who decided the link. See linear_link_source values below
timezone	varchar(64)	Time zone name from the IANA list, such as America/New_York. Only valid zone names are saved
work_start	time	Start of the work day
work_end	time	End of the work day. Never equal to work_start; the app checks this
setup_nudged_at	timestamptz(3)	When the one-time setup DM was attempted. Set before sending and kept even if the DM fails. Members who already have a zone and both work times are never nudged
left_at	timestamptz(3)	When the member left the server; null while present

A member is eligible for channel reminders only when left_at is null and timezone, work_start and work_end are all set.

linear_link_source value	Meaning
null	Nobody has decided. The member is open to automatic name matching
auto	The bot linked the member by exact name
manual	An admin linked or unlinked the member with /linear-admin. Name matching never touches the row again

The bot never stores a Linear user’s email or name, only the id. See Linking members to Linear.

channels

Channels and public threads the bot knows about, plus the backfill cursors. Rows are never deleted: when Discord deletes a channel or thread, the row and its team mapping stay and only its messages are emptied.

Column	Type	Meaning
discord_channel_id	text, not null, unique	Discord channel id
parent_channel_id	text	Set for threads only
name	varchar(100), not null	Channel name
kind	varchar(16), not null	See kind values below
newest_message_id	text	Backfill cursor: newest message seen
oldest_message_id	text	Backfill cursor: oldest message seen
backfill_complete	boolean, not null, default false	True once history was read back to the start
linear_team_id	varchar(64)	Linear team this channel is mapped to for summaries. Set by /summary-admin map-channel, cleared by unmap-channel

kind values: text, announcement, forum (media channels count as forum), thread (public and announcement threads; private threads are not stored). Only the backfill moves the cursors, and they only ever widen. Live events never do. Threads and forum posts count as mapped when their parent channel is mapped to that team.

messages

Messages from the text, announcement, forum and public thread channels the bot can read. Deleting a message keeps the row but empties its text. Rows are never removed.

Column	Type	Meaning
discord_message_id	text, not null, unique	Discord message id
channel_id	text, not null	Channel id; the thread id for thread messages
author_id	text, not null	Author’s Discord id
author_name	varchar(100), not null	Display name at the time, cut to 100 characters
author_is_bot	boolean, not null	True for bot authors
type	smallint, not null	Discord message type; only normal messages (0) and replies (19) are stored
content	text, not null, default empty	Message text, with NUL characters removed. For a forwarded message with no text of its own, the forwarded text. Empty after a delete
has_attachments	boolean, not null, default false	True when there is at least one attachment
attachment_count	smallint, not null, default 0	Attachments, including those on a forwarded message
reply_to_message_id	text	Set only for replies
sent_at	timestamptz(3), not null	When it was sent
edited_at	timestamptz(3)	When it was last edited
deleted_at	timestamptz(3)	When it was deleted; the row stays
search	tsvector, generated and stored	to_tsvector('english', coalesce(content, '')), kept up to date by the database

Files, embeds and reactions are not stored, only the attachment count.

A delete also empties the rows that quote the message: the draft of a ticket prompt for it, the question of a tag ask that used it, and the trace of an ask that replied to it. Deleting a channel or thread empties every message in it and in its threads. A delete the bot missed while it was offline is caught when the backfill rescans the last 7 days of a channel.

Indexes:

Index	On
messages_channel_sent_idx	channel_id, sent_at descending, discord_message_id descending
messages_sent_idx	sent_at descending
messages_author_sent_idx	author_id, sent_at descending
messages_reply_to_idx	reply_to_message_id, only where it is not null
messages_search_idx	GIN on search

Edit rule: a deleted row is never brought back, an older edit never overwrites a newer one, and a row with no change is not rewritten. After the first insert only content, the attachment fields and edited_at change.

message_embeddings

One vector per embedded message. Empty unless embeddings are active.

