* feat(ai): support Cursor CLI login for subscription quota
Allow the managed Cursor agent to authenticate via local `agent` CLI
login as an alternative to CURSOR_API_KEY, so Catty can use subscription
Auto quota. Keep API key and CLI login mutually exclusive, isolate
resume sessions by auth mode, and expose selectable CLI models with
auto as the default.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(ai): harden Cursor CLI login path after review
Address multi-agent review findings: soft-cancel abort without late
stream errors, prefer cursor-agent over ambiguous agent binaries,
mode-aware availability without wiping API keys or sticky-disabling,
and concurrent MCP merge refcount restore. Add unit coverage for the
new contracts.
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: bincxz <16399091+binaricat@users.noreply.github.com>
* fix(ai): stop OpenCode process spawn storms when unused (#2184)
OpenCode model catalog loads were spawning full servers for discovered
agents, skipping main-process cache, and not aborting on timeout—so
idle Netcatty could accumulate many opencode processes.
Only load catalogs for the selected agent, cache/coalesce list-models,
abort and tear down on timeout, and share a short-lived list-models
server pool per binary.
* fix(ai): address OpenCode list-models review findings
Restore the main-process env immediately after spawning OpenCode so an
idle list-models server cannot leave PATH/OPENCODE_BIN on a temp shim.
Key model caches and the list-models pool by catalog-affecting env so
profiles that share a binary do not receive each other's catalogs.
* fix(ai): tighten OpenCode model-cache invalidation
Include profile-affecting env in the renderer runtime model cache key so
OpenCode config switches do not reuse a stale catalog. Skip caching empty
list-models results in the main process so transient OpenCode failures can
recover before TTL expiry.
Store probed --version on ExternalAgentConfig from settings resolve-cli
and discovery enable, and prefer that when gating GPT-5.6 so custom
paths not on PATH still unlock Sol/Terra/Luna when compatible.
Netcatty is BYO-CLI: upgrading codex-sdk does not replace the user's
binary. Hide Sol/Terra/Luna when discovery reports Codex CLI < 0.144.0
so old installs do not auto-default to an unsupported model.
Address PR #2170 Codex review: stop auto-selecting the last thinking
level (which defaulted Sol to ultra), add defaultThinkingLevel from the
official catalog, and keep gpt-5.2 selectable so stored model IDs do not
silently fall back to Sol.
Expose snippets and automation scripts to Catty, MCP, and CLI with full run lifecycle, vault bridge CRUD, and clickable tool-result cards in the chat message list.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Improve Catty context budget, step guards, and compaction reliability.
Anchors session state across compactions and surfaces compaction to the UI so long tool-heavy turns stay within context limits.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix Codex review: gate reinjection and match sftp_read_file.
Only reinject session context after actual compaction adjustments, and recognize the generated sftp_read_file tool name in stale read pruning.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Scope stale pruning by session for SFTP reads and terminals.
Include sessionId in read fingerprints and keep the last two terminal outputs per session instead of globally across the conversation.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Cap output reserve for small context windows in budget math.
Prevent default 4096-token output reserve from collapsing compaction thresholds on 4k/8k models into near-zero values.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Gate terminal stale pruning on budget pressure and clear old assistant status.
Only omit older terminal_execute output when compaction budget is tight, and clear step-compaction statusText from the prior assistant when a new assistant message starts after tool results.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Gate SFTP read supersession on budget pressure like terminal pruning.
Preserve earlier file reads during read-edit-reread compare flows until context compaction is actually needed.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Use compaction summary budget for pre-summarize pruning.
Reserve the 1600-token summary output when fitting messages for LLM compaction, and skip superseding successful reads when a later read failed.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Prune oldest messages first when fitting compaction summarize input.
Keep the portion of pre-tail history closest to the current task instead of dropping recent summarize candidates from the end.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Detect JSON tool errors before superseding stale reads.
Treat structured error payloads as failed reads so a later SFTP failure does not replace the last successful file contents.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Show compaction as a transient inline shimmer instead of a banner.
Emit compaction_start/compaction events and render a single ephemeral status line in the message list that clears when compaction finishes.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(ai): unify capability registry, harness tools, and agent runtime
Centralize MCP/CLI/Catty tool surfaces in the capability catalog with codegen,
move Catty-only harness tools into catalog with local execution, converge
mcpServerBridge builtin dispatch via registry, and wire AgentRuntime with
vault bridge, permission grants, and Round 2 catalog parity.
