Issue #1101 follow-up. When a third-party Anthropic-compat backend
streams thinking deltas without first emitting a `reasoning-start`
content-block signal (DeepSeek's \`-v4-flash\` is the canonical
offender), the Vercel AI SDK has nothing registered for the incoming
\`part.id\` and enqueues an \`error\` chunk on \`fullStream\` with the
text \`reasoning part <id> not found\` — once per orphan delta. The
analogous error exists for text parts.
These are internal SDK bookkeeping noise, not user-facing errors. Our
\`case 'error'\` handler was treating each one as a real error: it
appended an empty assistant message carrying the SDK string. The
visible damage was a wall of red banners in the chat panel — but the
worse damage was protocol-level. On the next turn we rebuild
\`sdkMessages\` from local history; the placeholder assistants slot
in between the assistant that holds the parallel \`tool_use\` blocks
and the subsequent \`role: 'tool'\` messages with their results. The
Anthropic SDK's \`groupIntoBlocks\` only merges *consecutive* tool
messages into a single user block, so the tool_results no longer
immediately follow their tool_use parent — and the backend rejects
the request as
\`400 messages.N: tool_use ids were found without tool_result blocks
immediately after\`. \`messages.N\` was literally counting our 12-ish
phantom assistants.
Adds \`isSdkStreamStateError(error)\` to
\`infrastructure/ai/shared/streamStateErrors.ts\` (matches
\`/^(reasoning|text)\\s+part\\s+\\S+\\s+not\\s+found$/i\` against
string / Error / \`{ message }\` shapes), and skips placeholder
emission for matching errors in the streaming hook. Drops a
\`console.warn\` so the noise is still visible in devtools when
debugging.
5 unit tests cover positive matches, the Error/object wrappings,
non-matches that should stay surfaced, and non-string inputs.
Refs #1101.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): serialize Catty terminal_execute calls per session
Issue #1101 problem 3. When the LLM emits multiple tool_use blocks in
one assistant turn (DeepSeek's Anthropic-compat happily does this for
parallel info-gathering), the Vercel AI SDK dispatches every tool
through `Promise.all(toolCalls.map(execute))`. All of them then race
into the same per-session mutex inside the main-process bridge
(`mcpServerBridge.reserveSessionExecution`). One wins; the rest get
`{ ok: false, error: "Session already has another command in
progress..." }` synthesized back as the tool result. The LLM sees a
turn full of synthetic errors instead of the answers it asked for, the
UI cards look stuck on "executing", and the Anthropic API has
sometimes rejected the resulting trace as
`tool_use ids were found without tool_result blocks`.
Adds an in-renderer Promise-chain queue keyed by
`${chatSessionId}:${terminalSessionId}` so the actual
`bridge.aiExec()` calls are issued one at a time per terminal. The LLM
can still parallel-call as many tool_use blocks as it wants — they
just resolve sequentially with real output. Approval prompts are kept
*outside* the queue: three parallel tool_use blocks still surface
three approval cards together, so the user can dispatch them in any
order rather than waiting for each prior command to finish before the
next prompt appears. The bridge-side mutex stays as defense-in-depth
for non-LLM paths (terminal_start, MCP, etc.).
`chainBySessionKey` is a self-contained helper with cleanup logic so
the queue map doesn't leak across many short-lived sessions:
- Each task awaits the previous tail via `.then(task, task)` so a
failure doesn't poison the chain.
- A non-rejecting wrapper is stored as the new tail to keep that
contract regardless of how the task settles.
- The cleanup `finally` only deletes the map entry when the current
tail is still ours, so a caller that arrived between
`set` and `finally` doesn't have its tail evicted.
Four unit tests cover dispatch ordering, rejection isolation,
per-key parallelism, and the cleanup path. Full suite 1255/1255.
Refs #1101.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): preserve LLM emission order + honor abort in serialized queue
Codex local review on the first cut of this branch flagged two real
issues plus a weak cleanup test:
1. **Approval order vs queue order** — the previous version awaited
approval *before* reserving a queue slot. With three parallel
tool_use blocks the user could approve B's prompt first, and B
would slip into the queue ahead of A, running out of LLM emission
order. Reserving the slot synchronously up front fixes that:
`Promise.all`-dispatched executes each grab a slot at the same
instant in their dispatch order, then wait on approval while
holding it, then await `slot.ready` and run.
2. **Abort not honored** — once approvals were settled, queued-but-
not-started commands ran to completion even after the user hit
Stop. Their results were ignored by the SDK but the side effects
still happened. Two `abortSignal` checks now short-circuit before
the queue wait and again after.
3. **Cleanup test was vacuous** — the previous "drains" test would
pass even if the cleanup logic were removed. Exposed
`getSessionExecutionQueueSizeForTests` and assert the Map is empty
after the only queued task settles.
Queue API now has two entry points: high-level `chainBySessionKey`
for run-it-when-it's-your-turn callers, and lower-level
`reserveSessionSlot` for callers (like `terminal_execute`) that
need to do non-blocking pre-work in parallel with the queue wait.
Two new tests cover the reservation-order-vs-prework-order guarantee
and the skip-without-serialized-work path.
Refs #1101.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): re-issue cancel on abort to close IPC-transit race
Codex local review (round 2) flagged a small remaining race: between
the abort check after \`slot.ready\` and the main-process bridge
registering the new exec into \`activePtyExecs\`, the user can hit
Stop. \`handleStop\` does issue \`aiCattyCancelExec\`, but if that
cancel IPC arrives at main *before* the exec has finished
registering, the cancel finds nothing to cancel and the exec keeps
running. Result: the user already cancelled, but the command runs
anyway.
Plug the gap by re-issuing the cancel from the tool's own abort
listener. Once the exec has registered into \`activePtyExecs\` (a
synchronous step on the main side once the IPC lands), the duplicate
cancel finds the entry and cancels it. \`cancelPtyExecsForSession\`
is already idempotent — it iterates the live tracker and skips
entries with mismatched \`chatSessionId\` — so double-firing from
both \`handleStop\` and the tool is safe.
Extends \`NetcattyBridge\` with the optional
\`aiCattyCancelExec(chatSessionId)\` method. Already exposed in
\`electron/preload.cjs\` so the runtime call works on the live
bridge; the type addition just makes it visible to the tool.
