Readable release notes from GitHub, designed for humans.
Sep 8, 2026
Fixed Grok sessions now scroll cleanly on desktop and mobile. Grok used to open in its fullscreen mode, which has no scrollback and ignores the mouse wheel, so scrolling was erratic on desktop and impossible on phones and tablets.
opendray updateFixed Grok sessions now scroll cleanly on desktop and mobile. Grok used to open in its fullscreen mode, which has no scrollback and ignores the mouse wheel, so scrolling was erratic on desktop and impossible on phones and tablets.
Added Grok account switching can now carry the conversation across the switch, matching Claude.
Added Grok accounts get a full UI: a Grok accounts panel in the Providers page (list, import local, enable or disable, remove) and live account switching from the session header, matching Claude and Antigravity.
Grok catches up. You can pool several xAI accounts, bind one to a session, and switch a running grok session from one account to another without losing the session , the multi account parity Claude and Antigravity already had.
PATCH /sessions/{id}/grok-account)
without dropping it. Each account is an isolated GROK_HOME; MCP trust is
per-account and the heavy install/cache dirs are shared by symlink.
(#533, #534, #535)--rules. (#530)Grok catches up. You can pool several xAI accounts, bind one to a session,
and switch a running grok session from one account to another without
losing the session, the multi-account parity Claude and Antigravity
already had. Each account is an isolated GROK_HOME; the heavy install and
cache directories are shared across accounts by symlink, so a second
account costs kilobytes, not gigabytes.
PATCH /sessions/{id}/grok-account)
without dropping it. Each account is an isolated GROK_HOME; MCP trust is
per-account and the heavy install/cache dirs are shared by symlink.
(#533, #534, #535)--rules. (#530)The one line installer runs to the end again, as root, on a fresh Proxmox LXC or VPS, which is exactly how it is usually run.
run_priv -u postgres ran a literal -u), and
migrate ran with HOME=/root so its backup keyfile was unreadable. The
one-liner now runs to the end. (#529)doc_read's framing is no longer written into pages. (#517)The one-line installer runs to the end again, as root, on a fresh Proxmox LXC or VPS, which is exactly how it is usually run. And Cortex, the knowledge layer, grows a lifecycle: operator edits stick, a page can be handed to an in-session agent, proposals are reviewed as a diff, and the sweep respects a page's approval gate.
run_priv -u postgres ran a literal -u), and
migrate ran with HOME=/root so its backup keyfile was unreadable. The
one-liner now runs to the end. (#529)doc_read's framing is no longer written into pages. (#517)The Vault stops being a flat pile of markdown. It holds a folder structure you maintain, files each project under its own name, and renders HTML documents as carefully as it renders markdown, from an in memory string in
The admin UI works on a phone and a tablet. It was built for a desktop three-pane layout, which on a phone meant a horizontal scrollbar and two columns you could not reach. Every page now has a narrow form: below 1024px the third pane becomes a slide-over, below 768px the list becomes the page until you pick something and the navigation tree follows it into a drawer. Tables, spacing and the topbar were sized to match.
The drawer is one component rather than a pattern re-implemented per page, and it is reachable from a labelled control in the header, never from an edge handle alone, which is unreachable in practice.
The Vault syncs from the phone. Repository status, a manual commit / push / pull, and the auto-sync settings are all on the phone now; previously the mobile app could edit documents but not see whether they had reached the remote, so the answer to "did that save get anywhere" was only available on a desktop.
The mobile Vault has the knowledge layer the web has had. It
could create, read, edit, rename and delete a document, but nothing
that makes a vault more than a folder. It now browses by #tag,
shows a document's backlinks and its tags, follows [[wiki links]],
completes them while you type, jumps by heading, and opens today's
daily note from the same template the web writes.
Two constraints worth knowing. The preview runs with JavaScript off,
a document can arrive by git pull, so it is not the operator's
own writing by default, which means wiki-links are anchors on a
private scheme intercepted at navigation, and a heading jump lands
in the source view rather than scrolling the rendered page. And the
path rules are Unicode-aware on mobile: the web's ASCII-only
sanitiser turns 笔记.md into --.md.
The Vault stores and renders HTML documents, not just markdown.
Project documentation increasingly ships as HTML, exported from
Notion or Word, generated by typedoc, asciidoc or Sphinx, and a doc
library that cannot open those is a markdown library. .html /
.htm are now first-class alongside .md: listed in the tree,
creatable, movable, and rendered.
HTML is never served from opendray's own origin. There is no
endpoint returning a document as text/html; that would execute it
same-origin with the admin session, so a file pulled from a git
remote could take over the account. The body travels as a string
through the existing JSON read and is rendered from memory in a
sandboxed frame with an opaque origin. Scripts are off by default,
exported documentation is static markup and renders identically
without them, with a per-document opt-in for pages that genuinely
run code, remembered locally per file.
Mobile gains a rendered view for the first time. It previously showed only a document's source, which HTML would have made unreadable, so preview covers both kinds: markdown is converted and rendered through the same webview rather than drawn as native widgets, so the two formats look like one product.
Deliberately not symmetric: [[wiki links]] are still scanned and
rewritten in markdown only. An HTML document can be the target of
one, but its own <a href> links are left alone. Rewriting those on
a move is a different job, with relative paths, anchors and assets to
get wrong. Auto-derived paths (daily, project, personal notes) stay
markdown. And when a filename and a template disagree, guide.html
from the markdown Blank template, the filename wins, since the
operator chose the extension and the template was probably a default.
In-document links work: clicking a table-of-contents entry scrolls to that heading, and only links that actually leave the document open a new tab. This needed more than it sounds like. See Fixed below.
Vault auto-sync settings could not be saved. Whatever interval you set, the dialog snapped back to "every 10 minutes": the form refetched every 8 seconds and overwrote the draft mid-edit, so Save was never enabled and no setting had ever reached the database. The interval is now a free-text Go duration with presets, and the gateway rejects an unparseable one with a message instead of quietly substituting a default.
The mobile Vault's rename and delete were unfindable. They existed, behind a long press on a row, with nothing on screen saying so, reported in testing as missing outright, which for a hidden gesture amounts to the same thing. Each row now carries a visible actions button in place of a chevron that did nothing the row did not already do, and the open editor offers rename and delete where the web has always had them.
Creating guide.html on the phone produced guide.html.md. The
new-document path appended .md unconditionally, so an HTML
document could not be created from mobile at all.
Six hints rendered <prefix> as literal text instead of
<prefix>.
A table-of-contents link no longer loses the document. In the web
viewer, clicking an in-document #anchor navigated the frame away
and rendered the opendray app inside the document view. A srcdoc
frame's document URL is about:srcdoc while its base URL is
inherited from the parent page, so href="#install" resolved against
the gateway and became a real navigation. The base is now pinned to
the frame's own document, which also stops a pulled document's
relative URLs from resolving against the gateway's origin.
The first attempt at "external links should open a new tab" used
<base target="_blank">, which retargets every link, so a
contents entry opened a blank tab instead of scrolling. Anchors are
now rewritten individually: fragment links are left in place, an
author's explicit target is respected, and only links leaving the
document get target="_blank". Generated documentation is precisely
where this mattered: Sphinx, typedoc, asciidoc and Notion exports
all ship a fragment contents list, and markdown footnotes render as
fragment links too.
