94 releases 94 tags

RSS feed

• v5.4.0 722e49b7c7

  Compare

  v5.4.0 Stable

  joeyr released this 2026-10-11 16:54:10 -05:00 | 0 commits to main since this release

  The debug:True footer names the model and settings that answered, and
  Make Public posts it with the answer.

  Added

  • The footer ends with the model, effort and thinking setting:
    -# Model: claude-sonnet-5-5 · effort high · thinking on. The model is
    the one the API says answered. Effort and thinking are what the request
    sent. Thinking is adaptive, so "on" means the model was free to think, not
    that it did. Sonnet 5.5 returns its thinking blocks empty and also uses
    them for progress notes between tool calls, so a block in the response
    does not show that it reasoned.

  Changed

  • Make Public posts the debug footer under the answer. The public post
    now matches what the asker saw. Before, the footer was dropped on publish.
    The footer is still kept out of the saved conversation turn and out of the
    shared context that later questions read, so turning debug on changes
    nothing about the next answer.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v5.3.7 5ed1ce4b3a

  Compare

  v5.3.7 Stable

  joeyr released this 2026-10-10 13:04:51 -05:00 | 1 commits to main since this release

  Two small fixes found while fixing 5.3.6, and the dependencies brought up to
  date.

  Fixed

  • A fetched page that shows an HTML entity reached Claude with it decoded
    twice. web_fetch decoded & first and then the other entities, so
    the & the first step produced started a second one: a page writing about
    HTML that showed <p> reached Claude as the tag <p>. Each entity is
    now decoded once.
  • The migration rollback logged the migrations it undid as fields named
    0, 1, …. They now go in one migrations field, as they already did
    when migrating.

  Changed

  • @anthropic-ai/sdk 0.131 → 0.133, @modelcontextprotocol/sdk 1.32.1,
    dotenv 18.0.7, and on the dev side Prettier 3.9.10, typescript-eslint
    8.71.1, @types/node 24.19.2 and the js-yaml override 5.4.3. None of the
    SDK changes touch what the bot uses. TypeScript stays on 6.0: TypeScript 7
    is out, but the newest typescript-eslint (8.71.1) still accepts only
    <6.1.0 and the newest ts-jest (29.4.14) only <7.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v5.3.6 770c72a5b5

  Compare

  v5.3.6 Stable

  joeyr released this 2026-10-10 12:48:38 -05:00 | 2 commits to main since this release

  The last 22 findings from the whole-repo review, each minor or rare. LoL
  results survive a restart and a watchlist change, registering checks the
  region and Riot ID first, failed buttons say so, a broken mappings file no
  longer wipes names, and the logs keep their details and never show the
  database password. A setting that did nothing, D2EMU_POLL_INTERVAL_MS, is
  gone.