Refs codex review on #1101 problem 3 fix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): register pending-cancel marker before SSH exec-channel opens
Codex local review (round 3) caught a remaining race in
`execViaChannel`: the cancellation marker is only registered inside
`sshClient.exec`'s open callback, which is async. If a cancel arrives
in the window between `sshClient.exec(...)` being dispatched and the
callback firing, `cancelPtyExecsForSession` finds nothing for the
session and is a no-op. The channel then opens, the marker
registers, and the command runs to completion — exactly the "user
already cancelled, but the command still ran" failure mode.
Plug the window by registering a *pending* marker synchronously
before `sshClient.exec`. The pending marker carries a `cancel()` that
just latches a `cancelled` flag and a `cleanup()` that is a no-op.
When the open callback fires it removes the pending marker and
checks the latch: if it tripped, close the just-opened stream and
resolve with `{ ok: false, error: "Cancelled" }`. If not, the normal
post-open registration takes over with the real `execStream` close.
The PTY (`execViaPty`) and raw-serial (`execViaRawPty`) paths
already register their marker before any async wait, so they don't
share this race.
Two regression tests in `ptyExec.test.cjs`:
- The pending marker is present immediately after `execViaChannel`
returns, before the open callback runs.
- A cancel during the pre-open window short-circuits the eventual
callback with `{ ok: false, error: "Cancelled" }` and closes the
now-unwanted stream.
Refs codex review on #1101 problem 3 fix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): clean up pending-cancel marker when sshClient.exec throws sync
Codex local review (round 4) noted that if `sshClient.exec(...)`
itself throws synchronously — e.g. because the underlying ssh2 client
was destroyed between the session lookup and the actual `.exec`
call — the pending-cancel marker stays in `activePtyExecs`
indefinitely and the surrounding Promise rejects instead of
resolving with the normal `{ ok, error }` shape the tool layer
expects.
Wraps the `sshClient.exec` invocation in `try/catch` so:
- Pending marker is removed when the throw escapes.
- The Promise resolves cleanly with `{ ok: false, error: err.message }`
so the tool layer gets the same shape it does for every other
failure mode (closed stream, timeout, cancellation).
Adds a regression test covering this exact path: a fake client whose
`.exec` throws synchronously; the result is a clean failure and the
cancellation map is empty afterwards.
Refs codex review on #1101 problem 3 fix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(ai): provider switcher chip in the Catty Agent chat input
Catty Agent currently has no model chip in the chat input — modelPresets
is empty for the catty agentId so `hasModelPicker` is false. The
provider/model the chat ends up using comes from the global
`activeProviderId` / `activeModelId`, which means switching providers
requires opening Settings → AI → Providers. Closes the gap surfaced in
the original feedback on #1101 (problem 2) and partially addresses
#986 (per-context model switching) by going per-agent.
State layer adds an `agentProviderMap` (`Record<agentId, providerId>`)
alongside the existing `agentModelMap`. Both are written together when
the user picks from the new dropdown; together they form a per-agent
override that beats the global active provider/model when set. Cross-
window sync mirrors the agentModelMap pattern.
ChatInput grows a `providerSwitcher` prop carrying the enabled
ProviderConfigs, the selected (providerId, modelId), and an onSelect
callback. When supplied, the model chip switches from the Cpu glyph
to the provider's ProviderIconBadge + `providerName · modelId`, and
the popover renders a two-column layout — providers on the left
(icon + name + default model caption), that provider's known model(s)
on the right. ACP agents (Claude/Codex) are untouched because they
plumb their provider through the CLI, not the Vercel AI SDK.
AIChatSidePanel resolves `effectiveActiveProvider` /
`effectiveActiveModelId` for the catty agent from the per-agent map,
falling back to the global selection. handleSend now passes those
through to sendToCattyAgent so a provider picked in one Catty session
applies everywhere Catty runs.
For v1, each provider's right column shows just its configured
`defaultModel`. The two-level structure is in place to absorb richer
listings (cached /models, manual additions) later without UI
surgery — left a note in the popover code.
Refs #1101, #986.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* polish(ai): flatten provider picker to a single list
Two-level dropdown was theatre — each ProviderConfig exposes exactly
one model (its `defaultModel`), so the right column ended up showing
the same string already captioned on the left row. Picking a provider
implicitly picks its model; one click is enough.
Also drops the `p.enabled !== false` filter on the picker list. The
user's expectation, confirmed by feedback on the first cut, is that
the chat-input list mirrors what Settings → AI → Providers shows. The
per-provider `enabled` toggle is an "active-ish" flag, not a
visibility gate; hiding disabled providers in the picker silently
made everything but the one currently active disappear, which is what
the reviewer hit.
Pares the popover down to a vertical list — provider icon + name +
default-model caption (mono if set, italic "configure default model"
hint if not) + a Check on the bound row. Removes the `hoveredProviderId`
state along with the right-column rendering since it's no longer
driving anything.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* polish(ai): tune provider icon sizes in the chat input
Chip icon was sitting at sm (20px badge) and crowded the h-6 toolbar
chip alongside the truncated label; popover rows used the same sm, so
nothing differentiated the picker from the chip visually.
Adds an xs size to ProviderIconBadge (16px badge, 10px glyph, 9px
letter fallback) and uses it on the chip. Popover rows step up to md
(32px badge, 16px glyph) and the row padding grows accordingly so the
larger brand mark has room to breathe — the picker now reads like a
list of choices, the chip reads like a status line.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): keep per-agent override honest in model fallback + chip label
Two codex P1/P2 findings on #1107:
P1 — When Catty has a per-agent provider override but no per-agent
model (agentModelMap['catty'] empty), cattyAgentModelId fell through
to the global `activeModelId`. That id belongs to whichever provider
was globally active, not the one Catty is now bound to, so e.g.
overriding to DeepSeek without a defaultModel happily sent gpt-4o
(global) to DeepSeek and produced a wrong-model error. The fallback
is now: stored agent model → override provider's defaultModel → empty.
Only the no-override path keeps the activeModelId tail; the
no-provider send guard catches the empty case for everyone else.
P2 — ChatInput's selectedSwitcherProvider used `?? providers[0]` so
the chip always displayed *some* provider even when none was actually
bound (selectedProviderId missing). The rest of the pipeline (send
guard, agentProviderMap) treats that as "no provider", so the chip
was lying. Dropped the fallback; when nothing is bound the chip now
shows a generic Cpu glyph + "Select provider" label until the user
picks one from the popover. Adds `ai.chat.selectProvider` in en /
zh-CN / ru.