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(ai): vault host tools, artifact cards, and external agent vault guidance
Unify vault host create/import with capability RPC routing, add chat artifact cards with Vault navigation, and align external MCP agents with Catty rules so host creation never falls back to Vault Notes.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(ai): address Bugbot findings for streaming, vault bridge, and grants
Clear Catty streaming by chat sessionId, read latest vault state via getters on agent mutations, and stack grant persisters so unmounting one AI panel does not break others.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(vault): clear pending openNoteId after chat navigation focus
NotesManager notifies VaultView once the one-shot openNoteId request is applied so the id does not stick for the rest of the session.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(vault): serialize vault agent IPC mutations to prevent lost writes
Queue vault agent bridge handlers so concurrent hosts.create or note.create tool calls cannot overwrite each other via stale read-modify-write snapshots.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(vault): keep synchronous vault snapshot for agent IPC reads
Mirror hosts, notes, and groups in a ref updated immediately on agent writes so serialized mutations see the latest state before React re-renders.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(ai): sync Catty MCP sessions and defer runtime cleanup on delete
Register terminal sessions with the main process before Catty tool calls, and clear ToolOutputStore only after stopping an active turn when deleting a chat session.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(ai): avoid deleting a newer abort controller on late stop
Only remove the session abort-controller entry when stop finishes if it still refers to the controller that was stopped.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(ai): sync vault snapshot on render and await active turn on stop
Refresh vault agent reads when React props change without clobbering in-flight agent writes, and wait for runTurn to finish before clearing ToolOutputStore on session delete.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(ai): reload permission grants from storage before auto-approve
Re-read persisted grant rules on each approval check so revocations in Settings take effect without requiring a chat hook remount.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(ai): sync permission grants to main process immediately on write
Push MCP permissionGrantsSnapshot as soon as Always allow persists a rule so the next tool call in the same agent turn sees the grant without waiting for useEffect.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Route Netcatty context through body.system, filter echoed user parts,
keep OpenCode env/shim alive for server lifetime, and cap model preset dropdown height.
Co-authored-by: Cursor <cursoragent@cursor.com>
Fixes#1382
Deleting the active AI chat session left panelViewByScope pointing at a removed session, causing the side panel UI to blank. Sync panel view cleanup with session deletion, stop in-flight streams on delete, and return to draft with the deleted session's agent preserved.
Replace immersive instant-switch with animated active chrome theme sync so
top tabs match terminal sessions immediately on tab click, and clamp
autocomplete popups to the active pane so they stay anchored to the cursor
in split workspaces.
Co-authored-by: Cursor <cursoragent@cursor.com>
* perf(terminal): smooth layout drags and faster tab switching
Defer xterm refit during split, sidebar, and host-tree drags while keeping pane containers in sync with live layout measurements. Refactor TerminalLayer into focused sections with TabBridge/memo optimizations and add the terminal host tree sidebar.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(terminal): keep side panels alive and guard session attach races
Prevent terminal boot unmount from leaking backend sessions, keep SFTP/scripts/theme/AI state when switching side tabs, and defer heavy SFTP UI mount so first entry stays responsive.
Co-authored-by: Cursor <cursoragent@cursor.com>
* perf(terminal): reduce tab switch jank
---------
Co-authored-by: Cursor <cursoragent@cursor.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>
* Fix#954: unify Tooltip styling + replace native selects
Replace native HTML title= tooltips and native <select> dropdowns
with the existing Radix-based Tooltip / Select components so they
share the app's rounded styling, theme tokens and i18n pipeline.
Adds a global TooltipProvider in AppWithProviders so every
descendant Tooltip works without a per-file Provider wrapper.
Scope (driven by the issue #954 examples and "全部都处理" follow-up):
- TerminalLayer toolbar: Add Terminal / Split View / SFTP / Scripts
/ Theme / AI Chat / Move panel / Close panel.
- TopTabs middle bar: quick switcher, more tabs, AI assistant, theme
toggle, settings; window-control buttons (min/max/close), tray
close and hotkey reset/disable have their native title dropped per
the user's explicit opt-out ("可以不用Tooltip,直接全局禁用
原生title 属性").
