The terminal store keyed its directories by trimming trailing slashes,
while the new sidebar activity indicator and the whole-server session
grouping keyed the same folder with the project-action normalizer, which
also rewrites backslashes. On Windows that split one project into two
namespaces: the sidebar reconciled server sessions under "C:/repo" while
the terminal panel and the project actions button worked under "C:\repo",
so the running-action indicator never lit up, adopted tabs were duplicated
and a stop recorded in one namespace did not guard reconciliation in the
other. The project actions button also read the store map directly with a
path normalized somewhere else, missing its own directory entry.
All terminal directory keys now come from normalizeTerminalDirectory in
lib/pathNormalization, and store reads go through getDirectoryState.
The 256 KiB budget for an oversized tool diff preview counts UTF-16 units, so
the cut can land between the two halves of an astral character - an emoji in a
patched string, for instance. The preview then ended in a lone surrogate that
renders as the replacement glyph.
Step back one unit when the boundary splits a pair.
The IME fix covered the diff/file comment input and the browser annotation
overlay, but two other places where a user writes a comment still acted on
the Enter that confirms an IME candidate: the chat quote comment in the text
selection menu, and the in-place comment editor on a composer context chip.
Confirming a candidate there attached or committed the half-typed reading and
closed the input. Escape, which abandons a candidate, had the same problem.
Both handlers now return early on a composing keystroke, and a single test
asserts the guard across every comment input so the next one added does not
quietly skip it.
The add-folder dedup normalized the drive letter to uppercase on one side and compared it with Uri.fsPath, which VS Code lowercases again. On Windows an already-open folder therefore never matched and was re-added. Uses pathsEqualWithNormalizedDriveLetter so both sides normalize.
The inline-comment feature added its manifest and runtime strings to English and French only, so Turkish users saw English command titles, thread labels and warnings. Adds the ten missing keys and a test that fails whenever a locale bundle drifts from the English key set or its placeholders.
isBunInstalled() spawned <BUN_INSTALL>/bin/bun, which does not exist on
Windows where the installer ships bun.exe. spawnSync adds no extension to
an explicit path, so the probe failed and the CLI quietly started the
server under Node even with Bun installed. The probe now names bun.exe
on win32.
Claude-Session: https://claude.ai/code/session_01VqV56Hez25hTxXH4ipJfzH
Child discovery now pages through the whole session list, so one pull can
take longer than the 15s watchdog interval. Two overlapping pulls each read
the store before either commits, so both appended the same newly found child
and the directory store ended up with duplicate rows. Keep one pull per
directory in flight, mirroring the status-poll guard.
The web server read XDG_CONFIG_HOME through a redundant typeof guard on a
value that is already string|undefined, which tripped the anti-slop lint on
its own new line and diverged from the VS Code helper. Both now read the
same way. The VS Code provider test also still hard-coded ~/.config/opencode,
so it silently stopped asserting whenever XDG_CONFIG_HOME was set; it now
uses the shared constant, like the web test already does.
dev-web-full.mjs is started by node and spawned bare `bun`, so on Windows
machines where npm exposes only bun.cmd both children died with ENOENT and
`dev:web:full` never started. It now uses the resolver dev-web-hmr.mjs got,
and hides the console window like the other Node-spawned children.
The web route accepts the OpenAI Chat Completions, OpenAI Responses and
Anthropic Messages adapters, but the module doc still described the write
path as OpenAI-compatible only. The VS Code doc already names all three.
The legacy `providers` block is deleted whole when its last entry migrates
to `provider`. Nothing covered the case where other legacy entries remain,
so a regression there would silently drop unrelated providers. Adds the
case to both the web server and VS Code parity suites.
A comment went to the active session tab, and when there was none it opened
a new one, even while the user was chatting in the sidebar. That is not how
the other capture flows behave. The comment now takes the same route as
Add to Context: the active session tab when one exists, otherwise the
sidebar, revealed if it is closed.
Claude-Session: https://claude.ai/code/session_01VqV56Hez25hTxXH4ipJfzH
A comment written with no session tab open opens one, and that panel knows
its directory long before it has loaded the session list and selected its
session. The webview filed the draft on the first snapshot that had a
directory, with the session key falling back to "draft" because no session
was current yet. The panel's composer reads the session's key, so the chip
never appeared, while the editor thread saw the draft in the store snapshot
and reported it attached.