Refs codex review on #1107.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): purge per-agent binding when the bound provider is deleted
Codex local review (round 2) flagged that a saved Catty model id could
outlive its provider. The chain:
- catty bound to provider A (DeepSeek): agentProviderMap['catty']='A',
agentModelMap['catty']='deepseek-v4-flash'
- user deletes provider A
- cattyAgentProvider falls back to the global active provider (B, OpenAI)
- cattyAgentModelId still returned the saved 'deepseek-v4-flash'
- send dispatched the DeepSeek model id to OpenAI → wrong-model error
Two layers of fix:
1. **Defensive resolution** (AIChatSidePanel) — cattyAgentModelId now
looks at the override provider before trusting the stored model id.
If the override is stale (provider deleted), the stored model id is
treated as orphan and the resolution falls back to the global active
selection just like cattyAgentProvider already does.
2. **Cleanup on remove** (useAIState.removeProvider) — when a provider
is deleted, any agent whose agentProviderMap entry pointed at it
gets both maps cleared (provider override + saved model). Same
rationale: that model id is now meaningless against every other
provider. Mirrors the existing activeProviderId cleanup.
Adds a small agentProviderMapRef so removeProvider can snapshot which
agents to clean up without relying on the captured-at-callback-create
state (which would be stale).
Refs codex local review on #1107.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): include agentProviderMap in settings sync + block empty-model send
Codex local review (round 3) found two issues:
P2 — `agentProviderMap` was missing from the cross-device sync surface.
Added it to the sync key list, the build path, the apply path, and the
SyncPayload type alongside the existing `agentModelMap`. Sync tests in
`syncPayload.test.ts` now cover the new field on both the build and
apply sides.
P2 — Provider rows with no `defaultModel` were still clickable in the
chat-input picker, so selecting one would save a binding with empty
model id; the send path then dispatched an empty model name and the
SDK would surface a vague backend error. Two changes:
1. ChatInput disables the row (\`disabled\`, \`aria-disabled\`, dimmed
styling, tooltip pointing the user to Settings) when the provider
has no \`defaultModel\`. Click is suppressed.
2. AIChatSidePanel \`handleSend\` adds a model-required guard mirroring
the existing no-provider guard — surfaces \`ai.chat.noProviderModel\`
as an assistant message. Catches stale bindings (e.g. user edited
the provider's \`defaultModel\` to empty after binding) before they
reach the SDK.
Refs codex local review on #1107.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): trim whitespace-only model ids + reconcile orphan bindings on sync
Codex local review (round 4) found two more leaks:
P2 — Whitespace-only `defaultModel` passed the picker's `disabled` gate
(which trims) but every other layer (cattyAgentModelId resolution,
the send-guard `!sendActiveModelId` check, the SDK call) compared the
raw string. A provider whose default model was " " could still reach
the SDK and surface a vague backend error. Normalized: trim at the
resolution boundary (catty model fallback chain) and trim before the
send-guard's existence check.
P2 — Sync apply did not reconcile per-agent bindings against the
incoming provider set. A payload could change `providers` without
shipping a fresh `agentProviderMap`, leaving local overrides pointing
to provider ids the synced set no longer includes — the same ghost-
binding bug that `removeProvider` already handles for explicit user
deletes. Added `pruneOrphanPerAgentBindings()` that runs after every
AI settings apply: it walks `agentProviderMap`, drops any entry whose
provider id isn't in the current `providers` list, and clears the
saved model id for those agents alongside it (mirroring removeProvider
cleanup). A new test in `syncPayload.test.ts` exercises the ghost-
binding case end-to-end.
Refs codex local review on #1107.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): notify open AI state of cross-device sync apply in the same window
Codex local review (round 4) flagged that `applySyncPayload` writes
straight to localStorage but `useAIState` only re-reads on browser
`storage` events. Those events only fire in *other* windows — the
window doing the apply (the one with the active chat panel) keeps
showing pre-sync providers and per-agent bindings until reload.
Concrete leak: cloud sync swaps in a new provider set, pruning runs,
localStorage is correct, but the open Catty chip still references the
ghost provider binding until the user reloads.
Extracts the existing `AI_STATE_CHANGED_EVENT` constant +
`emitAIStateChanged` helper out of `useAIState` and into a new
`application/state/aiStateEvents.ts` so non-React call sites (sync
apply, future IPC handlers) can fire it without pulling in the hook.
After AI settings apply, `syncPayload.ts` walks every AI key it
touched (including the agentProviderMap / agentModelMap that the
reconcile step may have mutated even when the payload didn't ship
them) and emits a same-window nudge. `useAIState`'s existing local
listener routes unknown keys back through its storage handler, so
each affected piece of React state rehydrates from localStorage.
Adds a regression test in `syncPayload.test.ts` that asserts the
expected `netcatty:ai-state-changed` keys are dispatched for the
providers + per-agent map keys.
Refs codex local review on #1107.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The icon badge next to the Display Name input opens the icon picker
but had no visual cue — users had no reason to suspect it was
clickable, and reviewer feedback flagged it as too hidden.
Adds a primary-tinted ring on hover/focus plus a small pencil glyph
overlaid on the bottom-right corner of the badge (also hover/focus
gated). The badge itself doesn't change shape, so the layout stays
stable; the affordance is purely additive.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(ai): customizable provider name, icon, and protocol style
Issue #1101 reported that the only way to wire DeepSeek through the
Catty Agent was to add an "Anthropic" provider with a DeepSeek base URL,
which (a) wasn't relabelable except on the "custom" providerId and
(b) implicitly conflated wire-protocol style with providerId, locking
"custom" to the OpenAI-compatible client.
ProviderConfig now carries three optional fields — `style`
(anthropic/openai/google), `iconId` (built-in brand glyph key), and
`iconDataUrl` (user upload). createModelFromConfig routes on the
resolved style first, falling back to providerId only for per-vendor
quirks (ollama's throwaway apiKey and openrouter's baseURL). Display
name becomes editable for every provider, not just custom.
