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(models): 补全模型导入同步并按原生协议启用新型号 (#5063)

Chris Zhang061568865f85
diff --git a/apps/desktop/src/main/maker-host/__tests__/activeCatalogDiscovery.test.ts b/apps/desktop/src/main/maker-host/__tests__/activeCatalogDiscovery.test.ts
index 74e5e53690..1a5ae1d6c0 100644
--- a/apps/desktop/src/main/maker-host/__tests__/activeCatalogDiscovery.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/activeCatalogDiscovery.test.ts
@@ -90,6 +90,17 @@ function legacyCatalog(): Catalog {
}
describe('active-catalog discovered augment', () => {
+ it('keeps a future subscription model inherited window unverified', () => {
+ setActiveCatalog(structuredClone(BUNDLED_CATALOG));
+ setDiscoveredCodexModels([{ ...fake('gpt-7-sol'), discoveredMetadata: {}, contextWindowVerified: false }]);
+ for (const agent of ['codex', 'claude-code'] as const) {
+ const models = getActiveCatalog().providers.find(provider => provider.id === 'openai')!.models[agent]!;
+ const model = models.find(model => model.id === 'gpt-7-sol' || model.id === 'chatgpt/gpt-7-sol');
+ expect(model?.contextWindow).toBeGreaterThan(0);
+ expect(model?.contextWindowVerified).toBe(false);
+ }
+ });
+
it('keeps Claude subscription models on Claude Code only, even when the source catalog declares Codex / Pi', () => {
const catalog = bundledWithoutRegistry();
const builtin = catalog.providers.find((provider) => provider.id === 'anthropic')!;
@@ -840,7 +851,7 @@ describe('anthropic 发现条目的 modelRegistry 元数据基线', () => {
.filter((model) => model.defaultEnabled !== false)
.map((model) => model.id)
.sort(),
- ).toEqual(['claude-fable-5-1', 'claude-haiku-4-5', 'claude-opus-5-5', 'claude-sonnet-5']);
+ ).toEqual(['claude-fable-5', 'claude-fable-5-1', 'claude-haiku-4-5', 'claude-mythos-5', 'claude-opus-4-8', 'claude-opus-5', 'claude-opus-5-5', 'claude-sonnet-5']);
// Claude 订阅只供 Claude Code:不投影 Codex bridge,也不给 Pi。
expect(anthropicList('codex')).toEqual([]);
expect(anthropicList('pi')).toEqual([]);
diff --git a/apps/desktop/src/main/maker-host/__tests__/activeCatalogXdAvailability.test.ts b/apps/desktop/src/main/maker-host/__tests__/activeCatalogXdAvailability.test.ts
index 47ca262b96..96e17b98af 100644
--- a/apps/desktop/src/main/maker-host/__tests__/activeCatalogXdAvailability.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/activeCatalogXdAvailability.test.ts
@@ -879,7 +879,7 @@ describe('Anthropic 权威模型清单注入', () => {
});
describe('gateway cross-harness defaults', () => {
- it('projects reviewed Gateway defaults through native harness policy and live capability changes', () => {
+ it('preserves Gateway model defaults across capability changes while keeping compatibility opt-in', () => {
setActiveCatalog(BUNDLED_CATALOG);
const candidates = [
['deepseek/deepseek-v4-pro', ['text']],
@@ -902,10 +902,7 @@ describe('gateway cross-harness defaults', () => {
xdModels('pi')
.filter((model) => model.defaultEnabled !== false)
.map((model) => model.id);
- expect(enabled().sort()).toEqual([
- 'deepseek/deepseek-v4-flash-vision-exp',
- 'tencent/hy4-preview',
- ]);
+ expect(enabled().sort()).toEqual(candidates.map(([id]) => id).sort());
for (const agent of ['claude-code', 'codex'] as const) {
expect(xdModels(agent).every((model) => model.defaultEnabled === false)).toBe(true);
}
@@ -917,7 +914,11 @@ describe('gateway cross-harness defaults', () => {
modalities: { input: ['text'], output: ['text'] },
})),
);
- expect(enabled().sort()).toEqual(['deepseek/deepseek-v4-pro', 'tencent/hy4-preview']);
+ expect(enabled().sort()).toEqual(candidates.map(([id]) => id).sort());
+ setXdGatewayModels(live.map((model, index) => ({
+ ...model, agents: [...model.agents], defaultEnabled: index !== 0,
+ })));
+ expect(enabled().sort()).toEqual(candidates.slice(1).map(([id]) => id).sort());
});
it.each([1, 2, 3] as const)(
@@ -951,24 +952,21 @@ describe('gateway cross-harness defaults', () => {
})),
);
for (const [id, api] of examples) {
- const everyday = !['deepseek/deepseek-v4-flash-vision-exp', 'tencent/hy4-preview'].includes(
- id,
- );
const pi = xdModels('pi').find((m) => m.id === id)!;
expect(pi).toMatchObject({
nativeApi: api,
piApi: api,
- defaultEnabled: everyday,
+ defaultEnabled: true,
contextWindow: 123_456,
});
expect(resolveXdPiGatewayApi(id)).toBe(api);
expect(xdModels('claude-code').find((m) => m.id === id)).toMatchObject({
nativeApi: api,
- defaultEnabled: everyday && api === 'anthropic-messages',
+ defaultEnabled: api === 'anthropic-messages',
});
expect(xdModels('codex').find((m) => m.id === id)).toMatchObject({
nativeApi: api,
- defaultEnabled: everyday && api === 'openai-responses',
+ defaultEnabled: api === 'openai-responses',
});
}
setXdGatewayModels([]);
@@ -1053,7 +1051,7 @@ describe('gateway cross-harness defaults', () => {
expect(xdModels('codex')).toHaveLength(2);
});
- it('applies per-agent policy to the selected models while keeping old generations opt-in', () => {
+ it('applies per-agent policy without hiding older discovered generations', () => {
setActiveCatalog(BUNDLED_CATALOG);
const entries = ['codex/gpt-6', 'claude-opus-5', 'anthropic-claude/claude-opus-4-8'].map(
(id) => ({
@@ -1065,9 +1063,9 @@ describe('gateway cross-harness defaults', () => {
}),
);
setXdGatewayModels(entries.map((e) => ({ ...e, agents: [...e.agents] })));
- expect(xdModels('claude-code').map((m) => m.defaultEnabled)).toEqual([false, true, false]);
+ expect(xdModels('claude-code').map((m) => m.defaultEnabled)).toEqual([false, true, true]);
expect(xdModels('codex').map((m) => m.defaultEnabled)).toEqual([true, false, false]);
- expect(xdModels('pi').map((m) => m.defaultEnabled)).toEqual([true, true, false]);
+ expect(xdModels('pi').map((m) => m.defaultEnabled)).toEqual([true, true, true]);
setXdGatewayModels(
entries.map((e) => ({
...e,
@@ -1075,7 +1073,7 @@ describe('gateway cross-harness defaults', () => {
perAgent: { 'claude-code': { defaultEnabled: true }, codex: { defaultEnabled: true } },
})),
);
- expect(xdModels('claude-code').map((m) => m.defaultEnabled)).toEqual([true, true, false]);
- expect(xdModels('codex').map((m) => m.defaultEnabled)).toEqual([true, true, false]);
+ expect(xdModels('claude-code').map((m) => m.defaultEnabled)).toEqual([true, true, true]);
+ expect(xdModels('codex').map((m) => m.defaultEnabled)).toEqual([true, true, true]);
});
});
diff --git a/apps/desktop/src/main/maker-host/__tests__/custom-provider-store.test.ts b/apps/desktop/src/main/maker-host/__tests__/custom-provider-store.test.ts
index 890156b7cb..f496e86deb 100644
--- a/apps/desktop/src/main/maker-host/__tests__/custom-provider-store.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/custom-provider-store.test.ts
@@ -400,7 +400,7 @@ describe('validateCustomProviderConfig (per-runtime)', () => {
).toBe(false);
});
- it('accepts only explicit, non-empty, valid Pi reasoning effort capabilities', () => {
+ it('accepts inherited reasoning capabilities while validating explicit effort lists', () => {
const config = (model: Record<string, unknown>, agent: 'pi' | 'codex' = 'pi') => ({
id: 'reasoning-provider',
name: 'Reasoning provider',
@@ -422,7 +422,8 @@ describe('validateCustomProviderConfig (per-runtime)', () => {
}),
),
).toEqual({ ok: true });
- expect(validateCustomProviderConfig(config({ reasoning: true })).ok).toBe(false);
+ expect(validateCustomProviderConfig(config({ reasoning: true })).ok).toBe(true);
+ expect(validateCustomProviderConfig(config({ reasoning: true, reasoningEfforts: [] })).ok).toBe(true);
expect(
validateCustomProviderConfig(
config({
@@ -1646,3 +1647,24 @@ it('persists and reloads Google runtime/model routes without converting them to
await createCustomProvider(config);
expect((await getCustomProvider(config.id))?.runtimes).toEqual(config.runtimes);
});
+
+
+describe('imported capability persistence', () => {
+ it('round-trips every shared field and explicit false through storage for all engines', async () => {
+ mountDb();
+ const model = { id: 'gpt-7-sol', name: 'Seven', nativeApi: 'openai-responses' as const,
+ contextWindow: 128000, contextWindowMax: 1000000, maxOutputTokens: 32000,
+ supportsFastMode: false, supportsImageInput: false, supportsToolCalls: false,
+ reasoning: false, efforts: [], defaultEffort: null, reasoningRequired: false,
+ defaultEnabled: false, modalities: { input: ['text'], output: ['text'] },
+ discoveredMetadata: { supportsFastMode: true, efforts: ['high'] as ['high'] },
+ };
+ const runtime = { baseUrl: 'https://relay.example/v1', wireProtocol: 'openai-responses' as const, models: [model] };
+ await createCustomProvider({ id: 'all-fields', name: 'All fields', runtimes: {
+ 'claude-code': runtime, codex: runtime, pi: runtime,
+ } });
+ const stored = await getCustomProvider('all-fields');
+ for (const agent of ['claude-code', 'codex', 'pi'] as const)
+ expect(stored?.runtimes[agent]?.models[0]).toEqual(model);
+ });
+});
diff --git a/apps/desktop/src/main/maker-host/__tests__/modelDefaultSelection.test.ts b/apps/desktop/src/main/maker-host/__tests__/modelDefaultSelection.test.ts
index 529bfe6c6a..fced9d8fb1 100644
--- a/apps/desktop/src/main/maker-host/__tests__/modelDefaultSelection.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/modelDefaultSelection.test.ts
@@ -2,149 +2,36 @@ import { describe, expect, it } from 'vitest';
import { isModelVisible } from '@cindy/model-providers';
import { selectDefaultModels } from '../model-default-selection.js';
-const model = (id: string, extra = {}) => ({ id, mode: 'chat', ...extra });
-
-describe('small provider model defaults', () => {
- it('accepts sparse metadata and the Responses chat mode without removing any model', () => {
- const models = [{ id: 'new-bare-model' }, { id: 'gpt-6-astra', mode: 'responses' },
- { id: 'google/gemini-3.8-flash' }, { id: 'google/gemini-3.7-flash' }];
+describe('discovered model defaults', () => {
+ it.each([undefined, 'xd', 'openai', 'openrouter'])('keeps new families and variants visible for %s', provider => {
+ const models = ['gpt-6-astra', 'gpt-7-astra', 'gpt-7-new-variant', 'claude-mythos-5',
+ 'google/gemini-4-preview', 'deepseek/deepseek-v4-flash-vision-exp', 'private-new-model',
+ 'gpt-6-sol[1m]'].map(id => ({ id, defaultEnabled: true }));
const before = structuredClone(models);
- expect([...selectDefaultModels(models, 'xd')]).toEqual([
- 'new-bare-model', 'gpt-6-astra', 'google/gemini-3.8-flash',
- ]);
+ expect([...selectDefaultModels(models, provider)]).toEqual(models.map(model => model.id));
expect(models).toEqual(before);
});
- it('selects reviewed DS4 vision and HY4 routes only with live Gateway capabilities', () => {
- const text = { input: ['text'], output: ['text'] };
- const vision = { input: ['text', 'image'], output: ['text'] };
- const models = [
- model('deepseek/deepseek-v4-pro', { modalities: text }),
- model('deepseek/deepseek-v4-flash', { modalities: text }),
- model('deepseek/deepseek-v4-flash-vision-exp', { modalities: vision }),
- model('tencent/hy3', { modalities: text }),
- model('tencent/hy4-preview', { modalities: text }),
- ];
- expect([...selectDefaultModels(models, 'xd')]).toEqual([
- 'deepseek/deepseek-v4-flash-vision-exp', 'tencent/hy4-preview',
- ]);
- // Identical IDs from direct/custom providers do not inherit Gateway recommendations.
- for (const provider of [undefined, 'deepseek', 'openrouter']) {
- expect([...selectDefaultModels(models, provider)]).toEqual(['deepseek/deepseek-v4-pro', 'tencent/hy3']);
- }
- const missingEvidence = models.map(({ id, mode }) => ({ id, mode }));
- expect([...selectDefaultModels(missingEvidence, 'xd')]).toEqual(['deepseek/deepseek-v4-pro', 'tencent/hy3']);
- const changed = models.map(model => ({ ...model, modalities: text }));
- expect([...selectDefaultModels(changed, 'xd')]).toEqual(['deepseek/deepseek-v4-pro', 'tencent/hy4-preview']);
- const unavailable = models.map(model => model.id.endsWith('exp') || model.id.endsWith('preview')
- ? { ...model, availability: 'requires_payment' } : model);
- expect([...selectDefaultModels(unavailable, 'xd')]).toEqual(['deepseek/deepseek-v4-pro', 'tencent/hy3']);
- const disabled = models.map(model => ({ ...model, defaultEnabled: false }));
- expect([...selectDefaultModels(disabled, 'xd')]).toEqual([]);
- });
-
- it('keeps only the latest Gemini, even when older generations have larger discounts', () => {
- const models = [
- model('google/gemini-3.5-flash', { costDiscount: 0.9 }),
- model('google/gemini-3.8-flash'),
- model('google/gemini-3.7-flash'),
- ];
- expect([...selectDefaultModels(models)]).toEqual(['google/gemini-3.8-flash']);
- expect([...selectDefaultModels([...models, model('google/gemini-4-flash')])]).toEqual([
- 'google/gemini-4-flash',
- ]);
- });
-
- it('automatically enables a newly delivered GPT flagship and retains Sol, Terra and Luna independently', () => {
+ it('retains explicit off, paid locks and media boundaries', () => {
const models = [
- model('openai/gpt-5.6-sol'),
- model('openai/gpt-5.6-terra'),
- model('openai/gpt-5.6-luna'),
- model('openai/gpt-6-astra'),
+ { id: 'off', defaultEnabled: false },
+ { id: 'locked', availability: 'requires_payment' },
+ { id: 'image', mode: 'image_generation' },
+ { id: 'new', mode: 'responses' },
+ { id: 'sparse' },
];
- expect([...selectDefaultModels(models)]).toEqual(['openai/gpt-6-astra', 'openai/gpt-5.6-sol', 'openai/gpt-5.6-terra', 'openai/gpt-5.6-luna']);
- const updated = [...models, model('openai/gpt-7-astra')];
- expect([...selectDefaultModels(updated)]).toEqual(['openai/gpt-7-astra', 'openai/gpt-5.6-sol', 'openai/gpt-5.6-terra', 'openai/gpt-5.6-luna']);
- expect(isModelVisible(true, false)).toBe(true); // Explicitly selected older models remain visible.
- expect(isModelVisible(false, true)).toBe(false); // Explicitly hidden new models stay hidden.