Mobile deliberately does not pin a base: loadData makes the
document's own URL and its base both about:blank, so fragments
already resolve in place, and pinning would have introduced the bug
rather than fixed it.
CI no longer runs the whole suite twice on every branch commit.
A feat/** branch with an open PR matched both the push and
pull_request triggers, and the concurrency group keys on the ref,
which differs between the two events, so neither run cancelled the
other. push is now main-only; pull_request already covered branch
work.
A project's documents are filed under the project's name. The
Vault is a project documentation library, and it filed every project
under projects/, a folder naming what the whole library already
is. One level of nesting that told the reader nothing, on every path,
in every listing, and at the top of the git repository the Vault syncs
to. The operator's own notes were split off further still: agent docs
at projects/<name>/…, the human scratchpad at personal/<name>.md,
the same project's material in two distant places, sorted by who
wrote it rather than by what it is about.
New vaults now use <name>/… with that project's personal.md
inside it. Existing vaults are not rearranged. The layout is
decided once, at first start, and written into config.toml as
[vault] layout. Recording it is the point: the alternative, work
the shape out from what is on disk each time, is the bug that put
one install's entire document library behind a notes/ directory,
because a probe asking "does this folder have content?" changes its
answer the moment someone puts content there.
The doc library offers the conversion, on web and on mobile, to any vault that still nests projects. A migration that ships only as a CLI command is one that only the people who wrote it ever run. Every existing vault would have stayed nested with its owner never learning there was a choice. The offer is dismissible and never appears for a vault that is already flat or has nothing to move, and the gateway says the same thing once in the startup log for anyone who never opens the UI.
Nothing moves without a preview. Web and mobile both run the
migration as a dry run and show the real list, including what it
refuses to touch and why, before anything is renamed.
opendray notes flatten does the same from a terminal. It defaults
to a dry run, drives the same rename the UI uses so [[wiki links]]
are repointed as it goes, repoints per-cwd project overrides, and
never overwrites: a destination that already exists is reported
and skipped, leaving both copies for you to reconcile.
Where a project's notes live is now answered by the gateway,
/notes/info reports the layout and the project mapping carries
personal_path, instead of being re-derived by web, mobile and the
CLI. Three implementations guessing is three chances to disagree, and
the CLI's notes project was already guessing wrong.
Reserved names step aside rather than collide: a project called
daily files under daily-docs, since daily/YYYY-MM-DD.md belongs
to the whole vault. _- and .-prefixed names are reserved too.
The conversion records the resulting layout itself and applies it to
the running gateway, rather than leaving both to the next restart.
Otherwise a converted vault keeps deriving projects/<name> and
personal/<name>.md against the directories it just emptied. The
directories the migration empties are removed; one still holding
anything is left exactly where it is.
Documents save when you say so, not on a timer. The editor wrote after every pause in typing, so a long document rewrote itself to disk over and over while it was being worked on. Saving is now a button, or ⌘S / Ctrl+S, on web and on mobile. Leaving a document still flushes unsaved text, that safety net costs nothing while typing, and closing the browser tab with unsaved work asks first.
Autosave was not the whole story behind the typing lag, so two things that were: the editor streamed every keystroke to the page, which re-rendered the whole vault tree and recomputed the outline per character; and the tree re-rendered along with it. The stream is now paced to what a sidebar can use, and the tree only re-renders when the note list or the selection changes.
Line numbers in the source view. On by default for HTML, off for markdown, and toggleable either way. Numbers turn wrapping off, the way every code editor does it: a wrapped line covers several rows, so a gutter counting 1..N drifts on exactly the long lines an HTML document is full of.
The doc library can rename and delete a document. It could create
and edit one, and that was all. A daily note, or anything not bound
to a project, could be made but never moved or removed from the one
surface that lists everything. Rename goes through the move endpoint,
so the [[wiki links]] pointing at the old path are repointed rather
than left dangling, and a partial rewrite is reported instead of
being folded into a plain "renamed". Mobile gains rename too; it
already had delete.
"New" and "Today" are reachable once a document is open. They
were in the header the whole time, pushed off the right edge by the
vault path: a truncate element in a flex row still needs
min-width: 0, or it refuses to shrink below its content. Anyone
with a long vault path could only reach the two actions from the
empty state, which is exactly when they had no documents to leave.
A folder holding the selected document can be collapsed again. The tree re-applied "open the ancestors of the selection" on every render, so the collapse landed and the next render undid it, and "Collapse all" left that one branch open. Revealing the selection is a response to the selection changing, not a rule about what must stay open, so it now runs once per selection.
opendray notes accepts flags after the subcommand. Go's flag
package stops parsing at the first non-flag argument, so notes flatten --apply left apply false, performed a dry run, and advised
re-running with --apply, which is what had just been typed.
notes list --prefix=x silently listed everything for the same
reason. Both orders now mean the same thing.
The Vault is your documents. Agent skills and the MCP registry moved
out. One root held three tenants with nothing in common: the
operator's markdown, the skills opendray injects at spawn, and the MCP
registry, which is why the settings page could only describe it as
"notes, skills and git-versioned root", a sentence that parses only if
you already know the implementation. Opening the Vault showed
skills/ and mcp/ sitting among your folders, and because Vault
Sync commits that same directory, they went to your remote: on one
install a private docs repo had picked up opendray's own
skills/secretary/SKILL.md, and a git clean -fd there would have
deleted the gateway's skills.
New installs get ~/.opendray/vault for documents, ~/.opendray/ skills and ~/.opendray/mcp beside it, and a Vault repo holding
writing and nothing else. Existing installs are not moved: any
root with content still in the old place keeps being used, the
settings page prints where everything actually resolved, and says
plainly when it is still the shared layout. vault.notes and
vault.skills keep working; vault.root now means the documents
directory, and [skills].root is the new spelling. Machinery
directories that do sit inside the Vault are hidden from the doc
library and added to an opendray-managed .gitignore block, so
nothing new gets carried to your remote. Anything already committed
needs git rm --cached. opendray will not rewrite your repo.
Path resolution used to be reimplemented in four places (the gateway
plus each of opendray notes|skill|mcp) with different precedence in
each, so the CLI could read a different directory than the running
gateway. There is now exactly one resolver.
Verifying a git credential now checks that it can push, not just
read. A GitHub fine-grained token's Contents permission has three
levels: No access, Read-only, Read and write, and every check the
verification made passed identically for a Read-only one. So it went
green, the Vault pulled happily, and the first push came back
remote: Write access to repository not granted ... 403, the same
message a token with no Contents produces on a plain fetch, which
sends you looking at everything except the token.
Verification now also probes git's receive-pack advertisement:
literally the first request git push makes, asked with the same
credential over the same protocol, and a plain GET that changes
nothing. If it is refused, push will be refused. Because it is git's
wire protocol rather than a forge API, one probe covers GitHub, Gitea
and GitLab, and it cannot disagree with what git actually does. A
read-only credential now reports "CANNOT push (read-only)" and names
the setting to change. A forge answer that is neither a clear yes nor
a clear no is reported as nothing at all. Telling someone their
working token is read-only is the same mistake pointed the other way.