The icon picker exposes a curated lobe-icons subset (MIT) covering the
common Anthropic/OpenAI-compatible third parties — DeepSeek, Moonshot,
Kimi, Qwen, Zhipu, Doubao, Mistral, Cohere, Grok, Perplexity, Groq,
Hugging Face — plus the existing six built-ins. Uploads are downsampled
to 64×64 WebP on a canvas so localStorage doesn't blow up. See
public/ai/providers/NOTICE.md for attribution.
Closes part of #1101 (problem 2). PR2 will reuse the new name/icon
plumbing in the Catty Agent chat input to surface a provider switcher.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* polish(ai): bigger labelled icon tiles and auto-fill display name
The first pass shipped the icon picker as an 8-column grid of bare
20px badges. Reviewer feedback: the icons are too small to read at a
glance, every preset should announce its brand name, and a second
click on a selected preset should let the user back out.
Switches the grid to auto-fill 120px tiles with a 32px icon plus a
truncated label so the picker reads like a brand list, not a sprite
sheet. Selecting a preset now also writes the brand's canonical
English display name into the Display Name field — the picker labels
keep their "/ 中文" suffix as a bilingual hint, but a new `name` field
on each catalog entry supplies a clean string for autofill (e.g.
"Qwen / 通义" → "Qwen"). Re-clicking the selected tile clears
iconId/iconDataUrl but leaves the typed name alone, so users who
already edited the name don't lose their edit on accidental toggle.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* polish(ai): add explicit close button to icon picker
The picker only opened via the icon badge above the Display Name field,
which left no obvious affordance to collapse it once a preset was
chosen — users either had to scroll past or click back up to the icon.
Drops a ghost Close button in the bottom-right of the picker's action
row (next to Upload/Reset) so dismissing the panel is one click and
visible without hunting.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): wire model discovery to the resolved provider style
Codex review on #1105 flagged that ModelSelector still derives auth
headers from `providerId` (`x-api-key` only when providerId is the
literal "anthropic"). Now that ProviderConfig lets the user override
`style` independently, the form persists the override but the model
discovery call ignored it — so picking an Anthropic providerId pointed
at an OpenAI-compatible backend (or vice versa) would send the wrong
header and fail to list models even though chat routing already used
the override.
Extracts `buildModelDiscoveryHeaders(style, apiKey)` as a pure helper
in infrastructure/ai/. ModelSelector now resolves the protocol family
via `resolveProviderStyle` with the explicit `style` prop taking
precedence, then asks the helper for headers. ProviderConfigForm
passes the resolved style through. The `needsApiKey = providerId !==
"ollama"` shortcut stays as-is — that's a per-vendor concession, not a
style decision.
Refs codex review on #1105.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): apply per-vendor URL fallbacks across every style override
Codex review on #1105 caught that the openrouter (and ollama) baseURL
fallback was sitting inside the `style === 'openai'` branch, so a user
who picked OpenRouter providerId + Anthropic/Google style with an
empty baseURL would skip the fallback entirely. The SDK then routed
to its own default endpoint (api.anthropic.com /
generativelanguage.googleapis.com) using the user's OpenRouter key —
silent misrouting plus auth failures.
Pulls the per-vendor quirks out into a new pure helper
`resolveProviderEndpoint(config, style, safeApiKey)`. The URL
fallback now fires for every style — the user picked openrouter for a
reason, even if they overrode the wire format. The ollama-only
`'ollama'` literal apiKey swap stays gated on `style === 'openai'`
because Anthropic/Google clients need a real key, not the throwaway
placeholder.
Refs codex review on #1105.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): send x-goog-api-key for google-style model discovery
Codex review on #1105 flagged that buildModelDiscoveryHeaders bucketed
google-style providers with the OpenAI-compat default, so a model
refresh against any google-routed endpoint was sending `Authorization:
Bearer …`. That's the wrong auth header — Google Generative AI
rejects Bearer entirely and expects `x-goog-api-key` (or `?key=`).
The runtime chat path already uses `createGoogleGenerativeAI`, which
authenticates with the Google header family, so discovery was the
odd one out.
Adds an explicit `google` branch returning `{ "x-goog-api-key":
<apiKey> }` and updates the unit test. The default google providerId
never hits this path today (PROVIDER_PRESETS["google"] has no
modelsEndpoint), but the style override surface created in this PR
lets users compose providerId + style pairs where discovery does
fire — e.g. an openai/anthropic-style providerId pointed at a Google
endpoint via baseURL override.
Refs codex review on #1105.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): pick discovery endpoint from resolved provider style
Codex review on #1105 pointed out the lopsided fix in ae7394c8: I
switched discovery *headers* to honor the resolved `style` but left
the URL path coming straight from `PROVIDER_PRESETS[providerId]`. So
the moment style and providerId disagree the request shape is half
correct — e.g. Anthropic providerId + style=openai sends Bearer
(right) at `/v1/models` (wrong; the OpenAI-compat backend exposes
/models). 404s either way.
Adds `STYLE_DEFAULT_MODELS_ENDPOINT` (`anthropic → /v1/models`,
`openai → /models`, `google → undefined`) and a
`resolveModelsDiscoveryEndpoint(style, presetEndpoint)` helper. The
style's convention wins; the caller-supplied `presetEndpoint` is the
fallback for styles with no listing convention (currently just
google).
Behavior for stock configs is unchanged — every preset's existing
modelsEndpoint matches its style's default. Mismatches now line up
(headers + path together), and `custom` providers gain a sensible
discovery attempt when the user sets a style. My earlier inline reply
ducking this was wrong; codex's call was right.
Added three unit tests covering style defaults, the override-on-flip,
and the google-style passthrough.
Refs codex review on #1105.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The group details panel rendered Startup Command as a single-line <Input>,
so users couldn't type newlines into it — only the first line ever made it
into the saved config, which then broke the multi-line sequencing behavior
shipped in #1096 for any host inheriting the command from its group.
Switch to a 3-row <Textarea> to match the per-host details panel, so a
multi-line command typed on the group is preserved end to end.
