Skip to content

Releases

v0.1.0-alpha.10

v0.1.0-alpha.10 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 17 Nov 14:03

The GitHub organization FreshLabDev is now asterfield-dev, and the bot moves
with it. Nothing a person sees in Telegram changes.

Changed

  • Images are published as ghcr.io/asterfield-dev/framy. The organization
    was renamed, and GitHub Container Registry does not redirect the old name:
    ghcr.io/freshlabdev/framy answers 403. Point the deployment's image at the
    new path; every earlier digest is there under it too.
  • tg moves from v0.2.0-alpha.3.0.20261031203103-b5a82887d903 to
    v0.4.0-alpha.2, whose module path is github.com/asterfield-dev/tg.
    SendMediaGroup takes tg.MediaGroupOptions instead of a thread id; albums
    are sent as before.
  • Links, the module path and the docs point at github.com/asterfield-dev.
    Old links still redirect.

Container image

ghcr.io/asterfield-dev/framy@sha256:913c63180908b767c93320d07d9baf8c8a4b0ca54b543e3fc3a28aae9c9b81d7

v0.1.0-alpha.9

v0.1.0-alpha.9 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 16 Nov 14:39

A fix for several frames on the free Recraft model, and the older Recraft
Flash marked outdated. No migrations and no changes on the host.

Fixed

  • Several frames on Recraft V4.1 Flash V2 paid with your own key were shown
    at $0 each, but on Create framy joined them into one run of n images,
    which the free variant does not take. The run went to the paid route, and
    the site stopped it with "the price is now $0.028 instead of $0" before
    anything was charged. Each frame on a free variant is now a run of its
    own at $0, in the quote and on Create alike.

Changed

  • Recraft V4.1 Flash is outdated by V4.1 Flash V2, its update, and the
    picker shows it among the older variants. Recraft V4.1 stays the model the
    family opens on.

Container image

ghcr.io/freshlabdev/framy@sha256:41fdd0ee33c11f225313f22c5116efea81b79b97c261fed67bf6872c51edd82c

v0.1.0-alpha.8

v0.1.0-alpha.8 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 16 Nov 13:18

Images from words are free and do not count against the 5 a day. New models
from OpenRouter: Recraft V4.1 Flash V2, Seedream 5.5 and MiniMax H3 Image.
No migrations and no changes on the host.

Added

  • An image from words alone goes to recraft/recraft-v4.1-flash-v2:free,
    which OpenRouter serves at $0. It takes no run of the free allowance and no
    cent of the day's budget, and there is no cap per person. Telegram uses it
    for /img and plain prompts without a photo; the site opens on it while no
    photo is attached and moves to MAI 2.6 Flash when one is, unless you picked
    a model yourself. The aspect ratio is yours to choose. A photo, two results
    or a video go to the paid models as before.
  • On your own OpenRouter key the same run goes to the same free variant at $0,
    so free images from words keep working when framy's free requests are used
    up.
  • Seedream 5.5 from ByteDance, with a 2K variant.
  • MiniMax H3 Image as one model with a Fast switch for H3 Image Highspeed, as
    Wan 3.0 has. Its price before the run is worked out from MiniMax's own
    token examples.

Changed

  • The price line, the preview card in Telegram and the site say
    "free · not counted" for these runs.
  • A new fuse, free_requests, closes only the free variant when OpenRouter's
    daily requests for free models run out (50 a day below $10 of credits
    bought, 1,000 above), until 00:00 UTC. Images from words then go to MAI and
    count as before. The operator chat is not told about it, and the
    free_closed refusal now carries the fuse that closed it.

Known Limitations

  • Tencent HY World 2.1 Panorama stays hidden: OpenRouter gives no price for
    it before the run, and every model framy shows has one.

Container image

ghcr.io/freshlabdev/framy@sha256:582255335f13dec9bb631ea66cd9ebe4203ec0eeb25504e63b35e9f5306ce673

v0.1.0-alpha.7

v0.1.0-alpha.7 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 14 Nov 20:46

A Go security update and a quieter catalog. No migrations and no changes on
the host.