"Reset to remote" no longer silently destroys unpushed work. It
ran git reset --hard plus git clean -fd behind a confirmation
that named no quantity, survivable when the remote is ahead of you,
and not survivable when the remote is empty. A vault whose pushes
have all been failing is exactly that, and the two faults compose:
one operator lost 354 documents when a read-only token made every
push 403, the local commits piled up unpushed, a pull hit a rebase
conflict, and "reset to remote" looked like the way out of the
conflict.
The endpoint now counts what exists only locally (unpushed commits,
modified files, untracked files) and refuses with a 409 and that
breakdown unless the caller explicitly confirms. The dialog quotes
the numbers and names example files instead of asking "are you
sure?". And confirming is no longer final: opendray parks the
unpushed commits on an opendray-rescue/<timestamp> branch and
stashes the working tree (--include-untracked, since clean -fd
is what destroys untracked files and no ref can hold those) before
resetting, then names the rescue branch in the success toast. A tree
that is already level with its remote loses nothing and still resets
in one click. A confirmation that fires on no-ops is one people
learn to dismiss.
Git credentials are scoped per host and owner, so one forge can
hold several identities. One row per hostname assumed one identity
per forge, which breaks the moment you touch a personal repo and an
org repo on the same host: a fine-grained GitHub token is granted per
repository, so the token that reaches github.com/<you>/… generally
cannot reach github.com/<org>/…, and there was nowhere to put the
second one. Git host entries now take an optional Owner;
resolution prefers the owner-scoped credential and falls back to the
host-wide one, so existing setups keep working untouched. Vault sync
resolves the same way, which it previously could not: it only ever
looked up by hostname, and its auth panel now names the credential it
resolved to, saying plainly when the remote's owner has none of its
own and the host-wide one is standing in.
Git hosts is now the authority for HTTPS git auth. A session's
push went out with whatever the machine offered. Xcode ships
credential.helper = osxkeychain enabled, so a stale keychain entry
answered silently and failed with an error describing a token nobody
remembered configuring. Pushes now authenticate with the configured
credential, and inherited helpers are blanked for HTTPS remotes even
when opendray has nothing registered: failing as "no credentials"
beats quietly succeeding as an identity you never chose. SSH remotes
are untouched: the agent is a deliberate, visible configuration.
Disabling a git host entry now actually disables it. The toggle changed nothing: credential resolution returned disabled rows and every caller (vault sync, PR and issue listing, remote detection) used the token regardless, so an entry switched off kept authenticating. The check now lives in the resolver, which also makes disabling compose properly: turn off an owner-scoped entry and its host falls back to the host-wide one, exactly as if the row were absent.
Git host entries can be verified against the forge. A stored token
was a claim nobody checked, and the forges hide the mistake: a GitHub
fine-grained token keeps "which repositories" and "which permissions"
in separate sections of one form, with permissions defaulting to
none, so granting all repositories and stopping there produces a
token that authenticates perfectly and cannot read a single repo. Git
then reports Write access to repository not granted on a plain
fetch, naming the wrong permission on the wrong operation. Verify
now asks the forge who the token belongs to, warns when that differs
from the entry's owner, and optionally checks a specific repo, with a
hint that says where to look.
[[wiki-links]].README.md gets a control that opens
it, so a directory can say what lives in it. Templates render
server-side (the title comes from the filename, the date from the
clock) so a doc started on the phone and one started on the web come
out identical rather than drifting. Dropping _templates/<id>.md in
the vault overrides a built-in or adds a new one, so a project can
change the shape of its docs without a gateway release./ with -, so
features/canvas.md became features-canvas.md and a folder could not
be created from the UI at all, while the backend had stored nested
paths the whole time and the Notes page already rendered them as a tree.
The project-docs lane now renders that tree (rooted at the project, with
a Recent toggle for "the one I just edited"), typing a path with slashes
files a doc in a folder, and a new move/rename repoints the
[[wiki-links]] that pointed at the old path. Without that, filing a
doc away silently stranded every reference to it, which is why nobody
reorganised. Web and mobile both; the rewrite skips code blocks, so a
fenced example of the syntax is never edited.The Vault stops being a flat pile of markdown. It holds a folder
structure you maintain, files each project under its own name, and
renders HTML documents as carefully as it renders markdown, from an
in-memory string in a sandboxed frame, with scripts off, because a
document that arrived by git pull is not your own writing.
The phone catches up with the browser. It can see whether your edits
reached the remote, browse by tag, follow [[wiki links]], show what
links back, and start today's note, instead of being somewhere you
could only read what you had written elsewhere. The admin UI itself
now fits a phone screen, which it never did.
The admin UI works on a phone and a tablet. It was built for a desktop three-pane layout, which on a phone meant a horizontal scrollbar and two columns you could not reach. Every page now has a narrow form: below 1024px the third pane becomes a slide-over, below 768px the list becomes the page until you pick something and the navigation tree follows it into a drawer. Tables, spacing and the topbar were sized to match.
The drawer is one component rather than a pattern re-implemented per page, and it is reachable from a labelled control in the header, never from an edge handle alone, which is unreachable in practice.
The Vault syncs from the phone. Repository status, a manual commit / push / pull, and the auto-sync settings are all on the phone now; previously the mobile app could edit documents but not see whether they had reached the remote, so the answer to "did that save get anywhere" was only available on a desktop.
The mobile Vault has the knowledge layer the web has had. It
could create, read, edit, rename and delete a document, but nothing
that makes a vault more than a folder. It now browses by #tag,
shows a document's backlinks and its tags, follows [[wiki links]],
completes them while you type, jumps by heading, and opens today's
daily note from the same template the web writes.
Two constraints worth knowing. The preview runs with JavaScript off,
a document can arrive by git pull, so it is not the operator's
own writing by default, which means wiki-links are anchors on a
private scheme intercepted at navigation, and a heading jump lands
in the source view rather than scrolling the rendered page. And the
path rules are Unicode-aware on mobile: the web's ASCII-only
sanitiser turns 笔记.md into --.md.
The Vault stores and renders HTML documents, not just markdown.
Project documentation increasingly ships as HTML, exported from
Notion or Word, generated by typedoc, asciidoc or Sphinx, and a doc
library that cannot open those is a markdown library. .html /
.htm are now first-class alongside .md: listed in the tree,
creatable, movable, and rendered.
HTML is never served from opendray's own origin. There is no
endpoint returning a document as text/html; that would execute it
same-origin with the admin session, so a file pulled from a git
remote could take over the account. The body travels as a string
through the existing JSON read and is rendered from memory in a
sandboxed frame with an opaque origin. Scripts are off by default,
exported documentation is static markup and renders identically
without them, with a per-document opt-in for pages that genuinely
run code, remembered locally per file.
Mobile gains a rendered view for the first time. It previously showed only a document's source, which HTML would have made unreadable, so preview covers both kinds: markdown is converted and rendered through the same webview rather than drawn as native widgets, so the two formats look like one product.