Refs #1083.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(autocomplete): add snippet completion source
* feat(autocomplete): merge snippets at the command position
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* test(autocomplete): place snippet tests where the test runner finds them
The npm test glob covers components/terminal/*.test.ts but not the
autocomplete/ subdirectory, so the snippet tests added in the previous two
commits weren't actually running in the suite. Move them up to
components/terminal/ (the existing convention for autocomplete tests) with
corrected import paths; the engine snippet cases go in a separate
completionEngineSnippets.test.ts to avoid colliding with the existing
completionEngine.test.ts.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(autocomplete): snippet ghost-exclusion, preview skip, and accept path
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(autocomplete): wire snippets into the terminal + preview snippet command
* fix(autocomplete): show only label in snippet popup row; keep snippet over colliding history
The popup row for a snippet now omits the inline command echo — the full
command lives in the detail preview only, matching the "label-only row"
design. The completion engine pushes snippet suggestions without the early
seen-text skip so that when a snippet's label collides with a history
entry's text, the higher-scored snippet survives the final dedup.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(autocomplete): broadcast snippet command so popup acceptance mirrors peers
In broadcast mode, accepting a snippet from the autocomplete popup cleared
peer input (the line-clear keystrokes flow through the broadcast-aware path)
but never sent the command, since executeSnippetCommand wrote only to the
active session. Broadcast the normalized snippet data (matching the snippet
shortkey path) so peers receive both the clear and the command, keeping all
sessions in sync.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(autocomplete): broadcast wrapped snippet bytes to preserve noAutoRun
Broadcasting the raw normalized command sent un-wrapped newlines to peers,
so a multi-line noAutoRun snippet was pasted-but-not-run on the active
session yet executed line-by-line on broadcast peers (handleBroadcastInput
writes bytes directly without re-wrapping). Broadcast the exact bytes the
active session receives instead — bracketed-paste wrapping plus the auto-run
\r — so peers mirror the active session and noAutoRun is preserved.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(terminal): add global startupCommandDelayMs setting
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(terminal): add startup-command line-split and delay-clamp helpers
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(terminal): run multi-line startup commands in sequence with configurable delay
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(terminal): expose startup command delay in Terminal settings
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* refactor(terminal): keep startup-command line content verbatim
splitStartupCommandLines now only drops blank/whitespace-only lines and
normalizes CRLF, but no longer trims each line's content. This keeps a
single-line startup command byte-identical to what the user typed (e.g. a
leading space for HISTCONTROL=ignorespace is preserved), while still
supporting multi-line sequencing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(sync): include startupCommandDelayMs in synced terminal settings
Terminal settings sync via the SYNCABLE_TERMINAL_KEYS allowlist; the new
startupCommandDelayMs preference was missing, so it wouldn't propagate across
devices. Add it (it's a user preference like keepaliveInterval, not
device-specific).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(ai): add Claude auth-presence detection helper
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(ai): surface actionable Claude auth errors and reap stuck agent processes
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(ai): add pure helpers for Claude config dir + env editor
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(ai): let users set Claude config dir and env vars in settings
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* polish(ai): harden Claude config env, de-dupe error text, label a11y
- buildClaudeEnv: drop managed keys (CLAUDE_CONFIG_DIR/CLAUDE_CODE_EXECUTABLE)
if a user types them into the free-text env editor, so they can't clobber
the config-dir field or the discovered executable path (+ regression test).
- bridge: only append error data fields not already shown as message/code,
so the actionable error text doesn't echo the same code/message twice.
- ClaudeCodeCard: associate the new config-dir/env labels with their inputs
via htmlFor/id.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(ai): make the Claude auth & config section collapsible
The optional "Authentication & config" section now has a collapsible
header (chevron toggle). Collapsed by default to keep the card tidy, but
auto-expanded when the user already has a config directory or env vars set
so existing config isn't hidden. Local UI state, not persisted.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): preserve raw env editor text; don't encourage plaintext secrets
P1: the env textarea was bound to the persisted value, which is the parsed
env re-serialized — so typing a key before its "=" was erased mid-entry
(buildClaudeEnv drops lines without "="). Keep the raw typed text in local
draft state and only resync from the persisted value on genuine external
changes (not our own parse→serialize round-trip).
P2: the env editor persists to localStorage in plaintext (no credential
encryption). Stop suggesting ANTHROPIC_API_KEY in the placeholder and warn
that values are stored in plaintext, steering credentials to the config
directory (a `claude` login) — consistent with keeping Claude auth
CLI/config-owned (#705).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): expand ~ in Claude config directory before passing to the agent
CLAUDE_CONFIG_DIR is handed to the spawned agent as an env var, which is not
shell-expanded — so "~/.claude" was treated as a literal "~" directory.
Expand a leading ~ at consume time (normalizeAgentEnv + getClaudeConfigDir)
rather than on save, so the stored value stays portable across machines
(cloud sync) and each device expands to its own home.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): recreate cached Claude provider when its env config changes
Claude ACP providers are cached per chat session and reused unless one of
the fingerprinted dimensions changes. authFingerprint was null for Claude,
so editing the config directory / env vars in Settings didn't take effect on
an already-running session. Fingerprint the Claude agent env so a config
change invalidates the cached provider and the next turn respawns with it.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(terminal): add per-mode follow-theme resolver and storage keys
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(terminal): persist per-mode follow-theme selections in settings state
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(sync): include per-mode follow terminal themes in cloud sync
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* i18n(terminal): add per-mode follow-theme picker strings
* feat(terminal): add type filter and auto option to theme picker
* feat(terminal): pick dark/light terminal theme when following app theme
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(terminal): don't flag the auto sentinel as a missing theme in the picker
The per-mode follow-theme pickers default to the 'auto' sentinel, which is
not a real theme id, so ThemeList's deletedSelectedTheme check classified it
as a deleted custom theme and rendered a spurious "Missing Theme" banner above
the Auto entry on first open. Exclude TERMINAL_THEME_AUTO from that check.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* i18n(terminal): align zh-CN dark/light wording with app convention
Use 深色/浅色 (matching the theme picker section headers and global
appearance settings) instead of 暗色/亮色 for the per-mode terminal
theme labels, so the picker modal reads consistently.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(terminal): match follow-theme preview fallback to runtime resolution
The per-mode preview memos fell back straight to TERMINAL_THEMES[0] when a
selection resolved to a deleted theme, while the runtime currentTerminalTheme
memo falls back to the manual terminalThemeId first. Mirror the runtime chain
so the Settings preview matches the actual terminal for users with a
non-default manual theme.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
fetchSuggestions ran the full completion pipeline (history scan, fig specs, remote path lookups) on the main thread even when both the popup and ghost text were disabled — the results were then discarded. Add a shouldQueryCompletions(settings) gate and bail out early (clearing any stale state) when neither display mode is on.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The autocomplete hook (useState) lived in Terminal, so every suggestion / selection / live-preview update re-rendered the whole ~2775-line Terminal component. Move the hook and its popup into a dedicated <TerminalAutocomplete> component so those frequent state updates re-render only that small subtree.
