The list-models pipeline already served LKG on failure; this wires the
rest of the LKG-first, manual-refresh policy end to end.
Main: the list-models IPC accepts a force flag — manual refresh always
skips the in-memory cache and the persisted entry and goes live, and a
forced success overwrites the persisted catalog.
Renderer: the panel passes force only from the retry banner and the
picker's refresh button; panel-visibility and send-preflight loads stay
non-force and replay the renderer cache, so with a persisted catalog
main answers them without spending the live-fetch budget.
Picker: a footer refresh icon-button next to the status chip, visible
whenever any fetched catalog (live or last-known-good) or the degraded
warning is present; the spinner row remains the in-flight feedback.
Tooltip localized as ai.chat.refreshModelList in en/es/ru/zh-CN/zh-TW.
Fresh installs are unchanged (no persisted entry → live fetch as
before); existing users stop seeing jitter warnings on panel open and
pick up new CLI models via the refresh button.
* feat: alert when an external MCP approval is waiting
Show a system notification and request taskbar or Dock attention so confirm-mode approvals are not missed before the existing timeout.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix: stop MCP approval attention when a later window focuses
Watch app browser-window-focus while a flash or dock bounce is active so a
window created after the alert starts can clear it.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
opencode was the only SDK agent whose picker rendered nothing until its
live catalog pull resolved: the #3584 suppression replaced presets with
just the user's custom ids while runtimeModelLoading was set. Align it
with every other backend — presets stay selectable during loading, the
live catalog replaces them on arrival, and #3585's retry/warmer
mechanics are untouched. Adds a regression guard against reintroducing
the opencode gating.
Implements the unified model catalog UX on top of the last-known-good
persistence layer (#3624):
- R1 source provenance end-to-end: list-models now returns
source: 'live' | 'last-known-good' plus fetchedAt (ISO) on live
successes and last-known-good serve-outs; the pre-flight cache replays
the original provenance on hits. The renderer cache persists
source/fetchedAt alongside the catalog, and the build-time preset
fallback derives 'presets'.
- R2 loading state for all backends: drop the opencode-only gate on
modelCatalogLoading so every SDK backend shows the composer loading
state while its live fetch is in flight (opencode behavior unchanged).
- R3 unified status chip: ComposerModelPicker renders a muted footer chip
for every backend - live / cached with relative last-success time /
built-in default list - replaced by the spinner row while loading.
Keys added to en/es/zh-CN/zh-TW (ru keeps its pre-existing ai.opencode
gap).
- R4 copilot headless wrapper: generated per-resolve wrapper (baked with
the platform exe the npm loader would spawn) is handed to
RuntimeConnection.forStdio, spawning with windowsHide:true so no
console window flashes in Electron. Falls back to the loader when the
exe cannot be resolved (window may flash - documented degradation).
* feat(ai): scope external MCP approvals to connection or host
* test(ai): keep MCP fixture in Netcatty temp directory
* fix(ai): refine external MCP approval UI and grant safety
* fix(ai): wrap long approval targets and support Node 22
* fix(ai): keep automatic scripts in observer mode
* fix(sftp): release preflight markers after transfer
* fix(sftp): retain preflight cap through conflict resolution
* Keep SFTP preflight cap across panel remount
* Update worker terminal metadata expectations
* Persist SFTP preflight skip state across restart
* Limit persistent host approvals to saved hosts
* Scope host-open approvals to the selected saved host
* Exclude serial hosts from persistent approval scopes
* feat(ai): add MiMo Code (mimo) agent support scaffolding
Registry/wiring only: discovery entry, managed-agent metadata, model presets, icon, i18n and Settings UI for Xiaomi's MiMo Code CLI (an OpenCode fork). The SDK driver is intentionally not wired yet - it depends on the pending @mimo-ai/sdk parity probe, so DRIVER_REGISTRY keeps a TODO and run-turn/list-models reject the mimo backend until then.
* feat(ai): add MiMo Code (mimo) agent support
Adds Xiaomi's MiMo Code CLI as a managed agent backend, mirroring the existing
OpenCode integration (discovery, managed-agent metadata, model presets, icon,
i18n, Settings UI) and adding the SDK driver that was left as a TODO in the
scaffolding commit.
Driver notes:
- New electron/bridges/aiBridge/sdk/mimoDriver.cjs reuses the pure translation
helpers exported by opencodeDriver.cjs (config/prompt building, event
translation, model mapping) because MiMo Code is an OpenCode fork and speaks
the same wire protocol.