- AI panels: AIChatSidePanel session history / new chat / delete,
ConversationExport, AgentSelector, ChatInput attach / expand /
permission, ModelSelector, ProviderCard, ai-elements/tool-call.
- SFTP: SftpSidePanel header, SftpBreadcrumb, SftpFileRow,
SftpPaneToolbar, SftpTabBar, SftpTransferQueue.
- Settings: SettingsPage close, SettingsAppearanceTab theme/accent
swatches, SettingsFileAssociationsTab edit/remove, SettingsSystemTab
crash-log paths and global hotkey reset.
- Host vault: HostDetailsPanel (clear / suggestions / show-password /
key path / browse key), GroupDetailsPanel, KnownHostsManager,
ConnectionLogsManager, KeychainManager, SyncStatusButton,
CloudSyncSettings, LogView, QuickSwitcher, ScriptsSidePanel,
Terminal status bar copy-host + broadcast/focus, ZmodemProgressIndicator.
- Terminal subcomponents: HostKeywordHighlightPopover, TerminalComposeBar,
TerminalConnectionDialog, TerminalSearchBar.
- Editor: TextEditorPane (subtitle, search, wrap, promote-to-tab).
- TrayPanel session rows and port-forwarding rows.
Native <select> migrated to custom Select component:
- SerialConnectModal (data bits, stop bits, parity, flow control)
- SerialHostDetailsPanel (same four fields)
- HostDetailsPanel backspace behavior
- GroupDetailsPanel backspace behavior
- SettingsTerminalTab local shell picker
- terminal/ThemeSidePanel font weight
Hardcoded English strings extracted to i18n. New keys for both
en and zh-CN: terminal.layer.*, topTabs.*, ai.chat.* (sessionHistory,
attach, collapse, expand, enableAgent), zmodem.*, settings.shortcuts.
resetToDefault. Inline help text on SnippetsManager package-name input
removed because the same hint is already shown in a visible <p> below
the input.
Existing per-file <TooltipProvider> wrappers (SnippetsManager,
ScriptsSidePanel, SelectHostPanel, RuleCard, HostDetailsPanel proxy
section) are left in place — they nest harmlessly under the global
provider and stay self-sufficient for component tests.
Tests:
- tsc clean for changed files (pre-existing repo-wide errors
unrelated to this PR).
- All 802 tests pass (3 skipped pre-existing).
- HostDetailsPanel.proxyProfile.test and TextEditorPane.test
updated to wrap with TooltipProvider, matching the runtime
context now needed by the migrated components.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Fix#954: wrap Settings + Tray windows with TooltipProvider
Settings and the tray panel mount as separate Electron windows with
their own React root in index.tsx, so they do not inherit the global
TooltipProvider added under AppWithProviders. After the unified
Tooltip migration, any settings tab that used a Tooltip (Appearance,
Application, FileAssociations, System, Shortcuts, Terminal, AI
ProviderCard, AI ModelSelector) — and TrayPanel — threw
"Tooltip must be used within TooltipProvider" and rendered nothing.
Wrap both branches with TooltipProvider at the same level as
ToastProvider in index.tsx.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Preload compact history on first turn after app restart (#753 hedge)
Symptom (confirmed on Copilot CLI, originally reported on Codex in
#753): after closing and reopening Netcatty, the AI chat UI still
shows the prior conversation but the agent responds "this is the
beginning of our conversation, no previous records". Earlier context
is lost entirely.
Root cause: the bridge relied on session/load throwing "not found" to
trigger the catch-block fallback that replays compact history. Some
ACP agents (Copilot CLI, some Codex builds) silently spawn a new
session when handed a stale id instead of erroring. The catch-block
never fires → historyReplayFallback stays false → the first turn
sends only the latest prompt → agent sees zero context.
Fix: when we're creating a new provider process AND telling it to
resume an existing session id AND the renderer gave us compact
history, preload historyReplayFallback=true as a hedge. If the agent
really did reload the session, the replay is ~3KB of redundant
context (small waste). If the agent silently started fresh, the
replay restores durable constraints + last few raw turns so the
first response is coherent.
After the first successful streamed turn clears the flag (the round-2
post-stream hook), steady state is back to sending only the latest
prompt. Cost is bounded to one replay per app-restart-and-prompt.