Security

  • Go moves from 1.27.3 to 1.27.4: a crafted HTTP/1 request could crash the
    server through a race on the MaxBytesReader header (GO-2026-6859).

Removed

  • Seedance 1.5 Pro is gone from the overlay. OpenRouter stopped serving it on
    11 November and the catalog disabled it that day, but the leftover overlay
    entry made every hourly sync log catalog overlay mismatch.

Container image

ghcr.io/freshlabdev/framy@sha256:575ef6d50fecbefc72140534d077bc5003d76d138ca90883c1b533dc45da1174

v0.1.0-alpha.6

v0.1.0-alpha.6 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 06 Nov 20:47

Published without CI. The owner asked not to wait for CI on this alpha. Locally green before the tag: go vet under every build tag, go test ./..., and the integration suite against Postgres 17 (on 93ee1ae and on the GIF commit); dev CI on 93ee1ae, which has everything but GIF, passed in full.

Animated GIFs from GPT Image 2.6 on the site, the Mini App and Telegram,
Gemini Omni 1.2 Flash in place of 1.1, GPT Image 2.6 Flare as the suggested
GPT model, and fixes to the library and to video lengths. No migrations and
no changes on the host.

Added

  • Gemini Omni 1.2 Flash, out on 5 November, on a Gemini key and through
    OpenRouter: videos of 3 to 15 s at a length you choose, so the price is
    exact rather than "up to", 30% below 1.1, and a clip to edit or extend of
    up to 15 s. It takes 1.1's place in the picker; 1.1 is marked as closing
    on 3 December, when Google moves its requests to 1.2.
  • Animated GIFs from GPT Image 2.6 on an OpenAI key. GIF under Format
    turns the form into frames (2 to 16, 9 by default) and frame length
    (50 ms to 1 s, 200 ms by default) and shows how long the loop runs,
    with a strip that lights a cell per frame at its pace. With a photo
    attached, one tap animates it. OpenAI publishes no price for a GIF, so the estimate counts
    one still of the same quality and size per frame and says it is an
    estimate; the cost is the usage OpenAI reports. The key's account needs
    OpenAI's GIF preview. OpenRouter takes no frame settings, so there is no
    GIF through it.
  • A GIF plays on the stage, in the library and on the whole screen, and a
    button stops it on its first frame. In Telegram it comes as an animation
    in a direct chat, a group, a guest answer and inline mode, several of
    them one by one, and /img asks for one under Settings.

Changed

  • GPT Image 2.6 Flare is suggested for an image and for many photos in
    place of 2.5 Flare, and the GPT Image row opens on 2.6. 2.5 stays in the
    row: at low and medium quality it costs a quarter to a third of 2.6.

Fixed

  • The library wrote "null" under its heading once the failed filter was
    turned off, and the settings wrote it in place of the stored files until
    they were known.
  • Choosing another kind, the failed filter or a search no longer shows a
    loading line that jumps: the list keeps what is read of it while the
    page comes, and an empty one says it is loading only when the wait is
    long enough to see.
  • A model that makes fewer lengths from frames, on a clip or when it goes
    on from its own video offers only those then, and a run that asks for
    another is refused before it starts: Omni 1.2 makes 3 to 10 s in those
    cases, where words go to 15 s. Extending a clip by a set length is priced
    up to the whole clip, which Google may bill.
  • The estimate of a video priced in tokens by resolution found no count
    for 4K, written in capitals in the catalog.

Container image

ghcr.io/freshlabdev/framy@sha256:b77013db5fe29dcffa0c02c9ea7cb8f165506a2b3c4511cb422ed80ef7acd3e0

v0.1.0-alpha.5

v0.1.0-alpha.5 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 06 Nov 16:30

The review of 5 November, worked through: about 490 findings across the
bot, the site, the Mini App, jobs, prices, storage and deployment. Six
migrations, and changes on the host, below, before this image runs on
either stack.