+ expect([...selectDefaultModels(models)]).toEqual(['new', 'sparse']);
+ expect(isModelVisible(false, true)).toBe(false);
+ expect(isModelVisible(true, false)).toBe(true);
});
- it('keeps Spark, long-context variants and unapproved Claude families opt-in', () => {
- expect([...selectDefaultModels([
- model('gpt-5.3-codex-spark'), model('gpt-5.4-mini'), model('gpt-5.4-nano'),
- model('gpt-5.6-sol[1m]'), model('claude-mythos-5'), model('gpt-6-astra'),
- ])]).toEqual(['gpt-6-astra']);
- expect([...selectDefaultModels([
- model('claude-fable-5-1', { availability: 'requires_payment' }),
- model('claude-fable-5'), model('claude-opus-5'), model('claude-opus-4-8'),
- model('claude-sonnet-5'), model('claude-haiku-4-5'),
- ])]).toEqual(['claude-fable-5', 'claude-opus-5', 'claude-sonnet-5', 'claude-haiku-4-5']);
- });
-
- it('prefers the larger discount among equivalent routes without merging or removing them', () => {
+ it('does not turn a preferred default, discount or alias into a visibility filter', () => {
const models = [
- model('openai/gpt-6-astra', { costDiscount: 0.4 }),
- model('codex/gpt-6-astra', { costDiscount: 0.85 }),
- model('x-ai/grok-4.6'),
- model('x-ai-grok/grok-4.6', { costDiscount: 0.9 }),
+ { id: 'z-ai/glm-5.3' },
+ { id: 'z-ai/glm-5.3-flash', newSessionDefault: ['pi'] },
+ { id: 'openai/gpt-6-astra', costDiscount: 0.4 },
+ { id: 'codex/gpt-6-astra', costDiscount: 0.85 },
];
- const before = structuredClone(models);
- expect([...selectDefaultModels(models)]).toEqual(['codex/gpt-6-astra', 'x-ai-grok/grok-4.6']);
- expect(models).toEqual(before);
- expect([...selectDefaultModels([...models].reverse())].sort()).toEqual(
- [...selectDefaultModels(models)].sort(),
- );
- });
-
- it('keeps all four Claude families, treating dates and parameter sizes as variants', () => {
- const models = [
- model('anthropic/claude-opus-4-8'),
- model('anthropic/claude-opus-5'),
- model('anthropic-claude/claude-fable-5-1'),
- model('claude-sonnet-5'),
- model('claude-haiku-4-5-20251001'),
- model('qwen/qwen3.8-27b'),
- model('qwen/qwen3.8-max'),
- ];
- expect([...selectDefaultModels(models)]).toEqual([
- 'anthropic-claude/claude-fable-5-1',
- 'anthropic/claude-opus-5',
- 'claude-sonnet-5',
- 'claude-haiku-4-5-20251001',
- 'qwen/qwen3.8-max',
- ]);
- });
-
- it('keeps experiments, paid locks, upstream-disabled models and media out of everyday defaults', () => {
- const models = [
- model('google/gemini-4-preview'),
- model('google/gemini-5', { availability: 'requires_payment' }),
- model('google/gemini-6', { defaultEnabled: false }),
- model('google/gemini-3.8-flash'),
- model('deepseek/deepseek-v4-flash-vision-exp'),
- model('openai/gpt-5.6-sol[1m]'),
- model('openai/gpt-image-2', { mode: 'image' }),
- model('meta/muse-spark-1.3'),
- ];
- expect([...selectDefaultModels(models)]).toEqual([
- 'google/gemini-3.8-flash',
- 'meta/muse-spark-1.3',
- ]);
- });
-
- it('honors a server-selected everyday model and works with subscription wire aliases', () => {
- expect([
- ...selectDefaultModels([
- model('z-ai/glm-5.3'),
- model('z-ai/glm-5.3-flash', { newSessionDefault: ['pi'] }),
- ]),
- ]).toEqual(['z-ai/glm-5.3-flash']);
- expect([
- ...selectDefaultModels([
- model('chatgpt/gpt-6-astra'),
- model('chatgpt/gpt-5.6-sol'),
- model('chatgpt/gpt-5.6-luna'),
- ]),
- ]).toEqual(['chatgpt/gpt-6-astra', 'chatgpt/gpt-5.6-sol', 'chatgpt/gpt-5.6-luna']);
+ expect([...selectDefaultModels(models)]).toEqual(models.map(model => model.id));
});
});
diff --git a/apps/desktop/src/main/maker-host/__tests__/modelDiscoveryPages.test.ts b/apps/desktop/src/main/maker-host/__tests__/modelDiscoveryPages.test.ts
new file mode 100644
index 0000000000..4bd9c48417
--- /dev/null
+++ b/apps/desktop/src/main/maker-host/__tests__/modelDiscoveryPages.test.ts
@@ -0,0 +1,47 @@
+import { describe, expect, it, vi } from 'vitest';
+import { collectModelDiscoveryPages } from '../model-discovery-pages.js';
+
+const endpoint = 'https://models.example/v1/models';
+describe('model list pagination', () => {
+ it('merges later metadata for repeated IDs without dropping earlier fields', async () => {
+ const models = await collectModelDiscoveryPages({ data: [{ id: 'one', name: 'Named model',
+ supports_fast_mode: true, max_output_tokens: 4000 }], next: '?page=2' }, endpoint,
+ async () => ({ data: [{ id: 'one', context_window: 64000, supports_fast_mode: false,
+ efforts: [], defaultEffort: null }] }));
+ expect(models).toHaveLength(1);
+ expect(models?.[0]).toMatchObject({ name: 'Named model', contextWindow: 64000,
+ discoveredMetadata: { contextWindow: 64000, maxOutputTokens: 4000,
+ supportsFastMode: false, efforts: [], defaultEffort: null } });
+ });
+
+ it.each([
+ [{ has_more: true, last_id: 'one' }, 'after_id=one'],
+ [{ nextPageToken: 'two' }, 'pageToken=two'],
+ [{ next_page_token: 'two' }, 'page_token=two'],
+ [{ next_cursor: 'two' }, 'cursor=two'],
+ [{ next: '?page=2' }, 'page=2'],
+ ])('follows cursors without losing metadata: %j', async (pagination, query) => {
+ const fetch = vi.fn(async (url: string) => {
+ expect(url).toBe(`${endpoint}?${query}`);
+ return { data: [{ id: 'one' }, { id: 'two', native_api: 'openai-responses', supports_fast_mode: false }] };
+ });
+ const models = await collectModelDiscoveryPages({ data: [{ id: 'one', context_window: 128000 }], ...pagination }, endpoint, fetch);
+ expect(models?.map(model => model.id)).toEqual(['one', 'two']);
+ expect(models?.[0]?.discoveredMetadata?.contextWindow).toBe(128000);
+ expect(models?.[1]?.discoveredMetadata?.nativeApi).toBe('openai-responses');
+ });
+
+ it.each(['https://attacker.example/v1/models?page=2', '/other?page=2', 'https://user:[email protected]/v1/models'])('never sends credentials to an untrusted next link: %s', async next => {
+ const fetch = vi.fn();
+ expect(await collectModelDiscoveryPages({ data: ['one'], next }, endpoint, fetch)).toHaveLength(1);
+ expect(fetch).not.toHaveBeenCalled();
+ });
+
+ it('retains completed pages on failure and stops repeated cursors', async () => {
+ const page = { data: ['one'], next_cursor: 'repeat' };
+ const fetch = vi.fn(async () => page);
+ expect(await collectModelDiscoveryPages(page, endpoint, fetch)).toHaveLength(1);
+ expect(fetch).toHaveBeenCalledOnce();
+ expect(await collectModelDiscoveryPages(page, endpoint, async () => { throw new Error('offline'); })).toHaveLength(1);
+ });
+});
diff --git a/apps/desktop/src/main/maker-host/__tests__/modelPlane.test.ts b/apps/desktop/src/main/maker-host/__tests__/modelPlane.test.ts
index 1e85b96d19..1bce0d9145 100644
--- a/apps/desktop/src/main/maker-host/__tests__/modelPlane.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/modelPlane.test.ts
@@ -82,20 +82,10 @@ function models(providerId: string, agent: 'claude-code' | 'codex' | 'pi'): Cata
return p?.models[agent] ?? [];
}
-function withNativeMetadataAndDefaults(
+function withNativeMetadata(
providerId: string,
models: readonly CatalogModel[] = [],
): CatalogModel[] {
- const defaults: Record<string, readonly string[]> = {
- xai: ['grok-4.6'],
- anthropic: ['claude-fable-5-1', 'claude-opus-5', 'claude-sonnet-5', 'claude-haiku-4-5'],
- openai: [
- 'chatgpt/gpt-6-astra',
- 'chatgpt/gpt-5.6-sol',
- 'chatgpt/gpt-5.6-terra',
- 'chatgpt/gpt-5.6-luna',
- ],
- };
return models.map((model) => {
const nativeApi = resolveModelNativeApi(BUNDLED_CATALOG.modelRegistry, providerId, model.id);
return {
@@ -107,7 +97,6 @@ function withNativeMetadataAndDefaults(
nativeApi === 'google-generative-ai'
? { nativeApi }
: {}),
- ...(defaults[providerId]?.includes(model.id) ? {} : { defaultEnabled: false }),
};
});
}
@@ -548,7 +537,7 @@ describe('registry presence 实体化', () => {
expect(models('anthropic', 'claude-code').map((m) => m.id)).toEqual(['claude-next']);
expect(models('anthropic', 'codex')).toEqual([]);
expect(models('anthropic', 'pi')).toEqual(
- withNativeMetadataAndDefaults(
+ withNativeMetadata(
'anthropic',
BUNDLED_CATALOG.providers.find((provider) => provider.id === 'anthropic')?.models.pi,
),
@@ -1057,7 +1046,7 @@ describe('本地 override(local 永远最高)', () => {
});
describe('cross-harness defaults', () => {
- it('keeps curated native/Pi defaults and makes the Claude Code bridge opt-in', () => {
+ it('keeps discovered native/Pi models enabled and makes the Claude Code bridge opt-in', () => {
setActiveCatalog(baseCatalog([gpt6Entry({ defaultEnabled: true })]));
expect(models('openai', 'codex').find((m) => m.id === 'gpt-6')?.defaultEnabled).toBe(true);
expect(
@@ -1065,7 +1054,7 @@ describe('cross-harness defaults', () => {
).toBe(false);
expect(models('openai', 'pi').find((m) => m.id === 'chatgpt/gpt-6')).toBeUndefined();
expect(models('openai', 'pi')).toEqual(
- withNativeMetadataAndDefaults(
+ withNativeMetadata(
'openai',
BUNDLED_CATALOG.providers.find((p) => p.id === 'openai')?.models.pi,
),
diff --git a/apps/desktop/src/main/maker-host/__tests__/piNativeProviders.test.ts b/apps/desktop/src/main/maker-host/__tests__/piNativeProviders.test.ts
index f28e06006b..851f10c5a3 100644
--- a/apps/desktop/src/main/maker-host/__tests__/piNativeProviders.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/piNativeProviders.test.ts
@@ -2763,7 +2763,7 @@ it('carries every unambiguous portable catalog model into the Pi descriptor', ()
} },
}], () => 'test-key');
expect(providers[0]?.models[0], `${row.upstream} ${row.id}`).toMatchObject({
- api, input: row.modalities.input.filter(value => value === 'text' || value === 'image'),
+ api, input: (row.modalities?.input ?? ['text']).filter(value => value === 'text' || value === 'image'),
contextWindow: row.contextWindow, maxTokens: row.maxOutput,
});
checked++;
@@ -2879,3 +2879,34 @@ it('materializes a future GPT generation with inherited parameters in native Pi'
thinkingLevelMap: { low: 'low', medium: 'medium', high: 'high' } });
expect(result.providers[0]?.models[0]?.cost).toBeUndefined();
});
+
+
+it.each([
+ ['cloudflare-ai-gateway', true],
+ ['cloudflare-ai-gateway', false],
+ ['github-copilot', true],
+ ['github-copilot', false],
+] as const)('keeps the %s adapter for future generations (preset=%s)', (adapter, withPreset) => {
+ const row = PROVIDER_MODEL_CATALOG.providers[adapter].find(model => model.id === 'claude-opus-5')!;
+ const baseUrl = row.upstream.replace('{CLOUDFLARE_ACCOUNT_ID}', 'fixture-account')
+ .replace('{CLOUDFLARE_GATEWAY_ID}', 'fixture-gateway');
+ const readKey = vi.fn(() => 'current-connection-key');
+ const result = buildPiNativeProvidersFromConfigs([{ id: 'independent-connection', name: 'Independent',
+ auth: { method: 'apiKey' }, runtimes: { pi: { baseUrl, wireProtocol: 'anthropic-messages',
+ ...(withPreset ? { catalogPresetId: adapter } : {}),
+ models: [{ id: 'claude-opus-9', api: 'anthropic-messages' }],
+ } } }], readKey);
+ expect(result.providers[0]).toMatchObject({ id: 'independent-connection', adapterProvider: adapter,
+ baseUrl, models: [{ id: 'claude-opus-9' }] });
+ expect(readKey).toHaveBeenCalledExactlyOnceWith('independent-connection', 'pi');
+ expect(Object.values(result.env)).toEqual(['current-connection-key']);
+ expect(JSON.stringify(result.providers)).not.toContain('current-connection-key');
+});
+
+it('does not borrow a provider adapter for a future generation on an unrelated relay', () => {
+ const result = buildPiNativeProvidersFromConfigs([{ id: 'unrelated-relay', name: 'Relay',
+ runtimes: { pi: { baseUrl: 'https://relay.example/v1', wireProtocol: 'anthropic-messages',
+ models: [{ id: 'claude-opus-9' }] } } }], () => 'fixture-key');
+ expect(result.providers[0]?.models[0].id).toBe('claude-opus-9');
+ expect(result.providers[0]?.adapterProvider).toBeUndefined();
+});
diff --git a/apps/desktop/src/main/maker-host/__tests__/piProviderTransport.test.ts b/apps/desktop/src/main/maker-host/__tests__/piProviderTransport.test.ts
index e19ca6145e..c26ac17f86 100644
--- a/apps/desktop/src/main/maker-host/__tests__/piProviderTransport.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/piProviderTransport.test.ts
@@ -104,7 +104,7 @@ describe('Pi-owned transport for Cindy harnesses', () => {
it.each(['ant-ling', 'qwen-token-plan', 'zai', 'together'])('sends the actual %s thinking dialect and model limits', async providerId => {
const row = PROVIDER_MODEL_CATALOG.providers[providerId].find(row => row.reasoning && row.execution.pi.api === 'openai-completions')!;
- const effort = row.efforts.includes('high') ? 'high' : row.efforts[0];
+ const effort = row.efforts?.includes('high') ? 'high' : row.efforts?.[0];
let sent: Record<string, unknown> | undefined;
const send = createPiProviderFetch({ row, providerId, apiKey: 'fixture-provider-key', fetchImpl: async (_url, init) => {
sent = JSON.parse(String(init?.body));
diff --git a/apps/desktop/src/main/maker-host/__tests__/xaiSyncImport.test.ts b/apps/desktop/src/main/maker-host/__tests__/xaiSyncImport.test.ts
index 1143371949..c2510e1f52 100644
--- a/apps/desktop/src/main/maker-host/__tests__/xaiSyncImport.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/xaiSyncImport.test.ts
@@ -185,14 +185,20 @@ describe('xAI API sync through the actual picker import', () => {
expect(row.defaultEnabled).toBe(false);
});
- it('does not infer capabilities or prices from a Fast suffix, description, or adjacent model', () => {
- const candidate = build({ data: [...account.data, { ...account.data[0], id: 'grok-4.8-build-fast',
+ it.each([
+ ['grok-4.8-build-fast', true, false],
+ ['grok-unrelated-build-fast', undefined, undefined],
+ ] as const)('inherits only matching previous-variant capabilities, never prices: %s', (id, vision, fastToggle) => {
+ const candidate = build({ data: [...account.data, { ...account.data[0], id,
name: 'Grok 4.8 Fast', description: 'Fast variant. 2x the price.' }] });
+ const reported = candidate.catalog.modelRegistry!.baseModels!.find(model => model.id === `xai/${id}`)!.defaults;
+ expect(reported.supportsFastMode).toBeUndefined();
+ expect(reported.supportsImageInput).toBeUndefined();
const report = inspectXaiImport(candidate);
- const fast = report.rows.find(row => row.modelId.endsWith('build-fast'))!;
+ const fast = report.rows.find(row => row.modelId === `xai/${id}`)!;
for (const row of Object.values(fast.harnesses)) {
- expect(row.supportsFastMode).toBeUndefined();
- expect(row.supportsImageInput).toBeUndefined();
+ expect(row.supportsFastMode).toBe(fastToggle);
+ expect(row.supportsImageInput).toBe(vision);
expect(row.apiReferencePrice).toBeNull();
expect(row.defaultEnabled).toBe(row.protocol?.mode === 'matching');
}
diff --git a/apps/desktop/src/main/maker-host/codex-proxy-host.ts b/apps/desktop/src/main/maker-host/codex-proxy-host.ts
index f23d56ab56..9fa4bbb9ff 100644
--- a/apps/desktop/src/main/maker-host/codex-proxy-host.ts
+++ b/apps/desktop/src/main/maker-host/codex-proxy-host.ts
@@ -1189,7 +1189,7 @@ function createChatBridgeDecision(
const nativeHeaders = overrideHeadersCaseInsensitive(withChatBridgeUserAgent(Object.fromEntries(Object.entries(headers).filter(([name]) =>
!['authorization', 'x-api-key'].includes(name.toLowerCase())))), resolveConversationSessionHeaders(ctx?.headers));
if (standard.execution.pi.api === 'anthropic-messages' && (actualModel.endsWith('[1m]')
- || (isOfficialAnthropicUpstream(standard.upstream) && standard.contextWindow >= 1_000_000))) {
+ || (isOfficialAnthropicUpstream(standard.upstream) && (standard.contextWindow ?? 0) >= 1_000_000))) {
appendCommaSeparatedHeaderToken(nativeHeaders, 'anthropic-beta', 'context-1m-2025-08-07');
}
const nativeFetch = createPiProviderFetch({ row: { ...standard,
diff --git a/apps/desktop/src/main/maker-host/custom-provider-store.ts b/apps/desktop/src/main/maker-host/custom-provider-store.ts
index f87c6830a5..761d966850 100644
--- a/apps/desktop/src/main/maker-host/custom-provider-store.ts
+++ b/apps/desktop/src/main/maker-host/custom-provider-store.ts
@@ -1,4 +1,4 @@
-import { pickModelMetadata } from '@cindy/model-providers';
+import { pickModelMetadata, MODEL_METADATA_FIELDS } from '@cindy/model-providers';
import {
mergeDiscoveredRuntimeModels,
validModelMetadata,
@@ -107,12 +107,12 @@ function parseStoredReasoningCapability(
model: Record<string, unknown>,
): Partial<ProviderRuntimeModelConfig> {
if (model.reasoning === false) return { reasoning: false };
+ if (model.reasoning === true && model.reasoningEfforts === undefined) return { reasoning: true };
if (model.reasoning !== true || !Array.isArray(model.reasoningEfforts)) {
return {};
}
const efforts = model.reasoningEfforts.filter((item) => isReasoningEffortForAgent(agent, item));
if (
- efforts.length === 0 ||
efforts.length !== model.reasoningEfforts.length ||
new Set(efforts).size !== efforts.length
) {
@@ -241,7 +241,7 @@ function validateRuntime(agent: string, rt: unknown): ValidationResult {
if (typeof mm.name !== 'string' || mm.name.trim().length === 0) {
return invalid(`runtime '${agent}' model.name required`);
}
- for (const field of ['mode', 'modalities', 'officialDocs'] as const) {
+ for (const field of MODEL_METADATA_FIELDS) {
if (mm[field] !== undefined && !validModelMetadata({ [field]: mm[field] })) return invalid(`runtime '${agent}' model.${field} invalid`);
}
if (mm.discoveredCost !== undefined && (
@@ -305,19 +305,19 @@ function validateRuntime(agent: string, rt: unknown): ValidationResult {
return invalid(`runtime '${agent}' model.reasoning must be a boolean`);
}
if (mm.reasoning === true) {
- if (!Array.isArray(mm.reasoningEfforts) || mm.reasoningEfforts.length === 0) {
- return invalid(`runtime '${agent}' model.reasoningEfforts must be a non-empty array`);
+ if (mm.reasoningEfforts !== undefined && !Array.isArray(mm.reasoningEfforts)) {
+ return invalid(`runtime '${agent}' model.reasoningEfforts must be an array`);
}
if (
- mm.reasoningEfforts.some((effort) => !isReasoningEffortForAgent(agent, effort)) ||
- new Set(mm.reasoningEfforts).size !== mm.reasoningEfforts.length
+ (mm.reasoningEfforts as unknown[] | undefined)?.some((effort) => !isReasoningEffortForAgent(agent, effort)) ||
+ new Set(mm.reasoningEfforts as unknown[] | undefined).size !== ((mm.reasoningEfforts as unknown[] | undefined)?.length ?? 0)
) {
return invalid(`runtime '${agent}' model.reasoningEfforts invalid`);
}
if (
mm.reasoningDefaultEffort !== undefined &&
(!isReasoningEffortForAgent(agent, mm.reasoningDefaultEffort) ||
- !mm.reasoningEfforts.includes(mm.reasoningDefaultEffort))
+ !(mm.reasoningEfforts as unknown[] | undefined)?.includes(mm.reasoningDefaultEffort))
) {
return invalid(`runtime '${agent}' model.reasoningDefaultEffort invalid`);
}
@@ -587,7 +587,7 @@ function normalizeRuntime(
.map((m) => ({
id: m.id.trim(),
name: m.name.trim(),
- ...pickModelMetadata({ mode: m.mode, modalities: m.modalities, officialDocs: m.officialDocs }),
+ ...pickModelMetadata({ ...m, name: m.name.trim() }),
...(m.discoveredMetadata ? { discoveredMetadata: m.discoveredMetadata } : {}),
...(m.discoveredCost ? { discoveredCost: m.discoveredCost } : {}),
...(m.nameExplicit === true ? { nameExplicit: true } : {}),
@@ -607,10 +607,10 @@ function normalizeRuntime(
...(typeof m.supportsImageInput === 'boolean'
? { supportsImageInput: m.supportsImageInput }
: {}),
- ...(m.reasoning === true && m.reasoningEfforts?.length
+ ...(m.reasoning === true
? {
reasoning: true,
- reasoningEfforts: [...m.reasoningEfforts],
+ ...(m.reasoningEfforts ? { reasoningEfforts: [...m.reasoningEfforts] } : {}),
...(m.reasoningDefaultEffort
? { reasoningDefaultEffort: m.reasoningDefaultEffort }
: {}),
@@ -808,7 +808,7 @@ function parseRuntimes(raw: string): Partial<Record<AgentKind, CustomProviderRun
return {
id: String(m.id),
name: String(m.name ?? ''),
- ...pickModelMetadata({ mode: m.mode, modalities: m.modalities, officialDocs: m.officialDocs }),
+ ...pickModelMetadata(m),
...(isPiModelApi(m.api) ? { api: m.api } : {}),
...(agent === 'pi' && isPiModelApi(m.piApi) ? { piApi: m.piApi } : {}),
...(route ? { route } : {}),
diff --git a/apps/desktop/src/main/maker-host/generic-oauth.ts b/apps/desktop/src/main/maker-host/generic-oauth.ts
index de38a0ac74..4fbbf4eace 100644
--- a/apps/desktop/src/main/maker-host/generic-oauth.ts
+++ b/apps/desktop/src/main/maker-host/generic-oauth.ts
@@ -1,5 +1,6 @@
+import { collectModelDiscoveryPages } from './model-discovery-pages.js';
import { providerOAuthContract } from '@cindy/model-providers';
-import { parseModelsListResponse, isOpenRouterModelsUrl, type DiscoveredModel } from '@cindy/model-providers';
+import { isOpenRouterModelsUrl, type DiscoveredModel } from '@cindy/model-providers';
/**
* generic-oauth —— 目录 `auth.oauth` 描述符驱动的通用 OAuth Runner。
*
@@ -871,7 +872,14 @@ export async function discoverGenericOAuthModels(
} catch {
return null;
}
- return parseModelsListResponse(json, url);
+ const pageDeadline = Date.now() + 30_000;
+ return collectModelDiscoveryPages(json, url, async nextUrl => {
+ if (Date.now() >= pageDeadline) throw new Error('catalog deadline exceeded');
+ const page = await io.fetchImpl(nextUrl, { headers, redirect: 'error',
+ signal: AbortSignal.timeout(Math.max(1, Math.min(REFRESH_FETCH_TIMEOUT_MS, pageDeadline - Date.now()))) });
+ if (!page.ok) { await page.body?.cancel(); throw new Error('catalog page failed'); }
+ return page.json();
+ });
}
diff --git a/apps/desktop/src/main/maker-host/model-default-selection.ts b/apps/desktop/src/main/maker-host/model-default-selection.ts
index 2a6a5b5417..b8911ea1b4 100644
--- a/apps/desktop/src/main/maker-host/model-default-selection.ts
+++ b/apps/desktop/src/main/maker-host/model-default-selection.ts
@@ -8,159 +8,16 @@ interface DefaultModelCandidate {
modalities?: { input: readonly string[]; output: readonly string[] };
}
-/** Small default selection over the current account's provider catalog, never an availability filter.
- * Older/duplicate models stay in settings; renderer visibility overrides still win afterwards.
+/** Discovery determines membership; explicit upstream defaults determine visibility.
+ * A new family, variant or preview must not wait for a client allowlist update.
+ * User visibility overrides are applied by the caller afterwards.