  Fixed

  • A LoL result still pending when the bot restarted was never posted.
    Riot can take a few minutes to publish a finished match, and the bot keeps
    checking for it in the background. A restart in that window, such as a
    deploy, stopped the checks and nothing started them again, so the post said
    "results unavailable" for good. After a restart the bot now picks up any
    result from the last half hour that never arrived.
  • A restart could drop a LoL game's result over one failed Riot call. On
    startup the bot checks whether each game it was following is still on. One
    timeout or busy answer from Riot at that moment wrote the game off, so if it
    had ended, its result never appeared. The game is now checked again on the
    next poll, as it already was outside a restart.
  • Removing a player from the watchlist mid-game could cost that game its
    result. A game is followed through the first watched player found in it.
    If that player was removed while it was on, nobody noticed it end, and it
    was written off hours later with no result. Another watched player in the
    game now takes it over, and the result shown is theirs.
  • A LoL game that came back from a false end could lose its ranks and
    result after the notification channel changed. Riot sometimes briefly
    reports a live game as over, and when it reappears the bot edits its
    original post back. If monitor-channel had moved notifications meanwhile,
    the bot then looked for that post in the new channel, so the ranks and the
    result never arrived. It now follows the post in the channel it is in.
  • The cleanup of stuck LoL games always reported none. Its count was
    always 0, so its log line never appeared, and the games it closed had no
    end time. It now reports how many it closed and records when.
  • A LoL player registered under a region they don't play on was accepted,
    and their games were never posted. Riot finds an account whatever region
    is typed, so the wrong region was stored and their games were looked for on
    the wrong server. Register and monitor-add now ask Riot where the player
    plays and register them there, and say so when it differs from what was
    typed. If Riot can't be asked, the typed region is used as before.
  • A mistyped LoL region or a name without its #TAG got "try again
    later". Neither was checked until Riot was asked, so input that could
    never work failed as if Riot were down, and was logged as an error.
    Register, unregister, monitor-add and monitor-remove now reply at once with
    the valid regions or the Name#TAG format.
  • A connection to Riot that dropped mid-answer was logged as a bug in the
    bot. A cut-off response surfaced as a bare TypeError rather than a
    network error. It is now reported, and handled, like any other network
    failure.
  • Every query-chat answer said "(with context: last 1 hours)". The
    default one-hour window was named on every answer, private and public,
    though nobody chose it, and a follow-up reads what was said since the last
    question rather than an hour anyway. The header now names a window only when
    the asker typed one, and says "1 hour", not "1 hours".
  • /aufhocker espnff help said the fantasy server wasn't connected when it
    was. When the server was up but none of its tools were open to the person
    asking, help said "not connected" while espnff mcp said connected. Help
    now says the same as espnff mcp, and no longer retries a server that is
    already up.
  • A web page with one bad character code couldn't be read. A numeric
    character code past the last valid one, such as �, failed the
    whole fetch, so Claude answered without the page. The code is now left as
    written and the rest of the page comes through.
  • When chat history was too long to send in full, part of a multi-line
    message could lose its author. If the cut fell inside a message several
    lines long, its last lines were kept without the name and time, so Claude
    read them as nobody's, or as the next person's. A message is now either
    kept whole or left out.
  • Buttons and menus that hit an error showed only "This interaction
    failed". The bot logged the error but never answered the click. The
    clicker now gets a private error message, and the message the button sits
    on is left alone.
  • The role sign-up board couldn't be posted with one role for both games,
    and the bot blamed a permission. Each button needs its own role, so
    Discord refused the board, and the bot told the admin to check its Send
    Messages permission. The command now asks for a different role per game up
    front, and a post that fails for any other reason no longer blames
    permissions unless Discord said so.
  • Reloading name mappings from a broken file erased every real name. If
    user-mappings.json had a typo or was missing, /aufadmin reload-mappings
    dropped every mapping, reported success, and the bot fell back to Discord
    usernames everywhere. A file that can't be read now leaves the current names
    in place, and the command says the reload failed.
  • The startup log could show part of the database password. A password
    holding an @ followed by /, ? or # (or digits followed by # or
    ?), pasted into DATABASE_URL unencoded, confused the address parser, and
    pieces of it were logged as the database host, port or name on every boot.
    When the address can't be read with certainty, the line now says
    (unparseable) and shows nothing from it.
  • The startup checks said only "fetch failed" when a service couldn't be
    reached. The real reason (address not found, connection refused, a bad
    certificate) was dropped, so DNS, firewall and TLS problems all looked the
    same. The check now names it. Any token or password is removed from the
    line first, including one the error text quoted back, which could already
    happen before.
  • A crash caused by an unusual error could skip the clean shutdown or log
    no reason. When code threw something that wasn't a normal error, the crash
    handler either crashed itself, exiting without closing the database or
    Discord, or logged a line with no reason in it. It now logs what was thrown
    and always shuts down properly, and normal errors keep their type and cause.
  • A log line whose text contained a % sign could lose its details,
    including the error. The logging library treats %s, %o, %% and the
    like as fill-in markers, and a message with one dropped everything logged
    with it; a user typing 50%off was enough. The details are now always kept,
    and the text is left exactly as written.
  • A number setting written as 10_000, 1e4 or 30s was quietly cut
    short. Only the leading digits were read, so CLAUDE_MAX_TOKENS=10_000
    ran as 10 tokens and RATE_LIMIT_INTERVAL_MS=30s as 30 milliseconds, and
    the startup check passed both. Number settings must now be plain digits, and
    anything else stops the bot at startup with an error naming the setting.
    Every shipped config file and the Swarm stack already use plain digits.
  • The connection that runs migrations ignored
    DATABASE_SSL_REJECT_UNAUTHORIZED. With DATABASE_SSL=true, the bot's own
    connection verified the database server's certificate, but the one that
    applies schema changes at startup accepted any. Both now use the same SSL
    settings. The live stack runs with DATABASE_SSL=false, so it is unaffected.
  • CI's dependency audit passes again. handlebars 4.7.9 had newly
    published critical advisories. It is dev-only, reached through
    ts-jest, so it was never in the production image, and 4.7.10, within the
    range ts-jest asks for, fixes it.