Deliberately not symmetric: [[wiki links]] are still scanned and
rewritten in markdown only. An HTML document can be the target of
one, but its own <a href> links are left alone. Rewriting those on
a move is a different job, with relative paths, anchors and assets to
get wrong. Auto-derived paths (daily, project, personal notes) stay
markdown. And when a filename and a template disagree, guide.html
from the markdown Blank template, the filename wins, since the
operator chose the extension and the template was probably a default.
In-document links work: clicking a table-of-contents entry scrolls to that heading, and only links that actually leave the document open a new tab. This needed more than it sounds like. See Fixed below.
Vault auto-sync settings could not be saved. Whatever interval you set, the dialog snapped back to "every 10 minutes": the form refetched every 8 seconds and overwrote the draft mid-edit, so Save was never enabled and no setting had ever reached the database. The interval is now a free-text Go duration with presets, and the gateway rejects an unparseable one with a message instead of quietly substituting a default.
The mobile Vault's rename and delete were unfindable. They existed, behind a long press on a row, with nothing on screen saying so, reported in testing as missing outright, which for a hidden gesture amounts to the same thing. Each row now carries a visible actions button in place of a chevron that did nothing the row did not already do, and the open editor offers rename and delete where the web has always had them.
Creating guide.html on the phone produced guide.html.md. The
new-document path appended .md unconditionally, so an HTML
document could not be created from mobile at all.
Six hints rendered <prefix> as literal text instead of
<prefix>.
A table-of-contents link no longer loses the document. In the web
viewer, clicking an in-document #anchor navigated the frame away
and rendered the opendray app inside the document view. A srcdoc
frame's document URL is about:srcdoc while its base URL is
inherited from the parent page, so href="#install" resolved against
the gateway and became a real navigation. The base is now pinned to
the frame's own document, which also stops a pulled document's
relative URLs from resolving against the gateway's origin.
The first attempt at "external links should open a new tab" used
<base target="_blank">, which retargets every link, so a
contents entry opened a blank tab instead of scrolling. Anchors are
now rewritten individually: fragment links are left in place, an
author's explicit target is respected, and only links leaving the
document get target="_blank". Generated documentation is precisely
where this mattered: Sphinx, typedoc, asciidoc and Notion exports
all ship a fragment contents list, and markdown footnotes render as
fragment links too.
Mobile deliberately does not pin a base: loadData makes the
document's own URL and its base both about:blank, so fragments
already resolve in place, and pinning would have introduced the bug
rather than fixed it.
CI no longer runs the whole suite twice on every branch commit.
A feat/** branch with an open PR matched both the push and
pull_request triggers, and the concurrency group keys on the ref,
which differs between the two events, so neither run cancelled the
other. push is now main-only; pull_request already covered branch
work.
A project's documents are filed under the project's name. The
Vault is a project documentation library, and it filed every project
under projects/, a folder naming what the whole library already
is. One level of nesting that told the reader nothing, on every path,
in every listing, and at the top of the git repository the Vault syncs
to. The operator's own notes were split off further still: agent docs
at projects/<name>/…, the human scratchpad at personal/<name>.md,
the same project's material in two distant places, sorted by who
wrote it rather than by what it is about.
New vaults now use <name>/… with that project's personal.md
inside it. Existing vaults are not rearranged. The layout is
decided once, at first start, and written into config.toml as
[vault] layout. Recording it is the point: the alternative, work
the shape out from what is on disk each time, is the bug that put
one install's entire document library behind a notes/ directory,
because a probe asking "does this folder have content?" changes its
answer the moment someone puts content there.
The doc library offers the conversion, on web and on mobile, to any vault that still nests projects. A migration that ships only as a CLI command is one that only the people who wrote it ever run. Every existing vault would have stayed nested with its owner never learning there was a choice. The offer is dismissible and never appears for a vault that is already flat or has nothing to move, and the gateway says the same thing once in the startup log for anyone who never opens the UI.
Nothing moves without a preview. Web and mobile both run the
migration as a dry run and show the real list, including what it
refuses to touch and why, before anything is renamed.
opendray notes flatten does the same from a terminal. It defaults
to a dry run, drives the same rename the UI uses so [[wiki links]]
are repointed as it goes, repoints per-cwd project overrides, and
never overwrites: a destination that already exists is reported
and skipped, leaving both copies for you to reconcile.
Where a project's notes live is now answered by the gateway,
/notes/info reports the layout and the project mapping carries
personal_path, instead of being re-derived by web, mobile and the
CLI. Three implementations guessing is three chances to disagree, and
the CLI's notes project was already guessing wrong.
Reserved names step aside rather than collide: a project called
daily files under daily-docs, since daily/YYYY-MM-DD.md belongs
to the whole vault. _- and .-prefixed names are reserved too.
The conversion records the resulting layout itself and applies it to
the running gateway, rather than leaving both to the next restart.
Otherwise a converted vault keeps deriving projects/<name> and
personal/<name>.md against the directories it just emptied. The
directories the migration empties are removed; one still holding
anything is left exactly where it is.
Documents save when you say so, not on a timer. The editor wrote after every pause in typing, so a long document rewrote itself to disk over and over while it was being worked on. Saving is now a button, or ⌘S / Ctrl+S, on web and on mobile. Leaving a document still flushes unsaved text, that safety net costs nothing while typing, and closing the browser tab with unsaved work asks first.
Autosave was not the whole story behind the typing lag, so two things that were: the editor streamed every keystroke to the page, which re-rendered the whole vault tree and recomputed the outline per character; and the tree re-rendered along with it. The stream is now paced to what a sidebar can use, and the tree only re-renders when the note list or the selection changes.
Line numbers in the source view. On by default for HTML, off for markdown, and toggleable either way. Numbers turn wrapping off, the way every code editor does it: a wrapped line covers several rows, so a gutter counting 1..N drifts on exactly the long lines an HTML document is full of.
The doc library can rename and delete a document. It could create
and edit one, and that was all. A daily note, or anything not bound
to a project, could be made but never moved or removed from the one
surface that lists everything. Rename goes through the move endpoint,
so the [[wiki links]] pointing at the old path are repointed rather
than left dangling, and a partial rewrite is reported instead of
being folded into a plain "renamed". Mobile gains rename too; it
already had delete.
"New" and "Today" are reachable once a document is open. They
were in the header the whole time, pushed off the right edge by the
vault path: a truncate element in a flex row still needs
min-width: 0, or it refuses to shrink below its content. Anyone
with a long vault path could only reach the two actions from the
empty state, which is exactly when they had no documents to leave.
A folder holding the selected document can be collapsed again. The tree re-applied "open the ancestors of the selection" on every render, so the collapse landed and the next render undid it, and "Collapse all" left that one branch open. Revealing the selection is a response to the selection changing, not a rule about what must stay open, so it now runs once per selection.
opendray notes accepts flags after the subcommand. Go's flag
package stops parsing at the first non-flag argument, so notes flatten --apply left apply false, performed a dry run, and advised
re-running with --apply, which is what had just been typed.
notes list --prefix=x silently listed everything for the same
reason. Both orders now mean the same thing.