A session panel now stamps its session on every comment it delivers, and
the webview waits until it actually shows that session before filing. The
sidebar files on its current session or an open new-session draft, and no
longer falls back to "draft" merely because nothing is selected yet. The
resolver is a pure module with tests.
Claude-Session: https://claude.ai/code/session_01VqV56Hez25hTxXH4ipJfzH
Both failures showed up as "Test timed out in 30000ms" followed by "Hook
timed out in 30000ms" on loaded PR runners, in files nobody had touched.
relay host-client: the scripted client dialed its WebSocket before awaiting
its WebCrypto setup and only then attached the open/message listeners. On a
loaded runner the loopback socket opened while key generation was still
queued on the threadpool, the open event found no listener, no hello was
sent, and the client waited forever. The socket is now dialed after the key
material is ready, in the same tick as the listeners. The fake relay also
terminates leftover sockets on stop so a failure is reported once, not
twice.
walkthrough routes: each test slept 20 ms and assumed the request had
reached the route by then. The tests now wait until the service has been
asked to generate one more time than before the request, and afterEach
closes idle keep-alive connections so server.close() cannot hang on a
response a failed test never received.
Claude-Session: https://claude.ai/code/session_01VqV56Hez25hTxXH4ipJfzH
OpenCode's command route accepts file parts only. The server queue was
attaching captured context as text parts to POST /session/:id/command, so
a slash command queued with a comment, quote, or PR diff was rejected with
400 and retried forever.
A command queued without context still takes the command route with its
files. One queued with context now takes the prompt route the way the
composer does: the command's template is expanded with its arguments, a
skill keeps its text and gets the explicit skill-invocation instruction,
and the context rides along.
Claude-Session: https://claude.ai/code/session_01VqV56Hez25hTxXH4ipJfzH
When the server delivered the last item of a session's queue it deleted
the session entry and broadcast the update with an empty directory. The
UI keys its projection by directory, could not build the key, and dropped
the event, so the delivered message stayed in the "Queued messages" card
until the next full hydration.
The server now remembers a session's directory beyond the emptying of its
queue so that broadcast names it. The UI additionally treats an empty
session without a directory as "this session's queue is done" for every
projection keyed under that session id, which covers clients talking to
an older server.
Claude-Session: https://claude.ai/code/session_01VqV56Hez25hTxXH4ipJfzH
Exports the release package list so release prep stages the same versioned files as the bump script.
Validates that CHANGELOG.md contains the target release section before preparing a release.
Updates the suggested release steps to commit only release files and push the tag explicitly.
Settings → Integrations offered install cards for the Claude Code and
Cursor provider plugins. They are gone: the section, its plugin catalog,
its own i18n module and tests, the settings search entries, the page
keywords, and the two sprite icons only it used. The page now holds the
built-in GitHub and Linear cards, so it is hidden in VS Code where neither
applies; its title and description live in the settings dictionaries.
Docs follow: the Integrations page in every locale now documents GitHub
(pointing at its own page) and Linear in full, the GitHub page names
Settings → Integrations as the place to connect and covers linking an
issue or PR to a message, and the Providers pages no longer promise Claude
or Cursor subscriptions.
Claude-Session: https://claude.ai/code/session_01HB9wdLQoZX2vfyDjwv6Rso
Queueing captured only the text and files. Context chips (inline comments,
terminal selections, browser annotations, PR comments and checks, quotes,
linked issue/PR/Linear references, pending synthetic parts) stayed in the
composer and only left with the next manual send, so a queued message the
server delivered went out without them and the chips rode an unrelated
message later.
A queued message now carries what the composer would have sent: the text
with its agent mention stripped and file mentions resolved into
attachments, the attached context as structured parts, and the skill
instruction derived from the text. The server delivers those parts in the
composer's order, the VS Code auto-send does the same, and editing a queued
message puts the chips and linked references back. A failed queue restores
the composer completely. Snapshots and broadcasts omit the captured
context like attachment payloads; a take returns it.
Claude-Session: https://claude.ai/code/session_01HB9wdLQoZX2vfyDjwv6Rso
OpenCode marks an MCP server failed when it does not come up at startup
or when a live connection drops, and never retries. A new managed plugin
reconnects those servers with a per-server backoff (1s doubling to a 30s
cap), wakes early on mcp.tools.changed, and stops on dispose. Only the
failed state is retried; disabled and auth states stay untouched.
