Skip to content

armgit / cindy

Public Imported from GitHub
0

Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。

fix(settings): accept exact context tokens and unit suffixes

zqchris8a50f67d65b1
diff --git a/apps/desktop/src/renderer/components/settings/ModelAdvancedDrawer.tsx b/apps/desktop/src/renderer/components/settings/ModelAdvancedDrawer.tsx
index c52e7f2206..1d0937de3b 100644
--- a/apps/desktop/src/renderer/components/settings/ModelAdvancedDrawer.tsx
+++ b/apps/desktop/src/renderer/components/settings/ModelAdvancedDrawer.tsx
@@ -40,6 +40,7 @@ import { ClaudeMark } from '@/components/icons/ClaudeMark';
import { CodexMark } from '@/components/icons/CodexMark';
import { PiMark } from '@/components/icons/PiMark';
import { useModelContextLimit } from '@/hooks/useModelContextLimit';
+import { parseContextTokenInput } from '../../../shared/modelContextLimit';
import { ModelCompactionSettings } from './ModelCompactionSettings';
import { modelPriceDetailRows, type ModelPricePresentation } from '@/lib/modelPriceFormat';
import {
@@ -88,12 +89,6 @@ const AGENT_LABEL: Record<AgentKind, string> = {
pi: 'Pi',
};
-// Editing uses whole decimal K, rounded down to avoid suggesting a value above the upstream
-// maximum. Merely opening/blurring the field never writes this display approximation back.
-function editableContextK(tokens: number): string {
- return String(Math.max(1, Math.floor(tokens / 1000)));
-}
-
const AGENT_MARK: Record<AgentKind, (size: number) => ReactNode> = {
'claude-code': (size) => <ClaudeMark size={size} />,
codex: (size) => <CodexMark size={size} />,
@@ -356,16 +351,28 @@ export function ModelAdvancedDrawer({
const [ctxDraft, setCtxDraft] = useState('');
const ctxDirtyRef = useRef(false);
- const defaultWindow = contextAgent === 'codex' && ctx.codexContext
- ? ctx.codexContext.contextWindow : contextModel?.contextWindow ?? 0;
+ const defaultWindow =
+ contextAgent === 'codex' && ctx.codexContext
+ ? ctx.codexContext.contextWindow
+ : (contextModel?.contextWindow ?? 0);
// The editor guard is not a native request-capacity limit. Keep small/local
// model windows selectable and never derive this floor from a saved override.
// Local runtimes own their loaded window; their catalog maximum cannot set an editor floor.
- const modelWindows = chatAgents.map((agent) => row?.byAgent[agent]?.contextWindow)
- .filter((window): window is number => typeof window === 'number' && Number.isFinite(window) && window > 0);
+ const modelWindows = chatAgents
+ .map((agent) => row?.byAgent[agent]?.contextWindow)
+ .filter(
+ (window): window is number =>
+ typeof window === 'number' && Number.isFinite(window) && window > 0,
+ );
const minimumContextK = isLocalRuntimeBetaProviderId(provider.id)
- ? 1 : Math.max(1, Math.floor(Math.min(100_000, ...modelWindows) / 1000));
- const routeWindow = supportsMillionContext ? 1_000_000 : primaryModel?.contextWindowMax ?? primaryModel?.contextWindow ?? 0;
+ ? 1
+ : Math.max(1, Math.floor(Math.min(100_000, ...modelWindows) / 1000));
+ // A working default (including a heuristic fallback) is not a declared capacity.
+ // Codex's native maximum may already contain the user's override, so it
+ // cannot establish the provider's declared capacity.