Column	Type	Meaning
message_id	uuid, not null, unique	The message, with ON DELETE cascade
embedding	halfvec(1536), not null	Half-precision vector
model	varchar(64), not null	Embedding model id
content_md5	text, not null	MD5 of the message content at the time it was embedded
embedded_at	timestamptz(3), not null, default now()	When it was embedded

Index: message_embeddings_embedding_hnsw_idx, HNSW on embedding with cosine distance. Only messages that are not deleted, not written by a bot, not blank and of a stored type are embedded. Such a message needs embedding when it has no row, its content_md5 no longer matches md5(content), or the stored model differs from the current one. Deleting a message, or editing it to blank text, removes its vector.

reminder_rules

Column	Type	Meaning
name	varchar(50), not null, unique	Format ^[a-z0-9-]{1,50}$
kind	varchar(16), not null	See kind values below
channel_id	text	Channel to watch; null for follow-ups and unset rules
anchor	varchar(16)	See anchor values below
grace_minutes	integer	Minutes to wait after the anchor (start or end of the member’s work day) before the reminder is due. Null counts as 0. A rule made with /reminder-rules create and no grace gets 120 for a start-of-day rule and 0 for an end-of-day rule
require_attachment	boolean, not null, default false	Needs an attachment to count as posted
message	text, not null	Message template with placeholders
prompt	text	Not read by anything today
linear_team_id	varchar(64)	Not read by anything today
enabled	boolean, not null, default true	Rules are disabled, never deleted

kind values: channel_watch, follow_up. anchor values: start_of_day, end_of_day. The app lets a follow-up rule change only its message. Check reminder_rules_grace_minutes_check: grace_minutes is between 0 and 1440 (null passes).

Migration 0003 adds three rules and sets no channel on any of them.

Rule	Kind	Anchor	Grace	Attachment	Enabled	Channel
standup	channel_watch	start_of_day	120	no	yes	not set
art-eod	channel_watch	end_of_day	0	yes	no	not set
follow-up	follow_up	none	none	no	yes	none

Migrations 0011 and 0012 reword three rules, each only where the message still holds its seeded text, and bump updated_at on that row. A rule an admin has reworded is left alone. The default message of a rule created with /reminder-rules create is the new standup text.

• standup (0011), from Morning! You have not posted in {channel} yet. {previous} to {greeting} {name}, looks like you haven't posted in {channel} yet. This is a friendly reminder to do so whenever you get the chance. {previous}.
• follow-up (0012), from Reminder to follow up: {link} {note} to {greeting} {name}, this is a friendly reminder to follow up on {link}. {note}.
• art-eod (0012), from End of day: please post a short summary and an image of what you worked on in {channel}. {previous} to {greeting} {name}, looks like you haven't posted your end-of-day update in {channel} yet. This is a friendly reminder to share a short summary and an image of what you worked on whenever you get the chance. {previous}.

The placeholders {greeting} and {name} are filled in by the bot when it sends, see Reminder rules.

A watch rule with no channel is ignored. On each check of the reminder engine (a tick), if standup has no channel, the bot sets it once to the text or announcement channel named stand-up (any letter case, hyphen required, so standup does not match) that it stored first, and never overwrites a channel that is set. art-eod gets no such help.

member_reminder_preferences

A member’s personal switch and time for one rule.

Column	Type	Meaning
member_id	uuid, not null	The member, with ON DELETE cascade
rule_id	uuid, not null	The rule, with ON DELETE cascade
enabled	boolean, not null, default true	False means the member turned this rule off
custom_time	time	The exact time the reminder fires, with no wait added; null follows work hours

Unique constraint member_reminder_preferences_member_rule_unique on (member_id, rule_id). No row means the rule is on.

reminder_runs

One row per reminder the bot sent, tried or scheduled.

Column	Type	Meaning
rule_id	uuid, not null	The rule; no cascade
member_id	uuid, not null	The member; no cascade
scheduled_for	timestamptz(3), not null	The due time
status	varchar(16), not null	See status values below
sent_at	timestamptz(3)	Set only when the status is sent
failure_reason	text	Why it failed or was skipped
related_message_id	text	Follow-ups only: channelId/messageId. Null for channel reminders
note	varchar(200)	Follow-ups only: the optional note
shift_date	date	Channel reminders only: the member’s local shift day

status values: scheduled, claimed, sent, failed, skipped.