Test: "replays compact history on the first turn after app restart
even when session/load 'succeeds'" — mocks createACPProvider to
behave like Copilot CLI (no error thrown, no real resume), asserts
the first streamText call carries history+latest (length 2) and the
second only latest (length 1).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Fix AI session resume and agent switching
* Preserve hidden draft when switching agents
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Fix ACP history replay and compaction
* Fix PR keyword importance matching
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address codex review on #754: preserve short constraints + cancel-clear
Two recovery-path regressions flagged by codex review:
1. Compact ACP history dropped short load-bearing user constraints
(acpHistory.ts:55). The blanket length<10 rule treated short
non-trivial messages like "Use ssh2" or "中文输出" as filler,
while longer generic follow-ups still ate the budget. After
stale-session recovery the fresh ACP session would resume without
constraints that were present in the original chat. Removed the
length heuristic; the TRIVIAL_USER_MESSAGE_PATTERNS regex already
filters actual filler ("ok", "yes", "继续", "thanks").
2. historyReplayFallback was only cleared on non-aborted streams
(aiBridge.cjs:2837). If the user stopped the first turn after
stale-session recovery, the flag stayed set. The next turn would
then trigger shouldResetProviderForHistoryReplay, discard the
freshly recovered ACP session (resumeSessionId is forced to
undefined in that path), and re-spend tokens on another compact
replay — breaking the cancel-preserves-session contract. Now we
also clear on abort; the empty-but-not-aborted retry path in the
if-branch above is unchanged.
Tests:
- New test in acpHistory.test.ts asserts "Use ssh2" / "中文输出"
survive when pushed outside the recent raw window
- New test asserts "ok" / "继续" still drop (sanity check that the
trivial regex still does its job without the length backstop)
- Updated "does not treat pr inside ordinary words as important" to
no longer assert that approach/improve/prepare are absent — the
test's real intent (priority-2 line still wins) is preserved by
the 不要提交 assertion
- New test in aiBridge.test.cjs simulates a user cancelling the first
turn after recovery and verifies the next turn reuses the
recovered session (no extra provider creation, no re-replay)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Address codex re-review: preserve replay flag across orthogonal recreation + keep tool output in raw window
Two more P2 regressions flagged on the second review pass:
1. historyReplayFallback was only carried over in the reset-for-replay
branch of the provider recreation path. An orthogonal change between
an empty recovered turn and its retry — a permission-mode toggle,
MCP scope/fingerprint flip, or auth rotation — would flip
shouldReuseProvider to false, enter the !shouldReuseProvider branch,
and drop the flag because preserveHistoryReplayFallback only covered
the shouldResetProviderForHistoryReplay case. The next turn then
sent only the latest prompt and lost the recovered conversation.
Now the flag is preserved on any recreation where a replay is still
pending.
2. Tool messages didn't flow through toRawHistoryMessage at all, so on
stale-session recovery they only survived as the 500-char compact
summary in summarizeToolMessage. Any follow-up referencing the last
tool output ("use that output", "what did cat show?") lost the
actual bytes when they exceeded the compact cap. Now tool results
travel through the recent raw window up to MAX_RAW_MESSAGE_CHARS
(2000), flattened to the "assistant" role since ACP only accepts
user/assistant.
Tests:
- aiBridge.test.cjs: new "preserves history-replay across provider
recreation caused by permission-mode / MCP / auth change" —
exercises the gap via a permission-mode toggle between an empty
recovered turn and its retry. Extends mock to support a dynamic
getPermissionMode.
- acpHistory.test.ts: new "preserves recent tool results verbatim" —
pushes a ~1500-char tool output through the pipeline and asserts the
replay still contains enough bytes to exceed the 500-char compact
cap.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Address codex round 3: inline tool_call context + bound durable scan
Two findings from the third codex review pass, both legitimate:
1. [P2] When the raw window starts mid-tool-interaction, the preceding
assistant tool_call message can fall outside the 6-item slice while
the tool_result stays in. Without the call's name+arguments, the
result was opaque bytes and follow-ups like "use that output" had
no provenance. The compact pass only preserved calls that matched
IMPORTANT_PATTERNS, so read_file / grep / terminal_exec were
silently dropped.