Before you deploy

  • Each stack reads its database URL from secrets/database-url (one line,
    framy_core's URL with its password, uid 101, mode 0400) instead of
    building it from FRAMY_CORE_PASSWORD: a variable shows in
    docker inspect, a file does not.
  • TRUSTED_PROXIES is cloudflared's own address on the management
    network, no longer the network's range: any other container in that range
    could write CF-Connecting-IP.
  • The KEK file must be readable by its owner only (chmod 400); one that
    group or others can read stops the start and names the file. Keep it out
    of the host's backup of the stack directory, since the database dumps hold
    the keys it opens.
  • The container's own files are read-only, /metrics listens on the
    container's loopback only, and the stacks get four CPUs. The Bot API
    directory comes in as the volume bot-api, so docker inspect no longer
    shows the bot's token in a mount; after a new token, remove
    framy_bot-api with the stack down.
  • The test bot goes through the local Bot API, as production does.
  • Optional: R2_INCOMING_BUCKET with a token of its own puts the site's
    clips in a bucket that token alone can write.

Migrations

The test bot applies them first, and production on alpha.4 keeps running on
them.

  • 031_site_press_end.sql: an index for the DM's sender of site presses.
  • 032_session_stack.sql: a session works only on the stack that made it.
    Every session before it counts as production's, so people signed in on the
    test bot sign in there once more. A release before it reads sessions of
    any stack.
  • 033_job_notice_and_index.sql: a job's notice names the job and nobody,
    since any role on core-postgres can listen; indexes for chat runs that
    point at a file and for the runs of the last month.
  • 034_deleted_provider_message.sql: clears what providers said about runs
    deleted before the code did.
  • 035_job_forget.sql: the jobs of a deleted account stop saying whose they
    were once the free allowance no longer counts them.
  • 036_oauth_ticket.sql: one-time tickets to connect OpenRouter from the
    Mini App.

Added

  • Seed and Leave out (a negative prompt) in the settings, where the model
    takes them: Veo, Kling, Wan and the image models with a seed.
  • Connect OpenRouter from the Mini App without signing in again in the
    browser: the Mini App hands the browser a five-minute ticket.
  • Sign out everywhere, in the settings.
  • The library can show only the runs that did not work, and delete them
    all.
  • A picture made goes on as a photo to change or as a video's first frame.
  • Google's search suggestions and sources show with a grounded result.
  • A visitor picks the language before signing in.
  • The operator chat hears a disk past 80% and jobs nothing takes, and
    /healthz turns unhealthy when calls to Telegram keep failing.
  • @framyin_bot a fox in a group framy is in makes a fox; framy greets a
    group it joins.

Changed

  • The bot's button under a preview is Create, as on the site, in every
    language; it was Generate.
  • A sign-in lasts 30 days after the last visit, not from the sign-in, and
    signing in again ends the browser's earlier session.
  • The Mini App uses Telegram's own feedback: haptics when a run ends, a
    closing confirmation while a run is being set up, the main button's
    progress.
  • Prices: FLUX.2 Pro and Max count the references' megapixels, Nano Banana
    counts its inputs, FLUX Video Upscale is priced by creativity and frame
    rate, Seedream 5.0 Pro at 1K costs its base rate, and an "up to" run holds
    the free budget for its top.
  • Money and lists follow each language's way of writing them.
  • The texts in all 16 languages were reread: one word per thing in each
    language, a run is a generation where the language says so, and French
    has its non-breaking spaces.

Fixed

  • A paid video is no longer lost when its key, River's last attempt, a slow
    copy to R2 or a moment without the database interrupts the check on it;
    pictures kept before a lost ending make the run done, with their cost.
  • A second press of Create is a repeat, not a used-up allowance, and a run
    on a key starts at the price the bar shows.
  • A result Telegram refuses ends its run instead of being sent again
    forever, and one that went out is not sent twice when its record failed.
  • A busy group is paced, and a 429 holds that chat until its wait.
  • The site works from the keyboard and a screen reader, in forced colours,
    with less motion and right to left in Arabic, and every control on a touch
    screen is a finger wide.