+ const routeWindow = supportsMillionContext ? 1_000_000
+ : (contextModel ?? primaryModel)?.contextWindowMax
+ ?? (primaryModel?.contextWindowVerified === true ? primaryModel.contextWindow : 0);
const effectiveLimit = ctx.limit ?? (defaultWindow > 0 ? defaultWindow : null);
useEffect(() => {
ctxDirtyRef.current = false;
@@ -374,13 +381,15 @@ export function ModelAdvancedDrawer({
}, [open, primaryAgent, primaryModel?.id, provider.id]);
useEffect(() => {
if (ctx.loading || ctxDirtyRef.current) return;
- setCtxDraft(effectiveLimit === null ? '' : editableContextK(effectiveLimit));
+ setCtxDraft(effectiveLimit === null ? '' : String(effectiveLimit));
}, [ctx.loading, effectiveLimit, open]);
- const parsedK = Number(ctxDraft.trim());
- const parsedTokens = parsedK * 1000;
+ const parsedTokens = parseContextTokenInput(ctxDraft);
const ctxInvalid =
- ctxDirtyRef.current && ctxDraft.trim() !== '' &&
- (!Number.isSafeInteger(parsedK) || parsedK < minimumContextK || parsedTokens > maximumContext);
+ ctxDirtyRef.current &&
+ ctxDraft.trim() !== '' &&
+ (parsedTokens === null ||
+ parsedTokens < minimumContextK * 1000 ||
+ parsedTokens > maximumContext);
const commitCtxDraft = useCallback(() => {
if (!ctxDirtyRef.current || ctxInvalid || ctx.loading || localConfigurationBlocked) return;
ctxDirtyRef.current = false;
@@ -389,7 +398,7 @@ export function ModelAdvancedDrawer({
const resetCtx = useCallback(() => {
if (localConfigurationBlocked) return;
ctxDirtyRef.current = false;
- setCtxDraft(defaultWindow > 0 ? editableContextK(defaultWindow) : '');
+ setCtxDraft(defaultWindow > 0 ? String(defaultWindow) : '');
void ctx.reset();
}, [ctx, defaultWindow, localConfigurationBlocked]);
@@ -456,7 +465,7 @@ export function ModelAdvancedDrawer({
? defaultWindow
: parsedTokens
: effectiveLimit;
- const editRouteWindow = contextModel?.contextWindowMax ?? defaultWindow;
+ const editRouteWindow = routeWindow;
const overRouteWindow =
editRouteWindow > 0 &&
displayedLimit !== null &&
@@ -794,7 +803,7 @@ export function ModelAdvancedDrawer({
disabled={paymentRequired || ctx.loading || localConfigurationBlocked}
onClick={() => {
ctxDirtyRef.current = false;
- setCtxDraft(editableContextK(preset.tokens));
+ setCtxDraft(String(preset.tokens));
void ctx.setLimit(preset.tokens);
}}
>{preset.label}</Button>
@@ -824,14 +833,15 @@ export function ModelAdvancedDrawer({
}
}}
aria-invalid={ctxInvalid || undefined}
- inputMode="numeric"
+ inputMode="text"
+ maxLength={32}
disabled={paymentRequired || ctx.loading || localConfigurationBlocked}
aria-label={t(
'settings.providers.models.advanced.contextLimitAria',
)}
- className="w-20 bg-transparent text-center text-13 tabular-nums text-[var(--text-primary)] outline-none"
+ className="w-28 bg-transparent text-center text-13 tabular-nums text-[var(--text-primary)] outline-none"
/>
- <span className="text-11 text-[var(--text-tertiary)]">K</span>
+ <span className="text-11 text-[var(--text-tertiary)]">tokens</span>
</span>
<span className="min-w-0 flex-1 text-11 tabular-nums text-[var(--text-tertiary)]">
{t('settings.providers.models.advanced.contextLimitRoute', {
@@ -856,7 +866,7 @@ export function ModelAdvancedDrawer({
<p role="status" className="mt-1.5 text-12 text-[var(--warning-fg)]">
{t(
`settings.providers.models.advanced.${ctxInvalid ? 'contextInvalid' : ctx.error ? 'contextWriteFailed' : 'contextMixed'}`,
- { min: minimumContextK },
+ { min: formatExactTokens(minimumContextK * 1000, locale), max: formatExactTokens(maximumContext, locale) },
)}
</p>
)}
diff --git a/apps/desktop/src/renderer/components/settings/__tests__/ModelAdvancedDrawer.test.tsx b/apps/desktop/src/renderer/components/settings/__tests__/ModelAdvancedDrawer.test.tsx
index 45897b399f..9eb8b6572c 100644
--- a/apps/desktop/src/renderer/components/settings/__tests__/ModelAdvancedDrawer.test.tsx
+++ b/apps/desktop/src/renderer/components/settings/__tests__/ModelAdvancedDrawer.test.tsx
@@ -226,11 +226,11 @@ describe('model advanced editor', () => {
render(<ModelAdvancedDrawer provider={source} row={{ id: primary.id, name: primary.name, avail: ['pi'], byAgent: { pi: primary } }} open onOpenChange={vi.fn()} pricePresentationOf={() => null} onDisable={vi.fn()} disabled={false} paymentRequired={false} />);