  Changed

  • D2EMU_POLL_INTERVAL_MS is removed. It looked like it set how often the
    D2 tracker polls, but nothing read it: Terror Zones are checked on the
    5-minute marks and DClone every 10–30 seconds. A stack that still sets it
    gets a startup warning saying it is ignored, so remove it from the stack.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v5.3.5 1c9c16ec5b

  Compare

  v5.3.5 Stable

  joeyr released this 2026-10-10 08:58:37 -05:00 | 3 commits to main since this release

  Nine rarer bugs, each needing bad timing or an unusual answer to show up. A
  cut-off answer now says so, Make Public and the fantasy Confirm buttons hold
  up under slow or failed clicks, a LoL game is posted once and keeps its
  result, a long watchlist lists, a stalled Data Dragon no longer holds up
  startup, and a new DClone board shows the latest grid.

  Fixed

  • An answer cut off at its length limit was shown as if complete.
    Thinking counts toward Claude's token limit, so an answer can stop half
    written, or before it starts. A half answer was shown and saved for
    follow-ups, a "let me look that up" before a cut-off tool call passed for
    the answer, and thinking alone failed as an API error. What was written is
    now shown with a note that it was cut off (or a message saying none of it
    was), and it isn't saved for follow-ups.
  • Make Public could post an answer twice. A long post can outlast
    Discord's 3 seconds to answer a click, which shows "This interaction
    failed" beside a button that still works, and a second click posted the
    answer again. A second click while posting now says so. A retry after a
    post failed partway also posted its start a second time; it now picks up
    at the part that failed.
  • A fantasy Confirm message could pass 2,000 characters and stop
    updating. Each line was clipped, but the message wasn't: five long
    changes with five long failures under them went over, and Discord refused
    the edit after the change had run, so its outcome never showed. Lines are
    now clipped shorter until the message fits.
  • A Confirm press that Discord didn't take left the change stuck. If
    answering the press failed, the change stayed "running" though nothing had
    been sent, and every later press said it had already been handled. It now
    goes back to waiting, so pressing again runs it.
  • Two watched players in one LoL game could get it posted twice. Their
    detections, a moment apart, both found no row for the game, and the step
    meant to reserve it succeeded for both. The game is now claimed with an
    insert that only one detection can win.
  • A game that ended while its ranked data was loading lost its result.
    Ranks and recent form can take minutes behind the Riot rate limit, and a
    game that ended meanwhile (a remake ends at the next poll) had its result
    replaced by the in-progress embed. The late edits now check that the game
    is still live first.
  • /aufhocker lol monitor-list failed past about 30 summoners. The whole
    list went in one message, and Discord refuses one over 2,000 characters.
    It now continues in follow-up messages.
  • A stalled Data Dragon could hold up startup for minutes. Its two
    requests had no timeout, and every plugin after the LoL one, and the health
    endpoint, waited on them. Each now has 10 seconds, and champion names fall
    back to IDs, as they already did when Data Dragon was down.
  • A new DClone status message could show an old grid for hours. A poll
    that landed while /aufadmin d2 dclone-board was posting couldn't see the
    new message to update it, and later polls with the same data skipped it.
    It is now brought up to date as soon as it is recorded.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v5.3.4 dcfaf6a52b

  Compare

  v5.3.4 Stable

  joeyr released this 2026-10-10 00:34:15 -05:00 | 4 commits to main since this release

  Long answers get time to finish, a deploy no longer fails a Confirm click
  partway through, a stale DISCORD_GUILD_ID no longer keeps the bot down, and
  a DClone spawn reported as 6/6 stays up for 12 hours.