Fix: build a toolCallId → { name, arguments } index from every
assistant message and inline a `[from <name>(<args>)]` label next
to each Tool result line in the raw window. Args are truncated to
MAX_TOOL_CALL_LABEL_CHARS (200) so a verbose JSON payload can't eat
the entire raw budget.
2. [P3] buildCompactContext scanned messages.entries() over the full
transcript for durable-user/assistant candidates, even though
MAX_MESSAGES_TO_SCAN (20) suggested the path was meant to be
bounded. On a long ACP chat, every send did O(N) regex work plus
an O(N log N) sort — the very chat-length-dependent latency the
token-compaction PR was meant to address.
Fix: introduce MAX_DURABLE_SCAN_MESSAGES (200) and restrict the
durable scan to that tail. 200 is large enough to cover realistic
sessions (99th-percentile chats are << 200 turns) while giving a
constant-time worst case. Constraints older than the window age
out of the compact replay; the live ACP provider's own persisted
session still carries them when it can resume, which is the
common path.
Tests:
- "inlines tool_call name+args so tool_result is interpretable without
the preceding assistant turn" — pushes the tool_call out of the raw
window and asserts the result line carries [from <tool>(<args>)].
- "bounds the durable-candidate scan to avoid O(N) work per send on
long chats" — builds a 600+ message chat with an ancient priority-2
constraint outside the scan window and a recent one inside; asserts
only the recent one survives.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Address codex round 4: preserve short assistant decisions + provenance on older tool results
Two P2 findings from the fourth codex pass, both mirror-images of earlier
fixes on a different code path:
1. Short assistant decisions dropped from compact replay
(acpHistory.ts:75-83). isSubstantiveAssistantMessage required length
>= 40 OR a small English keyword match OR a numbered list. Short but
load-bearing replies like "Use ssh2", "rebase instead", "中文输出"
satisfied none of those and were silently dropped from the durable-
assistant compact section. Once they fell outside the 6-item raw
window, "do what you suggested earlier" would replay only the user
question without the assistant's actual decision.
Fix: mirror the user-side loosening — drop the length/keyword gate,
rely on TRIVIAL_ASSISTANT_MESSAGE_PATTERNS to filter actual filler
("ok", "ack", "got it", "明白").
2. Older tool results lost provenance (acpHistory.ts:108-114). The
raw-window fix (round 3) only covered the last 6 items. Once a tool
result fell into the compact section via summarizeToolMessage, the
paired assistant tool_call was usually gone too, so multiple older
outputs surfaced as indistinguishable "Tool result (callN): ...".
Follow-ups like "use the resolv.conf output" had no way to map to
the right call.
Fix: plumb the toolCallIndex through summarizeMessage →
summarizeToolMessage and inline `[from <name>(<args>)]` labels in
the compact section too, the same shape the raw window uses.
Tests:
- New: preserves short non-trivial assistant decisions that miss the
keyword heuristic (Use ssh2 / 中文输出 / rebase instead)
- New: still drops trivial assistant filler like 'ack' / 'ok' / '明白'
- New: inlines tool_call context on OLDER summarized tool results
- Updated earlier raw-window tool regex tests to match the [from X(Y)]
shape ([^)] was failing to cross the args JSON's closing paren)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Address codex round 5: de-dup raw ∩ compact + wire userSkills test into npm test
[P2] The scanned loop (last 20) overlaps with recentRaw (last 6), so
without a raw-window skip in the summarizeMessage path the same last-6
turns were summarized into the compact section AND appended verbatim
in the raw section. Important user turns and large tool output paid
the budget twice — eating into the 3k compact cap and crowding out
older durable context the replay is meant to preserve. Added the
same recentRawSourceIds skip the durable-user / durable-assistant
passes already use, and a regression test that asserts markers inside
the raw window don't surface in compact while still appearing in raw.
[P3] electron/bridges/ai/userSkills.test.cjs (added by this PR) sat
in a subdirectory that the default "npm test" glob
(electron/bridges/*.test.cjs) didn't pick up. The new routing /
index-budget regressions would never run locally or in CI until
someone noticed. Extended the glob to also match
electron/bridges/*/*.test.cjs; the userSkills tests are now included
in the 148-test run.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Address codex round 6: cancel+immediate-send race + tool-call id collision
Two P2 regressions in the recovery path:
1. If the user clicks Stop and immediately sends the next prompt, the
new stream handler's existingRun path unconditionally called
cleanupAcpProvider — destroying the fresh ACP session the cancel
IPC had just promised to preserve. The round-2 clear-on-abort
fix ran too late (in post-stream code) to help, because the new
stream can arrive before the aborted stream fully unwinds. In
that common timing window the follow-up still started from a
bare provider and lost all recovered conversation state.