Lifecycle	Steps
Channel reminder	Inserted as claimed, then sent or failed
Follow-up	scheduled, then claimed, then sent, failed or skipped. A claimed row not updated for 10 minutes goes back to scheduled, and the bot puts unprocessed claimed rows back when it shuts down.

Failure and skip reasons: dm_closed, discord_<code> for any other Discord error, a shortened error text of up to 300 characters and internal_error, which is written when anything fails after a run is claimed and before the result is recorded. Follow-ups can also end with rule_not_found, member_not_found, not_allowlisted (skipped; the member is not on REMINDER_ALLOWLIST) and left_guild (skipped).

A channel reminder that a crash leaves in claimed is never retried, and the shift index stops a second one for that shift day. A follow-up is the only run the bot picks up again.

Indexes that enforce the guarantees:

Index	On	Purpose
reminder_runs_recurring_unique	rule_id, member_id, scheduled_for, where related_message_id is null	Lets the bot claim a channel reminder only once
reminder_runs_follow_up_unique	rule_id, member_id, related_message_id, scheduled_for, where related_message_id is not null	The same follow-up is never stored twice
reminder_runs_shift_unique	rule_id, member_id, shift_date	At most one reminder per rule, member and shift day, so changing your hours after being reminded earns no second one. Follow-ups have no shift date, so they never clash.
reminder_runs_due_idx	scheduled_for, where status is scheduled	Finds due follow-ups quickly

summary_configs

One row per summary. Rows are never deleted, only disabled.

Column	Type	Meaning
name	varchar(50), not null, unique	Format ^[a-z0-9-]{1,50}$
kind	varchar(16), not null	See kind values below
description	varchar(200)	Given to the model as context
linear_team_id	varchar(64)	Linear team. Required when kind is project
linear_project_id	varchar(64)	Limits Linear tickets to one project; null means the whole team
prompt	text, not null	What to emphasise. Empty means none. Only the form limits it to 2000 characters
cadence	varchar(16), not null	See cadence values below
weekday	smallint	1 (Monday) to 7 (Sunday). Required for weekly
send_time	time, not null, default 09:00:00	Read in each recipient’s own timezone
ephemeral	boolean, not null, default true	Stored only. Nothing sets it other than the default, so it is true on every row, and no code reads it
review_status_names	text[], not null, default {In Review}	Linear state names that count as in review. No command changes it
review_flag_days	integer, not null, default 7	Tickets in review this many days or more are flagged. No command changes it
enabled	boolean, not null, default true	Summaries are disabled, never deleted

kind values: project, person. person rows are created by /interests, one per member, named interests-<discord user id>, with no linear_team_id, an empty prompt, a daily cadence until the member changes it and the member as the only recipient. They are never deleted, only disabled, and /summary-admin and /summary now do not list or change them. The prefix interests- is reserved: a project summary cannot use it. cadence values: daily, weekly, on_demand.

Checks the database enforces:

Check	Rule
summary_configs_weekly_weekday_check	cadence is not weekly, or weekday is set
summary_configs_weekday_range_check	weekday is null or between 1 and 7
summary_configs_project_team_check	kind is not project, or linear_team_id is set
summary_configs_review_flag_days_check	review_flag_days is 1 or more

Unique constraint: summary_configs_name_unique on name. The database allows a weekday on a non-weekly cadence; the bot does not. No migration adds any summary.

summary_recipients

Who gets a summary and may run /summary now for it. Each row is one person or one role.

Column	Type	Meaning
config_id	uuid, not null	The summary, with ON DELETE cascade
discord_user_id	text	A person’s Discord id
discord_role_id	text	A role’s Discord id. Everyone who holds it when a summary is sent is a recipient

Check summary_recipients_one_target_check: exactly one of the two ids is set.

Index	On	Purpose
summary_recipients_user_unique	config_id, discord_user_id, where the user id is not null	A person is added to a summary once
summary_recipients_role_unique	config_id, discord_role_id, where the role id is not null	A role is added to a summary once

summaries

One row per summary the bot tried to write for one person. Rows are kept, except a claim released after a short Discord lookup failure. After 30 days the text in content is cleared and the rest of the row stays, see Retention.

