Skip to content

Releases

Raynard v0.16.0

Choose a tag to compare

@github-actions github-actions released this 25 Oct 08:43

Raynard v0.16.0

This release fixes bookmark references, so an answer you saved can be brought
into a new question.

  • Picking a bookmark from the @ menu now attaches its saved question and
    answer to your message. Before, the menu inserted the bookmark's name, but
    sending only recognised it with a leading @. The bookmark never reached the
    model, which then said it couldn't see other chats.
  • A bookmark name you type or paste yourself, with or without @, is
    recognised too, including at the end of a sentence. The match must be exact,
    including the short code at the end of every bookmark name, so ordinary words
    never attach a bookmark by accident.
  • Bookmark matching now runs in the app's native core against the bookmark
    store it already keeps, so the window no longer loads every bookmark just to
    power the @ menu. Existing bookmark names are unchanged.

Raynard v0.15.1

Choose a tag to compare

@github-actions github-actions released this 07 Sep 00:14

Raynard v0.15.1

This release makes Telegram answers substantially easier to read and carries
visual results into scheduled alerts and remote conversations.

  • Literal and Markdown bullets now remain distinct, comfortably spaced list
    items instead of collapsing into a wall of text.
  • Small, narrow Markdown tables render as native Telegram rich tables. Wider
    or longer tables keep Raynard's phone-friendly labelled-row layout, and every
    native table carries that layout as a fallback if Telegram rejects the rich
    message.
  • Charts are rendered as branded PNGs and delivered after a concise standalone
    text summary, for both scheduled Telegram alerts and interactive Telegram
    replies. The text is always durable on its own, so a failed media upload never
    hides the actual result.
  • Telegram's durable queue now retries ordered text, rich-table, and chart
    parts independently, validates image payloads before persistence, preserves
    older text-only queue records, and discards delivered image bytes to keep the
    local store bounded.

Raynard v0.15.0

Choose a tag to compare

@github-actions github-actions released this 06 Sep 21:04

Raynard v0.15.0

This release adds a weekdays schedule to recurring tasks and fixes a release
pipeline bug that let real test failures slip through the Windows build
unnoticed.

  • Scheduled tasks now support a Weekdays frequency (Monday through
    Friday), shown next to Daily in the task editor's frequency picker. A
    request like "every weekday" or "Monday to Friday" now maps directly to
    this schedule instead of being approximated as Daily with a note explaining
    the substitution.
  • Fixed the release workflow's Windows job silently ignoring test failures:
    its "Run deterministic tests" step ran without an explicit shell, which
    defaulted to PowerShell — PowerShell does not stop a multi-line script when
    an earlier command fails, so a failing npm test was masked by the passing
    cargo test/cargo check that ran after it. The step is now pinned to
    bash, matching the macOS and Linux jobs, so a real test failure will fail
    the build going forward.
  • Fixing that surfaced two genuine Windows-only test bugs, both corrected:
    a catalog slug check that split a path on a literal forward slash (breaking
    on Windows' backslash-separated paths), and a workspace-containment test
    that compared resolved paths against hardcoded forward-slash strings
    instead of the platform's own path resolution.

Raynard v0.14.0

Choose a tag to compare

@github-actions github-actions released this 06 Sep 20:25

Raynard v0.14.0

This release fixes a startup delay on the Telegram channel, formats Telegram
replies more safely, and widens how much price history the Financial Modeling
Prep extension can pull in one call.

Faster startup

The Telegram channel's keychain check and poller resume used to run
synchronously during app setup, which could delay the window's first paint
when OS keychain authorization added latency (most noticeably in tauri dev,
where the ad-hoc signing ACL changes on every rebuild). That check now runs on
a background thread instead, so it can no longer hold up startup.

Safer Telegram replies

Telegram replies now use Telegram-safe HTML formatting with a persisted parse
mode, so both live and durably queued scheduled deliveries render consistently
and replies saved before this change keep their original plain-text delivery
behavior.