The OPENCODE_CONFIG_CONTENT merge that agent-tool and system-prompt each
carried is now one shared helper so the three managed plugins compose.
Claude-Session: https://claude.ai/code/session_01VqV56Hez25hTxXH4ipJfzH
Messages queued while a session is busy used to live in the browser tab and
were sent by that tab once the session went idle, so closing the tab (or
losing the connection) stranded them. The web server now owns the queue:
it persists to <data-dir>/message-queue.json, watches session.status on the
global event hub, re-verifies idleness against OpenCode before sending, and
delivers the head of the queue via prompt_async (or /command for slash
commands) with the model, agent, variant, attachments, and agent mention
captured at queue time. Failed sends stay queued and retry with backoff; a
user abort holds delivery briefly; every change is broadcast so all clients
see one queue.
The shared UI store becomes a projection of the server queue outside VS
Code (hydrate on connect, apply broadcasts, optimistic mutations settled on
the server's copy, one-time upload of locally queued messages from older
builds). Edit / send-now take the full message back from the server. A
UI-driven auto-review run asks the server to hold that session's queue.
VS Code keeps its local queue and foreground auto-send.
Claude-Session: https://claude.ai/code/session_01HB9wdLQoZX2vfyDjwv6Rso
Startup warm-up, opening the switcher and the refresh button can all probe at
once, and a relay host working through tunnel retries takes an order of
magnitude longer than a loopback one — so a slow older run landed last and
replaced a fresh "ok" with its own stale "unreachable". Each host now records
which run owns its status; a status from the switch flow outranks any probe
still running for it.
Each relay attempt is also capped by what is left of the 15s budget rather than
the full per-request timeout, so an attempt started just under the deadline can
no longer run the whole 8s past it.
MCP status and skills were cleared on a switch but their in-flight requests
were not, so a response for the previous instance could still write itself over
the new one's — the same race the quota store already guards. Both now carry a
generation.
Usage also claimed an instance as loaded before it had answered, so a load that
failed on a cold or briefly unreachable instance was never attempted again, and
the previous instance's display mode and provider selection survived a switch —
which decided what the new instance was even asked for.
The switcher held reachability in component state and replaced the whole map at
the end of a probe run. It ran once per open before the config had loaded — with
Local as the only host — so that pass wiped every other instance's status and
each open started on "Checking", including for the instance the app was
connected to and actively talking to.
Statuses move to their own module: startup warms them so the switcher opens on
real values, a re-probe replaces each value in place as it lands rather than
blanking them first, and stale entries are dropped against the loaded config
instead of a partial host list. The connected instance never reads "Checking" —
the live connection already answers what the probe would ask.
The tunnel rejects everything waiting on its channel the moment one connect
attempt fails, even though it has already scheduled the next one with backoff.
That is right for app traffic, which retries for itself, but it made the
one-shot probe report a durable red "Unreachable" for a host that answers when
the user presses refresh a second later — a cold start is exactly when that
first attempt loses.
The probe now spans the tunnel's own reconnects within a 15s budget, and ends
immediately on a terminal tunnel state (auth failed, duplicate client, limit),
which waiting cannot resolve.
Linear and GitHub logins, quotas, MCP status, skills and agent memory are
served by whichever instance is connected, but each was cached globally or
by directory alone — which two instances can share. Switching instances left
the previous instance's answers on screen and its Linear login usable against
a runtime that has no Linear.
Reset them all through runtimeEndpointReset, each store guarding its in-flight
requests with a generation so a response for the previous instance cannot land
in the new one. The Linear team filter is now persisted per instance: a team
belongs to one workspace, so carrying it across filtered the new instance's
issue list down to nothing.
Usage also waits for the instance to report itself initialised before loading.
Providers report themselves as configured only once the instance can read their
credentials, so a fetch fired at mount answered "nothing configured" for every
provider and cached it — which is why Usage stayed missing from the work-status
panel until Settings -> Usage forced a fresh fetch.
Accept browser https origins when TLS terminates before an HTTP proxy hop
Honor forwarded external host while rejecting mismatched origins
Add tests for proxy and host matching behavior
Prevents the project label area from shifting when actions appear
Matches spacing behavior with the collapse-toggle branch
Keeps hover and focus states aligned
Keeps idle readers from snapping back to the end while rows re-wrap
Releases the pin instead of scrolling if a resize moves an idle view off the end
Re-asserts the live edge after resize settle for active streaming sessions
Review comments, quotes, terminal selections and annotations were consumed
at send and never put back, so reverting pulled the message into the
composer without the context it was sent with.
