Repository navigation
Releases
v1.0.2
The GitHub organization FreshLabDev is now asterfield-dev, and the bot moves
with it. Nothing a person sees in Telegram changes.
Security
- Built with Go 1.26.11, which fixes GO-2026-6859 in
net/http.
govulncheckfinds the bot reaching it through its HTTP client and health
server.
Changed
- Images are published as
ghcr.io/asterfield-dev/quoto. The organization
was renamed, and GitHub Container Registry does not redirect the old name:
ghcr.io/freshlabdev/quotoanswers 403. Point the deployment's image at the
new path; every earlier digest is there under it too. - tg moves from
v0.1.1-alpha.1tov0.4.0-alpha.2, whose module path is
github.com/asterfield-dev/tg. A press on a button under an inline message
now arrives without a message; quoto sends none, so such a press only stops
the spinner. - Links, the module path and the docs point at
github.com/asterfield-dev.
Old links still redirect.
Known Limitations
- Released as a stable patch without a pre-release first, at the owner's
request: the change is the organization rename and the tg upgrade, checked
by the full test suite, not by a run on a test bot. v1.0.1was tagged but never published: its vulnerability scan stopped
the release on GO-2026-6859. This version is that one plus the Go update.
Container image
ghcr.io/asterfield-dev/quoto@sha256:3517ff65db0f3aa636f5f4b111cc008585dadb1068001c8165eacd175885c4c6
v1.0.0
Quoto now runs on Go. Groups get the same daily quote; the service behind it is
a single binary that uses a fraction of the memory, and about twenty bugs found
while porting are fixed.
In numbers
Measured on ws04 on 2026-10-22: the Python bot in production, the Go bot on the
test stack after a full day with photos, a sticker, a GIF and a Claude pick.
| v0.11.1 (Python) | v1.0.0 (Go) | |
|---|---|---|
| Container memory | 428 to 640 MiB | 8 to 10 MiB |
| Image on disk | 1.05 GB | 223 MB (61 MB to download) |
| Dependencies | 13 packages, 51 pinned | 4 Go modules |
| Settings read from env | 65 | 23 |
| Test functions | 300 | 645 |
| Media model answering | 404 every hour, then the fallback | first try |
On the test stack the quote went out 10 seconds after the cutoff; a photo or a
sticker took about 5 seconds to describe, a GIF 2.3.
What changed for groups
- Anonymous reactions count. Python read a field the reaction-count update
does not have, so the totals of anonymous reactions were never stored. - Media of a group that has not accepted the agreement goes to no AI
provider. It is kept as file type, size and duration until an admin accepts. - Messages older than 72 hours are deleted, and the message text copied into
the evaluation report is blanked at the same age. Before, days that were never
processed stayed forever. - A new Russian-speaking group starts in Europe/Kyiv, not Europe/Moscow.
Admins still pick any zone in the settings. - The welcome message shows the group's real cutoff instead of 21:00, and the
dashboard shows the real media-analysis state instead of always "on". - Voice messages are described in full. Python's silence filter cut a voice
note at its first pause. - Joins, leaves, pins, dice and locations are not stored and no longer count
towards the ten-message threshold. - The Details card drops the Reactions and Length bars. Neither one affected
the pick.
What changed for operators
- New compose manifest,
DATABASE_URLinstead ofDB_URL,/healthzinstead
of a heartbeat file. Rolling back to Python means its old compose file and its
old digest together. - Migrations are embedded and run at start; the Alembic database is adopted
after its tables and columns are checked, and the bot refuses to start if they
differ. - The first night after the deploy re-describes media once (prompt
v2tov3). - License: Apache-2.0 from this release. Releases up to v0.11.1 stay GPL-3.0.
Every change, with the reason for it, is in the v1.0.0-alpha.1 section below
and in docs/go-rewrite.md.