Fix: (a) cancel IPC now synchronously clears
historyReplayFallback on the preserved provider entry, so the
next stream can't trigger shouldResetProviderForHistoryReplay
and tear the session down via that path; (b) the existingRun
path skips cleanupAcpProvider when the prior run was already
cancelled via the cancel IPC (captured via existingRun.cancelRequested
before we overwrite it). True interrupt-and-restart without an
explicit cancel still falls back to the old clean-slate behavior.
2. The tool-call provenance index used raw toolCall.id as the key.
Nothing in ChatMessage or the ACP event path enforces per-chat
unique ids, so a provider reusing "call1" across turns would
overwrite the older entry and mis-label older tool results
(e.g., an /etc/hosts result annotated as /etc/resolv.conf in
the compact summary). That makes stale-session recovery
misleading whenever a follow-up refers back to an earlier tool
output.
Fix: key the index by `${toolResultMessageId}:${toolCallId}` and
walk the message stream in order, resolving each tool_result to
the most recent preceding assistant tool_call with matching id.
Each result keeps its own historically-correct label regardless
of later id reuse.
Tests:
- aiBridge: "preserves recovered ACP session when user cancels then
immediately sends the next prompt" — fires the next stream request
after cancel but BEFORE releasing the first stream's blocked read,
asserts providerCreationArgs.length stays at 2 (no third creation)
and the second turn sends only the latest prompt.
- acpHistory: "resolves tool_call provenance correctly when tool ids
are reused across turns" — two interactions sharing id "call1",
asserts each tool_result carries its own call's args label.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Address codex round 7: turn-based scan bound + single-pass history build
Two P2 regressions in long-chat / tool-heavy recovery paths:
1. MAX_DURABLE_SCAN_MESSAGES (200) bounded the scan by raw message
count. ACP tool interactions store the user turn, assistant
tool_call turn, and each tool_result as separate messages, so a
tool-heavy chat can produce 5+ messages per logical turn. 200
messages could be only 30-40 user turns — early constraints
like "不要提交" from turn 5 fell out of the compact replay long
before the turn count justified aging them out.
Fix: bound by MAX_DURABLE_SCAN_TURNS (100 user turns) instead.
Walk backwards from the end and stop after seeing 100 user
messages. Realistic tool-heavy 30-turn chats now keep their
early constraints alive, while true 100+ turn chats still
benefit from the bound.
2. buildToolCallIndex(messages) and messages.flatMap(...).slice(-6)
both walked the entire transcript on every send, even after the
bounded compaction window landed. Compaction's stated purpose
was to remove chat-length-dependent latency, but these per-send
linear passes kept it.
Fix: compute the scan start once via computeDurableScanStart,
then do all subsequent work over messages.slice(durableScanStart).
buildToolCallIndex walks only the window; the raw-6 flatMap also
runs over the window. On a 1000-message chat with 100-turn
window, send-time cost drops from O(1000) to O(~window_size).
Acceptable trade: if a tool_call's matching tool_result straddles
the window boundary (result inside, call outside), the single
surviving result loses its [from X(Y)] label. Tool_calls and their
results are almost always adjacent, so this affects at most the
first 1-2 messages of the window.
Tests:
- "preserves an early constraint in a tool-heavy chat where message
count balloons past the raw-count limit" — 35 turns × 6 msgs/turn =
212 messages. The old bound would have dropped the early
EARLY_CONSTRAINT_MARKER; with turn-based bound it survives.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: bincxz <16399091+binaricat@users.noreply.github.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix: correct terminal AI history resume behavior
The previous implementation plan mistakenly treated reopening an old terminal AI session in a fresh or reconnected SSH tab as a scope-retargeting feature.