The Vault is your documents. Agent skills and the MCP registry moved
out. One root held three tenants with nothing in common: the
operator's markdown, the skills opendray injects at spawn, and the MCP
registry, which is why the settings page could only describe it as
"notes, skills and git-versioned root", a sentence that parses only if
you already know the implementation. Opening the Vault showed
skills/ and mcp/ sitting among your folders, and because Vault
Sync commits that same directory, they went to your remote: on one
install a private docs repo had picked up opendray's own
skills/secretary/SKILL.md, and a git clean -fd there would have
deleted the gateway's skills.
New installs get ~/.opendray/vault for documents, ~/.opendray/ skills and ~/.opendray/mcp beside it, and a Vault repo holding
writing and nothing else. Existing installs are not moved: any
root with content still in the old place keeps being used, the
settings page prints where everything actually resolved, and says
plainly when it is still the shared layout. vault.notes and
vault.skills keep working; vault.root now means the documents
directory, and [skills].root is the new spelling. Machinery
directories that do sit inside the Vault are hidden from the doc
library and added to an opendray-managed .gitignore block, so
nothing new gets carried to your remote. Anything already committed
needs git rm --cached. opendray will not rewrite your repo.
Path resolution used to be reimplemented in four places (the gateway
plus each of opendray notes|skill|mcp) with different precedence in
each, so the CLI could read a different directory than the running
gateway. There is now exactly one resolver.
Verifying a git credential now checks that it can push, not just
read. A GitHub fine-grained token's Contents permission has three
levels: No access, Read-only, Read and write, and every check the
verification made passed identically for a Read-only one. So it went
green, the Vault pulled happily, and the first push came back
remote: Write access to repository not granted ... 403, the same
message a token with no Contents produces on a plain fetch, which
sends you looking at everything except the token.
Verification now also probes git's receive-pack advertisement:
literally the first request git push makes, asked with the same
credential over the same protocol, and a plain GET that changes
nothing. If it is refused, push will be refused. Because it is git's
wire protocol rather than a forge API, one probe covers GitHub, Gitea
and GitLab, and it cannot disagree with what git actually does. A
read-only credential now reports "CANNOT push (read-only)" and names
the setting to change. A forge answer that is neither a clear yes nor
a clear no is reported as nothing at all. Telling someone their
working token is read-only is the same mistake pointed the other way.
"Reset to remote" no longer silently destroys unpushed work. It
ran git reset --hard plus git clean -fd behind a confirmation
that named no quantity, survivable when the remote is ahead of you,
and not survivable when the remote is empty. A vault whose pushes
have all been failing is exactly that, and the two faults compose:
one operator lost 354 documents when a read-only token made every
push 403, the local commits piled up unpushed, a pull hit a rebase
conflict, and "reset to remote" looked like the way out of the
conflict.
The endpoint now counts what exists only locally (unpushed commits,
modified files, untracked files) and refuses with a 409 and that
breakdown unless the caller explicitly confirms. The dialog quotes
the numbers and names example files instead of asking "are you
sure?". And confirming is no longer final: opendray parks the
unpushed commits on an opendray-rescue/<timestamp> branch and
stashes the working tree (--include-untracked, since clean -fd
is what destroys untracked files and no ref can hold those) before
resetting, then names the rescue branch in the success toast. A tree
that is already level with its remote loses nothing and still resets
in one click. A confirmation that fires on no-ops is one people
learn to dismiss.
Git credentials are scoped per host and owner, so one forge can
hold several identities. One row per hostname assumed one identity
per forge, which breaks the moment you touch a personal repo and an
org repo on the same host: a fine-grained GitHub token is granted per
repository, so the token that reaches github.com/<you>/… generally
cannot reach github.com/<org>/…, and there was nowhere to put the
second one. Git host entries now take an optional Owner;
resolution prefers the owner-scoped credential and falls back to the
host-wide one, so existing setups keep working untouched. Vault sync
resolves the same way, which it previously could not: it only ever
looked up by hostname, and its auth panel now names the credential it
resolved to, saying plainly when the remote's owner has none of its
own and the host-wide one is standing in.
Git hosts is now the authority for HTTPS git auth. A session's
push went out with whatever the machine offered. Xcode ships
credential.helper = osxkeychain enabled, so a stale keychain entry
answered silently and failed with an error describing a token nobody
remembered configuring. Pushes now authenticate with the configured
credential, and inherited helpers are blanked for HTTPS remotes even
when opendray has nothing registered: failing as "no credentials"
beats quietly succeeding as an identity you never chose. SSH remotes
are untouched: the agent is a deliberate, visible configuration.
Disabling a git host entry now actually disables it. The toggle changed nothing: credential resolution returned disabled rows and every caller (vault sync, PR and issue listing, remote detection) used the token regardless, so an entry switched off kept authenticating. The check now lives in the resolver, which also makes disabling compose properly: turn off an owner-scoped entry and its host falls back to the host-wide one, exactly as if the row were absent.
Git host entries can be verified against the forge. A stored token
was a claim nobody checked, and the forges hide the mistake: a GitHub
fine-grained token keeps "which repositories" and "which permissions"
in separate sections of one form, with permissions defaulting to
none, so granting all repositories and stopping there produces a
token that authenticates perfectly and cannot read a single repo. Git
then reports Write access to repository not granted on a plain
fetch, naming the wrong permission on the wrong operation. Verify
now asks the forge who the token belongs to, warns when that differs
from the entry's owner, and optionally checks a specific repo, with a
hint that says where to look.
[[wiki-links]].README.md gets a control that opens
it, so a directory can say what lives in it. Templates render
server-side (the title comes from the filename, the date from the
clock) so a doc started on the phone and one started on the web come
out identical rather than drifting. Dropping _templates/<id>.md in
the vault overrides a built-in or adds a new one, so a project can
change the shape of its docs without a gateway release./ with -, so
features/canvas.md became features-canvas.md and a folder could not
be created from the UI at all, while the backend had stored nested
paths the whole time and the Notes page already rendered them as a tree.
The project-docs lane now renders that tree (rooted at the project, with
a Recent toggle for "the one I just edited"), typing a path with slashes
files a doc in a folder, and a new move/rename repoints the
[[wiki-links]] that pointed at the old path. Without that, filing a
doc away silently stranded every reference to it, which is why nobody
reorganised. Web and mobile both; the rewrite skips code blocks, so a
fenced example of the syntax is never edited.Your phone and your browser can reach the gateway again while its Mac sits idle. A machine that falls asleep takes its network with it, so the gateway simply stops answering, a failure that looks exactly like a flaky gat
Host power is now a setting you can see, on web and mobile. The
choice between "stay reachable" and "let the machine sleep" was only
reachable by hand-editing config.toml, which meant nobody found it.
Server settings gain a Host power section listing all four modes
with what each one costs (web renders them as annotated cards, mobile
as a labelled picker) plus a note that a deliberate sleep is never
blocked and that non-macOS hosts ignore the setting. Translated across
English, 中文 and Español.
