Heiryn Docs

Changelog

This page lists public, user-facing updates to the hosted Heiryn service. Older entries are summarized.

2026-10-08 #

  • Claude Haiku 5.5 replaces the retired Claude 3 Haiku behind the s:claude alias, and Mistral Large 4 is available as m:mistral. Claude Haiku 4.5 (m:claude) was removed; saved m:claude selections were migrated to s:claude. Other aliases moved to their newest same-family model within the same size category: Claude Sonnet 5.5, Claude Opus 5.5, Gemini 3.8 Flash, Grok 4.7, Muse Spark 1.3, MiniMax M3, MiMo V2.6 Flash and Pro (same prices and request profiles), and the dated Qwen3.8 Max snapshot.
  • The GPT aliases moved from GPT 5.6 to GPT 6: s:gpt is GPT 6 Luna (also the Auto Pro model), l:gpt is GPT 6.1 Sol, and z:gpt is GPT 6 Astra. GPT 6 has no Terra tier, and Sol's $2/$10 pricing fits Large, so xl:gpt was removed and saved xl:gpt selections were migrated to l:gpt, which maps to the same model. z:gpt has no price ceiling and now costs $10/$50 per million tokens.
  • Auto now promotes to Claude Haiku 5.5 instead of GPT 6 Luna for Rohan, dangerous tools, and retries after interrupted queries. Both cost $0.10/$0.50 per million tokens, but Haiku 5.5 performs better on agentic and tool-heavy work. Conversation compaction stays on GPT 6 Luna.
  • FLUX.3 Image is available for image generation as s:flux.
  • Scribe v2 is now also available through OpenRouter. It does not support non-verbatim cleanup, so ElevenLabs remains preferred in Auto mode when it is configured; see ElevenLabs and OpenRouter for the differences. Scribe v2 is also OpenRouter's default transcription model, and MAI-Transcribe was upgraded to version 2.

2026-09-01 #

  • GLM 5.3 Flash is now the default OpenRouter chat model with low reasoning effort. MiMo V2.5 provides its availability fallback, while GPT 5.6 Luna remains the Auto Pro model. The Medium GLM alias now maps to the full GLM 5.3 model with its provider-default maximum reasoning effort.
  • Qwen3.8 Flash and Max remain available as explicit choices. Stored versioned Qwen selections were migrated automatically. A bare rate limit on the Flash model also switches the active query to MiMo V2.5.
  • Any usable retry or reset header keeps the selected model and is honored normally; availability fallbacks do not alter saved model preferences.
  • Model aliases now explicitly represent service categories rather than a guaranteed upstream model or version. The model documentation defines each regular chat category's maximum base input and output token prices, and the Terms of Service describe how alias mappings may be updated.
  • Explicit image requests now recover from bare delivery claims such as “Here you go!” and from additional creation wording such as “redraw” and “generating.” Auto recovery uses the Pro model to require the image tool when the default model cannot force a specific function call. Rate-limit retries now preserve the loaded image-tool session and briefly pace the forced recovery request.
  • Mentioning the bot with a supported image or document attachment now starts an AI request even when the message contains no additional text.

2026-08-25 #

  • Ask interactions now work in direct messages when Discord omits the resolved channel type from an interaction payload.

2026-08-23 #

  • Missing provider-key reminders are now debounced only for repeated mentions. Existing reminder markers no longer block valid AI requests, and /ask always reports missing configuration instead of timing out.

2026-08-17 #

  • /ask now uses an available community provider key when invoked in that community, including when Discord resolves it from a user installation. Community keys take priority unless the selected model exceeds the community's configured size limit.
  • Language matching now uses only the active requester's current input instead of being influenced by conversation history, source messages, profiles, or server metadata.

2026-08-13 #

  • Short provider rate-limit bursts now retry using provider-specific metadata when available, with a bounded fallback delay when no reset hint is returned.