  Fixed

  • A long answer from Claude could be cut off, and the question failed.
    Each API call had 60 seconds, set before the models thought and before
    chat ran on Sonnet 5.5. An answer arrives whole once the model has
    finished, thinking included, so a hard question at high effort could run
    out of time, and the identical request was then sent three more times,
    about four minutes, before the question failed. Each call now has 3
    minutes, a request that runs out of time is sent once more rather than
    three times, and no call runs past 13 minutes from the question, so the
    reply or the error still reaches Discord inside its 15-minute window. The
    start-up key check keeps its 60 seconds.
  • A Confirm click during a deploy could report a failure after the change
    had gone through. Shutdown closed the fantasy server's MCP sessions
    before waiting for interactions in progress, so a Confirm click mid-call
    failed, and an espnff mcp question mid-loop lost its tools. The sessions
    now close after that wait. The LoL and D2 pollers still stop before it.
  • A stale or mistyped DISCORD_GUILD_ID kept the bot from starting.
    After registering its commands globally, the bot clears that guild's
    commands, and a refusal there (a 403 from a guild the bot has left) failed
    startup, so the container restarted into the same failure. It now logs a
    warning and starts. .env.example now says what the variable is for: a
    guild to clear leftover commands from, not one to register to.
  • A DClone spawn that the tracker reports as 6/6 was deleted at the
    reset. The bot counted a spawn only when a game dropped from a step
    straight back to idle. A game reported at 6/6 ("Diablo has invaded
    Sanctuary") pinged spawn subscribers under "DClone progress", and the drop
    to idle a poll later deleted that post instead of keeping it up for 12
    hours. The first arrival at 6/6 now counts as the spawn either way.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v5.3.3 5bb1c96036

  Compare

  v5.3.3 Stable

  joeyr released this 2026-10-10 00:13:38 -05:00 | 5 commits to main since this release

  A busy moment at Anthropic or a stalled web page no longer fails a question,
  and a LoL game that seemed to end twice still gets its result.

  Fixed

  • One overloaded or server-error answer from Anthropic failed the whole
    question. A 529 "overloaded", or a 500, 502 or 503, got no retry. The
    SDK's own retries are off because the bot runs its own, and the bot's
    matched only rate limits and network errors. Inside a tool loop, the
    failure also threw away every round already done. These are now retried
    up to 3 times, waiting 1, 2 and 4 seconds, or as long as Anthropic's
    Retry-After asks, up to a minute.
  • web_fetch could leave a question hanging with no answer. Its
    15-second timeout ended once the headers arrived, so a server that sent
    headers and then trickled the body (a live feed, a long-poll endpoint, a
    deliberately slow page) held the question open past Discord's 15-minute
    window, with no answer and no error. The 15 seconds now cover the body
    too, and web_search reads Brave's results inside its timeout the same
    way.
  • A LoL game that ended twice could lose its result. A spectator 404
    mid-game ends a game falsely and starts a result retry, the next poll
    brings the game back, and the real end starts another. The second
    orphaned the first, which went on calling Riot every minute for up to half
    an hour where shutdown couldn't stop it, then logged every minute until
    the next restart. When its attempts ran out it stopped the second retry
    instead, so a result Riot published late was never posted. The newer retry
    now replaces the older one, and each retry stops only itself.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v5.3.2 6e828fe586

  Compare

  v5.3.2 Stable

  joeyr released this 2026-10-09 23:54:54 -05:00 | 6 commits to main since this release

  Fixes a query that any user could send to crash the bot.

  Fixed

  • A query with ``` followed by about 2,000 characters without a
    space crashed the bot. The message splitter reads everything between the
    backticks and the first space or line break as a code block's language
    tag, and repeats the tag in every chunk of a long block. A tag close to a
    message long left no room for code, so the splitter looped until it ran out
    of memory. The reply header quotes the query, so any user could trigger
    it, and again after each restart. A block whose tag would take more than
    half of each chunk is now split as plain text. Before, a slightly shorter
    tag sent one message per character or two of code, and a block that was
    all tag was dropped from the reply. Both now split as text too.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v5.3.1 5c5cea9e49

  Compare

  v5.3.1 Stable

  joeyr released this 2026-10-09 23:50:26 -05:00 | 7 commits to main since this release

  Rate-limit bypasses last as long as they say, OCE and SEA players work with
  the LoL monitor, and LOG_LEVEL can no longer switch the logs off.