*/
export function selectDefaultModels(
models: readonly DefaultModelCandidate[],
- providerId?: string,
+ _providerId?: string,
): ReadonlySet<string> {
- const groups = new Map<string, DefaultModelCandidate[]>();
- const selected = new Set<string>();
- for (const model of models) {
- if (model.defaultEnabled === false || model.availability === 'requires_payment') continue;
- if (model.mode && model.mode !== 'chat' && model.mode !== 'responses') continue;
- // Alternate-context routes remain opt-in. Reviewed Gateway previews need live route
- // capabilities; shared Registry metadata must not enable a direct-provider experiment.
- if (/\[1m\]|:auto$/i.test(model.id)) continue;
- if (
- /(?:^|[-/:])(?:exp|experimental|preview)(?:$|[-/:])/i.test(model.id) &&
- !isReviewedGatewayPreview(model, providerId)
- )
- continue;
- const vendor = vendorOf(model.id);
- // Unrecognized bare IDs retain their upstream policy. Do not guess their manufacturer.
- if (!vendor) {
- selected.add(model.id);
- continue;
- }
- const group = groups.get(vendor) ?? [];
- group.push(model);
- groups.set(vendor, group);
- }
- for (const [vendor, candidates] of groups) {
- const sorted = [...candidates].sort((a, b) => compareCandidates(a, b, providerId));
- if (vendor === 'anthropic' || vendor === 'openai') {
- // Keep each approved family, so a new flagship never removes the everyday families.
- const roles = vendor === 'anthropic'
- ? ['fable', 'opus', 'sonnet', 'haiku']
- : ['flagship', 'sol', 'terra', 'luna'];
- for (const role of roles) {
- const winner = sorted.find((model) => roleOf(vendor, model.id) === role);
- if (winner) selected.add(winner.id);
- }
- } else {
- const winner = sorted.find((model) => model.newSessionDefault?.length) ?? sorted[0];
- if (winner) selected.add(winner.id);
- }
- }
- return selected;
-}
-
-function slug(id: string): string {
- return id.slice(id.lastIndexOf('/') + 1).toLowerCase();
-}
-
-function vendorOf(id: string): string | null {
- const [namespace] = id.toLowerCase().split('/');
- if (id.includes('/')) {
- const aliases: Record<string, string> = {
- 'anthropic-claude': 'anthropic',
- codex: 'openai',
- chatgpt: 'openai',
- 'x-ai-grok': 'xai',
- 'x-ai': 'xai',
- moonshotai: 'moonshot',
- };
- return Object.hasOwn(aliases, namespace!) ? aliases[namespace!]! : namespace!;
- }
- const families: Array<[string, RegExp]> = [
- ['anthropic', /^claude-/],
- ['openai', /^gpt-/],
- ['google', /^gemini-/],
- ['deepseek', /^deepseek-/],
- ['qwen', /^qwen/],
- ['z-ai', /^glm-/],
- ['moonshot', /^kimi-/],
- ['xai', /^grok-/],
- ['tencent', /^hy\d/],
- ];
- return families.find(([, pattern]) => pattern.test(id.toLowerCase()))?.[0] ?? null;
-}
-
-function roleOf(vendor: string, id: string): string | null {
- const name = slug(id);
- if (vendor === 'anthropic') {
- return /^claude-(fable|opus|sonnet|haiku)-/.exec(name)?.[1] ?? null;
- }
- const family = /(?:^|-)(sol|terra|luna)(?:-|$)/.exec(name)?.[1];
- if (family) return family;
- // Spark and small legacy variants remain opt-in. Bare numbered GPTs provide the
- // flagship fallback on accounts that do not yet have Astra.
- return /^gpt-\d+(?:[.-]\d+)*(?:-astra)?$/.test(name) ? 'flagship' : null;
-}
-
-function version(id: string): number[] {
- // Date suffixes identify snapshots, not model generations; 4-5 and 4.5 compare equally.
- return (
- slug(id)
- .replace(/\d+b(?:-|$)/g, '')
- .match(/\d+/g) ?? []
- )
- .filter((part) => part.length < 6)
- .slice(0, 3)
- .map(Number);
-}
-
-function variantRank(id: string): number {
- const name = slug(id);
- if (/(?:^|-)(fable|sonnet|sol|max|pro)(?:-|$)/.test(name)) return 0;
- if (/(?:^|-)(opus|terra)(?:-|$)/.test(name)) return 1;
- if (/(?:^|-)(haiku|luna|flash)(?:-|$)/.test(name)) return 2;
- return 1;
-}
-
-function discount(model: DefaultModelCandidate): number {
- const value = model.costDiscount;
- return typeof value === 'number' && Number.isFinite(value) && value >= 0 && value <= 1
- ? value
- : 0;
-}
-
-function hasImageInput(model: DefaultModelCandidate): boolean {
- return (
- model.modalities?.input.includes('image') === true && model.modalities.output.includes('text')
- );
-}
-
-/** Product-reviewed exceptions, scoped to live XD routes, never inferred from a model name. */
-function isReviewedGatewayPreview(model: DefaultModelCandidate, providerId?: string): boolean {
- if (
- providerId !== 'xd' ||
- model.mode !== 'chat' ||
- !model.modalities?.input.includes('text') ||
- !model.modalities.output.includes('text')
- )
- return false;
- if (model.id === 'deepseek/deepseek-v4-flash-vision-exp') return hasImageInput(model);
- return model.id === 'tencent/hy4-preview';
-}
-
-function compareCandidates(
- a: DefaultModelCandidate,
- b: DefaultModelCandidate,
- providerId?: string,
-): number {
- const av = version(a.id),
- bv = version(b.id);
- for (let i = 0; i < Math.max(av.length, bv.length); i++) {
- const difference = (bv[i] ?? 0) - (av[i] ?? 0);
- if (difference) return difference;
- }
- return (
- (providerId === 'xd' ? Number(hasImageInput(b)) - Number(hasImageInput(a)) : 0) ||
- variantRank(a.id) - variantRank(b.id) ||
- discount(b) - discount(a) ||
- a.id.localeCompare(b.id, 'en')
- );
+ return new Set(models.filter(model =>
+ model.defaultEnabled !== false && model.availability !== 'requires_payment' &&
+ (!model.mode || model.mode === 'chat' || model.mode === 'responses')
+ ).map(model => model.id));
}
diff --git a/apps/desktop/src/main/maker-host/model-discovery-pages.ts b/apps/desktop/src/main/maker-host/model-discovery-pages.ts
new file mode 100644
index 0000000000..446b0d7ef3
--- /dev/null
+++ b/apps/desktop/src/main/maker-host/model-discovery-pages.ts
@@ -0,0 +1,52 @@
+import { mergeModelMetadata, parseModelsListResponse, type DiscoveredModel } from '@cindy/model-providers';
+
+/** Follow catalog cursors only on the original endpoint. Never forward credentials to a next-link host. */
+export async function collectModelDiscoveryPages(
+ first: unknown,
+ endpoint: string,
+ fetchPage: (url: string) => Promise<unknown>,
+): Promise<DiscoveredModel[] | null> {
+ const origin = new URL(endpoint);
+ const visited = new Set([origin.toString()]);
+ const models = new Map<string, DiscoveredModel>();
+ let page = first;
+ let current = origin;
+ for (let count = 0; count < 100; count++) {
+ const parsed = parseModelsListResponse(page, endpoint);
+ if (!parsed) return models.size ? [...models.values()] : null;
+ for (const model of parsed) {
+ const previous = models.get(model.id);
+ models.set(model.id, previous ? {
+ ...previous,
+ ...model,
+ name: model.discoveredMetadata?.name ?? previous.discoveredMetadata?.name ?? previous.name,
+ contextWindow: model.contextWindow ?? previous.contextWindow,
+ discoveredMetadata: mergeModelMetadata(previous.discoveredMetadata, model.discoveredMetadata),
+ ...(previous.discoveredCost || model.discoveredCost
+ ? { discoveredCost: { ...previous.discoveredCost, ...model.discoveredCost } } : {}),
+ } : model);
+ if (models.size >= 10_000) return [...models.values()];
+ }
+ if (!page || typeof page !== 'object' || Array.isArray(page)) break;
+ const data = page as Record<string, unknown>;
+ const next = new URL(current);
+ const token = data.nextPageToken ?? data.next_page_token;
+ const cursor = data.next_cursor;
+ const link = data.next ?? (data.links as { next?: unknown } | undefined)?.next;
+ if (typeof token === 'string' && token) next.searchParams.set(typeof data.nextPageToken === 'string' ? 'pageToken' : 'page_token', token);
+ else if (typeof cursor === 'string' && cursor) next.searchParams.set('cursor', cursor);
+ else if (data.has_more === true && typeof data.last_id === 'string' && data.last_id)
+ next.searchParams.set('after_id', data.last_id);
+ else if (typeof link === 'string' && link) {
+ let resolved: URL;
+ try { resolved = new URL(link, current); } catch { break; }
+ if (resolved.origin !== origin.origin || resolved.pathname !== origin.pathname || resolved.username || resolved.password) break;
+ next.href = resolved.href;
+ } else break;
+ if (visited.has(next.toString())) break;
+ visited.add(next.toString());
+ try { page = await fetchPage(next.toString()); } catch { break; }
+ current = next;
+ }
+ return [...models.values()];
+}
diff --git a/apps/desktop/src/main/maker-host/pi-gateway-model-catalog.ts b/apps/desktop/src/main/maker-host/pi-gateway-model-catalog.ts
index 42158c818f..2db4f764bb 100644
--- a/apps/desktop/src/main/maker-host/pi-gateway-model-catalog.ts
+++ b/apps/desktop/src/main/maker-host/pi-gateway-model-catalog.ts
@@ -44,7 +44,7 @@ function piApiFromWireProtocol(protocol: ProviderWireProtocol | undefined): PiMo
}
function catalogModelApi(
- model: ProviderRuntimeModelConfig,
+ model: Pick<ProviderRuntimeModelConfig, 'piApi' | 'route'>,
defaultProtocol: ProviderWireProtocol | undefined,
): PiModelApi | undefined {
return model.piApi ?? piApiFromWireProtocol(model.route?.wireProtocol ?? defaultProtocol);
diff --git a/apps/desktop/src/main/maker-host/pi-host.ts b/apps/desktop/src/main/maker-host/pi-host.ts
index 6d6be04d7f..639466d8ab 100644
--- a/apps/desktop/src/main/maker-host/pi-host.ts
+++ b/apps/desktop/src/main/maker-host/pi-host.ts
@@ -1287,7 +1287,7 @@ export function buildPiNativeProvidersFromConfigs(
id: row.id, name: row.name, baseUrl: row.upstream,
...row.execution.pi, api: row.execution.pi.api as PiModelApi,
contextWindow: row.contextWindow, maxTokens: row.maxOutput,
- input: row.modalities.input.filter((value): value is 'text' | 'image' => value === 'text' || value === 'image'),
+ input: (row.modalities?.input ?? ['text']).filter((value): value is 'text' | 'image' => value === 'text' || value === 'image'),
reasoning: row.reasoning,
cost: row.cost?.input !== undefined && row.cost.output !== undefined
? { input: row.cost.input, output: row.cost.output,
@@ -1353,10 +1353,14 @@ export function buildPiNativeProvidersFromConfigs(
onSkip?.(cfg.id, 'native SDK requires an approved cloud endpoint');
continue;
}
- const adapterIds = new Set(rt.models.flatMap(model => {
- const row = providerModelRecord(model.id, model.route?.baseUrl ?? rt.baseUrl, model.api ?? model.piApi)
- ?? providerPresetModelRecord(rt.catalogPresetId, model.id, model.api ?? model.piApi);
- const adapter = row ? providerModelAdapterId(row) : undefined;
+ const adapterIds = new Set(rt.models.flatMap((model, index) => {
+ const api = modelApis[index];
+ // The native ChatGPT subscription transport owns its authentication separately.
+ if (!api || api === 'openai-codex-responses') return [];
+ const row = providerModelRecord(model.id, model.route?.baseUrl ?? rt.baseUrl, api)
+ ?? providerPresetModelRecord(rt.catalogPresetId, model.id, api)
+ ?? providerModelGenerationRecord(model.id, model.route?.baseUrl ?? rt.baseUrl, api, rt.catalogPresetId);
+ const adapter = row ? providerModelAdapterId(row, rt.catalogPresetId) : undefined;
return adapter ? [adapter] : [];
}));
providers.push({
diff --git a/apps/desktop/src/main/maker-host/pi-provider-transport.ts b/apps/desktop/src/main/maker-host/pi-provider-transport.ts
index 503de25350..ac4db06d05 100644
--- a/apps/desktop/src/main/maker-host/pi-provider-transport.ts
+++ b/apps/desktop/src/main/maker-host/pi-provider-transport.ts
@@ -167,13 +167,13 @@ function nativeInvocationModel(options: PiProviderTransportOptions, modelId: str
const destination = options.upstream ?? row.upstream;
assertHostCredentialEndpoint(row.execution.pi.api, destination, options.apiKey);
const model: Model<Api> = { id: modelId, name: row.name, provider: providerModelAdapterId(row) ?? options.providerId,
- api: row.execution.pi.api, baseUrl: destination, contextWindow: row.contextWindow,
- maxTokens: row.maxOutput ?? Math.min(4096, row.contextWindow), reasoning: row.reasoning,
+ api: row.execution.pi.api, baseUrl: destination, contextWindow: row.contextWindow ?? 200_000,
+ maxTokens: row.maxOutput ?? Math.min(4096, row.contextWindow ?? 200_000), reasoning: row.reasoning ?? (row.efforts?.length ?? 0) > 0,
input: row.supportsImageInput ? ['text', 'image'] : ['text'],
cost: { input: row.cost?.input ?? 0, output: row.cost?.output ?? 0,
cacheRead: row.cost?.cacheRead ?? 0, cacheWrite: row.cost?.cacheWrite ?? 0 },
thinkingLevelMap: { ...row.execution.pi.thinkingLevelMap, ...Object.fromEntries(PI_REASONING_EFFORTS.map(level =>
- [level, row.efforts.includes(level) ? row.execution.pi.thinkingLevelMap?.[level] ?? level : null])) },
+ [level, row.efforts?.includes(level) ? row.execution.pi.thinkingLevelMap?.[level] ?? level : null])) },
compat: row.execution.pi.compat,
samplingParams: row.execution.pi.samplingParams,
headers: row.execution.pi.headers,
@@ -276,7 +276,7 @@ export function createPiProviderFetch(options: PiProviderTransportOptions): type
['openai-responses', 'azure-openai-responses', 'openai-completions'].includes(model.api)
? { onPayload: (payload: unknown) => payload && typeof payload === 'object' && !Array.isArray(payload)
? { ...payload, service_tier: 'priority' } : payload } : {}),
- reasoning: reconcileOutboundReasoningEffort(request.reasoning?.effort, options.row.efforts) as ThinkingLevel | undefined,
+ reasoning: reconcileOutboundReasoningEffort(request.reasoning?.effort, options.row.efforts ?? []) as ThinkingLevel | undefined,
maxTokens: typeof request.max_output_tokens === 'number' ? Math.min(request.max_output_tokens, model.maxTokens) : model.maxTokens,
});
const iterator = events[Symbol.asyncIterator]();
diff --git a/apps/desktop/src/main/maker-host/provider-model-fetch.ts b/apps/desktop/src/main/maker-host/provider-model-fetch.ts
index 0da3f42b66..9fc796693d 100644
--- a/apps/desktop/src/main/maker-host/provider-model-fetch.ts
+++ b/apps/desktop/src/main/maker-host/provider-model-fetch.ts
@@ -1,3 +1,4 @@
+import { collectModelDiscoveryPages } from './model-discovery-pages.js';
import type { DiscoveredModel } from '@cindy/model-providers';
/**
* provider-model-fetch —— 供应商「获取模型列表」(自定义供应商表单消费)。
@@ -23,7 +24,6 @@ import {
import { classifyProviderError, type ProviderErrorCode } from '../../shared/providerErrors.js';
import {
deriveModelsDiscoveryUrl,
- parseModelsListResponse,
readCachedGenericOAuthAccessToken,
refreshGenericOAuthIfNeeded,
} from './generic-oauth.js';
@@ -303,7 +303,13 @@ export async function fetchProviderModels(
detail: 'models response is not JSON or exceeds the response limit',
};
}
- const models = parseModelsListResponse(json, url);
+ const pageDeadline = Date.now() + 30_000;
+ const models = await collectModelDiscoveryPages(json, url, async nextUrl => {
+ if (Date.now() >= pageDeadline) throw new Error('catalog deadline exceeded');
+ const page = await fetchImpl(nextUrl, { ...init, redirect: 'error', signal: AbortSignal.timeout(Math.max(1, Math.min(FETCH_TIMEOUT_MS, pageDeadline - Date.now()))) });
+ if (!page.ok) { await page.body?.cancel(); throw new Error('catalog page failed'); }
+ return JSON.parse(await readLimitedBody(page, spec.responseByteLimit ?? 8 * 1024 * 1024));
+ });
if (!models || models.length === 0) {
// 端点 200 但响应不是可识别的模型列表(或为空)——按「模型不存在」类引导用户手填。
return {
diff --git a/apps/desktop/src/main/maker-ipc/__tests__/providerHandlers.test.ts b/apps/desktop/src/main/maker-ipc/__tests__/providerHandlers.test.ts
index 6339f7c52a..77323a0a89 100644
--- a/apps/desktop/src/main/maker-ipc/__tests__/providerHandlers.test.ts
+++ b/apps/desktop/src/main/maker-ipc/__tests__/providerHandlers.test.ts
@@ -1286,13 +1286,13 @@ describe('provider:custom:* CRUD handlers', () => {
const saved = await listCustomProviders();
expect(saved).toHaveLength(1);
expect(saved[0]?.runtimes.codex?.models).toEqual([
- { id: 'fetched-model', name: 'Fetched Model' },
+ { id: 'fetched-model', name: 'Fetched Model', discoveredMetadata: { name: 'Fetched Model' } },
]);
});
it.each([
{ ok: false, models: [] },
- { ok: true, models: Array.from({ length: 257 }, (_, i) => ({ id: `m${i}`, name: 'Model' })) },
+ { ok: true, models: Array.from({ length: 10001 }, (_, i) => ({ id: `m${i}`, name: 'Model' })) },
{ ok: true, models: [{ id: 'x'.repeat(257), name: 'Model' }] },
{ ok: true, models: [{ id: 'model', name: 'x'.repeat(257) }] },
])('saves only the connection and key when discovery fails or exceeds bounds ($models.length models)', async (fetched) => {
diff --git a/apps/desktop/src/main/maker-ipc/providerHandlers.ts b/apps/desktop/src/main/maker-ipc/providerHandlers.ts
index 4048602653..df65b53407 100644
--- a/apps/desktop/src/main/maker-ipc/providerHandlers.ts
+++ b/apps/desktop/src/main/maker-ipc/providerHandlers.ts
@@ -18,6 +18,7 @@ import { setProviderPresentation, retainProviderPresentationAfterAuthChange } fr
import type { CodexContextWindowInfo } from '@cindy/maker-core';
import {
PI_MODEL_APIS,
+ mergeDiscoveredRuntimeModels,
isByokProviderId,
isLoopbackProviderUrl,
isProviderRequestPath,
@@ -1298,7 +1299,7 @@ export function registerProviderHandlers(
});
if (fetched.ok && fetched.models?.length) {
assertProviderImportModels(fetched.models);
- config.runtimes[agent] = { ...runtime, models: fetched.models };
+ config.runtimes[agent] = { ...runtime, models: mergeDiscoveredRuntimeModels([], fetched.models) };
} else modelsPending = true;
} catch {
modelsPending = true;
diff --git a/apps/desktop/src/main/maker-ipc/register.ts b/apps/desktop/src/main/maker-ipc/register.ts
index 514071bce5..11013458b5 100644
--- a/apps/desktop/src/main/maker-ipc/register.ts
+++ b/apps/desktop/src/main/maker-ipc/register.ts
@@ -70,6 +70,8 @@ import type {
} from '@cindy/maker-core';
import {
effectiveSourceIdForModel,
+ buildUserProvider,
+ mergeDiscoveredRuntimeModels,
findCatalogModel,
storedCustomProviderId,
isLocalOnlyProviderForAgent,
@@ -5925,21 +5927,17 @@ export function registerMakerIpc(maker: Maker, options: RegisterMakerIpcOptions)
setDiscoveredProviderModels(
providerId,
agent,
- models.map((m) => ({
- id: m.id,
- name: m.name,
- // 端点上报的窗口值优先,缺省才落 200K 保守默认(review P1):
- // 之前无条件写死 200K,发现的 1M 模型仍会显示并按 200K 压缩。
- contextWindow: m.contextWindow ?? 200_000,
- // 只有端点真给了才算已核实,可以拿去收敛运行期上报窗口;落 200K
- // 兜底的不标记 —— 否则 resolveVerifiedContextWindow 会拒收缺失
- // 标记的条目,inflate 的运行期值压不下来(review P1)。
- ...(m.contextWindow !== undefined ? { contextWindowVerified: true } : {}),
- efforts: [],
- defaultEffort: null,
- group: `custom:${providerId}`,
- defaultEnabled: false,