Wider Financial Modeling Prep history

fmp_price_history's range option topped out at one year even though the
underlying FMP endpoint already returns its full default history. It now
offers 2Y/5Y/10Y/MAX buckets as well.

Raynard v0.13.0

Choose a tag to compare

@github-actions github-actions released this 06 Sep 07:49

Raynard v0.13.0

This release adds private Telegram conversations and conditional scheduled
alerts, gives chats durable memory and bookmark references, expands equity
research, and makes result charts more useful to explore.

Telegram conversations and scheduled alerts

Connect a BotFather bot in Settings, explicitly pair one private Telegram
account, and chat with Raynard using the same selected model, installed
extensions, and locally saved history as the desktop app. Telegram supports
ordinary text questions plus /start and /new; configuration, credentials,
extension development, and other approval flows remain on the desktop.

Scheduled tasks now separate their always-saved Raynard history from their
notification channel. A task can notify the desktop or its fixed paired
Telegram recipient after every run, on every matching condition, or once when
a condition becomes true and then rearm after a non-match. Conditional checks
must finish with an explicit evidence-backed match result and fail closed when
the answer is unavailable or ambiguous. Telegram alerts are durably queued by
execution, resume at the first unsent chunk after temporary failures, and
expire before stale results can be delivered.

Bookmark references and agent memory

Type @ to reference saved bookmarks as well as installed extensions. A
bookmark mention supplies its complete question and answer to the model while
the chat bubble keeps a compact, clickable reading marker. Bookmark rows can be
renamed inline, show progress while Raynard generates a title, and recover from
title-generation attempts that spend their first token budget reasoning.

Raynard can also propose remembering, correcting, or forgetting a durable
fact. Every change requires in-chat confirmation, memories can be scoped to one
extension, and /memory provides a read-only overview. Only bounded, relevant
global and extension memories are added to a turn.

Richer charts

Bar charts can switch between grouped, stacked, and line views where the data
permits it. Charts appear at the point in an answer where the agent presented
them instead of always collecting at the end, and the renderer improves legend
spacing, dual-axis handling, series visibility, and highlighted categories.

Expanded Financial Modeling Prep research

The bundled Financial Modeling Prep extension adds historical price charts,
current TTM metrics, long-range valuation history, and a company screener.
Financial statements and revenue segments now accept inclusive fiscal-year
ranges and larger bounded histories, making requests such as an Apple analysis
from 2011 onward possible without manually paging periods.

Raynard v0.12.1

Choose a tag to compare

@github-actions github-actions released this 31 Aug 19:16

Raynard v0.12.1

A bugfix release for chart readability: legends now sit above the plot instead
of overlapping the x-axis.

Chart legends no longer cover x-axis labels

Multi-series charts previously used Recharts' default bottom legend. On line
charts with enough series or longer labels, the legend could occupy the same
space as the x-axis and obscure its labels.

Legends now use the chart's top alignment. Recharts reserves plot space for the
legend there, keeping the x-axis clear while preserving the existing controls
for hiding and restoring individual series.

Raynard v0.12.0

Choose a tag to compare

@github-actions github-actions released this 31 Aug 13:47

Raynard v0.12.0

A bugfix release: Kimi conversations work again after Moonshot decommissioned
the kimi-k2.5 model.

Kimi chats stopped working when Moonshot retired kimi-k2.5

Moonshot discontinued kimi-k2.5 on 2026-08-31 in favor of kimi-k2.6. Every
Kimi turn immediately failed with a 404 from Moonshot's API, since Raynard's
Moonshot provider defaulted to and stored the now-dead model id.

The provider preset's default chat model now points at kimi-k2.6. An
install that already had kimi-k2.5 saved as its active chat or coding model
is migrated automatically the next time its config loads — no need to notice
the failure and reselect a model in /models.

kimi-k3, the default coding model, is unaffected and unchanged.

