Show frequent commands, skills and plugins in matching visual and keyboard order.
Update discovery hints and computer-use labels across all five locales.
Unify skill, plugin and file search across the plus and mention menus.
Prioritize common slash commands and hide withdrawn managed skill packs.
Refresh capability candidates when reopening the composer menu.
Page transcript and trace reads, bound UI caches and retained task records,
and replace full-file background polling with incremental projections.
Preserve recovery and ownership semantics across pages and cancel stale work.
Add an execution-model picker row to the plan approval bar so users can
approve a plan and switch to a different model for implementation in one
step. Same-provider switches use the CLI's in-process set_model control
request (no restart); cross-provider switches approve → interrupt →
restart the CLI with the new env → auto-continue execution.
The "+" composer menu used to offer only file attachments and slash
commands, so skills, connectors, agents, agent teams, Computer Use and
workflows had no discoverable entry point beyond natural language. The
menu now opens a searchable panel — shared verbatim by both composers —
with drill-in sub-lists backed by the existing APIs: skills and
connected connectors insert mention badges, agents and workflows insert
their slash text, teams offer a create-via-chat prompt seed plus
workbench shortcuts, and Computer Use gets an inline global switch.
Slash commands that duplicate a permanent GUI surface (status, cost,
context, config, doctor, memory, plugin, help) no longer crowd the
empty-query listing; typing their names still matches and runs them.
The old panel led with a per-category breakdown (Messages / Skills /
System prompt) that answers a diagnostic question, not the ones users
actually watch: how much window is left, how fast the session is, and
how well the cache is doing.
Reorder the panel to match: the headline is now the remaining window,
a single segmented bar compresses the composition into one strip with
the per-category list behind a collapsed toggle, and the session stats
promote speed / cache hit / cost to a three-stat row.
The composer directory picker only showed ten recent projects with no
search and no way to name one. Load the full project list with a deep
session-history scan, filter it client-side with a fuzzy scorer, and
offer a "new project" entry that names a folder through the existing
ProjectEditorModal and selects it as the working directory.
Also key the recent-projects response cache by scan depth so a shallow
request can no longer truncate a deeper one.
The sidebar New Session button seeded the new session's directory with
currentSession.workDir first. Isolated-worktree sessions keep their workDir
pointing at .claude/worktrees/<slug> forever, so every follow-up session was
created inside the stale worktree: the branch panel then ran git there, main
showed as checked out elsewhere, and launching failed with 'Branch "main" is
already checked out in another worktree'.
Add getSessionSeedWorkDir() (projectRoot for worktree sessions, own workDir
otherwise, preserving intentional repo subdirectories) and use it at all
three seeding sites: the sidebar New Session button, the launch draft seed
for empty sessions, and the scheduled-task default folder.
A session with 44 linked subagent transcripts returned a 541,817,705-byte /
539,323,608-character /messages body — past V8's 536,870,888-character string
limit, where Chromium hands the renderer an empty string. The session opened to
"Unexpected end of JSON input" instead of its history.
HTTP now serves the root transcript only; an Agent card fetches its run's tool
stream from /subagents/by-tool when expanded, reusing the truncation the server
already applies to oversized runs. Rewind checkpoints, team task anchors and
workspace change attribution keep the merged view they depend on.
Measured on the session that failed: /messages 542 MB -> 40 MB; one run payload
45.8 MB -> 22.9 MB (the Activity projection is no longer sent twice when it is
the same array).
OpenCode Go binds the wire format to the URL path and translates nothing
between formats, so one provider record serves /chat/completions, /messages
and /responses depending on the model, each accepting a different credential
header. A record carries only one apiFormat, so presets can now declare
ordered per-model prefix rules (modelApiFormats) and the proxy resolves the
effective format from the request body. Only the exceptions are listed;
anything unmatched keeps the provider's format, which is the endpoint with the
broadest compatibility. A preset with rules is authoritative for apiFormat,
because a value written by a cc-switch import or the edit form would otherwise
silently disable every rule and point the CLI straight at the upstream.
The gateway also rejects any request without a stable per-conversation
x-opencode-session, so presets can declare upstreamHeaders with $SESSION_ID
and $VERSION placeholders. The id is the one the CLI already sends; it is
redacted in traces the same way the credential is, since it also names the
local transcript files.
Title generation builds its own upstream request and bypassed all of the
above, which left AI titles failing for every provider that needs local
request handling; it now goes through the same proxy the CLI uses.
Verified against the live gateway: glm/kimi reach /chat/completions,
minimax/qwen/union-alpha reach /messages, grok/gpt reach /responses, each
with the credential that endpoint accepts.