Security

  • A key pasted into a prompt or a text setting goes nowhere: the run is
    refused before anything keeps it.
  • Sessions: the login cookie is signed, a session ended anywhere is refused
    everywhere at once, and a Mini App token ends with the launch it came
    from.
  • Storage: no Telegram id in file names, a link to put a clip holds that
    clip and no more, and deleting a run or an account clears what the
    provider said about it and the DM message it went in.
  • The release publishes the image CI tested, built from pinned base images,
    and scans its packages.

Container image

ghcr.io/freshlabdev/framy@sha256:1435b4929c4c062bbe992ba9b83f5aadc9b74bf2f844c505956cfbe6b06f2d80

v0.1.0-alpha.4

v0.1.0-alpha.4 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 02 Nov 15:46

Fixes for what the first two days of alpha.3 showed in production's logs:
a refusal shown as a failure, a group /start read after downtime, and a
closed model in the catalog's overlay. No migrations.

Fixed

  • A picture Azure's moderation blocks for MAI Image is a refusal: the
    person reads that the model refused the request, not "try again", which
    would fail the same way. Like any refusal, a free one counts toward the
    daily free pictures.
  • A /start sent in a group while framy was down is no longer an error when
    framy reads it: Telegram takes an ephemeral answer for 15 seconds only, so
    framy leaves it and the person sends it again.
  • Grok Imagine Image Quality is out of the catalog's overlay. xAI shut it
    down on 2 November, and every catalog sync since warned that the overlay
    named a model no source has.

Container image

ghcr.io/freshlabdev/framy@sha256:2a1a23a1db4312322d18c7b95b9230cd3b841f92ed18eaeb6788a199dd88a850

v0.1.0-alpha.3

v0.1.0-alpha.3 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 31 Oct 21:45

Several results from one press, end to end: the site and the Mini App make
frames, each with its own model and settings, the library keeps a press as
one card, and the DM gets it as one album. The site carries framy's mark. No
migrations: production and the test bot take the same image in any order.

Known Limitations

  • The framy schema still has no backups (#28).
  • Switching the library between All, Images and Videos makes the page jump
    (#51).

Added

  • A press of several frames comes to the DM as one album, in the order of
    its frames, once its last job ended. The caption gives the prompt once and
    numbers the frames by model, so three models on one prompt read as a
    comparison. Videos keep their covers in the album.
  • Frames on the site and in the Mini App. Frame, over the stage on a
    computer and over the strip of frames on a phone, adds a result with its
    own model or settings, up to 6 pictures or 4 videos. All sets what every
    frame shares; a frame picked sets only itself and marks what is its own.
    On a phone a swipe in the settings goes to the next frame and the strip
    rolls with it. Create sends the frames as one press, at the price framy
    sums for the paid ones. The results stand in place of their frames, with
    their number and, when the models differ, the model, each with its own
    actions; a frame that failed leaves and the others go on.
  • The library shows a press as one card with its pictures together. Its
    viewer goes between the frames, and Again and Edit the prompt bring every
    frame back, a job of several results as that many frames.
  • The site carries framy's mark: in the browser tab, on an iPhone's home
    screen, and before the name in the header.

Changed

  • On a phone, files are added in the row under the prompt, for pictures as
    for video, and the plus in the row over the prompt is gone: in video it
    stood beside the slots' own pluses. The settings there are an icon only,
    with the number of frames beside it when there are several.

Fixed

  • On a phone 320 px wide the sign-in button no longer breaks onto two lines
    past its edges: the header shows the mark without the name there, and a
    sign-in line too long for the button ends in an ellipsis.