const input = screen.getByRole('textbox', { name: 'settings.providers.models.advanced.contextLimitAria' });
await waitFor(() => expect(window.electronAPI.maker.llamaCppStatus).toHaveBeenCalled());
- fireEvent.change(input, { target: { value: '64' } });
+ fireEvent.change(input, { target: { value: '64K' } });
expect(input.getAttribute('aria-invalid')).toBe('true');
fireEvent.blur(input);
expect(mocks.setLimit).not.toHaveBeenCalled();
- fireEvent.change(input, { target: { value: '32' } });
+ fireEvent.change(input, { target: { value: '32K' } });
fireEvent.blur(input);
expect(mocks.setLimit).toHaveBeenLastCalledWith(32000);
} finally { cleanup(); Object.defineProperty(window, 'electronAPI', { configurable: true, value: previous }); }
@@ -351,21 +351,21 @@ describe('model advanced editor', () => {
mocks.limit = 1_000;
draw();
const input = screen.getByRole('textbox') as HTMLInputElement;
- expect(input.value).toBe('1');
+ expect(input.value).toBe('1000');
expect(input.hasAttribute('aria-invalid')).toBe(false);
fireEvent.blur(input);
expect(mocks.setLimit).not.toHaveBeenCalled();
- fireEvent.change(input, { target: { value: '2' } });
+ fireEvent.change(input, { target: { value: '2K' } });
fireEvent.blur(input);
expect(input.getAttribute('aria-invalid')).toBe('true');
expect(mocks.setLimit).not.toHaveBeenCalled();
- fireEvent.change(input, { target: { value: '99' } });
+ fireEvent.change(input, { target: { value: '99K' } });
fireEvent.blur(input);
expect(mocks.setLimit).not.toHaveBeenCalled();
- fireEvent.change(input, { target: { value: '100' } });
+ fireEvent.change(input, { target: { value: '100K' } });
fireEvent.blur(input);
expect(mocks.setLimit).toHaveBeenLastCalledWith(100_000);
- fireEvent.change(input, { target: { value: '1' } });
+ fireEvent.change(input, { target: { value: '1K' } });
fireEvent.blur(input);
expect(mocks.setLimit).toHaveBeenCalledTimes(1);
});
@@ -373,10 +373,10 @@ describe('model advanced editor', () => {
it.each(['cindy-local-ollama', 'ollama', 'cindy-local-lmstudio'])('keeps %s editable at 1K even when the catalog advertises a large window', (id) => {
render(drawer(model, model.defaultEffort, model.efforts, { ...provider, id, source: 'user' }));
const input = screen.getByRole('textbox');
- fireEvent.change(input, { target: { value: '4' } });
+ fireEvent.change(input, { target: { value: '4K' } });
fireEvent.blur(input);
expect(mocks.setLimit).toHaveBeenLastCalledWith(4_000);
- fireEvent.change(input, { target: { value: '1' } });
+ fireEvent.change(input, { target: { value: '1K' } });
fireEvent.blur(input);
expect(mocks.setLimit).toHaveBeenLastCalledWith(1_000);
});
@@ -385,10 +385,10 @@ describe('model advanced editor', () => {
draw({ ...model, contextWindow: window, contextWindowMax: window });
const input = screen.getByRole('textbox');
const floor = Math.floor(window / 1000);
- fireEvent.change(input, { target: { value: String(floor - 1) } });
+ fireEvent.change(input, { target: { value: `${floor - 1}K` } });
fireEvent.blur(input);
expect(mocks.setLimit).not.toHaveBeenCalled();
- fireEvent.change(input, { target: { value: String(floor) } });
+ fireEvent.change(input, { target: { value: `${floor}K` } });
fireEvent.blur(input);
expect(mocks.setLimit).toHaveBeenCalledWith(floor * 1000);
});
@@ -444,8 +444,8 @@ describe('model advanced editor', () => {
const input = screen.getByRole('textbox', {
name: 'settings.providers.models.advanced.contextLimitAria',
});
- expect((input as HTMLInputElement).value).toBe('272');
- fireEvent.change(input, { target: { value: '500' } });
+ expect((input as HTMLInputElement).value).toBe('272000');
+ fireEvent.change(input, { target: { value: '500K' } });
fireEvent.blur(input);
expect(mocks.setLimit).toHaveBeenCalledWith(500_000);
fireEvent.change(input, { target: { value: '1.5' } });
@@ -474,7 +474,7 @@ describe('model advanced editor', () => {
/>,
);
const input = screen.getByRole('textbox');
- fireEvent.change(input, { target: { value: '1000' } });
+ fireEvent.change(input, { target: { value: '1000K' } });
fireEvent.blur(input);
expect(mocks.setLimit).toHaveBeenCalledWith(1_000_000);
expect(screen.getByText('1,050,000')).toBeTruthy();