The hook's handlers are surfaced back to Terminal via refs (the same refs already used to wire the xterm runtime), and the component is mounted unconditionally so the hook keeps recording command history and intercepting completion keys for the session's lifetime. No behavior change intended.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* perf(terminal): bound the connection log with a chunk ring buffer
The connection log kept the last 1,000,000 chars via `log += chunk; log = log.slice(-MAX)`. Once a session emits more than that, the slice flattens a ~1M-char string on every subsequent output chunk — on the render thread, for each echoed keystroke included — on long/busy sessions.
Replace the string with a small chunk-queue ring buffer that trims only the boundary chunk (amortized O(chunk) append) and materializes the full string once on read. Behavior is unchanged: it still retains exactly the last MAX_CONNECTION_LOG_DATA_CHARS characters.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* perf(terminal): coalesce connection log into bounded blocks (O(1) trim)
The first cut used one array entry per append and trimmed with chunks.shift(). For interactive output (many tiny chunks) the array grows toward the cap in entries, so once full, shift() reindexes ~N elements on every append — O(appends) per chunk, no better than the slice it replaced.
Coalesce appends into a small, bounded set of fixed-size blocks (~maxChars/blockSize). New data fills an open tail that seals into a block at blockSize; trimming only drops/slices a handful of blocks. Adds segmentCount() and a test asserting the segment count stays bounded across many tiny appends.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* perf(terminal): flush shell output on the event-loop turn, not a fixed 8ms timer
SSH/PTY output was coalesced and shipped to the renderer on a fixed 8ms timer. For interactive use that interval is pure added latency: every echoed keystroke waits out the timer before it can paint, so typing feels slightly behind.
Replace the timer with turn-based (setImmediate) coalescing in a single shared ptyOutputBuffer module, used by the SSH, local, telnet, and mosh paths. A single echoed keystroke is now forwarded almost immediately, while data arriving in the same turn still collapses into one IPC send, and a 16KB size cap still forces an immediate flush under heavy output.
Also de-duplicates two copies of the buffering logic (SSH had an inline copy; local/telnet/mosh shared another) and adds unit tests for the buffer.
Related to #1084.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(terminal): drop orphaned flushTimeout reference in SSH close handler
The SSH stream "close" handler still cleared `flushTimeout`, a variable that lived in the inline buffer removed when this path moved to the shared ptyOutputBuffer. Reading it now throws ReferenceError on every channel close, aborting the cleanup and exit signaling. The shared buffer's flush() cancels any pending flush internally, so the timer bookkeeping is removed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat #826: optional Option+←/→ word jump on macOS
Adds a Terminal → Keyboard toggle "Option+←/→ jumps by word" (off by default,
synced). When on, a bare Option+Left/Right sends Meta-b / Meta-f instead of
xterm's default ^[[1;3D / ^[[1;3C, so readline/zle moves by word without
per-host bindkey setup (Termius-style).
The key→sequence mapping is a tested pure function; the handler reads the
setting live (no reconnect) and runs after kitty mode + autocomplete so it
doesn't override them.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix#826: gate Option+←/→ word jump to macOS
The setting is syncable, so without a platform gate, enabling it on a Mac
would also rewrite Alt+←/→ to Meta-b/f on synced Linux/Windows devices,
breaking apps/shells that expect the default ^[[1;3D / ^[[1;3C. Pass
isMacPlatform() into the mapping so it only applies on macOS; add a test
for the non-macOS case.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix#1079: preserve remote file mode when rz overwrites a same-named file
#1070's overwrite path rm's the remote file and lets rz re-create it, which
writes with the remote umask and drops the original permission bits — e.g. a
0755 script became 0644 after choosing "replace". (It didn't happen before
because rz used to skip same-named files, leaving the original untouched.)
Capture each conflicting file's mode during the pre-upload probe
(stat -c %a, BSD stat -f %Lp fallback) and chmod it back once the transfer
finishes and the files are on disk. Restore is best-effort: any failure
silently falls back to today's behavior.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix#1079: probe file mode with `stat -- "$n"` for dash-prefixed names
Without `--`, `stat -c %a "-x.sh"` (and the BSD `-f %Lp` fallback) parse a
leading-dash filename as options, so the mode was never captured and overwrite
fell back to rz defaults — losing permission preservation for a valid filename
class. Mirrors the existing `rm -f --` handling. (chmod left as-is: its path is
always absolute, and BSD chmod doesn't accept `--`.)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
"Use Option as Meta key" was read into `altIsMeta` but only applied to the
mouse alt-click options (`altClickMovesCursor`). xterm.js's `macOptionIsMeta`
— the option that actually makes Option emit ESC-prefixed (Meta) sequences —
was never set, so on macOS Option kept producing layout characters (ƒ, ∫, …)
and readline/zle word shortcuts (Alt+f, Alt+b, Alt+Backspace) were dead.
Extract the altAsMeta→xterm mapping into one tested helper used by both the
terminal init path (createXTermRuntime) and the live settings sync
(Terminal.tsx) so the two can't drift again.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Drop the manualChunks 'vendor-react' entry: react/react-dom already land
in another chunk, so it only ever produced an empty chunk + a build
warning, with no caching benefit.
- Import domain/syncMerge statically in useAutoSync. It's already in the
eager graph via CloudSyncManager's static import, so the dynamic
`import()` couldn't be code-split anyway and only emitted a mixed
static/dynamic-import warning.
No behavior change; production build is warning-free.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Root cause of the persistent split-view 花屏: xterm's WebGL addon shares
ONE TextureAtlas across terminal instances with equal config (font / size
/ theme / DPR) — acquireTextureAtlas does `if (configEquals) { ownedBy.push;
return atlas }`. Two split panes then share an atlas, so the
clearTextureAtlas calls netcatty makes to recover from glyph corruption
(on resize / DPR / font change / tab show, from #1049 and #1066) clobber
the *other* pane's rendering. That's why the earlier redraw/clear-based
recovery attempts didn't help and only bounced the garble between panes.