Column	Type	Meaning
config_id	uuid, not null	The summary; no cascade
member_id	uuid, not null	The person it was written for; no cascade
period_start	timestamptz(3), not null	Start of the period it covers
period_end	timestamptz(3), not null	End of the period it covers
content	text, not null	The full summary text. Empty for failed and skipped rows, and after 30 days
status	varchar(16), not null	See status values below
delivered_at	timestamptz(3)	Set only when the status is sent
failure_reason	text	Why it failed or was skipped
trigger	varchar(16), not null	See trigger values below

status values: claimed, sent, failed, skipped. trigger values: scheduled, on_demand.

Lifecycle	Steps
Scheduled	Inserted as claimed, then sent, failed or skipped. A row still claimed 30 minutes after its last update becomes failed with reason interrupted. If a Discord lookup fails for a short while, the claim is deleted so the next tick can try again.
On demand	Inserted as claimed, then sent or failed. A crash can leave the row claimed for good: only scheduled rows are reclaimed

Failure and skip reasons: no_visible_channels (skipped), no_matches and no_interests (skipped, personal digests only), interrupted, shutdown, ai_not_configured, ai_<ErrorName> such as ai_AiOutputError, generation_failed, dm_closed, discord_<code>, reply_failed (on-demand only), or a shortened error text of up to 300 characters.

Index	On	Purpose
summaries_scheduled_unique	config_id, member_id, period_end, where trigger is scheduled	Lets the bot claim a scheduled summary only once. On-demand rows never clash
summaries_member_sent_idx	config_id, member_id, period_end, where status is sent	Finds a person’s previous sent summary
summaries_status_created_idx	status, created_at	Counts sent and failed summaries in the last 24 hours for the daily limit

Because of the first index, a failed or skipped scheduled summary is not tried again for the same period.

points_of_interest

What a member follows with /interests. Rows belong to one member and are hard-deleted when removed or when the member row is deleted.

Column	Type	Meaning
owner_member_id	uuid, not null	The member, with ON DELETE cascade
person_discord_user_id	text	The followed person’s Discord id
topic	varchar(200)	Free-text topic
channel_ids	text[], not null, default {}	Up to 3 channel ids
prompt	varchar(500)	Extra note for the digest

Check	Rule
points_of_interest_subject_check	A person, a topic or at least one channel is set
points_of_interest_channels_max_check	At most 3 channels

Index points_of_interest_owner_idx on owner_member_id. The cap of 10 per member is enforced by the bot, not the database.

A digest that matches nothing is stored in summaries as skipped. no_matches rows count toward the daily generation limit.

ticket_prompts

One row per message the bot acted on after checking it for a feature idea: a prompt sent, a prompt that failed, or a match to an existing Linear ticket. Rows are never deleted; only the draft text and the suggested team are cleared.

Column	Type	Meaning
discord_message_id	text, not null, unique	The message that was checked. One row per message
channel_id	text, not null	Channel id of that message
author_id	text, not null	The author’s Discord id
status	varchar(16), not null, no default	See status values below
draft_title	varchar(200)	The AI-written title, or the text the author last typed. Null once cleared
draft_description	text	The AI-written description, or the text the author last typed. Null once cleared. The app limits it to 4000 characters
draft_team_id	varchar(64)	The Linear team id the AI suggested, when the author was offered two or more teams and it clearly named one. The form opens with that team selected. Null when there is no suggestion or once cleared
linear_issue_id	varchar(64)	The Linear issue id. The bot picks a random uuid when the author submits the form and creates the issue in Linear under that id. It stays set after a failed create so a retry reuses it
linear_issue_identifier	varchar(32)	The Linear identifier. Set on created and duplicate
linear_issue_url	text	The Linear link. Set on created and duplicate
failure_reason	text	Why a prompt failed

status values: prompted, creating, created, dismissed, failed, duplicate. The database does not check them.