@@ -495,7 +495,7 @@ describe('model advanced editor', () => {
/>,
);
const input = screen.getByRole('textbox');
- fireEvent.change(input, { target: { value: '1000' } });
+ fireEvent.change(input, { target: { value: '1000K' } });
fireEvent.blur(input);
expect(mocks.setLimit).toHaveBeenCalledWith(1_000_000);
expect(screen.getAllByText('272K').length).toBeGreaterThan(0);
@@ -525,20 +525,20 @@ describe('model advanced editor', () => {
expect(screen.getByText(/^Messages/)).toBeTruthy();
expect(screen.getByText(/^Responses/)).toBeTruthy();
expect(screen.getByText(/^Chat Completions/)).toBeTruthy();
- expect((screen.getByRole('textbox') as HTMLInputElement).value).toBe('1048');
+ expect((screen.getByRole('textbox') as HTMLInputElement).value).toBe('1048576');
});
- it('shows whole K without rewriting the exact catalog value on untouched blur', () => {
+ it('shows exact tokens without rewriting the catalog value on untouched blur', () => {
draw({ ...model, contextWindow: 1_048_576, contextWindowMax: 1_048_576 });
const input = screen.getByRole('textbox') as HTMLInputElement;
- expect(input.value).toBe('1048');
+ expect(input.value).toBe('1048576');
expect(
screen.getByText('settings.providers.models.advanced.contextLimitRoute 1,048,576'),
).toBeTruthy();
fireEvent.focus(input);
fireEvent.blur(input);
expect(mocks.setLimit).not.toHaveBeenCalled();
- fireEvent.change(input, { target: { value: '1000' } });
+ fireEvent.change(input, { target: { value: '1000K' } });
fireEvent.blur(input);
expect(mocks.setLimit).toHaveBeenCalledWith(1_000_000);
});
@@ -605,7 +605,7 @@ describe('model advanced editor', () => {
mocks.limit = 1_048_900;
draw({ ...model, contextWindow: 1_048_576, contextWindowMax: 1_048_576 });
const input = screen.getByRole('textbox') as HTMLInputElement;
- expect(input.value).toBe('1048');
+ expect(input.value).toBe('1048900');
expect(
screen.getByText('settings.providers.models.advanced.contextLimitOverWindow'),
).toBeTruthy();
@@ -672,11 +672,11 @@ describe('model advanced editor', () => {
it('refreshes defaults and controls in an open drawer when the catalog changes', () => {
const { rerender } = draw();
- expect((screen.getByRole('textbox') as HTMLInputElement).value).toBe('272');
+ expect((screen.getByRole('textbox') as HTMLInputElement).value).toBe('272000');
rerender(
drawer({ ...model, contextWindow: 700_000, efforts: ['high', 'max'], defaultEffort: 'max' }),
);
- expect((screen.getByRole('textbox') as HTMLInputElement).value).toBe('700');
+ expect((screen.getByRole('textbox') as HTMLInputElement).value).toBe('700000');
expect(screen.queryByRole('radio', { name: 'effortLevels.low' })).toBeNull();
expect(
screen.getByRole('radio', { name: 'effortLevels.max' }).getAttribute('aria-checked'),
@@ -688,7 +688,7 @@ describe('model advanced editor', () => {
const input = screen.getByRole('textbox', {
name: 'settings.providers.models.advanced.contextLimitAria',
});
- fireEvent.change(input, { target: { value: '1200' } });
+ fireEvent.change(input, { target: { value: '1200K' } });
expect(
screen.getByText('settings.providers.models.advanced.contextLimitOverWindow'),
).toBeTruthy();
@@ -792,7 +792,7 @@ it('limits context and effort reads, writes and resets to the chat runtime', ()
fireEvent.click(screen.getByRole('button', { name: 'settings.providers.models.advanced.restoreDefault' }));
expect(clearProviderModelEffort).toHaveBeenCalledExactlyOnceWith('codex', 'private', 'shared');
const input = screen.getByRole('textbox');
- fireEvent.change(input, { target: { value: '128' } });
+ fireEvent.change(input, { target: { value: '128K' } });
fireEvent.blur(input);
expect(mocks.setLimit).toHaveBeenCalledWith(128_000);
expect(mocks.target).toHaveBeenLastCalledWith(expect.objectContaining({ agent: 'codex', relatedTargets: [] }));
@@ -927,3 +927,20 @@ it('retains the chat discount inferred from standard and effective prices', () =
const outputRow = screen.getByText('settings.providers.models.advanced.price.output').parentElement!;
expect(outputRow.nextElementSibling).toBe(screen.getByText('settings.providers.models.advanced.discount').parentElement);
});
+
+// Unit: real editor commits normalized tokens, preserving precision and rejecting fractional tokens.