- })),
+ buildUserProvider({
+ id: providerId,
+ name: provider.name,
+ runtimes: {
+ [agent]: {
+ baseUrl: provider.routing[agent]!.upstream,
+ wireProtocol: provider.routing[agent]!.wireProtocol,
+ models: mergeDiscoveredRuntimeModels([], models),
+ },
+ },
+ }, { modelRegistry: getActiveCatalog().modelRegistry }).models[agent] ?? [],
);
}
}
diff --git a/apps/desktop/src/main/model-access/__tests__/byokProvider.test.ts b/apps/desktop/src/main/model-access/__tests__/byokProvider.test.ts
index d4b652ffd7..16a7b0fe08 100644
--- a/apps/desktop/src/main/model-access/__tests__/byokProvider.test.ts
+++ b/apps/desktop/src/main/model-access/__tests__/byokProvider.test.ts
@@ -132,3 +132,51 @@ describe('buildByokProvider model types', () => {
expect(configs).toEqual([]);
});
});
+
+
+it.each([
+ ['openai-responses', ['codex', 'pi']],
+ ['anthropic-messages', ['claude-code', 'pi']],
+ ['openai-completions', ['pi']],
+] as const)('keeps managed model enablement separate from %s compatibility', (nativeApi, enabled) => {
+ const model = { ...chatModel, nativeApi, defaultEnabled: true,
+ maxOutputTokens: 32000, supportsFastMode: false,
+ agents: ['claude-code', 'codex', 'pi'] as Array<'claude-code' | 'codex' | 'pi'>,
+ perAgent: { 'claude-code': { wireProtocol: 'anthropic-messages' as const, defaultEnabled: true },
+ codex: { wireProtocol: 'openai-responses' as const, defaultEnabled: true }, pi: { wireProtocol: nativeApi, defaultEnabled: true } },
+ };
+ const provider = buildByokProvider({ provider: { id: 'byok-a', name: 'Enterprise',
+ connectionRevision: 1, models: [model] }, credential });
+ for (const agent of model.agents)
+ expect(provider.models[agent]?.[0]?.defaultEnabled).toBe((enabled as readonly string[]).includes(agent));
+ expect(byokNativeConfigs([provider])[0]?.runtimes.pi?.models[0]).toMatchObject({
+ nativeApi, maxOutputTokens: 32000, supportsFastMode: false,
+ });
+});
+
+
+it('keeps live text-only and Pi effort restrictions above inherited capabilities', () => {
+ const provider = buildByokProvider({ provider: { id: 'byok-a', name: 'Enterprise', connectionRevision: 1,
+ models: [{ ...chatModel, id: 'gpt-7-sol', nativeApi: 'openai-responses',
+ modalities: { input: ['text'], output: ['text'] }, efforts: ['high', 'ultra'], defaultEffort: 'ultra',
+ perAgent: { pi: { wireProtocol: 'openai-responses' } } }],
+ }, credential });
+ expect(provider.models.pi?.[0]).toMatchObject({ supportsImageInput: false, efforts: ['high'], defaultEffort: null });
+});
+
+it.each([undefined, 64000])('only verifies a future BYOK model window when explicitly declared: %s', (contextWindow) => {
+ const provider = buildByokProvider({ provider: { id: 'byok-a', name: 'Enterprise', connectionRevision: 1,
+ models: [{ ...chatModel, id: 'gpt-7-sol', contextWindow, nativeApi: 'openai-responses',
+ perAgent: { pi: { wireProtocol: 'openai-responses' } } }],
+ }, credential });
+ expect(provider.models.pi?.[0]?.contextWindow).toBeGreaterThan(0);
+ expect(provider.models.pi?.[0]?.contextWindowVerified).toBe(contextWindow !== undefined);
+});
+
+it.each(['byok-a/chat', 'gpt-6-sol'])('preserves an explicit per-engine null default above %s defaults', (id) => {
+ const provider = buildByokProvider({ provider: { id: 'byok-a', name: 'Enterprise', connectionRevision: 1,
+ models: [{ ...chatModel, id, efforts: ['low', 'high'], defaultEffort: 'high',
+ perAgent: { pi: { wireProtocol: 'openai-completions', defaultEffort: null } } }],
+ }, credential });
+ expect(provider.models.pi?.[0]).toMatchObject({ efforts: ['low', 'high'], defaultEffort: null });
+});
diff --git a/apps/desktop/src/main/model-access/byokProvider.ts b/apps/desktop/src/main/model-access/byokProvider.ts
index 5d0754862d..a0ee52ed4a 100644
--- a/apps/desktop/src/main/model-access/byokProvider.ts
+++ b/apps/desktop/src/main/model-access/byokProvider.ts
@@ -1,5 +1,12 @@
import {
isByokImageMode,
+ BUNDLED_CATALOG,
+ nativeModelAgents,
+ catalogModelMetadata,
+ pickModelMetadata,
+ resolveCatalogModelNativeApi,
+ resolveModelMetadata,
+ applyModelMetadata,
isOrganizationManagedProvider,
type AgentKind,
type CatalogModel,
@@ -46,11 +53,12 @@ export function buildByokProvider({ provider, credential }: ByokConnection): Pro
defaults.defaultEffort !== undefined ? defaults.defaultEffort : model.defaultEffort;
const price = (value: number | undefined) =>
value === undefined ? undefined : value * 1_000_000;
- return {
+ const projected: CatalogModel = {
id: model.id,
name: model.name ?? model.id,
mode: model.mode ?? 'chat',
- nativeApi: model.nativeApi === 'openai-images' ? undefined : model.nativeApi,
+ ...pickModelMetadata({ nativeApi: model.nativeApi !== undefined ? model.nativeApi
+ : resolveCatalogModelNativeApi(BUNDLED_CATALOG.modelRegistry, model.id) }),
description: model.description,
group: model.group,
icon: model.icon,
@@ -89,6 +97,23 @@ export function buildByokProvider({ provider, credential }: ByokConnection): Pro
cacheWrite: price(model.cacheCreationInputTokenCost),
},
};
+ const resolved = applyModelMetadata(projected, resolveModelMetadata(
+ BUNDLED_CATALOG.modelRegistry, provider.id, model.id,
+ { contextWindow: defaults.contextWindow ?? model.contextWindow,
+ maxOutputTokens: model.maxOutputTokens, efforts: defaults.efforts ?? model.efforts,
+ defaultEffort,
+ supportsFastMode: defaults.supportsFastMode ?? model.supportsFastMode,
+ modalities: model.modalities,
+ supportsImageInput: model.modalities ? model.modalities.input.includes('image') : undefined }, undefined, agent,
+ undefined, defaults.defaultEffort,
+ ));
+ if (agent === 'pi') {
+ resolved.efforts = resolved.efforts.filter(effort => effort !== 'ultra');
+ if (resolved.defaultEffort && !resolved.efforts.includes(resolved.defaultEffort)) resolved.defaultEffort = null;
+ }
+ resolved.defaultEnabled = (defaults.defaultEnabled ?? model.defaultEnabled ?? true) &&
+ nativeModelAgents({ id: provider.id, routing }, { [agent]: resolved }).includes(agent);
+ return resolved;
});
}
const imageModels = provider.models
@@ -130,7 +155,9 @@ export function byokNativeConfigs(providers: readonly Provider[]): CustomProvide
models: (provider.models.pi ?? []).map((model) => ({
id: model.id,
name: model.name,
- contextWindow: model.contextWindow,
+ ...catalogModelMetadata(model),
+ defaultEnabled: model.defaultEnabled,
+ ...(model.cost ? { discoveredCost: model.cost } : {}),
piApi: model.piApi,
...(model.route ? { route: model.route } : {}),
supportsImageInput: model.supportsImageInput,
diff --git a/apps/desktop/src/main/provider-import/__tests__/providerImport.test.ts b/apps/desktop/src/main/provider-import/__tests__/providerImport.test.ts
index 433afbb75f..8238079c56 100644
--- a/apps/desktop/src/main/provider-import/__tests__/providerImport.test.ts
+++ b/apps/desktop/src/main/provider-import/__tests__/providerImport.test.ts
@@ -3,6 +3,7 @@ import { afterEach, describe, expect, it, vi } from 'vitest';
import { buildUserProvider, type ProviderPreset, type ProviderView } from '@cindy/model-providers';
import {
+ assertProviderImportModels,
beginProviderImportConfirm,
cancelProviderImport,
clearProviderImportDraftsForTest,
@@ -11,6 +12,9 @@ import {
previewProviderImport,
} from '../providerImport.js';
+import { validateCustomProviderConfig } from '../../maker-host/custom-provider-store.js';
+import { applyRuntimeFillFields, type RuntimeFillDraft } from '../../../renderer/lib/customProviderRuntimeFill.js';
+
const SCOPE = { dataOwnerId: 'owner-a', generation: 1 };
function encodePayload(payload: unknown): string {
@@ -133,7 +137,7 @@ describe('provider import URL parsing', () => {
createDraft(customPayload({ endpoints: [{ protocol: 'openai-chat', baseUrl: 'https://api.acme.test/v1', modelsUrl: 'https://api.acme.test:443/catalog/models' }] }));
});
- it('keeps an explicit defaultEnabled override on imported compatibility routes', () => {
+ it('does not treat a model enabled in a link as compatibility-engine consent', () => {
const importId = createDraft(customPayload({
endpoints: [{
protocol: 'openai-chat',
@@ -146,8 +150,9 @@ describe('provider import URL parsing', () => {
const { draft } = beginProviderImportConfirm(importId, SCOPE, []);
expect(draft.kind === 'custom' && draft.config.runtimes['claude-code']?.models[0]).toMatchObject({
id: 'compat-model',
- defaultEnabled: true,
});
+ expect(draft.kind === 'custom' && draft.config.runtimes['claude-code']?.models[0]).not.toHaveProperty('defaultEnabled');
+
});
it('imports Google generateContent endpoints for all three engines', () => {
@@ -763,3 +768,114 @@ describe('provider import draft lifecycle', () => {
});
});
});
+
+
+it('imports complete metadata and explicit off into all three engines', () => {
+ const model = { id: 'gpt-7-sol', nativeApi: 'openai-responses', maxOutputTokens: 32000,
+ contextWindowMax: 1000000, supportsFastMode: false, supportsToolCalls: false,
+ supportsImageInput: false, reasoning: false, efforts: [], defaultEffort: null };
+ const importId = createDraft(customPayload({ endpoints: [{ protocol: 'openai-responses',
+ baseUrl: 'https://relay.example/v1', targets: ['claude-code', 'codex', 'pi'], models: [model],
+ }] }));
+ previewProviderImport(importId, SCOPE, []);
+ const { draft } = beginProviderImportConfirm(importId, SCOPE, []);
+ expect(draft.kind).toBe('custom');
+ if (draft.kind !== 'custom') throw new Error('wrong import kind');
+ const provider = buildUserProvider(draft.config);
+ for (const agent of ['claude-code', 'codex', 'pi'] as const) {
+ expect(draft.config.runtimes[agent]?.models[0]).toMatchObject(model);
+ expect(provider.models[agent]?.[0]).toMatchObject({ maxOutput: 32000,
+ supportsFastMode: false, supportsImageInput: false, supportsToolCalls: false,
+ efforts: [], defaultEnabled: agent !== 'claude-code' });
+ }
+});
+
+it('accepts a discovered catalog larger than the inline-link limit', () => {
+ expect(() => assertProviderImportModels(Array.from({ length: 1000 }, (_, index) =>
+ ({ id: `model-${index}`, name: `Model ${index}` })))).not.toThrow();
+});
+
+it.each([
+ [['minimal', 'low', 'high'], 'minimal', ['low', 'high'], undefined],
+ [['minimal', 'low', 'high'], 'high', ['low', 'high'], 'high'],
+ [['minimal'], 'minimal', [], undefined],
+] as const)('imports per-engine reasoning settings through the actual save validator: %j',
+ (efforts, defaultEffort, fixedEfforts, fixedDefault) => {
+ const importId = createDraft(customPayload({ endpoints: [{ protocol: 'openai-responses',
+ baseUrl: 'https://relay.example/v1', targets: ['claude-code', 'codex', 'pi'],
+ models: [{ id: 'new-model', reasoning: true, reasoningEfforts: efforts, reasoningDefaultEffort: defaultEffort }],
+ }] }));
+ previewProviderImport(importId, SCOPE, []);
+ const { draft } = beginProviderImportConfirm(importId, SCOPE, []);
+ if (draft.kind !== 'custom') throw new Error('wrong import kind');
+ expect(validateCustomProviderConfig(draft.config)).toEqual({ ok: true });
+ for (const agent of ['claude-code', 'codex', 'pi'] as const) {
+ const model = draft.config.runtimes[agent]!.models[0];
+ expect(model.reasoning).toBe(true);
+ expect(model.reasoningEfforts).toEqual(agent === 'pi' ? efforts : fixedEfforts);
+ expect(model.reasoningDefaultEffort).toBe(agent === 'pi' ? defaultEffort : fixedDefault);
+ }
+ });
+
+
+it.each([
+ { efforts: ['minimal', 'low', 'ultra', 'max'], defaultEffort: 'ultra', fixed: ['low', 'max'], fixedDefault: undefined },
+ { efforts: ['minimal', 'high'], defaultEffort: 'high', fixed: ['high'], fixedDefault: 'high' },
+ { efforts: [], defaultEffort: null, fixed: [], fixedDefault: null },
+ { defaultEffort: 'minimal', fixedDefault: undefined },
+])('filters canonical imported efforts at both metadata layers: %j', ({ fixed, fixedDefault, ...metadata }) => {
+ const importId = createDraft(customPayload({ endpoints: [{ protocol: 'openai-responses',
+ baseUrl: 'https://relay.example/v1', targets: ['claude-code', 'codex', 'pi'],
+ models: [{ id: 'new-model', ...metadata, discoveredMetadata: metadata }],
+ }] }));
+ previewProviderImport(importId, SCOPE, []);
+ const { draft } = beginProviderImportConfirm(importId, SCOPE, []);
+ if (draft.kind !== 'custom') throw new Error('wrong import kind');
+ expect(validateCustomProviderConfig(draft.config)).toEqual({ ok: true });
+ for (const agent of ['claude-code', 'codex', 'pi'] as const) {
+ const model = draft.config.runtimes[agent]!.models[0];
+ for (const layer of [model, model.discoveredMetadata!]) {
+ expect(layer.efforts).toEqual(agent === 'pi' ? metadata.efforts : fixed);
+ expect(layer.defaultEffort).toBe(agent === 'pi' ? metadata.defaultEffort : fixedDefault);
+ }
+ if (agent !== 'pi') {
+ const projected = buildUserProvider(draft.config).models[agent]![0];
+ expect(projected.efforts).not.toContain('minimal');
+ expect(projected.efforts).not.toContain('ultra');
+ }
+ }
+});
+
+
+it.each(['claude-code', 'codex'] as const)('saves Pi models filled into %s with target-compatible efforts', agent => {
+ const source: RuntimeFillDraft = { baseUrl: 'https://relay.example/v1', requestPath: '', apiKey: '',
+ wireProtocol: 'openai-responses', headers: [], modelsUrl: '', models: [
+ { id: 'future', name: 'Future', reasoning: true, reasoningEfforts: ['minimal', 'high'], reasoningDefaultEffort: 'minimal',
+ efforts: ['minimal', 'high', 'ultra'], defaultEffort: 'ultra',
+ discoveredMetadata: { efforts: ['minimal', 'high', 'ultra'], defaultEffort: 'ultra' } },
+ { id: 'empty', name: 'Empty', reasoning: true, reasoningEfforts: ['minimal'], reasoningDefaultEffort: 'minimal',
+ efforts: [], defaultEffort: null, discoveredMetadata: { efforts: [], defaultEffort: null } },
+ { id: 'sparse', name: 'Sparse', reasoning: true, reasoningEfforts: ['low'], efforts: ['low'] },
+ ] };
+ const target: RuntimeFillDraft = { ...source, models: [{ id: 'sparse', name: 'Old', reasoning: true,
+ reasoningEfforts: ['high'], reasoningDefaultEffort: 'high', efforts: ['high'], defaultEffort: 'high', defaultEnabled: true }] };
+ const before = structuredClone({ source, target });
+ const filled = applyRuntimeFillFields(target, source, ['models'], { sourceAgent: 'pi', targetAgent: agent });
+ const config = { id: 'filled-provider', name: 'Filled', runtimes: { [agent]: {
+ baseUrl: filled.baseUrl, wireProtocol: filled.wireProtocol, models: filled.models,
+ } } };
+ expect(validateCustomProviderConfig(config)).toEqual({ ok: true });
+ expect(filled.models.map(model => model.id)).toEqual(['future', 'empty', 'sparse']);
+ expect(filled.models[0]).toMatchObject({ reasoningEfforts: ['high'], efforts: ['high'], discoveredMetadata: { efforts: ['high'] } });
+ expect(filled.models[0].reasoningDefaultEffort).toBeUndefined();
+ expect(filled.models[0].defaultEffort).toBeUndefined();
+ expect(filled.models[0].discoveredMetadata!.defaultEffort).toBeUndefined();
+ expect(filled.models[1]).toMatchObject({ reasoningEfforts: [], efforts: [], defaultEffort: null,
+ discoveredMetadata: { efforts: [], defaultEffort: null } });
+ expect(filled.models[2]).toMatchObject({ reasoningEfforts: ['low'], efforts: ['low'], defaultEnabled: true });
+ expect(filled.models[2].reasoningDefaultEffort).toBeUndefined();
+ expect(filled.models[2].defaultEffort).toBeUndefined();
+ expect(buildUserProvider(config).models[agent]?.some(model => model.efforts.some(effort => effort === 'minimal' || effort === 'ultra'))).toBe(false);
+ expect({ source, target }).toEqual(before);
+ expect(applyRuntimeFillFields({ ...target, models: [] }, source, ['models'], { sourceAgent: 'pi', targetAgent: 'pi' }).models).toEqual(source.models);
+});
diff --git a/apps/desktop/src/main/provider-import/providerImport.ts b/apps/desktop/src/main/provider-import/providerImport.ts
index 2b2b310ddb..a3536682b4 100644
--- a/apps/desktop/src/main/provider-import/providerImport.ts
+++ b/apps/desktop/src/main/provider-import/providerImport.ts
@@ -8,7 +8,7 @@
import { randomUUID } from 'node:crypto';
import { validateHeaderName, validateHeaderValue } from 'node:http';
-import { isLoopbackProviderUrl } from '@cindy/model-providers';
+import { isLoopbackProviderUrl, MODEL_METADATA_FIELDS, pickModelMetadata, validModelMetadata } from '@cindy/model-providers';
import type {
AgentKind,
@@ -29,7 +29,7 @@ import {
MANAGED_LMSTUDIO_PROVIDER_ID,
MANAGED_OLLAMA_PROVIDER_ID,
} from '../../shared/localModelRuntime.js';
-import { configuredPresetAgents } from '../../shared/piRuntimeInitialization.js';
+import { adaptCustomProviderModelEfforts, configuredPresetAgents } from '../../shared/piRuntimeInitialization.js';
import { presetConnectionRuntime } from '../../shared/presetConnectionRuntime.js';
const IMPORT_TTL_MS = 10 * 60_000;
@@ -233,17 +233,26 @@ function parseModel(value: unknown, label: string): ProviderRuntimeModelConfig {
[
'id',
'name',
- 'contextWindow',
+ ...MODEL_METADATA_FIELDS,
'defaultEnabled',
'supportsImageInput',
'reasoning',
'reasoningEfforts',
+ 'reasoningDefaultEffort',
+ 'discoveredMetadata',
],
label,
);
const id = boundedString(model.id, `${label}.id`, 256);
const name = model.name === undefined ? id : boundedString(model.name, `${label}.name`, 256);
- const result: ProviderRuntimeModelConfig = { id, name };
+ for (const field of MODEL_METADATA_FIELDS) {
+ if (model[field] !== undefined && !validModelMetadata({ [field]: model[field] }))
+ fail(`${label}.${field} is invalid`);
+ }
+ if (model.discoveredMetadata !== undefined && !validModelMetadata(model.discoveredMetadata))
+ fail(`${label}.discoveredMetadata is invalid`);
+ const result: ProviderRuntimeModelConfig = { ...pickModelMetadata(model), id, name,
+ ...(model.discoveredMetadata ? { discoveredMetadata: pickModelMetadata(model.discoveredMetadata as object) } : {}) };
if (model.contextWindow !== undefined) {
if (!Number.isInteger(model.contextWindow) || (model.contextWindow as number) <= 0) {
fail(`${label}.contextWindow must be a positive integer`);
@@ -255,23 +264,29 @@ function parseModel(value: unknown, label: string): ProviderRuntimeModelConfig {
fail(`${label}.${field} must be a boolean`);
}
}
- if (typeof model.defaultEnabled === 'boolean') result.defaultEnabled = model.defaultEnabled;
- if (model.supportsImageInput === true) result.supportsImageInput = true;
+ // A link selects a model, not compatibility engines. Explicit off remains off.