2026-08-12 #

  • Grok Imagine Image 2.0 is now available as a category S OpenRouter image model, including support for up to three editing references.
  • Shared-channel conversations now identify each Discord participant more explicitly. Replies to older or out-of-history messages include an attributed copy of their target, without duplicating the latest history entry.
  • Replies whose target content is hidden by Discord now receive a direct hint to use Ask Heiryn or quote the content, without starting an AI inference.

2026-08-11 #

  • Auto image generation now also inherits the deployment default through a personal or community provider key. Image-specific instructions are exposed only when the MCP has actually loaded the generation tool.
  • Image generation now has explicit Auto and Disabled modes. Auto follows the personal, community, and configured deployment defaults; Reset restores Auto.
  • Available chat and image selectors are consistently capped at category S for system-funded use. Bring-your-own-key limits remain configurable where supported.
  • Deployment-funded voice transcription stays on the configured ElevenLabs route, while personal and community provider keys retain their own choices.
  • Image generation now receives validated image bytes directly when Heiryn can attach files, while preserving the hosted URL fallback in channels where it cannot.
  • External image-generation tools now have up to 120 seconds to complete.
  • Auto image generation now uses Nano Banana 2 for substantially faster everyday results.
  • Image generation now supports edits and iterative refinements. Images from the current message or its direct reply target are passed as references, and the latest generated image can be reused explicitly for a follow-up edit.
  • Explicit image requests now recover automatically when a chat model claims success without calling the image tool, preventing replies that announce an image without attaching one. The system prompt also names util_generate_image explicitly so compliant models call it on the first turn.
  • Generated-image references are now private harness metadata: chat and compaction models never receive them, while iterative edits can still reuse the latest image. History records now keep model content and internal state in separate deterministic JSON fields instead of model-visible markers.

2026-08-09 #

  • OpenRouter image generation now includes Qwen Image 3 and Qwen Image 3 Pro.
  • Muse Spark 1.2 is available as an additional multimodal chat model, and GPT Transcribe is available as an additional speech-to-text model.
  • Alibaba Cloud is no longer offered as a separate provider. Existing OpenRouter and ElevenLabs configuration controls now render entirely from the providers available in the current deployment.
  • Provider arguments now default to OpenRouter across key and model commands. Image model suggestions also display their XS, S, or L catalog category.

2026-07-31 #

  • Community Custom Instructions replace the former community description. Server managers can now define real instructions for tone, language, formatting, terminology, and local conventions. They take priority over both personal instructions and ordinary member input while Heiryn's fixed identity, safety, security, authority, permission, and tool rules remain protected.
  • /community_instructions and the server dashboard expose the new setting.
  • Personal Custom Instructions are available through /personal_instructions and the Account dashboard for behavior and style preferences that follow you across servers and DMs.
  • Conversation compaction keeps long channel discussions within a 128,000-token budget. It retains the recent exchanges, a reasonably sized first exchange, and a current reply target verbatim. Use the slash-only /compact command to compact eligible history early.
  • Automatic compaction now has a dedicated Discord waiting state. Cancelling the waiting query leaves compaction running safely in the background, and a later request or Retry can continue after the same task completes. Resetting history is the exception: it stops compaction internally and visibly cancels any persistent query that was using the cleared history.
  • Ask commands and message actions are explicitly one-shot and do not affect channel history. They remain independent when that channel's history is reset.
  • Cancelled queries and detected reasoning loops now retain the latest visible reasoning and any partial answer above the interruption message. Retry remains user-initiated, uses more permissive loop detection for that attempt only, and promotes an interrupted Auto request to the Pro model.
  • Reasoning-loop detection is isolated to each provider round, preventing tool activity and earlier reasoning rounds from causing false cancellations.
  • OpenRouter now applies explicit per-model inference profiles after Auto has selected the concrete model. Qwen3.7 Flash uses temperature 0.6, a 1,536-token reasoning budget, and an 8,192-token total output limit; Qwen3.7 Plus and Max use non-thinking mode to keep latency bounded.
  • Chat and compaction requests now share provider-aware rate-limit pacing and bounded HTTP 429 retries.
  • Auto now promotes only for an explicit --rohan, at least one loaded tool marked as dangerous, or Retry after an interrupted request. Staff status, long history, and ordinary tools no longer promote a request by themselves.
  • OpenRouter Pro now uses GPT 5.6 Luna.