Container image
ghcr.io/freshlabdev/quoto@sha256:69d764729553408b322091c95a6ba28e26eb14c46245e0ed627c84932646238a
v1.0.0-alpha.1
Quoto is rewritten in Go. For the people in the chats it does what it did, with
the changes listed below; what moves is the service around it: one static binary
and ffmpeg instead of a Python interpreter and its packages, and a license
that matches the rest of the family. The design, the compatibility contract and
every decision are in docs/go-rewrite.md.
Operators
Read this part before deploying.
- The new image does not run under the old manifest, and the old image does not
run under the new one.deploy/ws04/compose.yamlrenames the service from
apptobot(the container is stillquoto-bot), builds the database URL
with apostgres://scheme, and checks/healthzinstead of the heartbeat
file. To go back to a Python release, deploy that release'scompose.yaml
together with its image digest; setting onlyQUOTO_IMAGEback fails the health
check andws04 deployrolls back again. The database needs nothing for that. - Renamed:
DB_URLis nowDATABASE_URLand must be apostgres://URL
(postgresql+asyncpg://is refused at start);LOGS_PATHis now
AUDIT_LOG_DIR. The production manifest sets both, so the host.envis not
touched.BOT_TOKENandBOT_USERNAMEkeep their names. - New:
BOT_API_FILES_DIR(required only together with
TELEGRAM_BOT_API_BASE_URL),HTTP_ADDR(:8080, serves/healthz) and
LOG_LEVEL(info). - No longer read:
QUOTE_HOUR,QUOTE_MINUTE,QUOTE_MINUTE_JITTER,
TIMEZONE,MIN_MESSAGES_FOR_AUTO_REVIEW,MAX_MESSAGES_PER_DAILY_EVAL,
SCHEDULER_CONCURRENCY,WEIGHT_*,COLORS,EMOJIS,LENGTH_OPTIMAL_*,
ENABLE_DEVELOPERS_NOTIFY,HEARTBEAT_FILE, theMEDIA_*tuning variables,
OPENROUTER_DATA_COLLECTION,OPENROUTER_ZDR,OPENROUTER_BASE_URL,
OPENROUTER_APP_TITLE,OPENROUTER_HTTP_REFERER,OPENROUTER_MEDIA_{IMAGE,VIDEO,AUDIO}_MODEL
and the*_MAX_TOKENSceilings. Their old defaults are constants now. A name
that is still set is listed once at start (never its value) and does nothing.
The production.envsets only values equal to those defaults, so nothing
changes there. Alerts stay on whileDEVELOPER_IDSis not empty. - Migrations are embedded in the binary and applied at every start, under an
advisory lock, and recorded inschema_migrations. A database that Alembic
last touched (20261018_01) is adopted after its tables and columns are
compared with what the Go version expects; if they differ, the bot refuses to
start and names the table or column.alembic_versionstays.002_indexes
adds two indexes. No column is dropped in this release. - Core contract unchanged: app key
quoto,core.touch,core.set_language,
core.clear_language,core.effective_language; no direct writes tocore.*.
A new database (not production) also needsGRANT REFERENCES ON core.person, core.chat TO quoto_core, which core grants to other bots' roles but not to this
one yet. - The image runs as uid and gid 1000, the owner of the host
logs/
directory. The manifest has nouser:line. - The first night after the deploy costs more. The media description prompt
moves fromv2tov3, so the cache misses once and images and videos are
described again. - The first retention run deletes every message older than 72 hours, which
includes the messages of days that were never processed, and it cannot be
undone by going back to Python. Takepg_dump -n quotobefore the first start. - The first start needs more privileges than a restart. It creates
schema_migrationsand the two indexes of002, so the role needsCREATEon
schemaquotoand ownership ofmessagesandmessage_ai_scores. Python's
Alembic changed those tables as the same role, so it very likely has both; if
not, the start stops withpermission deniedand names theGRANT, and the
data is untouched. Later starts need neither. The catalog queries that check it
are indocs/go-rewrite.mdsection 5.5. - The manifest is hardened:
cap_drop: [ALL],no-new-privileges,
pids_limit: 256andmem_limit: 512m. The test stack publishes its port on
loopback only. - Memory, measured on macOS arm64 against the real binary and pipeline (the
Alpine image and a container were not measured; no Docker was available):
idle 8 to 20 MB, a day of 1500 messages 29 MB, a 50 MB video 135 to 145 MB for
the bot and ffmpeg together (the bot 20 to 26 MB), no growth over 30 videos in a
row. Python used 640 MiB in the container. A 16-megapixel picture sent as a
file takes 125 to 145 MB, a progressive 4:4:4 JPEG of that size up to 300 MB.