Status	Meaning
prompted	Inserted before the DM is sent. The author has not decided, or a failed create was released
creating	The author submitted the form and the ticket is being filed
created	The Linear ticket exists
dismissed	The author pressed No thanks
failed	The DM could not be sent, or the message was deleted during the check
duplicate	The bot found a matching Linear ticket and sent no DM

Lifecycle:

Step	Change
Prompt claimed	Inserted as prompted. The insert does nothing if the message already has a row, so each message is prompted at most once
DM not delivered, or message deleted mid-check	prompted to failed. failure_reason holds dm_closed, discord_<code>, message_deleted or an error text of up to 300 characters
No thanks	prompted to dismissed
Form submitted	prompted to creating. A creating row whose updated_at is more than 2 minutes old can be claimed again, for example after a crash
Ticket filed	creating to created
Create failed, or message deleted before filing	creating back to prompted
Match found	Inserted directly as duplicate

created, dismissed, failed and duplicate are final. A stale creating row is reclaimed with the same linear_issue_id, so a retry never makes a second Linear ticket.

Draft text and the suggested team are cleared when the ticket is created, the prompt is dismissed or fails, the source message or its channel is deleted, and 30 days after created_at (see Retention). The status, ids and failure reason stay.

Indexes and constraints:

Index	On	Purpose
ticket_prompts_discord_message_id_unique	discord_message_id	A message is checked and prompted at most once
ticket_prompts_author_created_idx	author_id, created_at	Counts one author’s prompts in the last 24 hours for the daily limit

Rows with status duplicate do not count toward the daily limit. All other statuses do. There are no foreign keys and no check constraints.

meetings

One row per uploaded transcript or voice meeting that the bot claimed. Rows are never deleted. Only the transcript text, in its own table, is deleted.

Column	Type	Meaning
title	text, not null	The meeting title. Empty until the summary is saved
held_at	timestamp(3)	The meeting date, only when the transcript stated it. Asking falls back to created_at when it is empty
linear_team_id	varchar(64)	The Linear team the meeting named, when exactly one public team matched, or the team a first approver picked when it named none. Used for every ticket from this meeting that the approver is offered that team for
suggested_team_id	varchar(64)	The public Linear team the AI suggested, only when no stated team resolved and the meeting has new work to file. Never copied into linear_team_id and never files anything: it only puts that team first in the team picker. Null when there is no suggestion. Not cleared
linear_project_id	varchar(64)	The Linear project the meeting named, when exactly one project of the resolved team matched by exact name. Tickets from this meeting are filed into it only while it is still one of the chosen team’s projects; otherwise they are filed without a project
decision_summary	text, not null, default empty	The decisions, one per line
discussion_summary	text, not null, default empty	The discussion points, one per line
uploader_id	text, not null	The Discord id of the member who uploaded the file, or who started the voice notes. A takeover by a later upload of the same file replaces it
source_message_id	text, unique	The Discord message that held the file. A takeover replaces it with the new upload’s message. Empty for a voice meeting, which the unique key allows many times
source	varchar(16), not null, default upload	upload or voice. The database does not check them
voice_channel_id	text	The voice channel of a voice meeting. Empty for an upload
started_at, ended_at	timestamp(3)	When a voice meeting started and ended. Empty for an upload
channel_id	text, not null	Where the summary was or will be posted. For an upload, the channel of its message. For a voice meeting, the shared summary channel or the voice channel’s own text chat
summary_message_ids	text[], not null, default {}	The ids of the bot’s posts in the channel, saved one by one as each is sent. The first one is linked from the direct messages. A re-run deletes these posts and starts the list again
content_sha256	varchar(64), not null, unique	SHA-256 of the file’s bytes, or of message:<id> for a file refused before it was read. For a voice meeting, the SHA-256 of the transcript, the voice channel id and the start time
status	varchar(16), not null	processing, posted or failed. The database does not check them
search	tsvector, generated and stored	to_tsvector of the title, decisions and discussion, kept up to date by the database
failure_reason	text	Why an upload failed: unsupported, too_large, not_text, empty, fetch_failed, ai_unavailable, daily_cap, uploader_cap, no_channel, recovery_failed, aborted or generic. aborted on a voice meeting means it was saved at shutdown and is finished at the next start

Index: meetings_search_idx, GIN on search.