Disable the sharing: remove the "reuse a matching atlas" loop so every
terminal creates its own atlas. The published bundle is minified, so this
is done with a small idempotent postinstall script (a patch-package patch
would be a ~550KB unreadable blob of the whole minified line). It
string-replaces the exact loop in the CJS + ESM builds, runs after
patch-package, and warns without failing if @xterm/addon-webgl changes.
Verified: split-view WebGL no longer garbles; script is idempotent
(patched=2 → already=2) and the production build is unaffected.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat #1064: add buildUploadPlan for rz overwrite/skip/cancel resolution
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat #1064: handle remote filename conflicts in rz handleUpload
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat #1064: SSH exec probe + remove for rz upload conflicts
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat #1064: IPC for rz overwrite-conflict prompt
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat #1064: renderer prompt for rz overwrite conflicts
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix#1064: repair sshBridge test mock (ipcMain.on) and i18n the overwrite dialog
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix#1064: make upload plan index-based to preserve per-file decisions
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix#1065: resolve terminal cwd through su/sudo for the SFTP locate
The SFTP "locate to terminal's current directory" feature kept showing the
login user's home (e.g. /root) after the user switched accounts with su /
sudo -s and cd'd elsewhere.
getSessionPwd walks the remote process tree from a sibling exec channel to
find the interactive shell's cwd, but it only followed children whose comm
is a shell name (bash/zsh/...). su and sudo are named "su"/"sudo", so the
walk stopped at the login shell and read its cwd. The actual shell the user
is typing in lives *under* su/sudo as the controlling tty's foreground
process group.
Rewrite the walk to pick the deepest foreground shell ("+" in stat) within
the login shell's whole process subtree, which transparently follows
through su/sudo to the active shell, falling back to the login shell when
no foreground shell is found.
Verified on a real server (root -> su user -> cd /tmp):
before: /root after: /tmp
and confirmed the no-su case is unchanged (cd /var -> /var).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Fall back to login shell cwd when the active shell's /proc is unreadable (Codex review)
When an unprivileged user runs `sudo -s` / `su root`, find_active_shell
correctly selects the root-owned foreground shell, but the exec channel
(running as the login user) cannot readlink another uid's /proc/<pid>/cwd
due to ptrace permissions. Without a fallback the script dropped straight
to the home directory, regressing user→root sessions.
Retry readlink on the same-uid login shell before falling back to home.
Verified live (user -> cd /var -> sudo -s -> cd /tmp): the root shell's
cwd is unreadable, and the result is now /var (login shell cwd) instead of
/home/<user>.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Select the interactive (tty-bearing) login shell deterministically (Codex review)
find_login_shell picked the first shell child of sshd and exited, but ps
output is unsorted, so when other exec channels (server-stats polls, etc.)
are running on the same connection their transient sh could be chosen,
making find_active_shell walk the wrong subtree.
Prefer the shell child that has a controlling tty: the interactive shell
has a pts, while non-PTY probe exec channels have tty "?". This is
deterministic regardless of ps order, in both the su and no-su cases (the
old "prefer foreground" heuristic was itself nondeterministic under su).
Falls back to any shell child if none has a tty.
Verified live with a concurrent no-tty `sh -c sleep` under the same sshd:
the pts/0 bash is selected and the result is /tmp, not the probe shell.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix#1062: treat SSH shell TMOUT auto-logout as a timeout, not a normal exit
A shell-level TMOUT idle auto-logout makes bash/csh exit cleanly (numeric
exit code, no signal), which is byte-for-byte indistinguishable from a
user-typed `exit` at the SSH protocol level. PR #1057 keyed the
close-vs-keep decision on `streamExited` (numeric code + no signal), so
TMOUT exits were reported as reason "exited" and the tab was auto-closed —
reintroducing the problem from #977.
Verified against a real server that bash TMOUT exits with code 0 / no
signal and prints "timed out waiting for input: auto-logout" to the
channel before it closes. Since exit code/signal can't distinguish it from
an intentional exit, detect that banner in the session's existing rolling
output tail (_promptTrackTail) and report reason "timeout" instead, which
routes to the existing markDisconnected path (keep tab + reconnect). A
normal `exit`/`logout` (no "auto-" prefix) still auto-closes the tab, so
PR #1057's behavior is preserved.
zsh's TMOUT raises SIGALRM (a signal), so it already took the
keep-tab/reconnect path and is unaffected.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Anchor TMOUT auto-logout match to the banner's final line (Codex review)
The detector matched "auto-logout" as an unanchored substring within the
last 256 chars, so command output that merely mentions it (e.g. `grep
auto-logout /etc/profile` while investigating TMOUT) followed by an
intentional `exit` could be misclassified as a timeout and wrongly keep
the tab open. Anchor on the final non-empty line of output instead — the
banner the shell prints right before exiting — which loses no true
positives (verified against the real-server output shape) while rejecting
mid-stream mentions.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Hidden tabs stay mounted off-screen (visibility:hidden) so each keeps a
live WebGL context. Creating another terminal's WebGL context — or the GPU
dropping a non-composited off-screen canvas — leaves the hidden terminals'
drawing buffers corrupted ("花屏"). This reproduces on both Windows and
macOS: opening 2 tabs garbles the 1st, opening 3 garbles the 1st and 2nd,
while the just-created (visible) one is always fine. The DOM renderer is
immune because it uses real DOM nodes.
A window resize recovers the display because it triggers a full repaint
(clearTextureAtlas + RenderService._renderRows). A tab switch did not:
the visibility effect only calls safeFit, which early-returns when the
pane's dimensions are unchanged, so no redraw happened.