Changed
- Media of a group that has not accepted the agreement goes nowhere. Until an
admin accepts version 1.2, a photo, video or voice message reaches neither
Claude nor OpenRouter; it is described from its metadata, as when media
analysis is off. Before, only the Claude path was gated and OpenRouter still
received the media. Media that arrived before acceptance is not analysed
afterwards. - The media model default is
google/gemini-3.5-flash-lite, with
google/gemini-3.1-flash-litebehind it.google/gemini-2.5-flash-litehas no
zero-data-retention endpoint on OpenRouter, so every request carrying
zdr: truewas refused with404once an hour and fell through to the
fallback. - The Details card drops the Reactions and Length bars. Neither affected the
pick (the score is the AI's); the heart count stays. The columns stay and hold 0. - Media: video is never scaled up (a 320x240 clip stayed 960x720 before) and
has one duration cap of an hour; audio is trimmed only at the start, where a
pause in the middle used to cut a voice message to its first phrase; a file
over 100 MiB is not downloaded and keeps its metadata description; a network
error while downloading is retried like other temporary errors. - The per-message copies in the evaluation report are blanked after 72 hours.
message_ai_scoreskept the author name, text, caption and media description
of every scored message, andai_evaluation_runsthe model's reason text,
indefinitely, although the messages themselves are deleted. Scores, ranks, token
counts and media metadata stay. - Oversized files are refused before they are decoded. An image above 16
megapixels or a video with frames above 2560x1440 keeps its metadata
description instead of being decoded, and a file that makes the process crash
is given up after five tries instead of being claimed again every ten minutes.
ffmpeg runs with a minimal environment, may open files only and refuses
oversized frames itself. A large picture and a video are never processed at
the same time. - A message with no text, caption or file is not stored. A join, a leave, a
pin, a poll, a dice or a location was stored as<type>: сообщение без текста,
counted towards the day's minimum and sent to the model as noise, so a quiet
day with a few joins cost an AI call. The group is still recorded. Rows already
stored go with the 72-hour retention. - A day that cannot be closed waits 15 minutes before the next try instead of
trying and paying for the AI every minute for two hours. A day with no eligible
message is closed without another call. A run that is still in progress is not
started a second time by the catch-up after the agreement is accepted, and
that catch-up only runs once the day's cutoff has passed. - A day verdict that does not parse sends the day to the next model instead of
publishing nothing for that run. - Messages are deleted after 72 hours even when their day was never processed
(a group that did not accept the agreement, an outage, an*_unknownday). - Secrets are scrubbed in more forms. The redactor also removes
percent-encoded passwords and wholepostgres://URLs, the Claude and
OpenRouter clients no longer follow redirects (the key and the day's text would
go to the new address), and the audit log is pruned at start as well as on
write. - Logs are JSON on stdout. The per-module files and the emoji console format
are gone;logs/ai_audit.jsonlstays, with its seven days and mode 0600. - About shows "Scoring: Claude, OpenRouter" instead of the stale
"OpenRouter". - License: Apache-2.0 from this release, like the other Asterfield bots.
v0.11.1and earlier stay GPL-3.0. The user agreement and its version are
unchanged; the license appears in About only.