Container image

ghcr.io/freshlabdev/framy@sha256:905e531e2029e471e47cbefc5a5127cd18f46433e7ba50792dbe2f97a8b82d44

v0.1.0-alpha.2

v0.1.0-alpha.2 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 31 Oct 20:37

The test bot moves onto production's database, so a person has one history
on both sites, and each bot still works only its own jobs and chats. The
backend for several results from one press comes with it. Deploy production
first; the test bot's stack follows (Breaking).

Known Limitations

  • The framy schema still has no backups (#28).
  • The site does not send several results from one press yet: the API is
    there, the composer that uses it comes next.

Changed

  • The test bot and production share one database: the framy schema in
    core-postgres. A person sees the same history, settings and connected keys
    on framy.ws04.cc and framy-test.ws04.cc, and the free tier's daily budget
    is counted once over both bots instead of once on each.
  • Each bot works only its own jobs and tells only its own chats. A job runs
    on the bot whose chat or site started it, and a result made on a site goes
    to the DM from that site's bot. Inline mode, replies to a result, albums
    and Send to Telegram use only what the bot itself sent or saw, since a
    Telegram file id works only for the bot that got it.
  • The catalog sync, the currency rates, the sweep of uploads no job took,
    the thumbnails and the operator chat run on production only. The test bot
    shows production's catalog, so a catalog change reaches it with the
    release that takes it to production.

Added

  • FRAMY_STACK names the bot when two share one schema: empty for
    production, test for the test bot. A stack other than production puts its
    name before its River queues and its Telegram dedupe keys.
  • Several results from one press, each with its own model and settings
    (POST /api/batches). framy decides how many requests that is: frames
    that are the same run go together, as many as the model makes at once,
    so five pictures where a model makes four at once are two requests, four
    and one. A press is queued whole or not at all, and the history keeps it
    as one run. A quote with batch: true says which frames go together and
    what the press costs.

Breaking

  • Migration 029_stack.sql marks jobs, chat runs and album items with their
    bot and changes the album cache's key. v0.1.0-alpha.1 cannot write the
    album cache after it, so a rollback past this release loses albums sent
    until the next deploy.
  • Deploy production first. The test bot's stack then moves onto
    core-postgres: FRAMY_STACK=test, DATABASE_URL as framy_core, the
    network core_net, production's KEK in secrets/kek, no R2_PREFIX
    (deploy/ws04-test/compose.yaml). Its own PostgreSQL goes; its data is not
    moved, and the volume stays.
  • The test bot runs dev and applies its migrations to production's schema
    before production does, so every migration from now on is additive
    (AGENTS.md, Migrations).

Fixed

  • Switching an empty composer from Image to Video no longer flashes the
    hint of the row of files at the left edge of the page, one word to a
    line.

Container image

ghcr.io/freshlabdev/framy@sha256:8059abf692efd02e00764fd372dbf60b9bee30a72df9eee0632f6a917911b6bc

v0.1.0-alpha.1

v0.1.0-alpha.1 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 31 Oct 19:44

The first alpha of framy: images and videos from AI models on the site
https://framy.ws04.cc, in the Telegram Mini App and in chats with
@framyin_bot. A few images a day and a video a week are free; everything
else runs on the person's own OpenAI, Gemini or OpenRouter key. Everything
below is written and tested, and most of it has not met real traffic yet.

Known Limitations

  • The framy schema has no backups yet: the daily encrypted dump into the
    framy-backups bucket is not set up (#28). R2 keeps its own copy of the
    files.
  • /img in reply to a photo sent before framy joined the group gets no
    photo: the Bot API gives a bot no message it did not see.
  • Transparency in HEIF and AVIF photos is lost when they are normalized.