The card renders from the transcript, but it can only be answered through the
live permission request. Two ways that mismatch bit users:
An unanswered question whose request is gone — the renderer was away when it
arrived, and the bounded 30-minute pending-permission window then reclaimed the
CLI — rendered as a fully editable form whose Submit and "chat about this"
buttons were silently dead, with nothing saying why. The card now reports the
question as expired and delivers the answers as an ordinary message instead,
which is the only channel left once nothing is waiting.
Answers also lived only in component state, so switching tabs (ContentRouter
mounts only the active tab) or scrolling the card out of the virtualized window
threw them away. They now live in a store slice keyed by session and toolUseId,
along with the two terminal states only this renderer knows about, so a remount
cannot resurrect an answerable form and deliver the same answer twice.
While a question is pending the composer no longer takes a message at all: the
model is blocked inside the tool call, so a queued message cannot reach it until
the question resolves, and reading it as "I already replied" is how questions
got abandoned. The round button stays Stop, which aborts the question and
unlocks the composer.
Tested: cd desktop && bunx vitest run (5889 passed; the one failure is the
pre-existing computerUseWorker case that needs adapters/node_modules)
Tested: cd desktop && bun run lint && bun run build
The context panel treated every cache read as spend and measured tok/s
against decode time without TTFT, so a 270k-window DeepSeek session
read as 40M tokens at 512 tok/s.
A building or empty local index used to fall through to a full transcript scan, so opening the sidebar or traces page could sit on GET /api/sessions until the 120s client timeout. Serve partial SQLite rows instead, look up trace titles by session id, and treat client disconnects as 499.
Print-mode team leads were treating mailbox permission asks as ordinary chat, so desktop members hung forever with no approval UI. Forward those requests through the existing can_use_tool host prompt and label them with the teammate name.
The test-connection row is a flex row where the result text kept its default
`min-width: auto`, so a long upstream error forced every pixel of overflow onto
the button. Because `size="sm"` pins the height at h-6, the label wrapped and
the button rendered as a squashed two-line block.
Pin the button with `shrink-0 whitespace-nowrap` and let the error text take the
remaining width (`min-w-0 flex-1 break-words`) so long messages, including
unbreakable URLs, wrap inside the row.
The sidebar shell's permanent compositing layer dropped macOS app-region
hits around the traffic lights. The workspace header slot also marked its
leftover titlebar space as no-drag, so the empty strip could not move the
window on either platform.
The fold button was shown whenever a project was marked expanded, including
short lists auto-expanded by history paging. Keep it only above the 6-session
threshold. Replace the composer percentage chip with a compact ring.
PUBLIC_ACCESS_CONSENT_VERSION moved to 2 when remote provider management
landed, but the Electron IPC validator still accepted only the literal 1.
The renderer sends the constant, so publicAccessStart was rejected as an
invalid payload in the preload guard and again in the main-process handler,
before PublicAccessManager saw the request at all. That rejection bypasses
the manager's error classification, so the settings page could only show the
generic "operation failed" line while the state stayed disabled: public
access could not be enabled, and autoStart could never become eligible
because consent v2 was never persisted.
Validate against the shared constant instead of a copy of its value, and
stop pinning the stale literal in the tests that let this drift through:
capabilities.test.ts and electronHost.test.ts now send the constant and
reject the previous, next and non-numeric versions.
The skills tab mounted the connector catalog as a featured row above the
skill market, and that row was the only place the five curated packages
(frontend design, canvas design, generative art, webapp testing, MCP
builder) were offered. Stop mounting it, so the tab renders the market
alone and those packages have no entry point left in the UI.
The featured slot on Market and MarketHome stays in place, and the
catalog, bundle lock and installer under src/services/connectors are
untouched: restoring the row is a change to one branch of ExtensionMarket,
packages already installed keep working, and their mentions still resolve.
The ExtensionMarket test now asserts the tab does not mount the catalog
again.
The context panel could say what was in the window but nothing about what the
session had spent: no total token count, no cache hit rate, no generation speed.
The numbers were already on the wire — translateCliUsage picked four token
buckets out of the CLI's result message and dropped duration_ms, duration_api_ms
and num_turns with them — and nothing anywhere accumulated how long the model
spent emitting tokens rather than waiting on prefill.
Measure that span where it happens: a decode span opens at the first generated
delta and closes at message_stop, rides the stream_event up to QueryEngine, and
accumulates in cost-tracker beside totalAPIDuration, including the project
config restore path so it survives a CLI restart. Tokens/sec is output over that
span, never over wall clock, which would divide by tool execution time.
The totals needed fixing before they were worth showing. Claude Code persists one
JSONL line per content block of a reply and repeats the whole usage object on
every one, so summing lines overstated real transcripts by 2.2x — and
chooseRicherUsage prefers the larger of two snapshots, so the inspector actively
selected the inflated one. The transcript readers and the renderer's history
summary now deduplicate on usageAccounting's key, the rule stats.ts and the
activity index already use.