Fixed
- Every group runs in its own time zone. The zone database is embedded in the
binary, because the image has none; without it the cutoff and the day window
would have been computed in UTC. The bot checks the zones it names itself at
start. - Anonymous reaction totals are counted. The handler read a field the update
does not have, so everymessage_reaction_countupdate failed. - Unblocking the bot in a private chat no longer creates a group. It created a
group_settingsrow for the user id, and the scheduler then posted agreement
reminders into that chat. - The
sendRichMessageprobe survives a network error,429,5xxand409
by staying on HTML, and rich Markdown is switched off only when the method is
really gone, not by a deleted panel or a message that cannot be edited. - A failing
getChatMemberanswers the button with an alert instead of
leaving it spinning. - The agreement reminder is marked done only after it was sent, as the
v0.10.2changelog said. - The welcome message shows the group's real cutoff (it printed 21:00 for a
group whose minute is between 21:00 and 21:10), and the group dashboard shows
the real media-analysis state. - Editing a media message keeps its stored description.
- A person without a name is shown as "Anonymous" in the viewer's language
instead of a hard-coded Russia...
v0.11.1
Claude Opus 5.5 picks the quote of the day and Claude Sonnet 5.5 describes
images, with OpenRouter behind both; the user agreement is 1.2, and every group
accepts it again before Quoto publishes there.
This is v0.11.1-alpha.2 promoted unchanged. Both alphas ran on the production
bot from 18 October: the day was scored by Claude every evening, including
20 October, and the report fix in alpha.2 held. The OpenRouter media fallback
answered 402 Payment Required on 18 October until its balance was topped
up; Claude's own path was not affected. Operators coming from v0.11.0: read
the Operators section of v0.11.1-alpha.1 for the new variables, the
migration and the new dependency.
Container image
ghcr.io/freshlabdev/quoto@sha256:245c9d3ec7e6139d1bbef57067ed267bafb4009c68cde4eb40c0eecad0f32d1f
v0.11.1-alpha.2
One fix found on the first real days of alpha.1. Nothing changes for
operators: no new variables, no migration.
Fixed
- A day recomputed after a failed publish keeps the report of the run that
succeeded.ai_evaluation_runsholds one row per chat and day. When the
first attempt ended asai_failedand a later one picked the quote, the
insert hituq_ai_evaluation_run_group_day, the error was logged as "already
exists", and the table kept the failed row: 17 October's row said
OpenRouter had failed while Claude Opus 5.5 had published the quote, and the
scores, model and cost of the real run were lost. The failed row is now
replaced together with itsmessage_ai_scoresin one transaction. A report
that is alreadyparsedis still left alone.
Container image
ghcr.io/freshlabdev/quoto@sha256:60ebedaaf934dc773059841b0f30d31985674d03cdd09052e4f2c7f4badaf1bd
v0.11.1-alpha.1
Claude now picks the quote and describes images, and the user agreement moves
to 1.2 so that every group agrees to that first.
Quoto has no test bot, so this alpha runs on the production bot. It either
becomes 0.11.1 after a few real days or gets rolled back to 0.11.0.
Changed
- Claude Opus 5.5 scores the day through the Anthropic API, at
medium
effort with adaptive thinking and a structured answer. OpenRouter's
openai/gpt-6-lunaandgoogle/gemini-3.1-flash-litestay behind it: a
refusal, an empty balance, an outage or an answer that does not parse moves
the day to the next model. Replaying 10–17 October, Opus atmediumand
highpicked the same quote on six days of seven, while Sonnet 5.5's pick
changed from run to run; Opus costs about $0.11 a day there, Sonnet 5.5
about $0.055. The API reports tokens, not dollars, so Claude's cost is
computed from list prices ($4 / $20 per million tokens for Opus 5.5) and
stored with the run like OpenRouter's. - A safety refusal no longer costs the day. Every Claude request opts into
Anthropic's server-side fallback (fallbacks: "default"): a day the
classifiers decline is answered in the same call by the model Anthropic
recommends for that refusal category. Replaying 16 October, the cyber
classifier declined a whole day of banter on Opus 5.5 and Sonnet 5.5 at every
effort level; with the fallback, Claude Opus 4.8 picked the quote. The
declined attempt is counted in the stored cost, so such a day costs about
twice as much as an ordinary day of the same size. - Claude Sonnet 5.5 describes photos, images and stickers. Video, video messages,
voice messages and audio stay on OpenRouter, because Claude takes none of
them. A failed Claude call falls back to the OpenRouter media chain, and
OpenRouter's cooldown after a 402 no longer holds images back from Claude. - The user agreement is 1.2, and every group accepts it again. The AI
section now says that text, images and stickers go to Claude, that Anthropic
does not train on them and deletes them within 30 days, keeping longer only
what its safety systems flag, and that video and audio go through OpenRouter
to providers that keep nothing. 1.1 promised only the latter, so a group that
accepted it publishes nothing until an admin accepts 1.2; it gets a daily
notice saying what changed, and its images keep going to OpenRouter only. scripts/compare_eval_models.pyreplays audited days on Claude as well, and
reads records written by either provider.claude-sonnet-5-5@low,claude-sonnet-5-5@high
compares effort levels in one run. Claude replays send the same fallback as
the bot;--no-fallbacksshows the bare model's refusals.