+ if (model.defaultEnabled === false) result.defaultEnabled = false;
+ if (typeof model.supportsImageInput === 'boolean') result.supportsImageInput = model.supportsImageInput;
+ if (model.reasoning === false) result.reasoning = false;
if (model.reasoning === true) {
- if (!Array.isArray(model.reasoningEfforts) || model.reasoningEfforts.length === 0) {
+ if (model.reasoningEfforts !== undefined && !Array.isArray(model.reasoningEfforts)) {
fail(`${label}.reasoningEfforts is required when reasoning is enabled`);
}
- const efforts = model.reasoningEfforts.map((effort) => {
+ const efforts = model.reasoningEfforts?.map((effort) => {
if (typeof effort !== 'string' || !PI_EFFORTS.includes(effort as PiReasoningEffort)) {
fail(`${label}.reasoningEfforts contains an unsupported value`);
}
return effort as PiReasoningEffort;
});
- if (new Set(efforts).size !== efforts.length)
+ if (efforts && new Set(efforts).size !== efforts.length)
fail(`${label}.reasoningEfforts contains duplicates`);
result.reasoning = true;
- result.reasoningEfforts = efforts;
- } else if (model.reasoningEfforts !== undefined) {
+ if (efforts) result.reasoningEfforts = efforts;
+ if (model.reasoningDefaultEffort !== undefined) {
+ if (!efforts?.includes(model.reasoningDefaultEffort as PiReasoningEffort)) fail(`${label}.reasoningDefaultEffort is unsupported`);
+ result.reasoningDefaultEffort = model.reasoningDefaultEffort as PiReasoningEffort;
+ }
+ } else if (model.reasoningEfforts !== undefined || model.reasoningDefaultEffort !== undefined) {
fail(`${label}.reasoningEfforts requires reasoning=true`);
}
return result;
@@ -288,9 +303,9 @@ type ParsedEndpoint = {
apiKey?: string;
};
-/** Discovery is untrusted input too: apply the same count and core model field bounds. */
+/** Discovery uses catalog-sized limits, independently of the compact deep-link payload. */
export function assertProviderImportModels(models: readonly { id: string; name: string; contextWindow?: number }[]): void {
- if (models.length > MAX_MODELS_PER_ENDPOINT) fail('models must be a bounded array');
+ if (models.length > 10_000) fail('discovered models exceed the catalog limit');
for (const model of models) {
parseModel({ id: model.id, name: model.name, ...(model.contextWindow !== undefined ? { contextWindow: model.contextWindow } : {}) }, 'models');
}
@@ -618,11 +633,7 @@ function parsePayload(value: unknown): ProviderImportDraft {
) {
fail(`multiple endpoints compete for ${agent}; set explicit targets`);
}
- const models = selected.models.map((model) => {
- if (agent === 'pi') return { ...model };
- const { reasoning: _reasoning, reasoningEfforts: _efforts, ...portable } = model;
- return portable;
- });
+ const models = selected.models.map(source => adaptCustomProviderModelEfforts(source, agent));
runtimes[agent] = {
wireProtocol: selected.protocol,
baseUrl: selected.baseUrl,
diff --git a/apps/desktop/src/renderer/__tests__/addProviderWizardPresetEntry.test.tsx b/apps/desktop/src/renderer/__tests__/addProviderWizardPresetEntry.test.tsx
index 8b6b349fa2..8a29da0b85 100644
--- a/apps/desktop/src/renderer/__tests__/addProviderWizardPresetEntry.test.tsx
+++ b/apps/desktop/src/renderer/__tests__/addProviderWizardPresetEntry.test.tsx
@@ -543,7 +543,7 @@ describe('AddProviderWizard — preset 直达', () => {
catalogPresetId: 'explicit-pi',
baseUrl: 'https://explicit.example/pi',
wireProtocol: 'openai-chat',
- models: [{ id: 'pi-model', name: 'Pi Model', discoveredMetadata: {}, defaultEnabled: true }],
+ models: [{ id: 'pi-model', name: 'Pi Model', discoveredMetadata: {} }],
});
expect(keys.pi).toBe('sk-test');
});
@@ -563,7 +563,7 @@ describe('AddProviderWizard — preset 直达', () => {
expect(config.runtimes['claude-code']).toEqual({
catalogPresetId: 'explicit-pi',
baseUrl: 'https://explicit.example/anthropic',
- models: [{ id: 'claude-model', name: 'Claude Model', discoveredMetadata: {}, defaultEnabled: true }],
+ models: [{ id: 'claude-model', name: 'Claude Model', discoveredMetadata: {} }],
});
expect(config.runtimes.pi?.models).toEqual([{ id: 'pi-model', name: 'Pi Model', discoveredMetadata: {}, defaultEnabled: false }]);
expect(keys.pi).toBe('sk-test');
@@ -599,7 +599,7 @@ describe('AddProviderWizard — preset 直达', () => {
await waitFor(() => expect(createCustomProvider).toHaveBeenCalledTimes(1));
const [config, keys] = vi.mocked(createCustomProvider).mock.calls[0];
expect(config.runtimes['claude-code']?.models).toEqual([
- { id: 'claude-only-model', name: 'Claude Only Model', discoveredMetadata: {}, defaultEnabled: true },
+ { id: 'claude-only-model', name: 'Claude Only Model', discoveredMetadata: {} },
]);
expect(config.runtimes.pi).toBeUndefined();
expect(keys.pi).toBeUndefined();
@@ -741,7 +741,7 @@ describe('AddProviderWizard — preset 直达', () => {
codex: expect.objectContaining({
baseUrl: 'http://localhost:4100/v1',
requestPath: '/tenant/acme/infer',
- models: [{ id: 'local-model', name: 'local-model', discoveredMetadata: {}, defaultEnabled: true }],
+ models: [{ id: 'local-model', name: 'local-model', discoveredMetadata: {} }],
}),
},
}),
@@ -931,7 +931,7 @@ describe('AddProviderWizard — preset 直达', () => {
baseUrl: 'https://open.bigmodel.cn/api/coding/paas/v4',
wireProtocol: 'openai-chat',
});
- expect(runtime?.models.filter(m => m.defaultEnabled !== false)).toEqual([{ id: 'glm-5.2', name: 'GLM-5.2', discoveredMetadata: {}, defaultEnabled: true }]);
+ expect(runtime?.models.filter(m => m.defaultEnabled !== false)).toEqual([{ id: 'glm-5.2', name: 'GLM-5.2', discoveredMetadata: {} }]);
});
it('可编辑预设改为同源 endpoint 后继续合并 Responses 目录', async () => {
@@ -969,12 +969,11 @@ describe('AddProviderWizard — preset 直达', () => {
}),
);
expect(vi.mocked(createCustomProvider).mock.calls[0][0].runtimes.codex?.models).toEqual([
- { id: 'chat-model', name: 'Chat Model', discoveredMetadata: {}, defaultEnabled: true },
+ { id: 'chat-model', name: 'Chat Model', discoveredMetadata: {} },
{
discoveredMetadata: { name: 'Responses Model' },
id: 'responses-model',
name: 'Responses Model',
- defaultEnabled: true,
route: {
baseUrl: 'https://editable.example/api/v1',
wireProtocol: 'openai-responses',
@@ -1145,10 +1144,49 @@ it('imports the same discovered OpenRouter identity, capabilities and prices int
cost: { input: 0.75864, output: 1.51728 } });
expect(modelProtocolComparison(provider, { [agent]: model }).forAgent(agent)?.mode)
.toBe(agent === 'pi' ? 'matching' : 'compatibility');
- expect(model.defaultEnabled).toBe(true);
+ expect(model.defaultEnabled).toBe(agent === 'pi');
}
});
+describe.each([true, false])('native defaults with discovery available=%s', discoveryAvailable => {
+ it.each([
+ ['openai/gpt-6-astra', 'openai-responses', [false, true, true]],
+ ['anthropic/claude-fable-5', 'anthropic-messages', [true, false, true]],
+ ['deepseek/deepseek-v4-pro', 'openai-completions', [false, false, true]],
+ ['google/gemini-3.8-flash', 'google-generative-ai', [false, false, true]],
+ ] as const)('enables only matching engines when selecting %s', async (id, nativeApi, enabled) => {
+ const agents = ['claude-code', 'codex', 'pi'] as const;
+ const preset = structuredClone(BUNDLED_CATALOG.presets!.find(p => p.id === 'openrouter')!);
+ for (const agent of agents) {
+ const model = preset.runtimes[agent]!.models.find(m => m.id === id)!;
+ expect(model).toBeDefined();
+ preset.runtimes[agent]!.models = [{ ...model, name: 'Selected model', defaultEnabled: false }];
+ }
+ vi.mocked(window.electronAPI.maker.listProviderPresets).mockResolvedValue({ presets: [preset] });
+ vi.mocked(window.electronAPI.maker.fetchProviderModels).mockResolvedValue(discoveryAvailable
+ ? { ok: true, models: [{ id, name: 'Selected model' }] }
+ : { ok: false });
+ renderWizard('openrouter');
+ fireEvent.click(await screen.findByText('settings.providers.wizard.useApiKey'));
+ fireEvent.change(screen.getByPlaceholderText('sk-…'), { target: { value: 'sk-test' } });
+ fireEvent.click(screen.getByText('settings.providers.wizard.next'));
+ await waitFor(() => expect(screen.queryByText('settings.providers.wizard.fetching')).toBeNull());
+ fireEvent.click(await screen.findByText('Selected model'));
+ fireEvent.click(screen.getByText('settings.providers.wizard.finish'));
+ await waitFor(() => expect(createCustomProvider).toHaveBeenCalledOnce());
+ const saved = vi.mocked(createCustomProvider).mock.calls[0][0];
+ const provider = buildUserProvider(saved, { presets: [preset], modelRegistry: BUNDLED_CATALOG.modelRegistry });
+ agents.forEach((agent, index) => {
+ // Model selection must not be persisted as a per-engine compatibility override.
+ expect(saved.runtimes[agent]!.models.find(m => m.id === id)).not.toHaveProperty('defaultEnabled');
+ const model = provider.models[agent]!.find(m => m.id === id)!;
+ expect(model).toMatchObject({ nativeApi, defaultEnabled: enabled[index] });
+ expect(modelProtocolComparison(provider, { [agent]: model }).forAgent(agent)?.mode)
+ .toBe(enabled[index] ? 'matching' : 'compatibility');
+ });
+ });
+});
+
it('imports an OpenCode Go Responses model into all three engines', async () => {
const preset = structuredClone(BUNDLED_CATALOG.presets!.find(p => p.id === 'opencode-go')!);
for (const runtime of Object.values(preset.runtimes)) {
@@ -1173,7 +1211,7 @@ it('imports an OpenCode Go Responses model into all three engines', async () =>
expect(model).toMatchObject({ piApi: 'openai-responses', supportsImageInput: true });
const projected = buildUserProvider(saved, { presets: [preset] });
for (const agent of projected.agents) {
- expect(projected.models[agent]?.find(m => m.id === 'gpt-5.6-luna')).toMatchObject({ api: 'openai-responses', defaultEnabled: true });
+ expect(projected.models[agent]?.find(m => m.id === 'gpt-5.6-luna')).toMatchObject({ api: 'openai-responses', defaultEnabled: agent !== 'claude-code' });
expect(saved.runtimes[agent]?.models.find(m => m.id === 'gpt-5.6-luna')).not.toHaveProperty('route');
}
});
@@ -1296,6 +1334,7 @@ it('preserves OAuth-discovered prices for every engine when finishing model sele
for (const agent of ['claude-code', 'codex', 'pi'] as const) {
expect(saved.runtimes[agent]!.models[0].discoveredCost).toEqual(prices[agent]);
expect(projected.models[agent]![0].cost).toEqual(prices[agent]);
+ expect(projected.models[agent]![0].defaultEnabled).toBe(agent === 'pi');
}
});
@@ -1319,7 +1358,7 @@ it('imports the full Hermes inventory immediately, enables selected Pi models an
expect(provider.agents).toEqual(['claude-code', 'codex', 'pi']);
for (const agent of provider.agents) {
expect(provider.models[agent]?.map(m => m.id)).toEqual(['google/gemini-test', 'other']);
- expect(provider.models[agent]?.find(m => m.id === 'google/gemini-test')?.defaultEnabled).toBe(true);
+ expect(provider.models[agent]?.find(m => m.id === 'google/gemini-test')?.defaultEnabled).toBe(agent === 'pi');
expect(provider.models[agent]?.find(m => m.id === 'other')?.defaultEnabled).toBe(false);
}
});
diff --git a/apps/desktop/src/renderer/__tests__/providersSectionUnavailableModels.test.tsx b/apps/desktop/src/renderer/__tests__/providersSectionUnavailableModels.test.tsx
index 13ab64a536..460507a07f 100644
--- a/apps/desktop/src/renderer/__tests__/providersSectionUnavailableModels.test.tsx
+++ b/apps/desktop/src/renderer/__tests__/providersSectionUnavailableModels.test.tsx
@@ -200,7 +200,7 @@ vi.mock('@/components/settings/AddProviderWizard', () => ({
import { setModelVisibilities } from '@/state/modelVisibilityPrefs';
import { readCustomProviderKey, updateCustomProvider } from '@/lib/customProviders';
-import { providerPresetOAuth } from '@cindy/model-providers';
+import { buildUserProvider, providerPresetOAuth } from '@cindy/model-providers';
import { ProvidersSection } from '@/components/settings/ProvidersSection';
@@ -267,7 +267,9 @@ describe('ProvidersSection — 双栏管理', () => {
};
providerSnapshotState.order = [id, 'xd'];
const fetchModels = vi.fn(async () => ({ ok: true, models: [
- { id: 'old-model', name: 'Old' }, { id: 'new-model', name: 'New' },
+ { id: 'old-model', name: 'Old' },
+ { id: 'new-model', name: 'New', discoveredMetadata: { nativeApi: 'openai-responses' } },
+ { id: 'chat-model', name: 'Chat', discoveredMetadata: { nativeApi: 'openai-completions' } },
] }));
if (id === 'openrouter-retry') fetchModels.mockResolvedValueOnce({ ok: false, models: [] });
Object.assign(window.electronAPI.maker, { fetchProviderModels: fetchModels });
@@ -292,10 +294,22 @@ describe('ProvidersSection — 双栏管理', () => {
expect(updateCustomProvider).toHaveBeenCalledWith(expect.objectContaining({ id,
...(method === 'oauth' ? { auth: { method, oauth } } : {}),
runtimes: expect.objectContaining({ codex: expect.objectContaining({ models: expect.arrayContaining([
- expect.objectContaining({ id: 'new-model', defaultEnabled: false }),
+ expect.objectContaining({ id: 'new-model', discoveredMetadata: { nativeApi: 'openai-responses' } }),
...(hasModels ? [expect.objectContaining({ id: 'old-model', contextWindow: 64000, defaultEnabled: false })] : []),
]) }) }),
}), {});
+ const saved = vi.mocked(updateCustomProvider).mock.calls[0]![0];
+ // Refresh adds membership without freezing a visibility choice. Native
+ // projection decides defaults; compatibility still requires an override.
+ for (const modelId of ['new-model', 'chat-model']) {
+ const model = saved.runtimes.codex?.models.find(model => model.id === modelId);
+ expect(model).toBeDefined();
+ expect(model).not.toHaveProperty('defaultEnabled');
+ }
+ const projected = buildUserProvider(saved).models.codex!;
+ expect(projected.find(model => model.id === 'new-model')?.defaultEnabled).toBe(true);
+ expect(projected.find(model => model.id === 'chat-model')?.defaultEnabled).toBe(false);
+ if (hasModels) expect(projected.find(model => model.id === 'old-model')?.defaultEnabled).toBe(false);
expect(refetchProvidersSpy).toHaveBeenCalled();
});
it('dims GPT Image 2 without a ready image channel, independently of chat connection', async () => {
diff --git a/apps/desktop/src/renderer/components/settings/AddProviderWizard.tsx b/apps/desktop/src/renderer/components/settings/AddProviderWizard.tsx
index 5af9daa1ec..ab99901588 100644
--- a/apps/desktop/src/renderer/components/settings/AddProviderWizard.tsx
+++ b/apps/desktop/src/renderer/components/settings/AddProviderWizard.tsx
@@ -1323,7 +1323,9 @@ export function AddProviderWizard({
return {
id: m.id,
name: m.name,
- defaultEnabled: m.checked,
+ // Selecting a model follows native-engine defaults; it is not an
+ // explicit opt-in to every compatibility engine carrying the model.