+it.each([['1000000', 1_000_000], ['1M', 1_000_000], ['1.5m', 1_500_000], ['100万', 1_000_000], ['262144', 262_144], ['1.000001M', 1_000_001], ['262.144K', 262_144]])('accepts context %s as %s tokens', (value, tokens) => {
+ draw();
+ const input = screen.getByRole('textbox');
+ fireEvent.change(input, { target: { value } }); fireEvent.blur(input);
+ expect(mocks.setLimit).toHaveBeenLastCalledWith(tokens);
+});
+it('does not label an unverified working default as the upstream capacity', () => {
+ draw({ ...model, contextWindow: 200_000, contextWindowMax: undefined, contextWindowVerified: false });
+ const input = screen.getByRole('textbox');
+ fireEvent.change(input, { target: { value: '1M' } });
+ expect(screen.queryByText('settings.providers.models.advanced.contextLimitOverWindow')).toBeNull();
+ expect(screen.getByText('settings.providers.models.advanced.catalogMissing')).toBeTruthy();
+ fireEvent.blur(input);
+ expect(mocks.setLimit).toHaveBeenLastCalledWith(1_000_000);
+});
diff --git a/apps/desktop/src/renderer/i18n/locales/en/common.json b/apps/desktop/src/renderer/i18n/locales/en/common.json
index 20e794740b..487390ccda 100644
--- a/apps/desktop/src/renderer/i18n/locales/en/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/en/common.json
@@ -3782,16 +3782,16 @@
"defaultEffort": "Default reasoning effort",
"defaultEffortHint": "The initial effort for new tasks. You can still adjust it within a task.",
"contextLimit": "Context limit",
- "contextLimitHint": "Applies to all available engines for this model. Edit in whole K (1 K = 1,000 tokens); untouched values retain their precision. Saved changes apply to subsequent runs without interrupting a response. Actual capacity remains subject to the model service’s limits.",
+ "contextLimitHint": "Enter tokens directly, or use K or M (for example, 1000000 or 1M). Applies to all available engines for this model on subsequent runs. Actual capacity is still limited by the model service.",
"contextLimitLocalMillionHint": "Native 256K (262,144 tokens), with optional expansion to 1M (1,000,000). If the service is running, restart it under Manage Service to apply changes; if stopped, they apply on next use. Longer inputs need more memory and take longer to process.",
- "contextLimitAria": "Context limit (thousands of tokens)",
+ "contextLimitAria": "Context limit (tokens, K or M)",
"compaction": {
"title": "Automatic Compaction",
"hint": "Compact context at the selected usage percentage. Lower values compact earlier; engine limits still apply.",
"default": "{{pct}}% (Default)",
"timing": "Applies to all available engines for this model and source on subsequent runs, without interrupting a response."