Wake-on-demand: the host may sleep, and phone/web traffic wakes it.
[host] prevent_idle_sleep = "on_demand" lets the gateway's Mac sleep
whenever things are quiet instead of being held awake around the clock.
Incoming traffic dark-wakes the machine (macOS "Wake for network
access"); the gateway then seizes a power assertion the moment a
request, live stream or mid-turn session shows up, extends the wake for
as long as serving actually takes, and lets go after a two-minute
linger so the host can sleep again. Running sessions count as activity,
so a turn started from a phone survives the phone locking. On-demand
warns at startup if womp is disabled, since without it a sleeping
host can't be woken remotely at all. The existing "ac" (default),
"always" and "off" behaviours are unchanged.
Saving invalid server settings no longer bricks the next restart.
PUT /admin/settings wrote whatever it was given. Since the loader
validates on every startup, a bad value (an unparseable duration, an
unknown mode) saved cleanly and then stopped the gateway from coming
back up, with nothing naming the offending field. The write path now
validates exactly what is about to hit disk and answers 400 instead,
leaving the stored config untouched.
The gateway keeps its host awake, so phone and web stay connected.
A Mac left to itself idle-sleeps, and a sleeping host takes its network
with it: the gateway stops answering, a remote Postgres becomes
unreachable, and every phone or web request times out until someone
physically wakes the machine. Incoming traffic does dark-wake the host,
but the window is short and a user-session LaunchAgent isn't reliably
scheduled inside it, so the request has already timed out, which reads
as "opendray is flaky" and sends people reading gateway code instead of
pmset -g log. opendray now holds a power assertion for as long as it
serves. Default is wall-power only, so a laptop on battery still sleeps
normally; [host] prevent_idle_sleep takes "always" or "off". A
deliberate sleep (lid close, Apple menu → Sleep) is never blocked, and
the assertion is tied to the gateway's pid so even a SIGKILL can't
strand the host awake. macOS only; accepted and ignored elsewhere.
Your phone and your browser can reach the gateway again while its Mac
sits idle. A machine that falls asleep takes its network with it, so the
gateway simply stops answering, a failure that looks exactly like a
flaky gateway from every surface that talks to it, and sends people
reading gateway code instead of pmset -g log. opendray now holds the
host awake while it serves, and lets you choose how far to take that.
Host power is now a setting you can see, on web and mobile. The
choice between "stay reachable" and "let the machine sleep" was only
reachable by hand-editing config.toml, which meant nobody found it.
Server settings gain a Host power section listing all four modes
with what each one costs (web renders them as annotated cards, mobile
as a labelled picker) plus a note that a deliberate sleep is never
blocked and that non-macOS hosts ignore the setting. Translated across
English, 中文 and Español.
Wake-on-demand: the host may sleep, and phone/web traffic wakes it.
[host] prevent_idle_sleep = "on_demand" lets the gateway's Mac sleep
whenever things are quiet instead of being held awake around the clock.
Incoming traffic dark-wakes the machine (macOS "Wake for network
access"); the gateway then seizes a power assertion the moment a
request, live stream or mid-turn session shows up, extends the wake for
as long as serving actually takes, and lets go after a two-minute
linger so the host can sleep again. Running sessions count as activity,
so a turn started from a phone survives the phone locking. On-demand
warns at startup if womp is disabled, since without it a sleeping
host can't be woken remotely at all. The existing "ac" (default),
"always" and "off" behaviours are unchanged.
Saving invalid server settings no longer bricks the next restart.
PUT /admin/settings wrote whatever it was given. Since the loader
validates on every startup, a bad value (an unparseable duration, an
unknown mode) saved cleanly and then stopped the gateway from coming
back up, with nothing naming the offending field. The write path now
validates exactly what is about to hit disk and answers 400 instead,
leaving the stored config untouched.
The gateway keeps its host awake, so phone and web stay connected.
A Mac left to itself idle-sleeps, and a sleeping host takes its network
with it: the gateway stops answering, a remote Postgres becomes
unreachable, and every phone or web request times out until someone
physically wakes the machine. Incoming traffic does dark-wake the host,
but the window is short and a user-session LaunchAgent isn't reliably
scheduled inside it, so the request has already timed out, which reads
as "opendray is flaky" and sends people reading gateway code instead of
pmset -g log. opendray now holds a power assertion for as long as it
serves. Default is wall-power only, so a laptop on battery still sleeps
normally; [host] prevent_idle_sleep takes "always" or "off". A
deliberate sleep (lid close, Apple menu → Sleep) is never blocked, and
the assertion is tied to the gateway's pid so even a SIGKILL can't
strand the host awake. macOS only; accepted and ignored elsewhere.
The Canvas: the agent renders a page you can actually see, mark up and iterate on, instead of describing a screen in prose and hoping it matches.
canvas_render and it appears in
the operator's panel, where they pin a point or drag a region and send
those marks back as feedback. Marking reports the real element selector
and markup (and for a region, the components inside the frame) so the
agent iterates on the DOM instead of guessing from coordinates. Not only
UI mocks: ui, flow (flowchart), mindmap, graph (relationship
diagram) and doc (a formatted spec page), with diagrams authored as
inline SVG so they need no external assets and stay markable node by node.canvas_context, and a slug-less
canvas_render targets the focused canvas rather than creating a
duplicate.opendray status is now a systemctl-style health dashboard. On macOS
the command used to dump ~80 lines of raw launchctl print internals.
A live process says nothing about whether the gateway is serving HTTP or
can reach its database. The default output is now a compact checklist:
process, HTTP health + uptime, database reachability, listen address,
restart count (with the last non-zero exit surfaced as a hint) and the
config path. The raw launchd/systemd dump moved behind --raw.primary in a design system means the BRAND colour, and the gateway
now says so. shadcn/ui and the Tailwind templates built on it use
--primary for a near-black or near-white ink and keep the brand hue in
--accent, so an agent mapping token names one-to-one produced a palette
of greys with no brand colour in it. The contract is now stated wherever
an agent reads it (the extract prompt, the canvas_design tool schema,
the token docs and the catalog guidance) and saving a palette whose every
colour resolves to a grey returns an advisory warning saying what probably
went wrong.getComputedStyle().color only yields rgb(…) for
legacy sRGB notations (a modern colour function round-trips unchanged in
both WebKit and Chromium) so an all-oklch project got a grid of identical
grey fallbacks on the web, and mobile, which parsed nothing but #rrggbb,
got no swatches at all. The gateway now resolves hex / rgb / hsl / oklch /
oklab centrally and serves the result, which also lets the hex-only OS
colour picker edit an oklch theme without converting it one field at a
time.x/y
field meant frame percentages from the web and document percentages from
mobile.?1000/1002/1003/1006 at
startup yet does nothing on button-64/65). The web terminal now recognises
these wheel-ignoring providers and sends arrow keys for both wheel and
one-finger touch scroll, scoped by providerId so Claude/Codex/Antigravity
(which genuinely consume the wheel) are untouched./fs/download, /fs/zip
and /fs/upload. The root-scoped filesystem endpoints now validate the
resolved path with filepath.IsLocal and refuse any residual ..
sequence in the canonical path (which also means oddly-named entries like
notes..md are rejected inside these endpoints). Resolves all seven open
CodeQL go/path-injection alerts plus one allocation-size-overflow
finding.@babel/core 7.29.7. Lockfile-only; no manifest changes.The Canvas: the agent renders a page you can actually see, mark up and iterate on, instead of describing a screen in prose and hoping it matches.
canvas_render and it appears in
the operator's panel, where they pin a point or drag a region and send
those marks back as feedback. Marking reports the real element selector
and markup (and for a region, the components inside the frame) so the
agent iterates on the DOM instead of guessing from coordinates. Not only
UI mocks: ui, flow (flowchart), mindmap, graph (relationship
diagram) and doc (a formatted spec page), with diagrams authored as
inline SVG so they need no external assets and stay markable node by node.canvas_context, and a slug-less
canvas_render targets the focused canvas rather than creating a
duplicate.opendray status is now a systemctl-style health dashboard. On macOS
the command used to dump ~80 lines of raw launchctl print internals.