Raynard v0.11.1

Choose a tag to compare

@github-actions github-actions released this 30 Aug 11:51

Raynard v0.11.1

A bugfix release: citations, links, and inline code wrapped in bold or italic
text now render properly instead of appearing as raw markup.

Citations inside emphasis render as chips

Assistant answers cite their sources with [^n] markers, which the renderer
turns into clickable chips that open the observation behind the number. Any
marker the model wrapped in emphasis was rendered as literal text instead.

That is exactly the shape models reach for when attributing a block, so it hit
constantly. A research answer ending its sections with

*Sources: [^7], [^8]*

showed the raw [^7] and [^8] rather than chips, while bare markers
elsewhere in the same answer worked normally — leaving an answer with some
citations clickable and others not, for no reason visible to the reader.

Replaying one real 31-reference answer through the fixed renderer: all 13
cited markers now resolve, where 9 of them previously stayed literal.

The cause was that inline markdown is matched with a single flat pattern, and
the <strong>/<em> branches filled their node with textContent — which
discarded every nested span. Links and inline code inside emphasis were lost
the same way, so **See [the docs](https://example.com)** rendered its
markdown source rather than a link.

Emphasis now renders its own contents, bounded by a nesting cap. Because the
pattern is a shared stateful regex, spans are collected before any node is
appended, so a nested walk cannot disturb the position of the walk that
started it.

Internals

The inline renderer moved out of src/main.ts into src/inline-markdown.ts.
main.ts is the application entry point and runs side effects on import, so
nothing in it could be covered by a test; the extracted module is now tested
directly, including citations, links, and code nested inside emphasis.

Raynard v0.11.0

Choose a tag to compare

@github-actions github-actions released this 30 Aug 08:55

Raynard v0.11.0

A /new slash command and a fix for wide markdown tables that were silently
falling back to unreadable raw text.

/new slash command

Typing /new now runs the same flow as the sidebar's New Chat action:
switches to the Chats sidebar and opens a fresh conversation with the
pre-chat suggestions. It's registered in the slash menu alongside
/extensions, /models, /settings, and /status, so it's discoverable
without knowing it exists.

Wide markdown tables render correctly

The lightweight markdown renderer capped tables at 8 columns. Any table wider
than that — a routine shape for a multi-metric comparison, e.g. several
stocks across price, valuation, and margin columns — silently fell back to
unparsed, raw pipe-delimited text instead of an HTML table, with no
indication anything had gone wrong.

The cap is now 24 columns, and a table wide enough to exceed the message
width scrolls horizontally instead of squeezing every column down to a few
unreadable pixels.

Raynard v0.10.0

Choose a tag to compare

@github-actions github-actions released this 28 Aug 08:48

Raynard v0.10.0

Chart JSON that a model corrupts in transit now recovers through a general
repair pipeline instead of a short list of exact, previously-seen shapes.

General-purpose chart JSON repair

v0.9.0 fixed two exact malformations observed from Kimi K2.5: a chart fence
missing its chart language tag, and one stray trailing }/] glued onto an
otherwise well-formed spec. Further testing turned up a third shape — one or
two stray backticks glued directly onto the JSON with no newline before the
real fence close — which the narrow, bounded bracket-stripping from v0.9.0
didn't cover.

Rather than add a fourth special case, chart parsing now runs a small pipeline
of general-purpose repairs, each re-validated by the full chart schema before
use:

  • straighten smart/curly quotes left over from pasted text
  • strip a dangling trailing comma before a closing bracket or brace, tracking
    string state so a comma inside a title or label is never touched
  • extract the first balanced top-level JSON object, ignoring any narration
    glued onto its head or tail and any amount of trailing garbage after it
    closes

That balanced-object extraction alone covers every trailing-garbage shape
seen so far — and any future one shaped the same way — instead of only the
specific bracket and backtick sequences already observed. A genuinely
truncated or broken spec still falls back to an ordinary code block rather
than being misread as a chart.