+ ...(!m.checked ? { defaultEnabled: false } : {}),
discoveredMetadata,
...(m.discoveredCosts?.[agent] ? { discoveredCost: m.discoveredCosts[agent] } : {}),
...(presetModel?.mode ? { mode: presetModel.mode } : {}),
diff --git a/apps/desktop/src/renderer/lib/__tests__/customProviderRuntimeFill.test.ts b/apps/desktop/src/renderer/lib/__tests__/customProviderRuntimeFill.test.ts
index 2b18cea0c6..cb673ce46a 100644
--- a/apps/desktop/src/renderer/lib/__tests__/customProviderRuntimeFill.test.ts
+++ b/apps/desktop/src/renderer/lib/__tests__/customProviderRuntimeFill.test.ts
@@ -612,7 +612,7 @@ describe('custom provider runtime fill', () => {
expect(result.models[0]?.route).not.toBe(source.models[0]?.route);
});
- it('ignores Pi-only capabilities when comparing or filling a non-Pi target', () => {
+ it('preserves shared capabilities when filling a non-Pi target', () => {
const source = draft({
models: [
{
@@ -631,13 +631,13 @@ describe('custom provider runtime fill', () => {
targetAgent: 'codex',
}).find((diff) => diff.field === 'models');
- expect(modelDiff?.targetState).toBe('same');
+ expect(modelDiff?.targetState).toBe('conflict');
expect(
applyRuntimeFillFields(target, source, ['models'], {
sourceAgent: 'pi',
targetAgent: 'codex',
}).models,
- ).toEqual([{ id: 'model-a', name: 'Model A' }]);
+ ).toEqual(source.models);
});
it('uses the same model and header counting semantics as save', () => {
@@ -759,3 +759,17 @@ describe('custom provider runtime fill', () => {
expect(runtimeFillSelectedTargetChanged(previous, fresh, ['apiKey'], 'codex')).toBe(false);
});
});
+
+
+it('copies complete metadata without silently opting another engine into compatibility', () => {
+ const model = { id: 'gpt-7-sol', name: 'Seven', api: 'openai-responses' as const,
+ nativeApi: 'openai-responses' as const, defaultEnabled: true, supportsImageInput: false,
+ supportsFastMode: false, reasoning: false, maxOutputTokens: 32000,
+ discoveredMetadata: { supportsToolCalls: true }, discoveredCost: { input: 1, output: 2 } };
+ const source = draft({ models: [model] });
+ const result = applyRuntimeFillFields(draft(), source, ['models'], { sourceAgent: 'pi', targetAgent: 'claude-code' });
+ const { defaultEnabled: _enabled, ...metadata } = model;
+ expect(result.models).toEqual([metadata]);
+ const existing = draft({ models: [{ id: model.id, name: 'Old', defaultEnabled: true }] });
+ expect(applyRuntimeFillFields(existing, source, ['models'], { sourceAgent: 'pi', targetAgent: 'claude-code' }).models[0]?.defaultEnabled).toBe(true);
+});
diff --git a/apps/desktop/src/renderer/lib/__tests__/customProviders.test.ts b/apps/desktop/src/renderer/lib/__tests__/customProviders.test.ts
index bd81c966ed..9b8cd00496 100644
--- a/apps/desktop/src/renderer/lib/__tests__/customProviders.test.ts
+++ b/apps/desktop/src/renderer/lib/__tests__/customProviders.test.ts
@@ -314,7 +314,7 @@ describe('piCatalogProviderIdAfterRouteEdit', () => {
});
});
- it('reconstructs explicit Pi reasoning capability from catalog efforts only for Pi', () => {
+ it('preserves reasoning metadata when reconstructing every engine', () => {
const catalogModel = {
id: 'reasoner',
name: 'Reasoner',
@@ -325,6 +325,7 @@ describe('piCatalogProviderIdAfterRouteEdit', () => {
expect(customProviderModelConfigFromCatalogModel(catalogModel, 'pi')).toEqual({
id: 'reasoner',
name: 'Reasoner',
+ efforts: ['low', 'high', 'xhigh'], defaultEffort: 'xhigh',
reasoning: true,
reasoningEfforts: ['low', 'high', 'xhigh'],
reasoningDefaultEffort: 'xhigh',
@@ -332,6 +333,7 @@ describe('piCatalogProviderIdAfterRouteEdit', () => {
expect(customProviderModelConfigFromCatalogModel(catalogModel, 'codex')).toEqual({
id: 'reasoner',
name: 'Reasoner',
+ efforts: ['low', 'high', 'xhigh'], defaultEffort: 'xhigh',
});
});
});
@@ -456,7 +458,7 @@ describe('providerViewToCustomProviderConfig', () => {
requestPath: '/tenant/acme/infer?stream=1',
wireProtocol: 'openai-chat',
modelsUrl: 'http://127.0.0.1:4000/v1/models',
- models: [{ id: 'local-model', name: 'Local Model' }],
+ models: [{ id: 'local-model', name: 'Local Model', efforts: [], defaultEffort: null }],
},
},
});
@@ -501,6 +503,7 @@ describe('providerViewToCustomProviderConfig', () => {
{
id: 'glm-5.3',
name: 'GLM-5.3',
+ efforts: [], defaultEffort: null,
route: {
baseUrl: 'https://open.bigmodel.cn/api/v1',
wireProtocol: 'openai-responses',
@@ -580,6 +583,7 @@ describe('providerViewToCustomProviderConfig', () => {
{
id: 'reasoner',
name: 'Reasoner',
+ efforts: ['low', 'high', 'xhigh'], defaultEffort: 'high',
reasoning: true,
reasoningEfforts: ['low', 'high', 'xhigh'],
reasoningDefaultEffort: 'high',
@@ -640,7 +644,7 @@ describe('appendDiscoveredCustomProviderModels', () => {
.toContainEqual(expect.objectContaining({ id: 'anthropic/openai/new[1m]' }));
expect(existing[0].id).toBe('anthropic/openai/new[1m]');
});
- it('only appends unknown models and defaults them to hidden', () => {
+ it('appends new models without overriding native-engine defaults', () => {
const result = appendDiscoveredCustomProviderModels(
[{ id: 'kept', name: 'Kept' }],
[
@@ -653,7 +657,7 @@ describe('appendDiscoveredCustomProviderModels', () => {
expect(result).toEqual({
models: [
{ id: 'kept', name: 'Kept', nameExplicit: true, discoveredMetadata: { name: 'New name' } },
- { id: 'new', name: 'New', defaultEnabled: false, discoveredMetadata: { name: 'New' } },
+ { id: 'new', name: 'New', discoveredMetadata: { name: 'New' } },
],
addedIds: ['new'],
});
@@ -673,11 +677,10 @@ describe('appendDiscoveredCustomProviderModels', () => {
id: 'big',
name: 'Big',
discoveredMetadata: { name: 'Big', contextWindow: 1_000_000 },
- defaultEnabled: false,
},
- { id: 'plain', name: 'Plain', discoveredMetadata: { name: 'Plain' }, defaultEnabled: false },
+ { id: 'plain', name: 'Plain', discoveredMetadata: { name: 'Plain' } },
// 非法值不落盘,回落保守默认
- { id: 'bogus', name: 'Bogus', discoveredMetadata: { name: 'Bogus' }, defaultEnabled: false },
+ { id: 'bogus', name: 'Bogus', discoveredMetadata: { name: 'Bogus' } },
]);
});
});
diff --git a/apps/desktop/src/renderer/lib/customProviderRuntimeFill.ts b/apps/desktop/src/renderer/lib/customProviderRuntimeFill.ts
index d896db98b1..37baac75a9 100644
--- a/apps/desktop/src/renderer/lib/customProviderRuntimeFill.ts
+++ b/apps/desktop/src/renderer/lib/customProviderRuntimeFill.ts
@@ -3,7 +3,7 @@ import type {
ProviderRuntimeModelConfig,
ProviderWireProtocol,
} from '@cindy/model-providers';
-import { savedCustomProviderModelShape } from '@/../shared/piRuntimeInitialization';
+import { adaptCustomProviderModelEfforts, savedCustomProviderModelShape } from '@/../shared/piRuntimeInitialization';
export type RuntimeFillAgent = Extract<AgentKind, 'claude-code' | 'codex' | 'pi'>;
export interface RuntimeFillHeaderRow {
@@ -154,24 +154,19 @@ function modelsForTarget(
) {
const targetById = new Map(validModels(targetModels).map((model) => [model.id, model]));
return validModels(sourceModels).map((sourceModel) => {
- if (targetAgent !== 'pi') return savedCustomProviderModelShape(sourceModel, false);
- if (sourceAgent === 'pi') return savedCustomProviderModelShape(sourceModel, true);
-
- const portable = savedCustomProviderModelShape(sourceModel, false);
+ const portable = savedCustomProviderModelShape(sourceModel, targetAgent === 'pi');
const existing = targetById.get(portable.id);
- return {
- ...portable,
- ...(existing?.supportsImageInput === true ? { supportsImageInput: true } : {}),
- ...(existing?.reasoning === true && existing.reasoningEfforts?.length
- ? {
- reasoning: true,
- reasoningEfforts: [...existing.reasoningEfforts],
- ...(existing.reasoningDefaultEffort
- ? { reasoningDefaultEffort: existing.reasoningDefaultEffort }
- : {}),
- }
- : {}),
- };
+ if (sourceAgent === targetAgent) return adaptCustomProviderModelEfforts(portable, targetAgent);
+ // Copy all portable metadata, but an engine's opt-in is not another engine's opt-in.
+ const { defaultEnabled: sourceEnabled, ...metadata } = portable;
+ // Adapt after merging so a retained target default cannot outlive a replaced effort list.
+ return adaptCustomProviderModelEfforts({
+ ...existing,
+ ...metadata,
+ ...(existing?.defaultEnabled !== undefined
+ ? { defaultEnabled: existing.defaultEnabled }
+ : sourceEnabled === false ? { defaultEnabled: false } : {}),
+ }, targetAgent);
});
}
diff --git a/apps/desktop/src/renderer/lib/customProviders.ts b/apps/desktop/src/renderer/lib/customProviders.ts
index 7efdf18297..4cf74393a5 100644
--- a/apps/desktop/src/renderer/lib/customProviders.ts
+++ b/apps/desktop/src/renderer/lib/customProviders.ts
@@ -1,4 +1,4 @@
-import { mergeDiscoveredRuntimeModels, isOpenRouterModelsUrl, type DiscoveredModel } from '@cindy/model-providers';
+import { catalogModelMetadata, mergeDiscoveredRuntimeModels, isOpenRouterModelsUrl, type DiscoveredModel } from '@cindy/model-providers';
/**
* customProviders —— 自定义供应商「配置 + per-runtime 密钥」的 renderer 侧写入编排。
*
@@ -75,25 +75,11 @@ export function customProviderWireProtocolForSave(
* 推断(PR review P1);无标记的旧视图快照回退等值判断,行为不变。
*/
export function customProviderModelConfigFromCatalogModel(
- model: Pick<
- CatalogModel,
- | 'id'
- | 'name'
- | 'contextWindow'
- | 'contextWindowExplicit'
- | 'userModelConfig'
- | 'discoveredMetadata'
- | 'discoveredCost'
- | 'nameExplicit'
- | 'defaultEnabled'
- | 'supportsImageInput'
- | 'piApi'
- | 'route'
- > &
- Partial<Pick<CatalogModel, 'efforts' | 'defaultEffort'>>,
+ model: Pick<CatalogModel, 'id' | 'name' | 'contextWindow'> & Partial<CatalogModel>,
agent?: AgentKind,
): ProviderRuntimeModelConfig {
if (model.userModelConfig) return structuredClone(model.userModelConfig);
+ const { contextWindow: _window, name: _name, ...capabilities } = catalogModelMetadata(model);
const reasoningEfforts =
agent === 'pi'
? (model.efforts ?? []).filter((effort): effort is PiReasoningEffort =>
@@ -101,6 +87,8 @@ export function customProviderModelConfigFromCatalogModel(
)
: [];
return {
+ ...capabilities,
+ ...(model.api ? { api: model.api } : {}),
id: model.id,
name: model.name,
discoveredMetadata: model.discoveredMetadata,
@@ -169,7 +157,7 @@ export function providerViewToCustomProviderConfig(p: ProviderView): CustomProvi
};
}
-/** 刷新时只追加接口新发现的模型,并让新增模型默认隐藏。端点声明的 contextWindow 随发现带入(#386)。 */
+/** Refresh metadata and append new models; native-engine defaults decide visibility. */
export function appendDiscoveredCustomProviderModels(
existing: readonly ProviderRuntimeModelConfig[],
discovered: readonly DiscoveredModel[],
@@ -195,7 +183,7 @@ export function appendDiscoveredCustomProviderModels(
}
prior = [...normalized.values()];
}
- const models = mergeDiscoveredRuntimeModels(prior, discovered, true);
+ const models = mergeDiscoveredRuntimeModels(prior, discovered);
const known = new Set(existing.map((model) => model.id));
return {
models,
diff --git a/apps/desktop/src/shared/piRuntimeInitialization.ts b/apps/desktop/src/shared/piRuntimeInitialization.ts
index 903df0b845..c10f533f5a 100644
--- a/apps/desktop/src/shared/piRuntimeInitialization.ts
+++ b/apps/desktop/src/shared/piRuntimeInitialization.ts
@@ -21,26 +21,45 @@ export function configuredPresetAgents(preset: ProviderPreset): AgentKind[] {
export function savedCustomProviderModelShape(
model: ProviderRuntimeModelConfig,
- includePiCapabilities: boolean,
+ includePiApi: boolean,
): ProviderRuntimeModelConfig {
+ const { piApi, ...portable } = structuredClone(model);
return {
+ ...portable,
id: model.id.trim(),
name: model.name.trim(),
- ...(includePiCapabilities && model.piApi ? { piApi: model.piApi } : {}),
- ...(model.route ? { route: { ...model.route } } : {}),
- ...(model.contextWindow !== undefined ? { contextWindow: model.contextWindow } : {}),
- ...(model.defaultEnabled === false ? { defaultEnabled: false } : {}),
- ...(includePiCapabilities && model.supportsImageInput === true
- ? { supportsImageInput: true }
- : {}),
- ...(includePiCapabilities && model.reasoning === true && model.reasoningEfforts?.length
- ? {
- reasoning: true,
- reasoningEfforts: [...model.reasoningEfforts],
- ...(model.reasoningDefaultEffort
- ? { reasoningDefaultEffort: model.reasoningDefaultEffort }
- : {}),
- }
- : {}),
+ ...(!includePiApi && !model.api && piApi ? { api: piApi } : {}),
+ ...(includePiApi && piApi ? { piApi } : {}),
};
}
+
+/** Copy reasoning metadata into a runtime's supported vocabulary without mutating the source. */
+export function adaptCustomProviderModelEfforts(
+ source: ProviderRuntimeModelConfig,
+ agent: AgentKind,
+): ProviderRuntimeModelConfig {
+ const model = structuredClone(source);
+ // Adapt every representation before persistence; canonical metadata otherwise
+ // bypasses the legacy field validator. Keep Pi's shared endpoint record intact.
+ if (agent !== 'pi') {
+ const supported = new Set(['low', 'medium', 'high', 'xhigh', 'max']);
+ if (model.reasoningEfforts) {
+ model.reasoningEfforts = model.reasoningEfforts.filter(effort => supported.has(effort));
+ }
+ if (model.reasoningDefaultEffort != null &&
+ (!supported.has(model.reasoningDefaultEffort) ||
+ (model.reasoningEfforts && !model.reasoningEfforts.includes(model.reasoningDefaultEffort)))) {
+ delete model.reasoningDefaultEffort;
+ }
+ for (const metadata of [model, model.discoveredMetadata]) {
+ if (!metadata) continue;
+ if (metadata.efforts) metadata.efforts = metadata.efforts.filter(effort => supported.has(effort));
+ if (metadata.defaultEffort != null &&
+ (!supported.has(metadata.defaultEffort) ||
+ (metadata.efforts && !metadata.efforts.includes(metadata.defaultEffort)))) {
+ delete metadata.defaultEffort;
+ }
+ }
+ }
+ return model;
+}
diff --git a/docs/dev-rules/model-catalog-runtime.md b/docs/dev-rules/model-catalog-runtime.md
index d1d580609e..7ecc681066 100644
--- a/docs/dev-rules/model-catalog-runtime.md
+++ b/docs/dev-rules/model-catalog-runtime.md
@@ -233,9 +233,12 @@ Server 对应入口为 `model-access-server/src/routes/modelCatalog.ts`、`servi
到客户端容量;`context_window` 仍是工作窗口。只有最大容量时可用它作为窗口,二者都给出
时不以容量覆盖工作预算;不将 `contextWindowMax` 写入 Registry。刷新保留用户显式覆盖。
- Fast 不仅是目录字段:Claude 自定义桥按能力将用户选择转换成 `service_tier: priority`,
- 原生适配器保留它;Pi 直连在既有 `before_provider_request` 钩子读取每运行时请求偏好,
- 只对该连接明确支持的 OpenAI 协议模型设置 Fast,关闭时删除该参数。偏好文件只含开关和
- 模型身份,沿用运行时隔离、远端文件操作和退出回收;不改权限文件或凭证。
+ 原生适配器保留它;Pi 直连在既有 `before_provider_request` 钩子通过现有 RPC 只读查询
+ 宿主内存中的 Fast 开关。宿主按当前运行实例及精确 provider/model 校验,只对该连接明确
+ 支持的 OpenAI 协议模型启用 Fast,不接受查询方传入的开关值。关闭、实例失效、查询失败、
+ 超时或回复无效时删除该参数,使用普通档;不弹确认、不增加模型工具。
+ 不再读写 `runtime/request-prefs-*.json`,遗留文件不参与判定;不改权限文件、凭证或
+ Full Access 的原生 shell 权限。详见 [Pi Fast 偏好](pi-harness.md#原生请求-fast-偏好)。
- Mobile 与远控消费执行端已有目录和 Fast 设置入口;Pi 也接入原有设置保存/失败恢复流程。
不增加 IPC、凭证传输通道或独立模型配置。
diff --git a/docs/dev-rules/pi-harness.md b/docs/dev-rules/pi-harness.md
index aa9eddb842..65cff37626 100644
--- a/docs/dev-rules/pi-harness.md
+++ b/docs/dev-rules/pi-harness.md
@@ -68,6 +68,15 @@ Cindy 以 `pi --mode rpc` spawn pi 二进制(JSONL/stdio),`translator.ts` 把 pi
时从本机 session JSONL 校正(只打开启动时 `--session-dir` 真身内的普通文件,
并有字节/时间预算,超限回退 RPC);远端仍走 `get_entries`。
+### 原生请求 Fast 偏好
+
+支持 Fast 的原生模型在 `before_provider_request` 中经现有 RPC 通道发出
+`cindy:request-preferences` 只读查询。宿主按当前运行实例、精确 provider/model 与
+内存中的 Fast 开关返回结果,不接受查询传入的开关值;不弹用户确认,不增加模型工具。
+开关不再写入 `runtime/request-prefs-*.json`,遗留文件不参与判定。关闭实例、查询失败、
+超时或回复无效时使用普通档;已开启的正常请求仍发送 `service_tier: priority`。
+这修复偏好文件被改写或重放的路径,不改变 Pi Full Access 的原生 shell 权限边界。
+
## 2. 配置面:Cindy 显式设置 vs 放任 pi 默认
图片能力未声明时,Pi 的模型配置与发送校验默认允许图片输入,不因新型号缺少能力资料而
diff --git a/docs/dev-rules/provider-catalog-generation.md b/docs/dev-rules/provider-catalog-generation.md
index 55b3a4ed83..0e8ae8783c 100644
--- a/docs/dev-rules/provider-catalog-generation.md
+++ b/docs/dev-rules/provider-catalog-generation.md
@@ -36,7 +36,7 @@
- 在线刷新已知供应商:`pnpm sync:pi-model-catalog`。公开 Pi 端点可能拒绝请求;失败时不替换目录。
- 导入 Pi 生成器的完整导出:`pnpm sync:pi-model-catalog --input /path/to/providers.json --source-version <version>`。
- 输入是 provider ID 到完整 Pi 模型数组的映射;新供应商自动进入标准表,缺少已有供应商则报错,避免误删。
+ 输入是 provider ID 到完整 Pi 模型数组的映射;新供应商自动进入标准表,未提供的供应商保留旧记录。
- 从 Cindy 已固定版本的 Pi 二进制读取生成数据:
`pnpm sync:pi-model-catalog --pi-bundle /path/to/pi --source-version 0.85.1 --generated-at 2026-09-12T00:00:00Z`。
仅解析生成的字面量数据,不执行二进制。旧版本已有的 Astra/xAI 修正由具名函数保留,新版本不套旧版本修正。
@@ -80,7 +80,7 @@ authenticated online OpenCode and GUI acceptance remain unverified.