A live process says nothing about whether the gateway is serving HTTP or
can reach its database. The default output is now a compact checklist:
process, HTTP health + uptime, database reachability, listen address,
restart count (with the last non-zero exit surfaced as a hint) and the
config path. The raw launchd/systemd dump moved behind --raw.primary in a design system means the BRAND colour, and the gateway
now says so. shadcn/ui and the Tailwind templates built on it use
--primary for a near-black or near-white ink and keep the brand hue in
--accent, so an agent mapping token names one-to-one produced a palette
of greys with no brand colour in it. The contract is now stated wherever
an agent reads it (the extract prompt, the canvas_design tool schema,
the token docs and the catalog guidance) and saving a palette whose every
colour resolves to a grey returns an advisory warning saying what probably
went wrong.getComputedStyle().color only yields rgb(…) for
legacy sRGB notations (a modern colour function round-trips unchanged in
both WebKit and Chromium) so an all-oklch project got a grid of identical
grey fallbacks on the web, and mobile, which parsed nothing but #rrggbb,
got no swatches at all. The gateway now resolves hex / rgb / hsl / oklch /
oklab centrally and serves the result, which also lets the hex-only OS
colour picker edit an oklch theme without converting it one field at a
time.x/y
field meant frame percentages from the web and document percentages from
mobile.?1000/1002/1003/1006 at
startup yet does nothing on button-64/65). The web terminal now recognises
these wheel-ignoring providers and sends arrow keys for both wheel and
one-finger touch scroll, scoped by providerId so Claude/Codex/Antigravity
(which genuinely consume the wheel) are untouched./fs/download, /fs/zip
and /fs/upload. The root-scoped filesystem endpoints now validate the
resolved path with filepath.IsLocal and refuse any residual ..
sequence in the canonical path (which also means oddly-named entries like
notes..md are rejected inside these endpoints). Resolves all seven open
CodeQL go/path-injection alerts plus one allocation-size-overflow
finding.@babel/core 7.29.7. Lockfile-only; no manifest changes.Round Table members gain live, grounded memory, the mobile session screen is rebuilt around a tool dock, and the operator takes ownership of the four global knowledge pages so their curation finally sticks.
opendray-memory MCP, so
members ground their claims in the real store instead of guessing. A new
read-only mode (OPENDRAY_MEMORY_READONLY=1) exposes only the search / read
tools and refuses every write server-side (safe even with tool permissions
open) and the table prompt flips to "ground your claims via these tools".
A single per-provider attach point (catalog.AttachMemoryMCP) handles the
antigravity / codex quirks and is reusable for any spawn.kb_infrastructure / kb_conventions / kb_lessons / kb_reusable was
hardcoded in the drafter's prompts, so every consolidation sweep
re-manufactured the fixed fat skeleton and operator curation never stuck.
The drafter now honours each page's blueprint maintainer_mode (human →
the operator owns it outright, never drafted) and, for AI-maintained pages,
edits the operator's current structure in place (folding in only new,
on-topic evidence) instead of regenerating from a template. A per-page
prompt_hint lets the operator steer form and scope without a code change.
Set both via PUT /blueprint/{slug} (cwd __global__); a web/mobile toggle
is a follow-up. No migration. The four pages default to the previous
behaviour.sessions.dock.more / sessions.tools.* strings ship across en / zh /
es at 100% parity.updated_by='agent', flipping HumanLocked off). An
approval is an operator decision, so the result now stays operator-authored
and locked; and rejecting a proposal didn't stick: the drafter re-generated
and re-filed the identical refresh every consolidation cycle. The drafter now
skips a locked page whose feedstock signature matches an already-rejected
proposal, so a rejection holds instead of nagging forever.Round Table members gain live, grounded memory, the mobile session screen is rebuilt around a tool dock, and the operator takes ownership of the four global knowledge pages so their curation finally sticks.
opendray-memory MCP, so
members ground their claims in the real store instead of guessing. A new
read-only mode (OPENDRAY_MEMORY_READONLY=1) exposes only the search / read
tools and refuses every write server-side (safe even with tool permissions
open) and the table prompt flips to "ground your claims via these tools".
A single per-provider attach point (catalog.AttachMemoryMCP) handles the
antigravity / codex quirks and is reusable for any spawn.kb_infrastructure / kb_conventions / kb_lessons / kb_reusable was
hardcoded in the drafter's prompts, so every consolidation sweep
re-manufactured the fixed fat skeleton and operator curation never stuck.
The drafter now honours each page's blueprint maintainer_mode (human →
the operator owns it outright, never drafted) and, for AI-maintained pages,
edits the operator's current structure in place (folding in only new,
on-topic evidence) instead of regenerating from a template. A per-page
prompt_hint lets the operator steer form and scope without a code change.
Set both via PUT /blueprint/{slug} (cwd __global__); a web/mobile toggle
is a follow-up. No migration. The four pages default to the previous
behaviour.sessions.dock.more / sessions.tools.* strings ship across en / zh /
es at 100% parity.updated_by='agent', flipping HumanLocked off). An
approval is an operator decision, so the result now stays operator-authored
and locked; and rejecting a proposal didn't stick: the drafter re-generated
and re-filed the identical refresh every consolidation cycle. The drafter now
skips a locked page whose feedstock signature matches an already-rejected
proposal, so a rejection holds instead of nagging forever.Grok reaches parity with the other cloud agents, the Cortex knowledge base grows a real management surface, and the mobile knowledge experience is rebuilt for phones.