Operators
- New variables:
ANTHROPIC_API_KEY;ANTHROPIC_EVAL_MODEL
(claude-opus-5-5) andANTHROPIC_MEDIA_MODEL(claude-sonnet-5-5);
ANTHROPIC_EVAL_EFFORTandANTHROPIC_MEDIA_EFFORT, bothmedium;
ANTHROPIC_EVAL_MAX_TOKENS(32000) andANTHROPIC_MEDIA_MAX_TOKENS(8000).
WithoutANTHROPIC_API_KEYthe bot sends everything to OpenRouter as 0.11
did.OPENROUTER_API_KEYis no longer required on its own: one of the two
keys is. - Migration
20261018_01addsquoto.group_settings.agreement_version.
Every group that accepted before has NULL there, which the bot reads as a
version older than 1.2. - New dependency:
anthropic1.15.0, which bringshttpx2,jiter,
docstring-parser,opentelemetry-api,sniffioandtruststore.
pip-auditreports nothing.
Container image
ghcr.io/freshlabdev/quoto@sha256:5b731410a345b7c6cf9470a7fd182c051b08173f19db0ef1015a85327d3cdc06
v0.11.0
A group's messages now reach only model endpoints that keep nothing and train
on nothing.
Until now production scored the day with a discounted Meta tier whose prompts
and outputs may be used to improve Meta's products, and the built-in defaults
were free endpoints. Every request now asks OpenRouter for zero-retention
endpoints only, the default models are ones that have them, and the user
agreement says so. The dependency lock moves too: 87 known advisories in ten
packages, most of them in aiohttp and pillow, are down to none.
There was no pre-release. Every day this waited on an alpha was another day of
chat text going to a provider allowed to keep it. The request changes by one
field, covered by tests, and each new model was checked against OpenRouter's
zero-retention list before the tag.
Security
- Chat text only reaches endpoints that keep nothing and train on nothing.
Quoto sends a whole day of a group's messages, with author names, to a model.
The built-in defaults were free endpoints, and production scored the day with
meta/muse-spark-1.3-contributor, a discounted tier whose prompts and outputs
may be used to improve Meta's products. Every OpenRouter request now carries
provider: {"data_collection": "deny", "zdr": true}, so OpenRouter routes it
only to endpoints with a zero-data-retention policy. A model without such an
endpoint fails over to the fallback model instead of reaching one that keeps
the data. The bench scripts, which replay real audited days, send the same. - aiogram 3.25.0 → 3.31.0, which lets aiohttp go from 3.13.3 to 3.14.5, and
pillow 12.2.0 → 12.3.0, with anyio, click, idna, mako, multidict,
pydantic-settings, pygments, pytest and python-dotenv moved to fixed
releases.pip-auditwent from 87 known advisories in ten packages to none.
aiohttp carries all of Quoto's Telegram traffic, and Pillow decodes the
images people send to the chat.
Changed
- New default models, each with zero-retention endpoints on OpenRouter.
Evaluation:openai/gpt-6-luna, falling back to
google/gemini-3.1-flash-lite. Media:google/gemini-2.5-flash-lite,
falling back togoogle/gemini-3.1-flash-lite. At last month's volume the
evaluation costs about $0.003 a day per group, the same as before. - Operators: two new variables, one env change.