},
- "contextLimitRoute": "Current {{tokens}} tokens · 1 K = 1,000 tokens",
+ "contextLimitRoute": "Current {{tokens}} tokens",
"contextLimitOverWindow": "Above the provider’s declared window; requests may fail.",
"restoreDefault": "Restore default",
"engines": "Engine support",
@@ -3818,7 +3818,7 @@
"close": "Close",
"maxOutput": "Maximum output",
"catalogCapabilities": "Capabilities declared in the catalog",
- "contextInvalid": "Enter a whole number from {{min}} to 100,000 K.",
+ "contextInvalid": "Enter {{min}}–{{max}} tokens; K and M are also accepted.",
"contextWriteFailed": "Not saved. Please retry.",
"contextMixed": "Existing harness settings differ. Editing or resetting applies to all harnesses.",
"inputModalities": "Input types",
diff --git a/apps/desktop/src/renderer/i18n/locales/ja/common.json b/apps/desktop/src/renderer/i18n/locales/ja/common.json
index c9d64370c3..57c21b86d2 100644
--- a/apps/desktop/src/renderer/i18n/locales/ja/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/ja/common.json
@@ -3751,16 +3751,16 @@
"defaultEffort": "既定の推論強度",
"defaultEffortHint": "新しいタスクの初期推論強度です。タスク内でも変更できます。",
"contextLimit": "コンテキスト上限",
- "contextLimitHint": "このモデルのすべての利用可能なエンジンに適用されます。整数 K(1 K = 1,000 tokens)で編集します。未編集の値は精度を保持します。保存した変更は以降の実行に適用され、生成中の応答を中断しません。実際に利用できる長さはモデルサービスの制限に従います。",
+ "contextLimitHint": "トークン数、K、M で入力できます(例:1000000、1M)。このモデルの利用可能な全エンジンで、次回の実行から適用されます。実際の容量はモデルサービスの制限に従います。",
"contextLimitLocalMillionHint": "標準は 256K(262,144 tokens)で、1M(1,000,000)への拡張を選べます。サービスの実行中は「サービスを管理」で再起動すると変更が反映されます。停止中なら次回の使用時に反映されます。長い入力ほどメモリと処理時間が必要です。",
- "contextLimitAria": "コンテキスト上限(千 tokens)",
+ "contextLimitAria": "コンテキスト上限(tokens、K、M)",
"compaction": {
"title": "自動圧縮",
"hint": "使用量が選択した割合に達するとコンテキストを整理します。低い値ほど早く圧縮します。実際のタイミングはエンジンの制限に従います。",
"default": "{{pct}}%(既定)",
"timing": "この提供元のモデルで利用可能なすべてのエンジンに適用します。以降の実行で有効になり、生成中の応答は中断しません。"
},
- "contextLimitRoute": "現在 {{tokens}} tokens · 1 K = 1,000 tokens",
+ "contextLimitRoute": "現在 {{tokens}} tokens",
"contextLimitOverWindow": "プロバイダーが宣言した上限を超えています。リクエストが失敗する可能性があります。",
"restoreDefault": "既定に戻す",
"engines": "エンジン対応",
@@ -3787,7 +3787,7 @@
"close": "閉じる",
"maxOutput": "最大出力",
"catalogCapabilities": "カタログに記載された機能",
- "contextInvalid": "{{min}}~100,000 K の整数を入力してください。",
+ "contextInvalid": "{{min}}〜{{max}} tokens を入力してください。K、M も使用できます。",
"contextWriteFailed": "保存できませんでした。再試行してください。",
"contextMixed": "実行環境ごとに設定が異なります。変更やリセットはすべてに適用されます。",
"inputModalities": "入力形式",
diff --git a/apps/desktop/src/renderer/i18n/locales/ko/common.json b/apps/desktop/src/renderer/i18n/locales/ko/common.json
index b5c597adfe..aba6bb9e83 100644
--- a/apps/desktop/src/renderer/i18n/locales/ko/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/ko/common.json
@@ -3751,16 +3751,16 @@
"defaultEffort": "기본 추론 강도",