doc_read, cross-layer recall) just like Claude / Codex
/ Antigravity. Its spawn folder is marked trusted so grok actually starts
the injected memory server instead of silently skipping it. Grok and OpenCode
are also selectable in Discuss with AI and as Memory Worker agent
providers, and creating a grok session now offers the Bypass permissions /
YOLO toggle (--always-approve) the other agents already had.kb_* page's title,
one-line description, nature (foundational / emergent) and inject flag were
locked in at creation; they are now editable in place (web + mobile) on every
page except the classic four, including seeded pages like Integrations, so
you can flip a page between full-inject and on-demand retrieval.@mentionable
on the next turn with the full thread as context; a removed one stops
replying while its past messages stay.kb_* docs; the KB tab is now a grouped, searchable list (Foundational /
Emergent) that grows gracefully, and tapping a page opens a full-screen
reader/editor with its actions in an AppBar overflow menu. New page and the
Librarian move onto a FAB.<cwd>/.grok/config.toml, but grok refuses to start
repo-local MCP servers in an untrusted folder as a supply-chain guard, so the
server was configured but never started. opendray now trusts the operator's
own spawn folder (--trust), matching the other CLIs.Grok reaches parity with the other cloud agents, the Cortex knowledge base grows a real management surface, and the mobile knowledge experience is rebuilt for phones.
doc_read, cross-layer recall) just like Claude / Codex
/ Antigravity. Its spawn folder is marked trusted so grok actually starts
the injected memory server instead of silently skipping it. Grok and OpenCode
are also selectable in Discuss with AI and as Memory Worker agent
providers, and creating a grok session now offers the Bypass permissions /
YOLO toggle (--always-approve) the other agents already had.kb_* page's title,
one-line description, nature (foundational / emergent) and inject flag were
locked in at creation; they are now editable in place (web + mobile) on every
page except the classic four, including seeded pages like Integrations, so
you can flip a page between full-inject and on-demand retrieval.@mentionable
on the next turn with the full thread as context; a removed one stops
replying while its past messages stay.kb_* docs; the KB tab is now a grouped, searchable list (Foundational /
Emergent) that grows gracefully, and tapping a page opens a full-screen
reader/editor with its actions in an AppBar overflow menu. New page and the
Librarian move onto a FAB.<cwd>/.grok/config.toml, but grok refuses to start
repo-local MCP servers in an untrusted folder as a supply-chain guard, so the
server was configured but never started. opendray now trusts the operator's
own spawn folder (--trust), matching the other CLIs.Added Round Table: a cross vendor AI group chat (experimental). Seat several providers (Claude / Codex / Antigravity / Grok / OpenCode) plus the operator in one shared thread; @mention who should reply (or @all) and each
@all) and each member
answers in character after reading the whole conversation, so heterogeneous
foundation-model families react to each other in seat order. Summarize the
discussion on demand, or turn it into a role-assigned execution plan:
each step runs as a real session in a shared project (bind the project after
the fact if you started without one). Hand the whole thread off to a
working session to do the actual code changes. A chat can be closed and
reopened (close keeps the thread, just stops new messages). Available on
both the web admin and the mobile app, where Round Table gets its own
bottom-nav tab, per-agent bubble colours, and labelled action menus.
Fully self-contained and rollback-able (internal/roundtable/ROLLBACK.md).group-hover) or keyboard focus. Tailwind v4 gates group-hover behind
@media (hover: hover), so on a touch device (which can neither hover nor
focus a row) the icon stayed at opacity-0 and was impossible to tap. It
now pins visible under @media (hover: none), so touch users get a
permanently-shown download control while pointer users keep the clean
hover-reveal. (Follow-up to the v2.11.6 positioning fix, which addressed
where the icon sits but not whether it ever appears without a mouse.)[mcp_servers."…"] TOML
table and died with duplicate key at startup (before printing a byte, so
the session flipped straight to the read-only "[buffer unavailable]" view)
while Claude's map-based renderer silently dropped one of them. renderMCP
now rejects a duplicate name up front (for every provider, before any config
file is written), and the Plugins create/update endpoints return 409 when
a new or edited server would reuse a name already taken by a different id.
Grok's manifest gap and this collision are unrelated; a stray second Notion
entry sharing the name Notion API is what exposed it.grok manifest carried an empty npmPackage, and the whole update path is
npm-gated: CheckUpdate returned early (no latest version, no
"update available" flag) and Update hard-errored with "not updatable via
npm". Grok is published as @xai-official/grok (maintainer
[email protected]), so the manifest now names it.curl -fsSL https://x.ai/cli/install.sh | bash) drops
a symlink into the npm bin dir that npm does not own, and npm refuses to
clobber it: EEXIST: file already exists. Simply naming the package would
therefore have shipped a dashboard that advertises an update behind a button
that always fails. Update now preflights the bin path: an unmanaged
symlink is cleared so npm can take ownership (and the update output tells
the operator exactly which link was replaced), while a regular file is
never deleted: it is reported instead, mirroring the existing
ErrUpdatePrefixReadonly preflight. Grok's install note now recommends
npm install -g @xai-official/grok.@all) and each member
answers in character after reading the whole conversation, so heterogeneous
foundation-model families react to each other in seat order. Summarize the
discussion on demand, or turn it into a role-assigned execution plan:
each step runs as a real session in a shared project (bind the project after
the fact if you started without one). Hand the whole thread off to a
working session to do the actual code changes. A chat can be closed and
reopened (close keeps the thread, just stops new messages). Available on
both the web admin and the mobile app, where Round Table gets its own
bottom-nav tab, per-agent bubble colours, and labelled action menus.
Fully self-contained and rollback-able (internal/roundtable/ROLLBACK.md).group-hover) or keyboard focus. Tailwind v4 gates group-hover behind
@media (hover: hover), so on a touch device (which can neither hover nor
focus a row) the icon stayed at opacity-0 and was impossible to tap. It
now pins visible under @media (hover: none), so touch users get a
permanently-shown download control while pointer users keep the clean
hover-reveal. (Follow-up to the v2.11.6 positioning fix, which addressed
where the icon sits but not whether it ever appears without a mouse.)[mcp_servers."…"] TOML
table and died with duplicate key at startup (before printing a byte, so
the session flipped straight to the read-only "[buffer unavailable]" view)
while Claude's map-based renderer silently dropped one of them. renderMCP
now rejects a duplicate name up front (for every provider, before any config
file is written), and the Plugins create/update endpoints return 409 when
a new or edited server would reuse a name already taken by a different id.
Grok's manifest gap and this collision are unrelated; a stray second Notion
entry sharing the name Notion API is what exposed it.grok manifest carried an empty npmPackage, and the whole update path is
npm-gated: CheckUpdate returned early (no latest version, no
"update available" flag) and Update hard-errored with "not updatable via
npm". Grok is published as @xai-official/grok (maintainer
[email protected]), so the manifest now names it.curl -fsSL https://x.ai/cli/install.sh | bash) drops
a symlink into the npm bin dir that npm does not own, and npm refuses to
clobber it: EEXIST: file already exists. Simply naming the package would
therefore have shipped a dashboard that advertises an update behind a button
that always fails. Update now preflights the bin path: an unmanaged
symlink is cleared so npm can take ownership (and the update output tells
the operator exactly which link was replaced), while a regular file is
never deleted: it is reported instead, mirroring the existing
ErrUpdatePrefixReadonly preflight. Grok's install note now recommends
npm install -g @xai-official/grok.Fixed The download icon is reachable again in a deep or long file tree. The session inspector's Files tree renders inside a scroll area whose inner wrapper sizes to its content, so long filenames and deep nesting pushed
Operator polish, reconnect stability, memory improvements, and clearer docs.