- It does NOT use @mimo-ai/sdk's createOpencode(). On 0.1.15 that helper spawns
a binary named "opencode", waits for an "opencode server listening" banner,
and passes config via OPENCODE_CONFIG_CONTENT; the mimo CLI prints
"mimocode server listening on <url>", reads MIMOCODE_CONFIG_CONTENT, and ships
as "mimo"/"mimo.exe". So createOpencode() can only ever time out. The driver
runs "mimo serve" itself, parses the readiness URL, and attaches the SDK's
createOpencodeClient({ baseUrl }).
- Binary resolution honours binPath, MIMOCODE_BIN and MIMOCODE_BIN_PATH, then
falls back to PATH via resolveCliFromPath (the npm shim is mimo.cmd on
Windows, which Node cannot spawn directly).
- @mimo-ai/sdk ^0.1.15 added as a dependency.
Verified on Windows against mimo 0.1.15 + @mimo-ai/sdk 0.1.15:
- "mimo serve --port" boots and reports its URL; driver parses it.
- SDK symbols used all exist: createOpencodeClient, client.config.providers(),
client.global.event() (SSE stream), client.session.create/abort/promptAsync.
- Live provider catalog reads 2 providers (xiaomi, mimo); MIMO_MODEL_PRESETS
updated to the real lower-case provider-prefixed ids and its TODO removed.
- Server process is killed on close (taskkill tree kill), no orphan processes.
Not yet verified at runtime:
- A full end-to-end turn (no prompt was sent, so no tokens were spent): text /
reasoning / tool-part streaming and abort handling are untested against a
live model.
- MIMOCODE_CONFIG_CONTENT accepting the OpenCode config schema (permission,
mcp, model, share, autoupdate) end-to-end; only that the env var exists in
the binary has been confirmed.
- Whether the loopback server needs MIMOCODE_SERVER_PASSWORD when the API
client attaches.
- The placeholder public/ai/agents/mimo.svg still needs an official asset.
- Multi-turn resume (resumeSessionId) and MCP injection against the live CLI.
* feat(ai): inject MIMOCODE_BIN for discovered MiMo Code agents
Mirror the OpenCode OPENCODE_BIN handling so a discovered or re-detected
mimo binary path reaches the driver, and cover it with the same assertion
shape as the opencode managed-agent env test.
* fix(ai): use MiMo branding in the mimo driver
* test(ai): cover the mimo driver
Add a unit test file for the MiMo Code driver, matching the existing
opencode driver test in shape, runner (node:test via node --test) and
`.test.cjs` extension. It is the last driver in the registry without
one.
Coverage:
- spawn/launch: `mimo serve --hostname/--port` argv, binary resolution
(explicit binPath / MIMOCODE_BIN / MIMOCODE_BIN_PATH), pass-through
env and MIMOCODE_CONFIG_CONTENT; `child_process.spawn` is stubbed, so
no real `mimo` process is ever started.
- readiness: parsing of `mimocode server listening on <url>` from
stdout (including split chunks) and stderr, the OpenCode spelling,
and the MIMO_SERVE_TIMEOUT_MS timeout path (asserts the rejection and
that the child is killed).
- abort/cleanup: stopMimoProcess kill/idempotency/swallow, the Windows
taskkill tree kill, server.close() teardown, and abort mid-turn with
a throwing close().
- branding: OpenCode -> MiMo Code for the busy/idle/status, failed and
tool-failure fallback strings, with tool output and non-brand
technical errors left verbatim.
- happy path: session create -> streamed event translation -> done,
plus listMimoModels catalog mapping and the empty/aborted paths.
* fix(ai): filter MiMo events by session and stop the server behind the npm shim
Address review on #3500:
- The global SSE stream carries every session on the server, and
runMimoTurn translated all of them, so another session's text could leak
into the reply and its session.idle could end the turn early. Skip events
whose session id is not ours, reusing the OpenCode driver's
getOpenCodeSessionIdFromEvent (now exported). Covered by a test with
interleaved sessions.
- On macOS/Linux the official @mimo-ai/cli `mimo` is a Node shim that runs
the native binary through spawnSync, so killing the shim left the server
listening. `mimo serve` is now spawned detached on POSIX and
stopMimoProcess signals its whole process group (SIGTERM, then SIGKILL
after a grace period). Windows keeps taskkill /T.