Claude-Session: https://claude.ai/code/session_01TwLFeTfBnWdvbg9XyezZQx
Clarify chat session switching is visually stable
Note GitHub account connection moved to Settings → Integrations
Update VS Code changelog wording for session switching
Adds a shared sidebar tooltip provider with delayed open and instant close behavior
Removes per-button tooltip delays so sidebar actions feel more consistent
Aligns session sidebar hover interactions with the opencode-style tooltip experience
Switching sessions moved the conversation on screen before the swap: the
composer and the status chip followed the live selection and re-shaped a
commit ahead of the timeline, so the pinned outgoing chat jumped; and the
reveal effect re-ran for the outgoing session when its waited flag flipped,
hiding it a few frames before the next one mounted. The chat column now
reads one deferred session, and the reveal runs once per opened session.
The pull-request rail icon now appears only while GitHub is connected
(OAuth or gh CLI), like Linear; Linear sits after the walkthrough in the
default rail order. The GitHub account avatar and switcher leave the
header for the pull-request panel, where the walkthrough, refresh, and
account controls share one row and one height, and the account stays
visible on the panel's empty state. A manual refresh keeps its spinner on
screen long enough to read as work done.
GitHub sign-in moves from Settings → Git to Settings → Integrations →
Built-in integrations as a card before Linear; search and the connect
buttons follow it.
The scoped theme entry is now parsed by one boundary parser with a
stated invariant instead of ad hoc typeof narrowing, and the runtime
keys that mean "no instance connected" live next to the code that
produces them so a new sentinel cannot miss the theme-storage guard.
Every project-config write re-serialized normalized tasks, so a server
that shares the config file but predates a field (goal, auto-accept)
stripped it the first time any task ran. Untouched tasks now go back to
disk verbatim, a state update swaps only `state`, and only a deliberately
replaced task is serialized from the normalized shape.
Opening a file from the Git panel while a nested repository was selected
created the diff tab under the repository path, a key the context panel
never displays. Tabs are keyed by the project root; the diff surface
resolves the selected nested repository itself.
The work-status Project section now reads branch, changes, and PR from the
same resolved repository as the Git tab and names the nested folder under
the branch so the reader knows which repository the readouts describe.
The server joins discovered repository paths with the platform separator
while every other git directory key in the UI is normalized, so on Windows a
discovered repository never matched its own selection or the root prefix the
picker strips. Parse the route's response at the boundary and normalize each
path. Note in the store docs that worktree bootstrap and session machinery
stay keyed on the project root while a nested repository is selected.
Text-to-speech picked one voice regardless of what language a reply was in.
A dependency-free language detector (script, marker letters, function words)
now decides the language of the whole message once; with the new
"Match the voice to the language of the text" setting the local provider
switches to a catalog model for that language (Kokoro zh/en and Piper models
for 12 languages, downloaded on first use like the existing model) and macOS
say switches to an installed voice whose locale matches. The local voice
picker lists voices of every installed model, and the settings show which
language models are on disk.
The Ukrainian Piper medium build is a character-level model that sherpa-onnx
turns into noise, so the espeak-based Lada build is used instead.
Claude-Session: https://claude.ai/code/session_017TK5JAYDfT3Fotc23UEg98
A turn that OpenCode stopped could end with nothing on screen: the
session.error event was only turned into a sidebar badge, its message was
dropped (the notification expected a different shape than OpenCode sends),
and a send that was accepted but never answered looked the same as success.
- The chat shows what OpenCode reported under the last message while that
turn is the latest one, and names a user message an idle session has left
unanswered for five seconds.
- The last 20 session errors are kept in memory and listed in the status
report (Ctrl/Cmd+Shift+L, also `__opencodeDebug.statusReport()`), next to
rejected sends, the managed OpenCode process's last error and stderr
tail, and the OpenCode and desktop log file locations.
- The OpenCode health probe hits /global/health instead of a route that
does not exist, and probe URLs resolve against the page for web runtimes.