- Vercel 读取 `max_tokens`、输入模态、`reasoning_options` 的明确 effort 取值及语言模型的 USD/token 价格;非语言模型保留类型,不把图片单价当 token 价格。单位依据官方 [模型发现文档](https://vercel.com/docs/ai-gateway/models-and-providers)。
- LiteLLM 的 `model_name` 保留为调用 ID,`model_info` 可补窗口、输出上限、视觉与工具能力;LM Studio 的 `key` / capabilities 结构可被解析。普通 OpenAI 兼容端点的 ID 列表也保持可用。
- 刷新仅返回 ID 时,保留此前发现的资料;明确 false 仍覆盖旧资料,用户字段仍最后覆盖。连接编辑与列表刷新均保留缺失字段。
-- 不按模型名模糊猜测厂商协议,也不把任意代理价格解释为美元。仅有 ID 且标准表无记录时,未知字段仍未知;这些事实不能靠静态代码补造。
+- 不按模型名模糊猜测厂商协议,也不把任意代理价格解释为美元。仅有 ID 且当前型号无记录时,按模型资料优先级使用同系列前代能力;API 实报与服务器修正优先。不得把继承值标记为实报,也不借用前代价格。
验证:公共 GET 的 OpenRouter 445 条、Vercel 375 条(其中语言模型 252 条)在三引擎投影中核对输入能力、窗口、输出与可解析价格。OpenCode 的匿名接口仍返回 403;本地服务没有本次真实实例验收,测试覆盖其响应结构,不等同于已访问用户的服务。没有生成调用。
@@ -242,3 +242,14 @@ OAuth 复用既有 Main 凭证存储、取消与刷新机制;新渠道登录
验证:兼容包 204、目录 934、三个桥接/代理包合计 788 项通过;Desktop 定向 400 项通过,另有既有跳过。受影响包类型检查通过。没有真实账号付费生成请求。本机许可证全量再生成缺 Cargo,已按两侧版本合并 Pi 0.85.1 与 OpenCodex MIT 声明并核对 SPDX 引用,无 dangling references。
新隔离测试版 `dev2-provider-four-api` 已取得 `DESKTOP_DEV_VERDICT=ready`,工作区为 `cindy-provider-native-defaults`。旧测试工作区保留;新档案需重新登录、导入渠道,真实账号首次导入由用户继续验收。没有推送、发布或合并原 PR。
+
+
+## 部分失败与新型号
+
+Pi 同步先读供应商索引,合并已有供应商;单个来源或型号失败保留最后有效记录,不中断其它来源。
+供应商清单完整且全部转换成功时按新清单移除下架型号;存在无效条目或来源失败时不据此删旧型号。
+同型号、同端点和同协议刷新缺失适配字段时保留上次值,明确提供的空映射或否定值仍替换旧值。
+成功输出通过临时文件原子替换。缺失能力不合成为明确 false / 空数组,运行时继续按字段继承。
+通用 API 发现读取分页;后续页失败保留已取得的型号,不向跨域或其它路径的 next 链接发送凭证。
+链接导入、配置保存、跨引擎复制和刷新共享 ModelMetadata,保留明确否定值。
+新型号不固化为 defaultEnabled:false;原生引擎默认和用户明确开关在目录投影时统一决定。
diff --git a/docs/product-rules/model-metadata-precedence.md b/docs/product-rules/model-metadata-precedence.md
index 40266e4b7d..f2f5881b13 100644
--- a/docs/product-rules/model-metadata-precedence.md
+++ b/docs/product-rules/model-metadata-precedence.md
@@ -5,6 +5,7 @@
裁决日期:2026-09-08。模型资料按字段合并,最终采用以下优先级(低到高):
+0. 同厂商、同系列和变体的前代能力:仅填补缺项,不继承 ID、名称、路由、凭证、成员、开关和报价。
1. `modelRegistry.baseModels[].defaults`:模型公共资料。
2. 供应商/运行时默认资料:匹配的预设模型默认(适用时)→ 接入条目顶层字段 → `routes[].defaults` → 条目 `perAgent`。每一步只覆盖明确填写的字段。
3. 供应商接口明确返回的资料。没返回的字段继续继承默认;不能把客户端合成值当成供应商事实。
@@ -17,7 +18,7 @@
标准档位和默认强度只在公共模型维护一套,三个 Harness 默认继承;只有经核实的来源或
Harness 差异才写稀疏覆盖,不为每个 Harness 复制完整默认配置。现有 route / perAgent、
-账号实报和用户显式覆盖继续承载这些例外,不能由客户端型号判断自动补档。
+账号实报和用户显式覆盖继续承载这些例外;新代型号缺项采用下文的前代继承。
`defaultEffort` 表达推荐意图,区别于 `efforts` 的可执行能力。因此它是上述实报优先级的
例外:已配置的模型默认(含明确的 route / perAgent 默认例外)优先于 Harness 实报的
@@ -42,35 +43,18 @@ Pi 明确声明的 `reasoningDefaultEffort` 属于 Harness 默认例外,高于
## 推理能力缺失时的统一规则
-裁决日期:2026-09-22;自定义供应商的新代际默认继承于 2026-09-24 按用户定稿修订,见下方专节。
-**能力资料与客户端兼容默认分开保存,不把默认值冒充供应商实报。**
-本节适用于 Claude Code、Codex、Pi 的模型目录、选择器和实际调用,不因入口或引擎不同另设通用档位。
-
-1. **未知不等于不支持。** 资料层保留以下区别;最终可选条目的兜底值不作为能力事实写回资料层。
-
- | 原始声明 | 含义 | 继承后的行为 |
- | --- | --- | --- |
- | 缺少推理能力字段 | 能力未知,允许继续继承 | 仍未知则不指定强度,不主动发送关闭思考的参数 |
- | 明确 `efforts: []` 或用户配置 `reasoning: false` | 明确无可调档位 | 保留明确声明,不用低优先级资料补回档位 |
- | 非空 `efforts` | 已声明支持的档位 | 仅提供和发送该范围内的强度 |
-
-2. **不提供通用档位套餐。** 缺少资料时不得补 `low / medium / high / xhigh / max`,
- 也不得统一指定 `medium` 或 `high`。兼容 OpenAI 等接口协议只证明请求格式可用,
- 不证明模型支持该协议的全部可选参数或档位。
-3. **按字段继承并保留来源。** 沿用本文的公共资料 → 接入默认 → 接口实报 → 有依据的强制修正
- → 用户显式配置顺序。缺字段继续继承;明确的 `false`、空数组和 `null` 保留各自语义,
- 不被低优先级资料补回。客户端回退值不得冒充供应商实报或用户配置。
-4. **默认值不能创造能力。** 非空默认档必须属于有效支持列表;列表已知而默认字段缺失时,
- 只按共享默认选择规则从已声明档位中选取。`defaultEffort: null` 保持不指定默认档。
- 列表仍未知时最终使用 `efforts: []`、`defaultEffort: null`,请求省略推理强度参数。
-5. **展示与请求遵守同一有效能力。** 普通任务、子 Agent、自动任务和辅助调用均按实际目标的
- 连接、模型与引擎校验档位。旧任务保存的选择不是能力声明,不能凭旧 `max` 绕过当前校验;
- 选择器占位值、父任务档位和 GPT 模板也不能补回未经确认的能力。支持列表为空时不发送强度,
- 非空时沿用共享协调规则处理不再受支持的选择;不借此覆盖已保存的用户偏好。
-6. **新资料自动生效,用户设置不被覆盖。** 除下述自定义供应商代际继承外,新型号先按未知能力处理;目录或接口补齐资料后,
- 沿用现有刷新和继承链恢复已声明能力。不得按型号前缀或名称相似度猜测能力;精确型号、明确
- alias 和既有协议 ID 归一仍按本文身份规则处理。临时回退不落成用户 override,刷新不覆盖
- 用户显式设置,也不因缺少推理资料删除模型。
+裁决日期:2026-09-24。**账号发现的新型号立即导入;接口优先补资料,缺项采用同系列前代预设,不等服务器逐型号补丁。**
+本节替代旧的“新型号一律空档位”策略,适用于 Claude Code、Codex、Pi 的目录与请求。
+
+1. 缺字段允许继承;明确 `false`、`efforts: []` 和 `defaultEffort: null` 保留原义,不用低优先级资料补回。
+2. 按厂商、系列、变体及数字代次匹配最近前代;私有命名空间不剥除,不把 Chat 协议本身当成 GPT 身份。
+ 参数继承不改变型号、账号成员资格、路由、协议或报价,不把继承值写成 API 实报或用户覆盖。
+3. 当前型号公共资料、渠道资料与 API 实报优先于前代;服务器 `forceOverrides` 可修正新型号的参数差异,用户覆盖最后生效。
+4. 默认思考强度必须属于合并后的有效档位。确无同系列前代或明确资料时才使用无档位兜底,模型本身仍导入可用。
+5. 模型默认开启与兼容引擎授权分开:按模型主 API 与实际执行协议判断原生路径;导入列表勾选、链接里的
+ 模型开启、同步发现和跨引擎复制,不构成开启 Claude Code / Codex 兼容模式的授权。
+6. 新型号、变体和 preview 不受客户端逐型号白名单限制;用户明确关闭与账号不可用状态保持有效。
+ 普通任务、子 Agent 和辅助调用共用合并后的能力及出站参数规则。
回归验证必须覆盖:缺资料时的三个引擎、无目录与旧版/当前目录、明确空值、接口实报与用户覆盖、
后续资料补齐、旧任务档位的出站参数,以及子 Agent 已有条目/新增模板两条路径。
@@ -83,9 +67,9 @@ Pi 明确声明的 `reasoningDefaultEffort` 属于 Harness 默认例外,高于
### 自定义供应商的新代际兼容默认(2026-09-24)
-用户明确要求新代模型无需逐次更新客户端,导入时自动继承前代配置,并否决额外的“未验证”提示。
-因此,自定义供应商(含 Sub2API)允许按同系列、同变体的数字代际继承较旧型号的缺失配置,
-不要求上游另行声明代际关系。这是上述第 2、4、6 条在该路径的明确例外,不是供应商能力事实。
+自定义供应商(含 Sub2API)按同系列、同变体的数字代际继承较旧型号的缺失配置,
+让新代模型无需逐次更新客户端即可导入使用,不要求上游另行声明代际关系。
+这是上述统一继承规则在自定义连接的应用;继承值不是供应商实报,也不增加额外提示。
- 按数字版本选择较近前代,同代际优先同连接配置;连接候选须保持端点、路径和协议一致。
Sol / Luna、不同私有命名空间、型号尺寸和日期后缀不得混用,不按任意名称相似度匹配。
diff --git a/packages/maker-core/src/agents/pi/__tests__/cindyBridgeSource.test.ts b/packages/maker-core/src/agents/pi/__tests__/cindyBridgeSource.test.ts
index 434fd87730..a96dc9dbbc 100644
--- a/packages/maker-core/src/agents/pi/__tests__/cindyBridgeSource.test.ts
+++ b/packages/maker-core/src/agents/pi/__tests__/cindyBridgeSource.test.ts
@@ -20,7 +20,7 @@ import { createInterface } from 'node:readline';
import { runInNewContext } from 'node:vm';
import ts from 'typescript';
-import { describe, expect, it } from 'vitest';
+import { describe, expect, it, vi } from 'vitest';
import {
CINDY_BRIDGE_EXTENSION_SOURCE,
@@ -146,11 +146,15 @@ function loadBashIsolationHelper(
home: string | undefined,
) => Record<string, string | undefined> {
const source = CINDY_BRIDGE_EXTENSION_SOURCE;
- const start = source.indexOf('function withoutPiSecrets');
+ const constantsStart = source.indexOf('const MANAGED_RG_PATH_ENV');
+ const constantsEnd = source.indexOf('const PI_PACKAGE_MANAGEMENT_TITLE');
+ const start = source.indexOf('const SECRET_ENV_NAMES');
const end = source.indexOf('function managedRipgrepPath');
- if (start < 0 || end <= start) throw new Error('bash isolation helper was not found');
+ if (constantsStart < 0 || constantsEnd <= constantsStart || start < 0 || end <= start) {
+ throw new Error('bash isolation helper was not found');
+ }
const executableSource = [
- "const SECRET_ENV_NAMES = new Set(['PI_CODING_AGENT_DIR', 'CINDY_PI_PACKAGE_MANAGEMENT', 'CINDY_PI_BASH_PACKAGE_HOME']);",
+ source.slice(constantsStart, constantsEnd),
source.slice(start, end),
'(globalThis as any).isolatedBashEnvironment = isolatedBashEnvironment;',
].join('\n');
@@ -160,7 +164,7 @@ function loadBashIsolationHelper(
target: ts.ScriptTarget.ES2022,
},
}).outputText;
- const context: Record<string, unknown> = { path: pathImpl };
+ const context: Record<string, unknown> = { path: pathImpl, process: { env: {} } };
runInNewContext(compiled, context);
return context.isolatedBashEnvironment as (
env: Record<string, string | undefined>,
@@ -1792,6 +1796,30 @@ describe('cindy-bridge extension source', () => {
expect(() => isolateWindows({}, 'relative\\home')).toThrow(/unavailable/);
});
+ it.each([
+ ['POSIX', path.posix, '/isolated/pi-home'],
+ ['Windows', path.win32, 'D:\\isolated\\pi-home'],
+ ] as const)('hides Fast preferences from %s shells without mutating the runtime', (_platform, pathImpl, home) => {
+ const isolate = loadBashIsolationHelper(pathImpl);
+ const runtimeEnv = {
+ CINDY_PI_MODEL_REQUEST_PREFS_FILE: pathImpl.join(home, 'request-prefs.json'),
+ CINDY_PI_FAST_MODELS: '[]',
+ PATH: 'ordinary-shell-path',
+ SHELL_CANARY: 'preserved',
+ };
+ const shellEnv = isolate(runtimeEnv, home);
+ expect(shellEnv).not.toHaveProperty('CINDY_PI_MODEL_REQUEST_PREFS_FILE');
+ expect(shellEnv).not.toHaveProperty('CINDY_PI_FAST_MODELS');
+ expect(shellEnv).toMatchObject({
+ PATH: runtimeEnv.PATH,
+ SHELL_CANARY: runtimeEnv.SHELL_CANARY,
+ PI_CODING_AGENT_DIR: home,
+ });
+ expect(runtimeEnv.CINDY_PI_MODEL_REQUEST_PREFS_FILE).toBe(
+ pathImpl.join(home, 'request-prefs.json'),
+ );
+ });
+
it('routes both Pi command names to the single host permission service', () => {
expect(CINDY_BRIDGE_EXTENSION_SOURCE).toContain(
"if (event.toolName === 'cindy_pi_extension' || event.toolName === 'cindy_pi_command') return;",
@@ -2329,23 +2357,32 @@ describe('Pi same-turn library native mapping', () => {
});
-it('applies native Fast only to the exact declared connection and follows preference changes', () => {
- const start = CINDY_BRIDGE_EXTENSION_SOURCE.indexOf('function nativeFastPayload(');
+it('applies native Fast only to the exact declared connection and fails closed on invalid host replies', async () => {
+ const start = CINDY_BRIDGE_EXTENSION_SOURCE.indexOf('async function nativeFastPayload(');
const end = CINDY_BRIDGE_EXTENSION_SOURCE.indexOf('export default async function cindyBridge');
const helpers = ts.transpileModule(CINDY_BRIDGE_EXTENSION_SOURCE.slice(start, end), {
compilerOptions: { module: ts.ModuleKind.ESNext, target: ts.ScriptTarget.ES2022 },
}).outputText;
- let fast = true;
- const read = () => JSON.stringify({ fast, models: [{ provider: 'relay-a', id: 'gpt-6-sol' }] });
- const adapt = new Function('readFileSync', 'process', `${helpers}; return nativeFastPayload;`)(
- read, { env: { CINDY_PI_MODEL_REQUEST_PREFS_FILE: '/fixture/prefs.json' } });
+ let reply: unknown = JSON.stringify({ fast: true });
+ const adapt = new Function('process', `${helpers}; return nativeFastPayload;`)({ env: {
+ CINDY_PI_FAST_MODELS: JSON.stringify([{ provider: 'relay-a', id: 'gpt-6-sol' }]),
+ } });
+ const ctx = { ui: { input: async () => { if (reply instanceof Error) throw reply; return reply; } } };
const original = { model: 'gpt-6-sol', reasoning: { effort: 'high' }, input: 'hello' };
const model = { provider: 'relay-a', id: 'gpt-6-sol', api: 'openai-responses' };
- expect(adapt(original, model)).toEqual({ ...original, service_tier: 'priority' });
- expect(adapt(original, { ...model, provider: 'relay-b' })).toBeUndefined();
- expect(adapt(original, { ...model, id: 'other-model' })).toBeUndefined();
- expect(adapt(original, { ...model, api: 'anthropic-messages' })).toBeUndefined();
- fast = false;
- expect(adapt({ ...original, service_tier: 'priority' }, model)).toEqual(original);
+ expect(await adapt(original, model, ctx)).toEqual({ ...original, service_tier: 'priority' });
+ expect(await adapt(original, { ...model, provider: 'relay-b' }, ctx)).toBeUndefined();
+ expect(await adapt(original, { ...model, id: 'other-model' }, ctx)).toBeUndefined();
+ expect(await adapt(original, { ...model, api: 'anthropic-messages' }, ctx)).toBeUndefined();
+ for (reply of [JSON.stringify({ fast: false }), JSON.stringify({ fast: 'true' }), undefined, 'broken', new Error('closed')]) {
+ expect(await adapt({ ...original, service_tier: 'priority' }, model, ctx)).toEqual(original);
+ }
+ vi.useFakeTimers();
+ try {
+ const pending = adapt({ ...original, service_tier: 'priority' }, model,
+ { ui: { input: () => new Promise(() => {}) } });
+ await vi.advanceTimersByTimeAsync(5000);
+ expect(await pending).toEqual(original);
+ } finally { vi.useRealTimers(); }
expect(original).not.toHaveProperty('service_tier');
});
diff --git a/packages/maker-core/src/agents/pi/__tests__/cindySubagentForeground.test.ts b/packages/maker-core/src/agents/pi/__tests__/cindySubagentForeground.test.ts
index 9215e1aeef..d5dc7a3a10 100644
--- a/packages/maker-core/src/agents/pi/__tests__/cindySubagentForeground.test.ts
+++ b/packages/maker-core/src/agents/pi/__tests__/cindySubagentForeground.test.ts
@@ -85,7 +85,7 @@ process.stdin.on('data', (chunk) => {
process.stdout.write(JSON.stringify({ type: 'tool_execution_start', toolName: 'write' }) + '\\n');
process.stdout.write(JSON.stringify({
type: 'message_end',
- message: { role: 'assistant', content: [{ type: 'text', text: 'foreground approved result:' + process.env.CINDY_PI_SESSION_TOKEN + (process.env.CINDY_PI_MODEL_REQUEST_PREFS_FILE ? ':prefs=' + require('node:fs').readFileSync(process.env.CINDY_PI_MODEL_REQUEST_PREFS_FILE, 'utf8') : '') }], usage: { input: 2, output: 3, cost: { total: 0.02 } } },
+ message: { role: 'assistant', content: [{ type: 'text', text: 'foreground approved result:' + process.env.CINDY_PI_SESSION_TOKEN }], usage: { input: 2, output: 3, cost: { total: 0.02 } } },
}) + '\\n');
process.stdout.write(JSON.stringify({ type: 'agent_end' }) + '\\n');
process.stdout.write(JSON.stringify({ type: 'agent_settled' }) + '\\n');
@@ -154,12 +154,8 @@ describe('Cindy PI Subagent foreground durable path', () => {
'waits for the runner and forwards Ask approval through the parent UI',
async () => {
const f = await fixture();
- const requestPrefsFile = path.join(f.root, 'request-prefs.json');
- const prefs = JSON.stringify({ fast: true, models: [{ provider: 'fixture', id: 'fixture-model' }] });
- await writeFile(requestPrefsFile, prefs);
const previous = { ...process.env };
Object.assign(process.env, {
- CINDY_PI_MODEL_REQUEST_PREFS_FILE: requestPrefsFile,
[CINDY_SUBAGENT_ENV.binary]: process.execPath,
[CINDY_SUBAGENT_ENV.depth]: '0',
[CINDY_SUBAGENT_ENV.runtimeFile]: f.runtimeFile,
@@ -199,14 +195,7 @@ describe('Cindy PI Subagent foreground durable path', () => {
const extension = require(f.extensionFile).default as (pi: { registerTool: (tool: unknown) => void }) => Promise<void>;
await extension({ registerTool: (tool) => Object.assign(registered, tool) });
expect(registered.execute).toBeTypeOf('function');
- const launchInput = hostInput(f);
- const input = vi.fn(async (title: string, placeholder: string) => {
- if (title === 'cindy:pi-subagent-runner' && JSON.parse(placeholder).action === 'launch') {
- // Model the parent closing before the durable runner starts.
- await rm(requestPrefsFile);
- }
- return launchInput(title, placeholder);
- });
+ const input = hostInput(f);
const result = await registered.execute!(
'tool-foreground',
{ agent: 'worker', task: 'write the fixture', model: 'claude-fable-5' },
@@ -221,9 +210,6 @@ describe('Cindy PI Subagent foreground durable path', () => {
const [runId] = await readdir(f.runRoot);
const configText = await readFile(path.join(f.runRoot, runId, 'config.json'), 'utf8');
const config = JSON.parse(configText);
- expect(config.requestPrefsFile).toBe(path.join(f.runRoot, runId, 'request-prefs.json'));
- expect(result.content[0].text).toContain(':prefs=' + prefs);
- expect(await readFile(config.requestPrefsFile, 'utf8')).toBe(prefs);
expect(config.tasks[0]).toMatchObject({
provider: 'anthropic',
model: 'claude-fable-5',
diff --git a/packages/maker-core/src/agents/pi/__tests__/cindySubagentRunner.test.ts b/packages/maker-core/src/agents/pi/__tests__/cindySubagentRunner.test.ts
index f4a7e71640..c356f39c1d 100644
--- a/packages/maker-core/src/agents/pi/__tests__/cindySubagentRunner.test.ts
+++ b/packages/maker-core/src/agents/pi/__tests__/cindySubagentRunner.test.ts
@@ -667,9 +667,14 @@ describe('Cindy durable PI Subagent runner', () => {
return null;
}
});
- const [closing] = await listPiSubagentRuns(fixture.root);
- expect(closing && closing.state !== 'completed' && closing.state !== 'failed').toBe(true);
- await expect(controlPiSubagentRuns(fixture.root, closing!.runId, 'follow_up', {
+ // The child marker is independent of the runner's status publication;
+ // listPiSubagentRuns may omit a temporarily unreadable snapshot on Windows.
+ const closing = await waitFor(async () => {
+ const [run] = await listPiSubagentRuns(fixture.root);
+ return run ?? null;
+ }, undefined, 'readable status after child RPC input closes');
+ expect(closing.state !== 'completed' && closing.state !== 'failed').toBe(true);
+ await expect(controlPiSubagentRuns(fixture.root, closing.runId, 'follow_up', {
message: 'too late for this generation',
})).resolves.toBe(0);
await writeFile(fixture.exitReleaseFile, '1');
@@ -976,9 +981,6 @@ describe('Cindy durable PI Subagent runner', () => {
'the first generation to complete',
);
await waitForClose(fixture.child, fixture.stderr);
- const prefs = JSON.stringify({ fast: true, models: [{ provider: 'fixture', id: 'fixture-model' }] });
- const liveRequestPrefsFile = path.join(fixture.runDir, 'current-parent-prefs.json');
- await writeFile(liveRequestPrefsFile, prefs);
const resumeTokenCanary = 'resume-parent-token-canary-1234567890';
const priorConfigPath = path.join(fixture.runDir, 'config.json');
const priorConfig = JSON.parse(await readFile(priorConfigPath, 'utf8')) as {
@@ -1002,7 +1004,6 @@ describe('Cindy durable PI Subagent runner', () => {
runnerFallbackFile: fixture.runnerFile,
env: {
...process.env,
- CINDY_PI_MODEL_REQUEST_PREFS_FILE: liveRequestPrefsFile,
CINDY_PI_SESSION_TOKEN: resumeTokenCanary,
CINDY_TEST_PI_ARGS: fixture.argsFile,
CINDY_TEST_PI_PROMPTS: fixture.promptsFile,
@@ -1023,10 +1024,6 @@ describe('Cindy durable PI Subagent runner', () => {
});
expect(resumed.tasks[0]?.sessionId).toBe(first.tasks[0]?.sessionId);
expect(resumed.runtimeOwnerId).toBe('resume-owner');
- const resumedConfig = JSON.parse(await readFile(path.join(fixture.root, resumedRunId!, 'config.json'), 'utf8'));
- expect(resumedConfig.requestPrefsFile).toBe(path.join(fixture.root, resumedRunId!, 'request-prefs.json'));
- await rm(liveRequestPrefsFile);
- expect(await readFile(resumedConfig.requestPrefsFile, 'utf8')).toBe(prefs);
await expect(readFile(path.join(fixture.root, resumedRunId!, 'permission.json'), 'utf8'))
.resolves.toContain('/current-parent');
await expect(readFile(path.join(fixture.root, resumedRunId!, 'pi-home', 'models.json')))
@@ -1056,7 +1053,6 @@ describe('Cindy durable PI Subagent runner', () => {
runnerFallbackFile: fixture.runnerFile,
env: {
...process.env,
- CINDY_PI_MODEL_REQUEST_PREFS_FILE: undefined,
CINDY_PI_SESSION_TOKEN: resumeTokenCanary,
CINDY_TEST_PI_ARGS: fixture.argsFile,
CINDY_TEST_PI_PROMPTS: fixture.promptsFile,
@@ -1076,12 +1072,6 @@ describe('Cindy durable PI Subagent runner', () => {
return runs.find((run) => run.runId === secondResumedRunId && run.state === 'completed') ?? null;
});
expect(secondResumed.tasks[0]?.sessionId).toBe(first.tasks[0]?.sessionId);
- const secondConfig = JSON.parse(await readFile(path.join(fixture.root, secondResumedRunId!, 'config.json'), 'utf8'));
- expect(secondConfig.requestPrefsFile).toBeUndefined();
- await expect(readFile(path.join(fixture.root, secondResumedRunId!, 'request-prefs.json')))
- .rejects.toMatchObject({ code: 'ENOENT' });
- // The old snapshot still exists, but is no longer authoritative for resume.