Verified against the real @mimo-ai/cli@0.1.15 npm shim: on Linux (WSL
Ubuntu 22.04) the listener is the native .mimocode child; before this change
it was still listening 3s after close(), now the port is released within 1s
and no mimo process remains. On Windows the .cmd shim path also releases the
port within 1s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix: harden MiMo CLI integration before merge
* fix: wait for MiMo events and handle setup cancellation
* fix: finish MiMo stream handling and scope skill access
* fix: preflight MiMo skill paths before one server startup
* fix: scope MiMo skill reads to trusted roots
* fix: allow MiMo skills from ancestor project roots
* fix: honor MiMo observer and confirm command permissions
* fix: clear MiMo approval when its service exits
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: bincxz <16399091+binaricat@users.noreply.github.com>
* feat(vault): let Skills + CLI operate notes with the same approval as MCP
Expose vault note list, read, write, and markdown import on the tool CLI so agents can save generated documents only after the user approves.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(vault): keep explicit empty note fields in the CLI
A supplied empty string was treated as an omitted flag, so note clears and empty markdown imports never reached the note service.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(vault): allow an explicit empty note body on CLI create
notes create --content "" was rejected as a missing flag, so it did not match vault_notes_create.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(vault): stop echoing imported note bodies and recognize leading CLI flags
A batch import resent every document over IPC, and a leading --json hid the notes command from artifact inference.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(vault): make note CLI import safe and fail writes cleanly
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: bincxz <16399091+binaricat@users.noreply.github.com>
Avoid rewriting thinking/provider maps when values are unchanged so
mounted AI panels do not re-render. Put permission title and short
description on one line and drop the Cursor CLI parenthetical.
Stop advertising Gemini 3.7 Flash minimal, GPT-5.x-chat reasoning, and
Claude 4.7+ extended thinking. Map GPT-5 Off to minimal, send adaptive
effort on Claude 4.6+, and keep composer Enter on an exact or custom id.
Codex P2: switching to a catalog model dropped context_length, so
Catty budgeted 128k instead of the advertised window. Keep the
metadata on picker models and write it to modelContextWindows.
Codex P2: models whose menu includes Off still showed Off for users
with no stored preference, while the request sent no override. Leave
the chip unselected until the user picks a level.
Codex P2: hiding the custom-model row still left Enter falling back
to the typed query, which send-time validation then discarded. Only
use the query as a custom ID when Catty offers that action.
Codex P2: opening the composer on Gemini Pro persisted Low for users
who had never chosen a thinking level. Only remap an explicit stale
value; leave an empty preference unset so the request uses the
provider default.
Codex P2: pinning or recents in one parked ChatInput only updated that
instance. Subscribe every mounted composer so sibling terminals see
the same recent/pinned list without remounting.
Codex P2: refreshed CLI catalogs still showed saved IDs that the
catalog no longer contains, then send-time validation swapped them.
Only Catty may keep a custom ID that is not in the live list.
Codex P2: external-agent pickers advertised a custom ID that send-time
validation ignored. Offer Use custom model only for Catty providers.
Codex P2: the thinking budget was added to both compaction max output
and reservedTokens, so it was subtracted twice. Keep it only on
compaction max output.
Codex P2: Claude 3 Haiku still offered Low/Medium/High and would send
an enabled thinking budget. Gate the chip and the provider option on
Claude 3.7 / 4+ model ids. Also memoize Catty thinking levels.
Codex P2: switching away from Gemini 3 Flash left `minimal` selected
on models that do not advertise it. Persist a valid level for the new
model.
Codex P2: Gemini 3 Pro and 2.5 Pro cannot disable thinking, so Off
was a lie. Those menus only offer low/medium/high.
Codex P2: catalog cache ignored API key rotation, so a replaced
credential could keep the previous account's models. Fingerprint the
stored secret in the cache key.
Codex P2: Gemini 3 Flash cannot turn thinking off; Off was sending
thinkingLevel=minimal. Surface Minimal on the chip instead of Off.
Codex P2: OpenAI-style models such as gpt-4o rejected reasoning_effort
when the chip always offered Off/Low/Medium/High. Gate the chip and
the provider option on known reasoning model ids, and send explicit
none when Off is selected on a reasoning model.
Remove the provider divider and search-field outline in the sidebar
model picker. Size Save/Test/Cancel to the same h-8 box without
changing the global Button styles.
Give Catty a two-column provider/model menu in the composer, add search
plus recent/pinned models, and move thinking effort onto its own chip
so Claude, Codex, Grok, and Catty share the same compact controls.
Closes#2137Closes#2637Closes#1884
Reconcile the app-lock feature with main's refactored renderer
architecture (publishers/hosts/AppShell) and main-process changes:
- index.tsx/App.tsx: AppLockGate keeps ownership of useSettingsState and
useAppLockState; new AppLockRuntimePublisher publishes the runtime via
appRuntimeBridge; SettingsPublisher now binds the gate-owned settings.
- AppSideEffects: defer deep links, tray port-forward toggles, global
hotkeys and window close commands while locked; flush on unlock.