The intended rule is draft-first:
- a fresh or reconnected terminal opens on a blank draft
- older chats remain available in history for manual access
- selecting history does not imply automatic scope transfer into the new tab
This change is a rule correction, not a conflict between product rules.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix: harden ai draft transitions
* fix ai session continuation from history
* fix: clear stale activeSessionIdMap entry when view resolves to draft
Addresses the Codex P2 review on aiPanelViewState.ts:38. When a terminal
scope mounts with a persisted activeSessionIdMap entry but no explicit
panelView and no draft, resolveDisplayedPanelView now returns the
default draft view (terminal fresh-start behavior). The sync effect
that writes into activeSessionIdMap is guarded by `if (!activeSession)
return`, so the old entry stays put. That stale entry then leaks into
activeTerminalTargetIds in every other scope, and
getSessionScopeMatchRank uses it to suppress host-matched history that
is actually resumable — so valid sessions vanish from the history
drawer until another action rewrites the map.
Add a dedicated effect that clears the scope's activeSessionIdMap
entry whenever the resolved panel view is draft but a persisted
session id is still present. This keeps the map an accurate record of
"which session each scope is currently showing" instead of a lagging
snapshot.
Also extend sessionScopeMatch.test.ts to cover the rank=2 exact-match
branch and the scope-type mismatch short-circuit, which were missing
from the original suite.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: track cross-terminal session ownership by session id, not targetId
Addresses the Codex follow-up review on commit 345244b2. When a user
resumes a session from history into a different terminal, the session's
`scope.targetId` still points at the original terminal. The previous
ownership tracking — which checked whether `session.scope.targetId`
appeared in `activeTerminalTargetIds` (derived from the keys of
`activeSessionIdMap`) — therefore:
- could not prevent the same session from being resumed in multiple
terminals simultaneously, because the resumed session's targetId
never matches the current scope's targetId; and
- let `pruneInactiveScopedSessions` treat a session as orphaned and
clear its `externalSessionId` the moment the original terminal
closed, even though another terminal was actively using it.
Switch ownership to be keyed on session id:
- `getSessionScopeMatchRank` now takes `activeTerminalSessionIds`
(a Set of session ids currently displayed by other terminal scopes)
and returns rank 0 when `session.id` is in that set.
- `AIChatSidePanel` derives `activeTerminalSessionIds` from the
*values* of `activeSessionIdMap`, excluding the current scope's key.
- `pruneInactiveScopedSessions` gains an `activeSessionIds` parameter;
sessions whose id is in this set are never reported as orphaned and
never have their `externalSessionId` cleared, regardless of their
stored `scope.targetId`.
- `cleanupOrphanedAISessions` computes the in-use set from the
pre-cleanup `activeSessionIdMap`, filtered to live scopes, and
passes it through. The map is read once and reused.
Tests cover the new id-based ownership, the rank-2 exact-match path,
the scope-type-mismatch short-circuit, and the
"resumed-elsewhere session must not be cleaned" invariant.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: bincxz <16399091+binaricat@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- persist drafts, panel views, and active sessions per terminal/workspace scope
- restore scoped AI session selection on reconnect and cold mount
- prefer unsent drafts over implicit history fallback
- avoid redundant active session map rewrites during scoped cleanup
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Address Codex review feedback on #708: the previous guard silently
returned on an empty-but-ok probe response, which left any previously
cached runtimeAgentModelPresets[currentAgentId] in place. That kept
Claude/Copilot pickers showing stale model IDs (and skipped currentModelId
reconciliation) instead of falling back to the hardcoded presets when the
backend no longer advertised a catalog.
Now we explicitly drop the cache entry so the agentModelPresets memo falls
through to getAgentModelPresets(...) via the `?? ` branch.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Claude agents now advertise their real model catalog via the ACP
initSession response, just like Copilot already does. Confirmed locally
that `claude-agent-acp` returns `models.availableModels` with full ids +
names + descriptions (default / sonnet / haiku on subscription; and would
return Bedrock/Vertex/custom-proxy ids when the user has configured those).
This closes the gap where the Claude picker was stuck on three hardcoded
entries from CLAUDE_MODEL_PRESETS regardless of what the underlying CLI
actually supports. If the probe fails or returns an empty list, we keep
the hardcoded presets as a fallback.
Codex keeps its existing path via `aiCodexGetIntegration` (reads
~/.codex/config.toml) — we deliberately do not probe codex-acp, since
probing would just return the stock OpenAI model list even when the
user has a custom model_provider set.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>