Added

  • The Go service skeleton. Configuration comes from the environment, and any
    secret can instead come from a file named by NAME_FILE, as Docker secrets
    do. PostgreSQL is used in the framy schema, with migrations embedded in
    the binary. Long polling records every sender, and every group framy is in,
    in core through core.touch. Preflight runs before polling starts.
    /healthz reports the version, the database and how fresh polling is.
    SIGTERM stops polling at once, but the updates already taken get up to 15
    seconds to finish, so a /img that comes in as framy stops still gets its
    preview instead of a draft cancelled halfway. The HTTP server stops within
    20 seconds.
  • /start and the family's panels, in the sixteen languages of the family
    and the shape every sibling bot uses (core/docs/panels.md): home, help,
    about and the language picker in a direct chat; in a group /start is
    ephemeral, so the panel reaches only the one who asked, with the group's
    language for its admins. The language is core's, shared with the other bots:
    a pick writes a manual observation and Follow Telegram withdraws it. The
    command menu and the profile descriptions are published per language at
    start, and a description is written only when it changed.
  • The home panel of a direct chat says where the person stands: how many of
    the free images (5 in any 24 hours) and the free video (1 in any 7 days)
    are left, which providers they have connected, their language, and the
    currency prices are shown in. A provider is named with a word beside it
    only when something is wrong with the key (invalid, needs billing, needs
    verification, no credits); no part of a key is ever shown. The currency is
    the person's choice, else the language's (Ukrainian, Belarusian and Polish
    map to UAH, BYN and PLN), else USD, since the bot has no country to go by.
    Each part is read when the panel is drawn, all under a 3-second deadline;
    one that cannot be read is left out and /start still answers. A group's
    panel shows none of this. With an https SITE_URL the panel gets an Open
    framy button that opens the site as a Mini App, and at start the menu
    button of every direct chat is set to open it, labelled framy. Without
    SITE_URL, or with one that is not https (Telegram refuses it), neither is
    set and the menu button is left as it was. The default models are not
    shown yet, since nothing stores them. The text that described the free
    allowance in words is gone from the panel, replaced by the numbers.
  • /metrics with framy's counters in the Prometheus text format, on its own
    listener (METRICS_ADDR, :9090 by default), apart from the public one.
  • The model catalog. Every hour framy reads OpenRouter's public catalogs: the
    image models with the price of every upstream route, the video models, and
    their expiration dates. OpenAI's and Gemini's models are written by hand in
    catalog/openai.yaml and catalog/gemini.yaml, and catalog/overlay.yaml
    joins the offers of one model, so Nano Banana 2.1 comes from Gemini and from
    OpenRouter. The overlay also gives display names, the featured and free
    flags, the effort setting, values not to offer such as OpenAI's auto,
    pass-through settings, reference counts, what a model needs from an
    attached video, continuation and the face policy. The video editors and the
    upscaler (Aleph 2, FLUX Video Edit, FLUX Video Upscale) are in the video
    list and run only on an attached video; FLUX Video Edit takes up to 15 s and
    50 MiB. Hidden: models within 14 days of leaving, Google's preview ids,
    wrappers, closed OpenAI models and the talking-head model, which needs a
    voice the composer does not take yet. A model gone
    upstream gets disabled_at and comes back when it returns. A failed sync
    changes nothing, and one that would disable more than half of OpenRouter's
    offers is refused. CATALOG_SYNC_INTERVAL (1h, or off) and
    OPENROUTER_API_BASE configure it.
  • The price estimate. Before a run framy works out its price from the
    catalog's rates and the run's settings: OpenRouter's price lines and video
    SKUs, OpenAI's token formula, Gemini's tokens per size, Seedance's video
    tokens, per-reference charges and Aleph's minimum. Each estimate says
    whether it is exact, about, up to, from, or known only after the run, and
    keeps the rates and quantities it used. It agrees with OpenAI's published
    gpt-image-2 prices and OpenRouter's 16:9 example. The overlay's price hints fill in what the rates leave out, and
    every sync logs the price lines the estimate cannot read. MAI counts an