  Fixed

  • A temporary rate-limit bypass expired as soon as it was granted. Its
    end time was stored as a TIMESTAMP, which keeps the bot's local time
    (America/Chicago) and drops the zone, then compared against NOW() in the
    database's UTC session. It read as five or six hours earlier than it was,
    so any bypass shorter than that was over before it started, and the
    longest ran 5–6 hours short. Migration 018 (applied at boot) turns every
    TIMESTAMP column into TIMESTAMPTZ, which stores the instant.
  • Channel history around the November clock change. Local time runs
    1 AM twice that night, and the same zoneless columns sorted messages from
    the second 1 AM before those from the first, so a follow-up asked in that
    hour missed what was said since the question before it. Fixed by the same
    migration.
  • OCE game results and recent form. Riot serves OCE match data from
    sea, but the bot asked americas, which answers 404, so every OCE game
    ended "results unavailable" after half an hour of retries.
  • Players on PH, SG, TH, TW and VN couldn't be registered. Their account
    lookups went to sea, which the account API doesn't serve. They now go to
    asia; any of the three account clusters returns any account.
  • LOG_LEVEL=INFO, warning, or any name winston doesn't know silenced
    every log line, errors included. winston drops every record at a level it
    doesn't know. Case is now ignored, warning reads as warn, and anything
    else logs at info, with a startup warning naming the value.
  • A LOG_LEVEL set in .env or .env.config never reached the logger
    under npm run dev or npm start: the logger read it before those files
    were loaded, while config/runtime reported the file's value. Docker, where
    the variables arrive in the environment, was unaffected.

  Deployment

  Migration 018 runs at boot and rewrites the messages table once. It reads
  the columns the bot writes in the bot's own time zone (the stack's TZ),
  which MigrationRunner now passes to every migration, so let the bot apply
  it rather than running npm run db:migrate from a machine in another zone.
  Rows those columns got before 1.10.0, when the bot ran in UTC, come out 5–6
  hours late; only spent bypasses and audit history are that old. Rolling the
  image back to 5.3.0 needs no database rollback: 5.3.0 reads and writes
  TIMESTAMPTZ correctly.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v5.3.0 521b082df7

  Compare

  v5.3.0 Stable

  joeyr released this 2026-10-09 12:21:45 -05:00 | 8 commits to main since this release

  query-chat and espnff mcp run on Claude Sonnet 5.5, and LoL lane
  detection thinks before it answers.

  Changed

  • query-chat, espnff mcp and the start-up key check run on Claude
    Sonnet 5.5 instead of Haiku 5.5. Everything else they send is
    unchanged: thinking on, high effort, 10,000 max tokens, a one-hour
    prompt cache. CLAUDE_MODEL now defaults to claude-sonnet-5-5. The
    live model is the stack's CLAUDE_MODEL, so set it to
    claude-sonnet-5-5 there. 5.2.0 runs fine on Sonnet 5.5 too, so the
    stack can go first. The cache belongs to the model, so each
    conversation's first follow-up after the switch reads its history at
    full price once, then caches again.
  • LoL lane detection thinks before it assigns lanes. It stays on Haiku
    5.5 at medium effort. Its token cap goes from 150 to 10,000, because
    thinking counts toward it, and a cap that thinking fills leaves no answer
    and every lane Unknown.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads
• v5.2.0 a4ce36a489

  Compare

  v5.2.0 Stable

  joeyr released this 2026-10-07 22:47:02 -05:00 | 9 commits to main since this release

  Finished walk posts no longer get stuck in the channel.

  Added

  • A Clone Hunters post that is never struck through goes 24 hours after
    its walk. Every post that names a time now goes on the delete list when
    it arrives, with the bot's reply under it, due a day after the last time
    in it. A post they forget to strike, or strike in a way the bot misreads,
    still goes. A strike still wins: struck through, the post goes in 30
    minutes as before. If they edit the walk time, the day counts from the
    new time. A post with no time in it ("Walk NOW EU") stays until it is
    struck, as before.

  Fixed

  • A walk post edited before its walk now goes once it is struck
    through. When Clone Hunters change a post first, crossing out the old
    region beside the new one (Region: ~~Asia~~ **Americas**), and strike
    the whole thing later, the strike marks pair up wrongly and Discord leaves
    Asia uncrossed. That one word made the bot read the post as still live,
    so it was never deleted. A post now counts as struck when it opens and
    closes with a strike mark, whatever is inside.

  Downloads

  • Source code (ZIP)
    0 downloads
  • Source code (TAR.GZ)
    0 downloads