"defaultEffortHint": "새 작업의 초기 추론 강도입니다. 작업 안에서도 변경할 수 있습니다.",
"contextLimit": "컨텍스트 상한",
- "contextLimitHint": "이 모델의 모든 사용 가능한 엔진에 적용됩니다. 정수 K 단위로 편집합니다(1 K = 1,000 tokens). 수정하지 않은 값은 정밀도를 유지합니다. 저장한 변경 사항은 이후 실행에 적용되며 생성 중인 응답을 중단하지 않습니다. 실제로 사용할 수 있는 길이는 모델 서비스의 제한을 따릅니다.",
+ "contextLimitHint": "토큰 수 또는 K, M 단위로 입력하세요(예: 1000000, 1M). 이 모델의 모든 사용 가능한 엔진에서 다음 실행부터 적용됩니다. 실제 용량은 모델 서비스의 제한을 따릅니다.",
"contextLimitLocalMillionHint": "기본 256K(262,144 tokens)에서 1M(1,000,000) 확장을 선택할 수 있습니다. 서비스가 실행 중이면 ‘서비스 관리’에서 다시 시작해야 변경 사항이 적용되며, 중지된 상태라면 다음 사용 시 적용됩니다. 긴 입력은 더 많은 메모리와 처리 시간이 필요합니다.",
- "contextLimitAria": "컨텍스트 상한(천 tokens)",
+ "contextLimitAria": "컨텍스트 한도(tokens, K 또는 M)",
"compaction": {
"title": "자동 압축",
"hint": "사용량이 선택한 비율에 도달하면 컨텍스트를 정리합니다. 값이 낮을수록 더 일찍 압축하며 엔진 제한이 적용됩니다.",
"default": "{{pct}}% (기본값)",
"timing": "이 공급 경로의 모델에서 사용 가능한 모든 엔진에 적용됩니다. 이후 실행부터 반영되며 생성 중인 응답은 중단하지 않습니다."
},
- "contextLimitRoute": "현재 {{tokens}} tokens · 1 K = 1,000 tokens",
+ "contextLimitRoute": "현재 {{tokens}} tokens",
"contextLimitOverWindow": "공급업체가 명시한 한도를 초과하여 요청이 실패할 수 있습니다.",
"restoreDefault": "기본값 복원",
"engines": "엔진 지원",
@@ -3787,7 +3787,7 @@
"close": "닫기",
"maxOutput": "최대 출력",
"catalogCapabilities": "카탈로그에 명시된 기능",
- "contextInvalid": "{{min}}–100,000 K 범위의 정수를 입력하세요.",
+ "contextInvalid": "{{min}}–{{max}} tokens를 입력하세요. K, M도 사용할 수 있습니다.",
"contextWriteFailed": "저장되지 않았습니다. 다시 시도하세요.",
"contextMixed": "실행 환경별 설정이 다릅니다. 변경하거나 초기화하면 모두에 적용됩니다.",
"inputModalities": "입력 형식",
diff --git a/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json b/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json
index 3d5d09cb94..7f4fa98fa9 100644
--- a/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json
@@ -3751,16 +3751,16 @@
"defaultEffort": "默认推理强度",
"defaultEffortHint": "新任务选中该模型时的初始档位,任务内仍可临时调整。",
"contextLimit": "上下文上限",
- "contextLimitHint": "用于此模型的所有可用引擎。以整数 K 编辑(1 K = 1,000 tokens);未编辑时保留精确值。保存后用于后续运行,正在生成的内容不受影响。实际可用长度仍受模型服务限制。",
+ "contextLimitHint": "可直接输入 token 数,或使用 K、M、万,例如 1000000、1M。应用于此模型的所有可用引擎,后续运行生效。实际可用长度仍受模型服务限制。",
"contextLimitLocalMillionHint": "原生 256K(262,144 tokens),可选扩展至 1M(1,000,000)。服务运行时需在「管理服务」中重启以应用更改;已停止时,下次使用自动生效。长输入会增加内存占用和等待时间。",
- "contextLimitAria": "上下文上限(千 tokens)",
+ "contextLimitAria": "上下文上限(tokens、K、M 或万)",
"compaction": {
"title": "自动压缩",
"hint": "占用达到所选比例时整理上下文。比例越低,越早压缩;实际时机仍受引擎限制。",
"default": "{{pct}}%(默认)",
"timing": "用于此来源下该模型的所有可用引擎。后续运行生效,不中断正在生成的内容。"
},
- "contextLimitRoute": "当前 {{tokens}} tokens · 1 K = 1,000 tokens",
+ "contextLimitRoute": "当前 {{tokens}} tokens",
"contextLimitOverWindow": "已超过上游声明的长度,可能导致请求失败。",
"restoreDefault": "恢复默认",
"engines": "引擎支持",
@@ -3787,7 +3787,7 @@
"close": "关闭",
"maxOutput": "最大输出",
"catalogCapabilities": "目录声明的能力",
- "contextInvalid": "请输入 {{min}}–100,000 K 的整数。",
+ "contextInvalid": "请输入 {{min}}–{{max}} tokens,也支持 K、M、万。",