• The two unique keys make an upload idempotent: a repeat of the file or of the message finds the existing row.
• A processing run refreshes updated_at every minute.
• A failed row, or a processing row not updated for 10 minutes, is taken over with a compare-and-set when the same file is uploaded again. The takeover sets uploader_id, source_message_id and channel_id to the new upload, and saving the draft again replaces the meeting’s participants and action items.
• A posted row whose items still have unsent messages gets them sent when the same file is uploaded again, 2 minutes after its last update.
• At the next start, a voice meeting that has a saved transcript and is failed with aborted, or processing and not updated for 10 minutes, is finished once. When Linear is configured, a posted voice meeting from the last 7 days whose action item DMs never went out gets them sent. See Restart and recovery.

meeting_transcripts

The transcript text, kept apart so it can be deleted without touching the summary.

Column	Type	Meaning
meeting_id	uuid, not null, unique	The meeting, with ON DELETE cascade. One transcript per meeting
transcript	text, not null	The text after reading the file, at most 400,000 characters. For a .vtt or .srt file, the cleaned text. For a voice meeting, the Display Name: words lines
transcription_models	text[], not null, default {}	The speech-to-text models that answered for a voice meeting, such as @playsafe/nvidia_parakeet_0_6B_v3_cluster_1. Empty for an upload. Deleted with the text
search	tsvector, generated and stored	to_tsvector of transcript, kept up to date by the database

Index: meeting_transcripts_search_idx, GIN on search.

Rows are deleted by age, see Retention. Saving the transcript again after a failed run resets created_at.

meeting_participants

The attendee list of a meeting. For an upload, the names as the model read them, at most 50. For a voice meeting, the people who spoke, by their display name. Saving the draft again replaces the list.

Column	Type	Meaning
meeting_id	uuid, not null	The meeting, with ON DELETE cascade
display_name	text, not null	The name as said in the transcript, or the speaker’s display name for a voice meeting
member_id	uuid	The member the name matched (for a voice meeting, the member who spoke), or null. ON DELETE set null

meeting_action_items

One row per action item, up to 15 per meeting. Saving the draft again after a failed run replaces them.

Column	Type	Meaning
meeting_id	uuid, not null	The meeting, with ON DELETE cascade
position	integer, not null	The order in the summary
title	text, not null	The item title
description	text, not null, default empty	The item description. Becomes the ticket description
owner_name	text	The owner as said in the meeting
assignee_member_id	uuid	The member the owner name matched, or null when none or several matched. ON DELETE set null
status	varchar(16), not null, default pending	See below
dm_status	varchar(16), not null, default pending	pending, sent or failed
kind	varchar(16), not null, default new_work	new_work or update_existing. An update changes a ticket that already exists instead of creating one
cycle_choice	varchar(8), not null, default none	none, current or next: the cycle picked in the direct message for the ticket this item files. The database does not check them. Updates never move a ticket into a cycle
update_kind	varchar(16)	status or comment for an update, null otherwise
target_status	varchar(16)	For a status update: backlog, todo, in_progress or in_review. There is no value for done or canceled
target_issue_identifier, target_issue_url, target_issue_title	varchar(32), text, varchar(255)	The existing ticket an update applies to. Null when none was found, and then the item is shown as a note and cannot be selected
possible_issue_identifier, possible_issue_url, possible_issue_title	varchar(32), text, varchar(255)	A ticket that may cover the item. Shown only in the direct message. A ticket that clearly covers it is shown in the channel post and is not stored
linear_issue_id	varchar(64)	The Linear issue id, fixed when the item is first claimed. It stays set after a failed create so a retry reuses it
linear_issue_identifier, linear_issue_url	varchar(32), text	The ticket the bot created, or for an update the ticket it changed. Set on created. An update never sets linear_issue_id
failure_reason	text	Why a direct message failed: not_allowlisted, dm_closed, discord_<code> or send_failed

status value	Meaning
pending	Offered, not selected
approved	Selected in the drop-down
rejected	Dismissed with Dismiss or Dismiss all
creating	Claimed, and being filed or applied. A creating row whose updated_at is more than 2 minutes old can be claimed again
created	The Linear ticket exists, or the update was applied

Indexes: meeting_action_items_meeting_position_idx on meeting_id, position, and meeting_action_items_assignee_status_idx on assignee_member_id, status. A retry reuses linear_issue_id, so it never makes a second Linear ticket. There are no check constraints.