2026-04-21 #

  • Public AI channels (server dashboard): choose which channels everyone may use the AI in. When the list is empty, nothing changes. When it is not, server staff can still use the AI anywhere; everyone else is limited to those channels (slash/context menus show a short error elsewhere; mentions in other channels are ignored).
  • Optional Allow AI use in all threads (/ai_allow_threads or dashboard): regular members can use the AI in any thread; Heiryn may also start a public thread from a blocked channel when permissions allow.
  • One Save applies the full channel list; suggestions help when adding channels.

2026-04-18 #

  • The Cancel button on AI replies now works reliably from /ask and the Ask Heiryn context menu (it previously did nothing in those entry points). The button also turns into a disabled Cancelling… label as soon as you click it, so it is obvious the cancellation has been registered, and a stuck request will be force-closed shortly after instead of hanging until the read timeout.
  • Heiryn now detects when a model gets stuck repeating the same templated reasoning line over and over and automatically cancels the query, freeing the slot instead of letting the loop run on.

2026-04-17 #

  • Documentation links are now canonical: trailing slashes and stray .md suffixes redirect to the short URL, and every page carries richer titles and link previews when shared on Discord.

2026-04-16 #

  • Heiryn now gives clearer, more helpful error messages when a provider call fails, without leaking internal details.
  • When you clear an AI conversation history for a channel from the dashboard, Heiryn posts a short notice in that Discord channel so other members know the context was reset.
  • Image generation is more reliable: the bot only tells you an image is ready when it really has a usable URL, and it retries in a more predictable way when a model misbehaves.
  • Ask the bot about its own features and it will look them up in these docs directly before answering.

2026-04-15 #

  • Server-settings, block list, and history pages now update in real time. If you or another admin changes something, every open dashboard tab for that server picks it up right away without a refresh.
  • History transcripts are rendered as Markdown with line breaks, sanitized HTML, and role-colored mentions for better readability.
  • The history viewer shows a loading spinner on first open and refreshes quietly afterward, so you are not distracted by flashing content.
  • Dashboard CSS and JavaScript are split into smaller, cacheable files, which makes page loads snappier on repeat visits.

2026-04-14 #

  • Added a personal AI opt-out: run /ai_opt_out or toggle the setting on the Account page to disable the AI feature for your account. While opted out, nobody can use Ask interactions on your messages. Opting out also removes your personal provider API keys.
  • Dashboard history page now shows per-channel context (channel name, category, expiry) and lets admins adjust how long each history is kept before it expires.
  • Heiryn now keeps a short-lived, per-server cache of role mention colors so history transcripts look closer to how they appear in Discord itself.
  • Error messages coming from AI providers are sanitized before being shown, so no internal details leak into chat.

2026-04-13 #

  • Conversation history is now encrypted at rest, in addition to provider API keys.
  • Server names are kept up to date whenever settings are changed, so the dashboard always shows the current name even if the bot is not currently in the server.

2026-04-12 #

  • More robust handling of malformed streaming responses from OpenRouter, which reduces the chance of a query stalling or truncating unexpectedly.

2026-04-10 #

  • Brand new web dashboard:
  • Sign in with Discord at https://heiryn.dis.gg/dashboard.
  • Manage your personal AI settings, image preferences, and API keys from the Account page.
  • Review and revoke active sessions with browser, device, and country information.
  • Configure servers you help run: community description, default chat model, image provider, report channel, and community API keys with an optional maximum model size.
  • Manage the block list with a visual duration picker and per-block reasons.
  • Browse each channel's conversation history and clear it with a single click.
  • Dashboard is responsive and usable on mobile.
  • Live usage statistics on the Account and server pages update in real time while you watch them.
  • Link previews (OpenGraph) for the landing page and dashboard look nicer when shared on Discord.
  • Normalized error pages, hardened OAuth flow, and trusted proxy handling for deployments behind Cloudflare.
  • Country information on sessions is looked up locally from an embedded IPInfo Lite dataset; your IP address is never sent to a third party for this.
  • Security headers are applied to every response.
  • Added a Privacy Policy and Terms of Service.