The panel polls a usageOnly inspection mode while it is open and stops when it
closes: one get_session_usage control, no skills-directory scan, no transcript
re-read, and no request stacked behind one that has not answered.
The @ and / menus each carried their own icon fallback — skills rendered as a
sparkle in one and a box in the other — so the same entry changed shape
depending on which menu opened it. Both now share one vocabulary, and the @
menu shows the source labels the slash menu already had.
Brand icons were resolved against the document root, which only worked while
`base` was `/`; the packaged renderer loads from file://, where
`/connectors/x.svg` points outside the bundle. Connector rows also opened the
detail view and started installing in the same click — installation now
happens only from the detail action.
The fallback command list no longer offers commands the headless CLI cannot
run (`clear`, `vim`, `commit`, `pr`, …): selecting one only produced
"Unknown skill". Of the 59 compiled commands, 16 support the headless path a
desktop session drives. An unprioritised menu also opened on the CLI's
bundled skills (`update-config`, `debug`, `batch`), so desktop-owned commands
now lead instead.
With the workspace panel open, the titlebar's right section was paper right
up to a 16px notch of sidebar grey welded to the window's top-right corner.
The paper was painted by the window header, but the drag gutter — and, on
Windows, the window controls — are its transparent siblings inside the same
frame, so whatever their parent showed came through. Closed, that is
invisible: the whole strip is the sidebar's ground and the gutter just
continues it. Open, it is a strip of trough next to a white panel.
The frame now owns the ground for everything above the panel, so the header,
the gutter and the native controls share one surface.
The trailing icon was the only hit area on an assistant output card, so
opening a generated file meant aiming at a 32px target on the far right while
the file name and path sat inert. The row body is now a real button — icon,
name, badge and path all open the target, with a pointer cursor, a row hover
fill and a keyboard focus ring, and a label that names the file.
The open-with control is untouched and still opens its menu without also
opening the file. Text inside the row is no longer drag-selectable, which is
how browsers treat buttons; the path stays copyable from the open-with menu.
A single click in the file tree opened a VSCode-style preview tab, and the
next click replaced it — picking four files left one tab, which read as a
hard limit. Every pick now opens a permanent tab (repeat picks of one path
still dedup in the store), and the blank-placeholder takeover extends from
blank browser tabs to the empty Files launcher, so the first pick fills the
launcher tab instead of stranding it.
Output cards and Open-with menus for files like README-拍摄大纲.md rendered
fine but died silently on click: card creation reads real paths from the turn
checkpoint, while the click route ran parseFilePathRef whose segment charset
was ASCII-only, classifying the href as `ignored`.
parseFilePathRef now matches on a Unicode-tolerant segment set — the whole
string is already delimited by the markdown span or href, so the prose-bleed
concern that keeps the prose scanner ASCII-only does not apply.
Multi-question prompts get a Next button at the end of the action bar,
disabled until the question on screen is answered. Picking a single-select
option advances on its own, matching ChatGPT's ask_user_input card;
multi-select, free text and the last question stay put, and nothing
auto-submits.
Add mobile provider and General settings while preserving desktop behavior.
Harden remote credential handling, ngrok session ownership and consent upgrades.
Use a native Electron menu with validated host actions and lifecycle cancellation. Preserve browser state, persistent annotations, and disabled page actions on blank tabs.
Replaces the two-mode right-side panel (files ↔ browser) with one controller
that owns layout, navigation and resource lifetime for four content kinds:
files, browser pages, Git review and terminals. Follows the Codex desktop
reference captured in the workspace-refactor plan.
The old panel conflated three things that have different lifetimes: a UI tab,
the content it shows, and the process behind it. That is why switching modes
destroyed a live page, why a second link overwrote the first, and why the
toolbar button reported "show workspace" while the workspace was already open.
Splitting them is the whole change:
- workspaceStore layout, docks, tab order, preview/pinned, focus
- workspaceContentStore / workspaceReviewStore file and diff data
- host services webContents and PTYs, keyed by resource id, never by tab id
Consequences that fall out of the split:
- Hiding the panel, switching tabs and switching tasks keep every page and PTY
alive; only closing a tab releases them.
- A terminal moves between the side and bottom docks without restarting.
- Pages live in a shared persistent partition with native navigation history;
popups become sibling tabs in the task that opened them.
- Review names both sides of every comparison and performs real index and
working-tree writes, guarded by a snapshot token and a backup-first revert.
Also adds versioned workspace persistence with a forward migration: terminals
come back stopped and restartable, pages reload lazily, and nothing holding
content, cookies or handles is written to storage.
The previous implementation is no longer reachable but is left in place: the new
file tab does not yet reproduce workspace search, quick-open or the changed-file
view, so removing it now would lose those. Real three-platform Electron
acceptance (plan item A10) has not been performed; all evidence here is
deterministic tests, type checks and a packaging smoke.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
程序员阿江(Relakkes)