Deleting the upload message or its channel changes none of these rows.

meeting_excluded_channels

Voice and stage channels where members may not start voice notes. Rows are removed by /notes-admin include.

Column	Type	Meaning
channel_id	text, not null, unique	The Discord id of the blocked channel
added_by	text, not null	The Discord id of the admin who blocked it

There are no foreign keys. A deleted channel keeps its row.

meeting_settings

One row per server, written by /notes-admin summary-channel set.

Column	Type	Meaning
guild_id	text, not null, unique	The server id
summary_channel_name	varchar(100)	The summary channel name an admin chose. Empty means the MEETING_CHANNEL_NAME default applies
updated_by	text, not null	The Discord id of the admin who last set it

The name is read once at startup and kept in memory. The command stores the value first and then updates the running bot.

ask_questions

One row per question claimed with /ask or a tag, created as pending at the claim. Rows hold the question, the tool arguments and the ids the tools returned, no message text and not the answer. Rows are deleted by age, see Retention.

The daily limits count every row from the last 24 hours. A crash can leave a row pending: it counts and is never retried.

Column	Type	Meaning
member_id	uuid, not null	The asker, with ON DELETE cascade
discord_user_id	text, not null	The asker’s Discord id
channel_id	text	The channel of the question. Null when it was not in a channel
via	varchar(16), not null	command or tag
source_message_id	text	The tagging message, for tag rows
context_message_id	text	The message the tagging message replied to, if any
question	text, not null	The question, at most 500 characters. Blank once the tagging message is deleted, or its channel or thread is, whatever the status. An edit does not blank it
status	varchar(16), not null	pending, answered, no_answer or failed
failure_reason	varchar(64)	The error class name, scope_unavailable or no_guild. Set only when the status is failed
trace	jsonb, not null, default []	The model’s tool calls, at most 30. Each entry holds tool, args (cut to 500 characters), resultIds (at most 200) and resultCount. Emptied when the tagging message, its channel or the replied-to message is deleted
cited_refs	text[], not null, default {}	The sources the answer cited, as discord:<id> or meeting:<id>
model	varchar(64)	The model that answered
latency_ms	integer	Milliseconds the answer took
input_tokens, output_tokens	integer	Token counts of the answer call
feedback	varchar(8)	up or down
answered_at	timestamp(3)	When the answer was finished

Indexes: ask_questions_member_created_idx on member_id, created_at; ask_questions_created_idx on created_at; ask_questions_source_message_idx on source_message_id.

Retention

The bot clears or deletes only these things. Everything else stays until you remove it by hand, including messages, vectors, reminder runs, ticket prompt rows, meetings and their action items.

What	Rule	When it runs
summaries.content	Emptied 30 days after created_at. Rows still claimed and rows already empty are skipped. The row stays. The 30 days are fixed	At every start, even with summaries off, and about once a day while the summary scheduler runs
ticket_prompts draft fields	draft_title, draft_description and draft_team_id set to null 30 days after created_at. The 30 days are fixed	At every start only, even with ticket prompts off. Never while the bot keeps running
meeting_transcripts rows	Deleted after MEETING_TRANSCRIPT_RETENTION_DAYS days from created_at. Whole number from 1 to 3650, default 90	At every start, then once a day, even with the feature off
ask_questions rows	Deleted after ASK_LOG_RETENTION_DAYS days from created_at. Whole number from 1 to 3650, default 30	At every start, then once a day, even with asking off

A failed clean-up logs a line like those under Migrations and the bot carries on.

Next

• Where these values are set: Commands.
• What the bot keeps about you in plain words: What the bot stores.