"contextWriteFailed": "未保存,请重试。",
"contextMixed": "已有引擎配置不同;修改或恢复默认将应用到全部引擎。",
"inputModalities": "输入类型",
diff --git a/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json b/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json
index cb08e66e30..310b02c748 100644
--- a/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json
@@ -3751,16 +3751,16 @@
"defaultEffort": "預設推理強度",
"defaultEffortHint": "新任務選中該模型時的初始檔位,任務內仍可臨時調整。",
"contextLimit": "上下文上限",
- "contextLimitHint": "套用至此模型所有可用引擎。以整數 K 編輯(1 K = 1,000 tokens);未編輯時保留精確值。儲存後用於後續執行,不影響正在產生的內容。實際可用長度仍受模型服務限制。",
+ "contextLimitHint": "可直接輸入 token 數,或使用 K、M、萬,例如 1000000、1M。套用至此模型的所有可用引擎,後續執行生效。實際可用長度仍受模型服務限制。",
"contextLimitLocalMillionHint": "原生 256K(262,144 tokens),可選擴展至 1M(1,000,000)。服務執行時需在「管理服務」中重新啟動以套用變更;已停止時,下次使用自動生效。長輸入會增加記憶體用量和等待時間。",
- "contextLimitAria": "上下文上限(千 tokens)",
+ "contextLimitAria": "上下文上限(tokens、K、M 或萬)",
"compaction": {
"title": "自動壓縮",
"hint": "占用達到所選比例時整理上下文。比例越低,越早壓縮;實際時機仍受引擎限制。",
"default": "{{pct}}%(預設)",
"timing": "套用至此來源下該模型的所有可用引擎。後續執行生效,不中斷正在產生的內容。"
},
- "contextLimitRoute": "目前 {{tokens}} tokens · 1 K = 1,000 tokens",
+ "contextLimitRoute": "目前 {{tokens}} tokens",
"contextLimitOverWindow": "已超過上游宣告的長度,可能導致請求失敗。",
"restoreDefault": "恢復預設",
"engines": "引擎支援",
@@ -3787,7 +3787,7 @@
"close": "關閉",
"maxOutput": "最大輸出",
"catalogCapabilities": "目錄宣告的能力",
- "contextInvalid": "請輸入 {{min}}–100,000 K 的整數。",
+ "contextInvalid": "請輸入 {{min}}–{{max}} tokens,也支援 K、M、萬。",
"contextWriteFailed": "未儲存,請重試。",
"contextMixed": "現有引擎設定不同;修改或恢復預設將套用至所有引擎。",
"inputModalities": "輸入類型",
diff --git a/apps/desktop/src/shared/modelContextLimit.ts b/apps/desktop/src/shared/modelContextLimit.ts
index 61e67681d5..f9f5e8689a 100644
--- a/apps/desktop/src/shared/modelContextLimit.ts
+++ b/apps/desktop/src/shared/modelContextLimit.ts
@@ -4,6 +4,18 @@ import type { AgentKind, ProviderView } from '@cindy/model-providers';
export const DEFAULT_MODEL_COMPACTION_PCT = 90;
+/** Decimal units, independent of locale. Bare numbers always mean exact tokens. */
+export function parseContextTokenInput(input: string): number | null {
+ const match = input.trim().match(/^(\d+(?:\.\d+)?)\s*(k|m|万|萬|tokens?)?$/i);
+ if (!match) return null;
+ const unit = match[2]?.toLowerCase();
+ const places = unit === 'k' ? 3 : unit === 'm' ? 6 : unit === '万' || unit === '萬' ? 4 : 0;
+ const [whole, fraction = ''] = match[1].split('.');
+ if (/[1-9]/.test(fraction.slice(places))) return null;
+ const tokens = Number(whole + fraction.slice(0, places).padEnd(places, '0'));
+ return Number.isSafeInteger(tokens) && tokens > 0 ? tokens : null;
+}
+
/** Match the model row identity, removing only the provider's declared bridge prefix. */
export function modelCompactionTarget(
provider: Pick<ProviderView, 'id' | 'routing'>,