Perform the same recovery a resize does when a tab becomes visible:
clear the texture atlas (no-op on the DOM renderer) and synchronously
repaint every row. Verified against xterm core that _renderRows draws
unconditionally, independent of dimension changes or dirty-row tracking.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat #1005: add live-preview keystroke calculator for popup autocomplete
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat #1005: live-render the selected popup suggestion on arrow navigation
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat #1005: free Tab for the shell; Enter runs the rendered line; Esc reverts
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat #1005: show key hint (→ expand / ↵ run) on the selected popup row
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat #1005: live-render full path while navigating sub-directory panels
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test #1005: move live-preview test into the npm test glob
The test runner only scans components/terminal/*.test.ts (not the
autocomplete/ subdir), matching where the other autocomplete-module tests
live (e.g. completionEngine.test.ts). Relocate so it actually runs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix#1005: center and refine the popup key-cap hint
Use inline-flex centering (the ↵ glyph was vertically off with line-height +
padding), softer color-mixed border/background, a system-sans font so the
glyph renders consistently regardless of the terminal font, and the more
balanced ⏎ return symbol.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix#1005: record the actual executed line on Enter, not the stale suggestion
Codex review (P2): the popup Enter handler recorded selected.text and
suppressed handleInput's recorder, so editing a previewed command (select
docker, type ' ps', Enter before the re-query) logged the stale 'docker'
instead of 'docker ps'. Delegate to handleInput's Enter path, which records
lastAcceptedCommandRef on a clean select and falls back to the live buffer
after an edit (typing nulls that ref).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix#1005: don't revert user edits when Escape closes the popup
Codex review (P2): previewActiveRef stayed true after the user edited a
previewed command, so Escape (before the debounced re-query reset state)
called renderPreviewSelection(-1) and rewrote the line back to the stale
baseline, dropping the edits. Clear previewActiveRef when the user types
(alongside the existing lastAcceptedCommandRef reset), so Escape only reverts
a pristine preview and otherwise just dismisses the popup.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Middle-clicking a tab (mouse wheel click) is a conventional "close tab"
gesture in browsers and editors. Wire it to every closeable tab strip:
the top session / workspace / log-view / editor tabs and the SFTP tab bar.
A small shared helper (lib/tabInteractions.ts) handles the gesture:
onAuxClick closes the tab when button === 1, and onMouseDown calls
preventDefault for the middle button so the Chromium/Electron autoscroll
overlay does not appear. Left-click activation and right-click context
menus are untouched.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Heavy full-screen TUIs (claude code / gemini cli / opencode), font changes,
and device pixel ratio changes can leave xterm.js's WebGL glyph texture atlas
in a corrupted state that persists for the life of the terminal — users see
persistent "garbled / 花屏" output that only clears when a brand-new terminal
is opened (most often on Windows with display scaling / multi-monitor setups).
Clear the texture atlas so glyphs re-rasterize at the correct scale instead of
forcing users to reopen the terminal:
- Add watchDevicePixelRatio() helper (TDD, unit-tested) that re-registers a
matchMedia listener across DPI changes and fires a repair callback.
- Wire it into createXTermRuntime: on devicePixelRatio change, clear the atlas
and refit; also clear the atlas on reflow (term.onResize). Watcher is torn
down on dispose.
- Expose clearTextureAtlas() on XTermRuntime and call it after font changes in
Terminal.tsx (xterm.js #3280). All calls are no-ops under the DOM renderer.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Inline (ghost-text) suggestions render suggestion.substring(trackedInput.length)
after the cursor, where trackedInput is a client-side reconstruction of the
command line (buffer heuristics + keystroke prediction, to mask SSH echo
latency). On hosts with non-standard echo — hardware bastion hosts / network OS
like `ecOS#` (#1013, previously #756 / #906) — that reconstruction drifts and
the ghost gets painted over characters the user already typed (`int` + ghost
`terface` -> `intterface`).
Add a fail-safe consistency check: on each post-echo render, if the real
terminal line before the cursor contains the tracked input followed by more
untracked, non-whitespace characters (reality is AHEAD of what we tracked),
hide the ghost instead of drawing it over real text. SSH echo latency is the
opposite case (the line is a prefix-behind of the tracked input) and is
deliberately not flagged, so the ghost stays responsive on slow links. The
check is ASCII-only (wide-char column mapping is ambiguous) and fail-open, so
it can only ever suppress a ghost that would otherwise corrupt — never change
correct behaviour.
This converts the recurring "ghost shows already-typed characters" bug into
"ghost simply doesn't show" on devices we can't track reliably.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
macOS keys the "Local Network" privacy permission on the main executable's
Mach-O LC_UUID (Apple TN3179). Electron's prebuilt binary is linked with LLD,
which derives the UUID from a content hash, so every app built from the same
Electron version ships the *same* LC_UUID even with a different bundle id. That
collision makes the grant unreliable: a user who enables Local Network for
Netcatty can still hit `connect EHOSTUNREACH` on LAN / VMware host-only
addresses, while loopback-forwarded connections work.
Add an electron-builder afterPack hook that rewrites the packaged macOS
executable's LC_UUID to a value derived deterministically from the appId —
stable across builds (so the grant survives updates) but distinct from every
other app. It runs before code signing, so signature/notarization cover the
patched binary. No-op on Windows/Linux.
Verified the rewrite on a copy of Electron's binary (LC_UUID changes, file
stays a valid Mach-O, deterministic) and added unit tests for the Mach-O
patcher (thin + fat) and the UUID derivation.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Electron's BoringSSL dropped several standard MODP groups from the named
crypto.createDiffieHellmanGroup() API — notably the 1024-bit Oakley Group 2
(modp2) that backs SSH's diffie-hellman-group1-sha1. ssh2 calls
createDiffieHellmanGroup('modp2') for that kex, so connecting to legacy
network devices that only speak group1-sha1 failed with "Error: Unknown DH
group".
The underlying DH math still works on BoringSSL via createDiffieHellman()
with an explicit prime, so add a compatibility shim that wraps
createDiffieHellmanGroup and falls back to the well-known prime constants
when (and only when) the runtime can't resolve a group by name. On OpenSSL
builds the original call succeeds and the fallback is never used.
The shim is installed in main.cjs before any ssh2-using bridge loads, since
ssh2 destructures createDiffieHellmanGroup at module load. Once installed,
the existing legacy-group probe detects modp2 as supported again and offers
group1-sha1, so affected devices actually connect (still gated behind the
per-host legacy-algorithms toggle).
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The fixed-DH-group support probe called crypto.createDiffieHellmanGroup()
for each MODP group to feature-detect runtime support. Under Electron's
BoringSSL, instantiating the large groups is pathologically slow
(modp18/8192-bit takes ~20s on first call), and the result is only cached
in-process, so the first connection after every app launch froze for ~24s.
The standard modern groups (modp14/16/18) are universally supported and
always pass the probe anyway, so treat them as supported without probing.
Only groups a runtime may genuinely drop (e.g. BoringSSL removed the weak
1024-bit group1/modp2) are still feature-detected; those fail instantly.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>