- main.cjs: keep fresh-session re-lock before main's reusable-window
fast path; before-quit guard delegates to handleBeforeQuit with main's
async plugin-shutdown commitQuit and pending-update cancellation.
- globalShortcutBridge: keep locked-state tray redaction on top of
main's deferred tray panel show logic.
- preload: adopt main's lastPayload replay for terminal popup config
(superset of the gate-delay replay fix).
- vite config: keep dev CSP plugin, adopt main's warn-only stale
optimize-dep handling; electron-builder: keep Windows Hello helper
extraResource with main's CLI-controlled arch targets.
Co-authored-by: Cursor <cursoragent@cursor.com>
A 300ms initial delay let ChatMessageList start the Streamdown import
while the user was still opening the panel to type. Restore the 4s
expand grace and drop a pending Shiki plugin setState if the message
unmounts before the idle callback lands.
Blocking hydrate on focus left history as raw markdown for as long
as the input was focused. Shiki still waits for a real idle gap.
Prose, lists, and inline code can mount after typing stops even if
the textarea still has focus.
messageResponse pulled @streamdown/code on first markdown import, so
opening a chat with history paid for the full highlighter even when
the reply was just prose.
Render with CJK only first. When a fence exists and the composer is
idle, lazily import the isolated Shiki plugin.
Typing a few characters then pausing still hitch because markdown
hydrate only checked the 800ms activity window. History rows would
mount Streamdown while the textarea still had focus.
Treat focused/composing as busy, wait 2s after the last key, and
keep idle IPC on the same predicate.
The first glyph still re-rendered the whole 1500-line ChatInput because
hasComposerText lived in the same component as the model picker. Token
usage snapshots also broke memo with a new object each time.
Move send-button emptiness onto a tiny external store so only that
control updates, and compare contextUsage by value in ChatInput's memo.
Clearing the box immediately after onSend overwrote the pending draft
while handleSend was still awaiting provider sync. A rejected sync or
stale-scope abort then left the prompt gone even though no user
message was created.
Leave the text in place until the parent clears the draft after the
send is accepted. Unmount commit still uses a ref so setter identity
changes cannot write the prompt back.
Typing still hitched because each IME commit started a 120ms
updateDraft, which re-ran the whole useAIState hook. Parent also
passed a new hosts array and providerSwitcher object on every
render, so ChatInput's memo never held. field-sizing-content
relayouted the textarea on every glyph.
Keep typed text in pending refs until send/hide/unmount. Memoize
mention hosts and the provider switcher. Use fixed field sizing
in the composer.
Send awaits provider sync before dropping the draft. During that
wait, ChatInput's onChange cleanup could fire on a new setter
identity and write the still-visible text back into the pending
draft, so the uncontrolled box kept the last prompt.
Clear the field right after onSend starts, and only commit on true
unmount via a ref so callback identity changes cannot restore it.
New Chat wrote an empty composer-only draft, which the text-only
equality treated as identical to the previous typed draft. The panel
skipped re-render, so the uncontrolled textarea kept the old string
and send could still submit it.
Treat emptying draft text as a lifecycle change so listeners wake, and
commit the live box on ChatInput unmount so mid-IME hide still flushes.
Every keystroke was re-rendering ChatInput and the first character
created a store draft that broadcast AI state, so the first 1-2 glyphs
painted and the rest dumped after a stall. IME composition hit the
parent on every pinyin letter, and Chromium spellcheck loaded on the
first word.
Keep the textarea uncontrolled, commit parent state on idle/blur/send,
and treat missing-to-composer-only drafts as text-only so store
listeners do not wake. Draft flush, send, and hide still persist the
current box text.
Codex P2: Agent selector, New Chat, and history stayed clickable while
provider/model IPC ran. Switching agent or starting a new chat then let
the stale send finish and wipe the replacement draft. Disable those
header transitions until preflight ends, and abort if the view still
changed.
Codex P2s: keep a stable pending draft identity so steer success can
clear the composer; take the send gate before awaiting provider sync;
and hydrate Streamdown after recent typing instead of waiting for blur.
Codex P1: first-launch send could race a 4s idle delay on
aiSyncProviders while main still had an empty provider list. Sync on
panel show and await it in handleSend.
Codex P2: slash-picker /stop and /compact called parent onChange('')
without updating local composerText, so the command stayed in the
textarea. Route those clears through commitComposerText.
Codex P2: focusing or typing cancelled the only scheduled callback, so
provider sync and discovery never ran if the user submitted without
blurring. Re-arm after activity and treat IPC as busy only while the
composer was recently used, not for the entire focused lifetime.