2026-04-07 #

  • Added a full automated test suite, which helps ship changes with fewer regressions.

2026-04-06 #

  • Added /personal_stats and /community_stats so you can view your own and your server's AI usage totals (queries, ask interactions, images generated).
  • When the AI returns a Markdown table, the bot now formats it as a code block so Discord renders it as intended.
  • Tool calls and query tracking are counted per user and per server.

2026-04-05 #

  • Reminder delivery (for apps integrating with the Dis.gg MCP) falls back to DMs when the bot cannot post in the target channel.
  • Fixed a few cases where duplicated reasoning would appear on long responses.
  • Landing page has fresh OpenGraph metadata and an updated tagline.

2026-04-02 #

  • /support and /invite short links route through https://heiryn.dis.gg/support and https://heiryn.dis.gg/invite.
  • Documentation assets are served directly from the same site.

2026-04-01 #

  • First public release of the documentation site (this site) and of the legal pages: Privacy Policy and Terms of Service. Docs are rendered by the same engine as Dis.gg MCP.
  • Introduced the public API service that hosts documentation and the dashboard.

2026-03-31 #

  • Added /block check to look up whether a user is currently blocked and see the details (author, reason, expiry).

2026-03-30 #

  • Reworked how chat models are organised into size categories (XS, S, M, L, XL, Z) and how the active model is resolved between personal preference, server default, and provider default.

2026-03-29 #

  • Added Bring Your Own Key: configure a personal or server-wide OpenRouter or Alibaba API key from /personal_provider_key, /community_provider_key, or the dashboard. Keys are encrypted at rest and never displayed again after being set.

2026-03-27 #

  • Added OpenRouter as a chat provider, giving access to a much wider catalog of chat models (Claude, GPT, Qwen, Gemini, Grok, and more).
  • Introduced the % and %% model override suffixes for switching models per-message or per-account without leaving the chat.

2026-03-26 #

  • Added a Move to Thread button on AI responses so you can branch a conversation into a thread without losing context.
  • Ask (Modal) now lets you choose a visibility mode: public, partial (response visible, reasoning hidden), or private.

2026-03-22 #

  • AI queries now stream reasoning and the final response live, so you see progress instead of waiting in silence.
  • Conversation history and follow-ups work more consistently across servers, DMs, and user app installs.
  • /community_description lets a server set a short prompt that helps the AI tailor replies to the community's topic and tone.
  • /report_channel lets a server designate a channel where AI safety reports are delivered.

2026-03-11 #

  • Added Alibaba (DashScope) as a chat and image provider, with Qwen models and Qwen Image 2.0 image generation.

2025-11-28 #

  • Friendlier error messages when an AI request fails, with less noise in chat.
  • Source link previews are handled more cleanly.

2025-08-18 #

  • Added Tenor link support so the bot can understand GIF links in a conversation.

2025-08-09 #

  • The AI can now read text file attachments included with your message (up to 20 MB per file).

2025-08-08 #

  • Standard AI queries can now be cancelled mid-way with a button.

2025-08-05 #

  • /ask gained an ephemeral option for private responses.
  • Ask (Modal) context menu now supports adding extra context to your question.

2025-08-03 #

  • The bot can mention you back when replying to an Ask interaction on your own message, making it easier to find replies in busy channels.
  • Added the first version of a server description command.

2025-08-01 #

  • Added the context menu entries that let you right-click any message to ask Heiryn about it.

2025-07-31 #

  • /ask improvements and higher default history limit.
  • AI now works in DMs and through user-app installs in servers where the bot is not a member.

2025-07-29 #

  • Conversation history now scales properly across multiple servers and handles replies more naturally.

2025-07-25 #

  • Added the /reset_history command.

2025-07-21 #

  • Initial public release of Heiryn, with the AI feature and a /block command set.

This changelog is intentionally concise. For setup and day-to-day usage, start with Quick Start and the AI Overview.