OPENROUTER_DATA_COLLECTION
(defaultdeny) andOPENROUTER_ZDR(defaulttrue); the bot warns at
startup when either is relaxed. A deployment that pins
OPENROUTER_EVAL_MODELor the media models to a model without a
zero-retention endpoint keeps working through the fallback, but every
request first fails against that model: remove the override or point it at
a model with a zero-retention endpoint. - The user agreement says that OpenRouter passes messages only to providers
that neither store them nor train models on them, in all sixteen languages.
Its version is 1.1; groups that accepted 1.0 are not asked again. - The language prompt listed four interface codes while the response schema
allowed sixteen. It now lists all sixteen. - README and README.ru are rewritten in the family's shape: what Quoto does,
how to use it, a privacy section, and the sixteen languages and GPL-3.0
licence it actually has. AGENTS.md said Apache-2.0 and four languages; fixed. - A release waits for CI.
release.ymlrunsci.ymlon the tagged commit
and publishes nothing unless it passed. When CI cannot run,release.ymlcan
be started by hand with the tag and a reason; that run skips CI and prints the
reason at the top of the release notes. Seedocs/releases.md. - The release notes end with the image digest to pin, which
docs/releases.md
already said they did. - CI audits
requirements.txtagainst known advisories. It reports on every run
and fails only the weekly one, so an advisory does not hold a release of
unrelated code. When it was added it found 87 advisories in ten packages,
most of them in aiohttp and pillow; the Security entries above clear them. - Dependabot opens grouped version updates against
dev.
Container image
ghcr.io/freshlabdev/quoto@sha256:5b492d98e70d504a87e36c22a87179d4e4d6d4ef088681b866e3c733dc8831a4
v0.10.7
Quoto speaks sixteen languages instead of four, and each of them writes a date
the way that language writes one.
Twelve new locales landed, and they broke a date that four languages had hidden:
Spanish needed a preposition that was baked into the month name, Chinese and
Japanese write the year first, German and Czech want a period after the day.
The order now lives in the locale beside the month names.
Every screen is also built by one helper, so the group panel stops printing the
bot's name above its own title, and no screen repeats in the message what its
buttons already say.
Changed
docs/releases.mdnames the real cause of the403a digest pull used to
answer: the account running the deploy could not read the private package,
not anything about the digest. Pulling the tag first was a workaround for
that, and is no longer required.
v0.10.7-alpha.1
Changed
- The panel now follows the Asterfield family contract, so a person who learns
one of the bots reads all of them. Every screen — both homes, language,
quote day, timezone, publishing, stats, About and the user agreement — is
built by one helper as title, hint, quote, and no screen repeats the bot's
name above its own title any more. - The language picker shows flag and native name in the family order
(en, ru, uk, de), two to a row, in the panel and in the agreement alike.
The list and its labels live inapp/i18n.py; quoto renders the languages it
has, so a new locale is a JSON file plus a code and its button appears in the
right place by itself. - Buttons carry colour where Telegram can show it: Close is destructive, the
current language, the chosen timezone and the open stats tab are marked as
the state you are in, and each screen with one obvious next step has exactly
one highlighted button — "Add to group" in a direct chat, "Stats" in a group,
"Accept" on the agreement. Screens that lead nowhere have none. - The timezone, language and publishing screens no longer list in the message
what their own buttons already say. The state lives on the buttons; the text
says only what a button cannot — where a language came from, and what
"context" means.
Fixed
- The quote-day time nudges said "ч" and "м" on every locale's panel, Russian
units on the English and German ones included. They now read−1:00and
+0:15, which is the same in every language and needs no translation.
v0.10.6
Quoto is GPL-3.0. The About card and the agreement signature said Apache-2.0.
Fixed
- The licence named in the About card and under the user agreement is GPL-3.0,
which is whatLICENSEhas always said. Apache-2.0 was written in by
assumption from the rest of the family, where it is correct — and it went
into two places a reader is meant to be able to rely on, one of them a
document people accept.
Use this section for changes that are merged but not released yet.