    input image by its size, about 1,100 tokens a megapixel as measured, so an
    edit's estimate includes its photos. Of two video SKUs that fit a run
    equally, one whose rate a broader SKU already has is left out, so Seedance
    with a clip and no sound is no longer dearer than with sound, and Kling
    with sound is exact. A model that takes clips as references bills every
    clip's seconds beside the output's, so the estimate counts them all,
    each rounded up to a hundredth of a second: Wan 3.0 and Hailuo 3 at the
    output's rate, HeyGen at its reference rate, Seedance as video tokens at
    the output's size. Two runs on the test bot showed it: Wan 3.0 with a
    15 s clip cost $0.85 against an estimate of $0.21, HeyGen $0.70 against
    $0.05; both estimates now come within a hundredth of a cent. GPT Image 2
    and 2.5 count an input image in 32 px patches, at most 1,536, as worked
    out on OpenAI's forum, so an edit's estimate includes its photos: a
    1280×853 photo took 1,080 tokens on the test bot, which the estimate had
    left out. Seedance is priced as it bills: at the size it makes (864×496
    at 480p 16:9 on Seedance 2.0, where OpenRouter lists 854×480), 24 frames
    a second and one more, and a reference clip as at least 3 s. Two paid runs
    of Seedance 2.0 Mini, 4 s of 480p with a 2 s and with a 3 s clip, cost
    $0.1485 each where the estimate said $0.121 and $0.141; it now matches
    both to the token. A size the overlay does not list, such as 9:21 or 4K,
    is "about".
  • OpenRouter's discounts come off the price. The sync reads them from
    /api/v1/models/{id}/endpoints, the only place that carries them, and the
    estimate takes them off the total and keeps the listed price beside it, so
    the composer can show it struck through. On 28 October that is 15% off Wan
    3.0 and 50% off HeyGen Video 1, the free weekly video: framy showed $0.50
    for 5 s of Wan 3.0 at 720p, where OpenRouter charges $0.425. A discount
    that could not be read counts as none, so a bad hour only makes a price
    dearer.
  • The job queue, on River in the framy schema. A job is a row in
    framy.job and a River job, created in one transaction, and it moves
    through draft, queued, running and done, failed, cancelled or expired. The
    rules about money: the mark that a paid request is out is written before the
    request goes, so a run cut short by a crash is never sent again; a
    provider's job id is saved the moment it comes back, so a retry asks
    instead of paying; only a 429 or a 5xx before the run started is retried,
    after the provider's wait. A repeated Telegram update starts nothing, one
    person runs at most two jobs at once and the rest wait in order, and a job
    keeps its estimate next to what the provider charged. A stop waits for runs
    in progress up to JOBS_STOP_GRACE (3m30s).
  • The OpenRouter adapter for images, on POST /api/v1/images. A job sends
    its settings, its pass-through settings under the routes' own options and
    its reference images, and goes only to the routes that take them, the
    cheapest first. What no route takes is refused before the call, so it costs
    nothing. Such a run, in every adapter, fails as invalid_request with
    framy's own words, so the site does not show them as the provider's. usage.cost becomes the job's actual cost. A moderated prompt ends
    as moderation with its reasons, apart from money, key, refused, timeout
    and provider failures, and an error inside a 200 is read too. OpenAI's own
    filter, which OpenRouter passes on as a bare 400, is told apart by its
    words. Only a 429, a
    provider that is down or overloaded, and a connection never made are
    retried. framy's own key for the free allowance is OPENROUTER_API_KEY
    (or OPENROUTER_API_KEY_FILE).
  • The OpenAI adapter for images, for a person's own key. A run is one call:
    POST /v1/images/generations, or /v1/images/edits when it has reference
    images, which go as the presigned R2 link OpenAI fetches. quality is
    always one of the model's levels, low to max, and never auto; size
    is a concrete WxH, any size up to 3840 px that OpenAI's rules allow. A run
    framy can tell is wrong is refused before the call, so it costs nothing.
    safety_identifier is a hash of the person's id, never the id. A
    moderation_blocked answer ends the run as moderation with its stage and
    categories, apart from money, key, refused, timeout and provider failures;
    only a 429 for the rate, a 503 and a connection never made are retried. The
    actual cost is the usage tokens times the catalog's rates for text, im...
Read more