A session opened from the sidebar could land above the bottom, or show a
frame sitting lower and then snap up. The viewport was pinned before the
content was final: the recap note renders once the session record arrives
and grew the footer under the pinned viewport, and on large sessions
subagent task cards grow when their child sessions load, moving everything
above the viewport.
- The recap note holds the timeline reveal until the session record is in
memory, so it is part of the first finished picture.
- The scroll hook holds the reveal until the viewport is pinned; the reveal
itself runs once the content height has held still for two frames, with
one exact pin against the final height (bounded at 300ms).
- Sitting on the end of a session that is not producing output is an
invariant: content growth re-pins from a MutationObserver in the same
frame the list writes its layout, so no frame paints with the end out of
view. Output growth keeps gliding through followEnd, which now glides only
while the session is working.
- ChatContainer swaps the timeline on a deferred copy of the selection so the
active row, URL, and tab commit first. (This part had been overwritten by a
concurrent edit before the previous commit and is restored here.)
- A session whose messages are not in memory keeps the previous conversation
on screen for up to 400ms while they load, instead of flashing a skeleton
between two conversations.
- The reveal fade runs only for a session the user waited for and is 100ms;
a session that was ready at the click appears in the same frame.
- The sidebar prefetches the two rows on either side of the open session
150ms after it settles, two at a time, so neighbouring switches are warm.
Commands and skills for the composer were read from the ambient store, which
follows the project selected in the sidebar; opening a managed chat or a chat
draft never changes that selection, so the previous project's commands,
skills, and (for drafts) files leaked into chats. The autocompletes now load
and read commands and skills for the effective directory, and a chat draft
resolves its effective directory to the prepared chat directory or the Chats
root instead of the last project.
Claude-Session: https://claude.ai/code/session_017TK5JAYDfT3Fotc23UEg98
Switching sessions ran as one synchronous commit: sidebar highlight, URL,
a full timeline remount with markdown re-parse, and around nine requests,
so nothing changed on screen for 150-250ms after the click.
- ChatContainer swaps the timeline on a deferred copy of the selection, so
the active row, URL, and tab commit first and the timeline renders behind
them; selection policy keeps reading the live store value.
- The message fetch starts before the selection is published.
- Sidebar rows stop re-rendering on a project switch: directory-scoped sync
hooks read the runtime context and a subscribable current-directory source
instead of the directory-bearing context; the grouping builder reads git
branches through a ref and section caches key the branches they use;
descendant ids are keyed by content. Rows per switch went from 73 to 8.
- Markdown skips the async re-render when the settled cached blocks are
already painted, and mounts synchronously once its lazy module is loaded;
the module is preloaded at boot.
- A timeline reveal gate holds a freshly opened session at opacity 0 while
any provisional markdown paint catches up (250ms cap), then fades the whole
timeline in once, so text, tools, and recap appear together.
- Switch fan-out trimmed: knowledge summary deduped, MCP status refreshed only
when stale, non-repo directories cached by the git repo check, OpenChamber
defaults cached briefly, agent memory reused for the same project, goal
text cached, PWA manifest rebuilt after the switch settles.
- Header tabs snap into the active state and keep the title at the same
height in both states.
- Prefetch on row press; composer focus moved off the commit.
`bun run profile:switch` records ack/content latency, longest task, and
requests per switch, cold and warm, and compares runs against a baseline.
Measured warm switch: ack 228ms to about 40-60ms, content 228ms to about
100-120ms.
The find_in_file keybind only opens the preview find bar when the event
target sits inside the preview container, but that container was a plain
div that never held focus, so the shortcut did nothing in read-only preview.
Both preview containers are focusable now and take focus when a Markdown
file opens in preview or is clicked, without stealing focus from a field
the user is typing in.
Claude-Session: https://claude.ai/code/session_017TK5JAYDfT3Fotc23UEg98
Context parts are marked synthetic, so a message made only of quoted
fragments was skipped as an injected message and could not be navigated
to. Such turns now appear with the same caption the bubble shows.
Opening or closing the context panel re-wrapped the timeline; the settle
handler then scrolled to LegendList's total content length, which still
carried pre-wrap sizes for unmounted rows, leaving a viewport-sized blank
tail under the last message.
Reverts #2687. In real use the injected small_model behaves poorly with
OpenCode: its internal small-model consumers and OpenChamber's own small
model are different things and must stay configured separately.
Bohdan Triapitsyn