- expect(await readFile(resumedConfig.requestPrefsFile, 'utf8')).toBe(prefs);
const resumedPrompts = (await readFile(fixture.promptsFile, 'utf8'))
.trim().split('\n').map((line) => JSON.parse(line) as string);
expect(resumedPrompts.at(-1)).toBe('continue for a second resumed generation');
diff --git a/packages/maker-core/src/agents/pi/__tests__/pi-agent.integration.test.ts b/packages/maker-core/src/agents/pi/__tests__/pi-agent.integration.test.ts
index fca30af879..a2a7176a5c 100644
--- a/packages/maker-core/src/agents/pi/__tests__/pi-agent.integration.test.ts
+++ b/packages/maker-core/src/agents/pi/__tests__/pi-agent.integration.test.ts
@@ -432,6 +432,46 @@ describe.skipIf(!piAvailable)('PiAgent integration (real pi binary + fake gatewa
};
}
+ it('reads native Fast from host RPC and ignores replayed preference files in Full Access', { timeout: 30_000 }, async () => {
+ const deps = buildDeps();
+ deps.resolvePiNativeProviders = async () => ({ providers: [{ id: 'fast-relay', name: 'Fast Relay',
+ baseUrl: endpoint, api: 'openai-responses', apiKeyEnvVar: 'CINDY_PI_API_KEY',
+ models: [{ id: 'gpt-6-sol', name: 'Fast fixture', contextWindow: 128000, supportsFastMode: true }],
+ }], env: {} });
+ const workingDir = mkdtempSync(path.join(tmpdir(), 'pi-fast-rpc-'));
+ let handle: AgentSessionHandle | undefined;
+ const before = seenRequests.length;
+ try {
+ handle = await new PiAgent(deps).startSession({ sessionId: 'fast-rpc', workingDir,
+ providerId: 'fast-relay', model: 'gpt-6-sol', permissionMode: 'bypassPermissions' });
+ const legacyFile = path.join(agentHome, 'runtime', 'request-prefs-forged.json');
+ writeFileSync(legacyFile, JSON.stringify({ fast: true, models: [{ provider: 'fast-relay', id: 'gpt-6-sol' }] }));
+ const send = async () => {
+ const done = (async () => { for await (const event of handle!.events()) {
+ if (event.type === 'error') throw new Error(JSON.stringify(event.data));
+ if (event.type === 'interaction_request') throw new Error('Fast lookup must not prompt');
+ if (event.type === 'done') return event;
+ } })();
+ await handle!.send({ type: 'user', content: 'Reply briefly.' });
+ expect((await done)?.data).toMatchObject({ status: 'completed' });
+ };
+ await send();
+ await handle.setFastMode!(true);
+ await send();
+ await handle.setFastMode!(false);
+ // Replaying an earlier enabled snapshot cannot restore host Fast intent.
+ writeFileSync(legacyFile, JSON.stringify({ fast: true, models: [{ provider: 'fast-relay', id: 'gpt-6-sol' }] }));
+ await send();
+ const bodies = seenRequests.slice(before).map(request => JSON.parse(request.body));
+ expect(bodies.map(body => body.service_tier)).toEqual([undefined, 'priority', undefined]);
+ expect(bodies.every(body => body.model === 'gpt-6-sol')).toBe(true);
+ expect(JSON.stringify(bodies)).not.toContain('cindy:request-preferences');
+ } finally {
+ await handle?.close();
+ rmSync(workingDir, { recursive: true, force: true });
+ }
+ });
+
it('applies configured windows on creation and same-history resume, then restores the default',
{ timeout: 60_000 }, async () => {
let limit: number | null = 80_000;
@@ -950,7 +990,7 @@ describe.skipIf(!piAvailable)('PiAgent integration (real pi binary + fake gatewa
providers: [{
id: 'xai', sourceProviderId: 'xai', name: 'xAI',
baseUrl: `${endpoint}/v1`, inheritModels: true,
- models: [{ ...row, api, input: row.input.filter((kind): kind is 'text' | 'image' => kind === 'text' || kind === 'image'),
+ models: [{ ...row, api, input: row.input?.filter((kind): kind is 'text' | 'image' => kind === 'text' || kind === 'image'),
cost: { ...row.cost, input: row.cost?.input ?? 0, output: row.cost?.output ?? 0,
cacheRead: row.cost?.cacheRead ?? 0, cacheWrite: row.cost?.cacheWrite ?? 0,
tiers: row.cost?.tiers?.map(tier => ({ ...tier,
diff --git a/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts b/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
index a73a18e9e5..2889c2371b 100644
--- a/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
+++ b/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
@@ -1,3 +1,4 @@
+import ts from "typescript";
import { createHash } from "node:crypto";
import {
mkdirSync,
@@ -21,6 +22,8 @@ const captured = vi.hoisted(() => ({
args: [] as string[],
env: {} as Record<string, string | undefined>,
requests: [] as Array<Record<string, unknown>>,
+ responses: [] as Array<Record<string, unknown>>,
+ onEvent: undefined as undefined | ((event: Record<string, unknown>) => void),
requestOptions: [] as Array<
| {
timeoutMs?: number;
@@ -88,12 +91,14 @@ vi.mock("../rpc-client.js", () => {
PiRpcProcess: class {
isClosed = false;
constructor(opts: {
+ onEvent: (event: Record<string, unknown>) => void;
onExit: (info: {
code: number | null;
signal: NodeJS.Signals | null;
}) => void;
}) {
captured.onExit = opts.onExit;
+ captured.onEvent = opts.onEvent;
}
async request(
command: Record<string, unknown>,
@@ -150,7 +155,7 @@ vi.mock("../rpc-client.js", () => {
}
return response;
}
- send(): void {}
+ send(command: Record<string, unknown>): void { captured.responses.push(command); }
async close(): Promise<void> {
this.isClosed = true;
captured.closes += 1;
@@ -217,6 +222,8 @@ describe("Pi provider-aware model routing", () => {
beforeEach(() => {
captured.args = [];
captured.requests = [];
+ captured.responses = [];
+ captured.onEvent = undefined;
captured.requestOptions = [];
captured.closes = 0;
captured.onExit = undefined;
@@ -234,8 +241,8 @@ describe("Pi provider-aware model routing", () => {
rmSync(cwd, { recursive: true, force: true });
});
- it("persists native Fast per runtime and updates it before later sends", async () => {
- let fast = true;
+ it("keeps native Fast in the host even when a Full Access shell can rewrite old preference files", async () => {
+ let fast = false;
const agent = new PiAgent({
auth: { getState: async () => ({ authenticated: true, authSource: 'api-key' as const }),
triggerLogin: async () => ({ authenticated: true }), logout: async () => {}, getAuthEnv: async () => ({}) },
@@ -246,18 +253,43 @@ describe("Pi provider-aware model routing", () => {
models: [{ id: 'private-sol', supportsFastMode: true }, { id: 'no-fast' }] }], env: {} }),
});
const handle = await agent.startSession({ sessionId: 'native-fast', workingDir: cwd,
- model: 'private-sol', providerId: 'relay', getPriceVariant: () => fast ? 'priority' : 'standard' });
- const file = captured.env.CINDY_PI_MODEL_REQUEST_PREFS_FILE!;
- expect(JSON.parse(captured.env.CINDY_PI_SECRET_ENV_NAMES!)).toContain('CINDY_PI_MODEL_REQUEST_PREFS_FILE');
- expect(JSON.parse(readFileSync(file, 'utf8'))).toEqual({ fast: true, models: [{ provider: 'relay', id: 'private-sol' }] });
+ permissionMode: 'bypassPermissions', model: 'private-sol', providerId: 'relay',
+ getPriceVariant: () => fast ? 'priority' : 'standard' });
+ expect(captured.env.CINDY_PI_MODEL_REQUEST_PREFS_FILE).toBeUndefined();
+ expect(readdirSync(path.join(agentHome, 'runtime')).some(name => name.startsWith('request-prefs-'))).toBe(false);
+ // Even a known/replayed legacy path and attacker-authored fast=true are inert.
+ const file = path.join(agentHome, 'runtime', 'request-prefs-attacker.json');
+ writeFileSync(file, JSON.stringify({ fast: true, models: [{ provider: 'relay', id: 'private-sol' }] }));
+ const source = CINDY_BRIDGE_EXTENSION_SOURCE.slice(
+ CINDY_BRIDGE_EXTENSION_SOURCE.indexOf('async function nativeFastPayload('),
+ CINDY_BRIDGE_EXTENSION_SOURCE.indexOf('export default async function cindyBridge'),
+ );
+ const js = ts.transpileModule(source, { compilerOptions: { module: ts.ModuleKind.ESNext } }).outputText;
+ const adapt = new Function('process', `${js}; return nativeFastPayload;`)({ env: {
+ ...captured.env, CINDY_PI_MODEL_REQUEST_PREFS_FILE: file,
+ } });
+ const query = (payload: unknown) => {
+ captured.onEvent!({ type: 'extension_ui_request', method: 'input', id: 'fast-query',
+ title: 'cindy:request-preferences', placeholder: JSON.stringify(payload) });
+ return captured.responses.at(-1)!.value as string;
+ };
+ const ctx = { ui: { input: async (_title: string, payload: string) => query(JSON.parse(payload)) } };
+ const model = { provider: 'relay', id: 'private-sol', api: 'openai-responses' };
+ const payload = { model: 'private-sol', input: 'hello' };
+ expect(await adapt({ ...payload, service_tier: 'priority' }, model, ctx)).toEqual(payload);
+ expect(JSON.parse(query({ provider: 'relay', model: 'private-sol', fast: true })).fast).toBe(false);
+ await handle.setFastMode!(true);
+ expect(await adapt(payload, model, ctx)).toEqual({ ...payload, service_tier: 'priority' });
+ expect(JSON.parse(query({ provider: 'other', model: 'private-sol' })).fast).toBe(false);
+ expect(JSON.parse(query({ provider: 'relay', model: 'no-fast' })).fast).toBe(false);
await handle.setFastMode!(false);
- expect(JSON.parse(readFileSync(file, 'utf8')).fast).toBe(false);
- fast = false;
+ expect(await adapt({ ...payload, service_tier: 'priority' }, model, ctx)).toEqual(payload);
+ fast = true;
await handle.send({ type: 'user', content: 'hello' });
- expect(JSON.parse(readFileSync(file, 'utf8')).fast).toBe(false);
+ expect(await adapt(payload, model, ctx)).toEqual({ ...payload, service_tier: 'priority' });
expect(captured.requests.some(request => request.type === 'set_fast_mode')).toBe(false);
await handle.close();
- await vi.waitFor(() => expect(readdirSync(path.join(agentHome, 'runtime'))).not.toContain(path.basename(file)));
+ expect(JSON.parse(query({ provider: 'relay', model: 'private-sol' })).fast).toBe(false);
});
it("uses providerId as the primary key when duplicate model ids exist", async () => {
diff --git a/packages/maker-core/src/agents/pi/cindy-bridge-source.ts b/packages/maker-core/src/agents/pi/cindy-bridge-source.ts
index d10c978522..a2f73808ea 100644
--- a/packages/maker-core/src/agents/pi/cindy-bridge-source.ts
+++ b/packages/maker-core/src/agents/pi/cindy-bridge-source.ts
@@ -113,6 +113,7 @@ const SECRET_ENV_NAMES = new Set<string>([
'CINDY_PI_SECRET_ENV_NAMES',
'CINDY_PI_PERMISSION_FILE',
'CINDY_PI_MODEL_REQUEST_PREFS_FILE',
+ 'CINDY_PI_FAST_MODELS',
'CINDY_PI_TURN_TOOL_POLICY',
PI_PACKAGE_MANAGEMENT_ENV,
PI_BASH_PACKAGE_HOME_ENV,
@@ -3638,19 +3639,26 @@ function astraResponsesPayload(payload, model) {
return out;
}
-function nativeFastPayload(payload, model) {
+async function nativeFastPayload(payload, model, ctx) {
if (!payload || typeof payload !== 'object' || Array.isArray(payload)) return undefined;
if (!model || !['openai-responses', 'azure-openai-responses', 'openai-completions'].includes(model.api)) return undefined;
+ let models;
+ try { models = JSON.parse(process.env.CINDY_PI_FAST_MODELS || '[]'); } catch { return undefined; }
+ if (!Array.isArray(models) || !models.some(item => item.provider === model.provider && item.id === model.id)) return undefined;
+ const out = { ...payload };
+ delete out.service_tier;
+ let timer;
try {
- const file = process.env.CINDY_PI_MODEL_REQUEST_PREFS_FILE;
- if (!file) return undefined;
- const prefs = JSON.parse(readFileSync(file, 'utf8'));
- if (!Array.isArray(prefs.models) || !prefs.models.some(item => item.provider === model.provider && item.id === model.id)) return undefined;
- const out = { ...payload };
- if (prefs.fast === true) out.service_tier = 'priority';
- else delete out.service_tier;
- return out;
- } catch { return undefined; }
+ // No UI is shown: Cindy answers this internal query from current host memory.
+ // Missing/closed hosts and malformed replies must not retain a premium tier.
+ const response = await Promise.race([
+ ctx.ui.input('cindy:request-preferences', JSON.stringify({ provider: model.provider, model: model.id })),
+ new Promise(resolve => { timer = setTimeout(() => resolve(undefined), 5000); }),
+ ]);
+ if (typeof response === 'string' && JSON.parse(response)?.fast === true) out.service_tier = 'priority';
+ } catch { /* A failed preference read falls back to the standard tier. */ }
+ finally { if (timer) clearTimeout(timer); }
+ return out;
}
${PI_NATIVE_PROVIDER_ADAPTER_SOURCE}
@@ -3659,9 +3667,9 @@ export default async function cindyBridge(pi: any) {
installTextOnlyTurnPolicy(pi);
await registerCindyNativeProviderAdapters(pi);
if (!currentPermissionState().reviewOnly) registerCindyQuestionTool(pi);
- pi.on('before_provider_request', (event, ctx) => {
+ pi.on('before_provider_request', async (event, ctx) => {
const payload = astraResponsesPayload(event.payload, ctx.model) ?? event.payload;
- return nativeFastPayload(payload, ctx.model) ?? payload;
+ return (await nativeFastPayload(payload, ctx.model, ctx)) ?? payload;
});
const mcpGateway = new CindyMcpGateway();
// bash 隔离 home 经 resolveBashPackageHome 解析(首次加载读删 + 防篡改 stash,
diff --git a/packages/maker-core/src/agents/pi/cindy-subagent-runner-source.ts b/packages/maker-core/src/agents/pi/cindy-subagent-runner-source.ts
index c3c9c0233d..6c54c5262d 100644
--- a/packages/maker-core/src/agents/pi/cindy-subagent-runner-source.ts
+++ b/packages/maker-core/src/agents/pi/cindy-subagent-runner-source.ts
@@ -772,7 +772,6 @@ function main() {
if (task.thinking) args.push('--thinking', task.thinking);
const childEnv = Object.assign({}, process.env, {
CINDY_PI_PERMISSION_FILE: config.permissionFile,
- ...(config.requestPrefsFile ? { CINDY_PI_MODEL_REQUEST_PREFS_FILE: config.requestPrefsFile } : {}),
PI_CODING_AGENT_DIR: config.childConfigHome,
});
const childRg = path.join(config.childConfigHome, 'bin', process.platform === 'win32' ? 'rg.exe' : 'rg');
diff --git a/packages/maker-core/src/agents/pi/cindy-subagent-source.ts b/packages/maker-core/src/agents/pi/cindy-subagent-source.ts
index 2790a59621..f5bececf24 100644
--- a/packages/maker-core/src/agents/pi/cindy-subagent-source.ts
+++ b/packages/maker-core/src/agents/pi/cindy-subagent-source.ts
@@ -794,7 +794,6 @@ async function launchDurableRun(binary, tasks, runtime, taskId, mode, context, d
const runnerFile = process.env[RUNNER_FILE_ENV];
const configHome = process.env[CONFIG_HOME_ENV];
const sourcePermissionFile = process.env[PERMISSION_FILE_ENV];
- const sourceRequestPrefsFile = process.env.CINDY_PI_MODEL_REQUEST_PREFS_FILE;
const parentSessionId = process.env[PARENT_SESSION_ENV];
const runtimeOwnerId = process.env[OWNER_ID_ENV];
if (!runRoot || !runnerFile || !configHome || !sourcePermissionFile || !parentSessionId || !runtimeOwnerId) {
@@ -822,7 +821,6 @@ async function launchDurableRun(binary, tasks, runtime, taskId, mode, context, d
childConfigHome: childConfigHome,
bridgeExtension: bridgeExtension,
permissionFile: permissionFile,
- requestPrefsFile: sourceRequestPrefsFile ? join(runDir, 'request-prefs.json') : undefined,
depth: readDepth() + 1,
mode: tasks.length > 1 ? 'parallel' : 'single',
context: context,
@@ -912,11 +910,6 @@ async function launchDurableRun(binary, tasks, runtime, taskId, mode, context, d
throw new Error('subagent: the parent task was deleted; this launch will not start.');
}
try { chmodSync(runDir, 0o700); } catch (err) { /* best effort on Windows */ }
- if (config.requestPrefsFile) {
- // Durable children must not depend on the parent's ephemeral control files.
- copyFileSync(sourceRequestPrefsFile, config.requestPrefsFile);
- try { chmodSync(config.requestPrefsFile, 0o600); } catch (err) { /* best effort on Windows */ }
- }
copyFileSync(sourcePermissionFile, permissionFile);
try { chmodSync(permissionFile, 0o600); } catch (err) { /* best effort on Windows */ }
mkdirSync(childConfigHome, { recursive: true, mode: 0o700 });
diff --git a/packages/maker-core/src/agents/pi/index.ts b/packages/maker-core/src/agents/pi/index.ts
index fee0db5245..8fb669c927 100644
--- a/packages/maker-core/src/agents/pi/index.ts
+++ b/packages/maker-core/src/agents/pi/index.ts
@@ -3230,31 +3230,16 @@ export class PiAgent extends BaseAgent {
runtimeDir,
`perm-${sid ?? `anon-${process.pid}-${Date.now()}`}-${runtimeInstanceId}${remote ? `-${permissionSnapshotHash}` : ''}.json`,
);
- const requestPrefsFile = joinRemotePosixPath(runtimeDir, `request-prefs-${runtimeInstanceId}.json`);
const fastModels = nativeProviders.flatMap(provider => provider.models
.filter(model => model.supportsFastMode === true)
.map(model => ({ provider: provider.id, id: model.wireId ?? model.id })));
- let requestPrefsWriteChain: Promise<void> = Promise.resolve();
- let requestPrefsSnapshot: string | undefined;
+ // Fast is host-owned state. The bridge reads it over RPC; no writable file
+ // (or replayable snapshot) can authorize a paid request tier.
let requestPrefsClosed = false;
let nativeFastEnabled = opts.getPriceVariant?.() === 'priority';
- const writeRequestPrefs = (fast = opts.getPriceVariant ? opts.getPriceVariant() === 'priority' : nativeFastEnabled): Promise<void> => {
- if (requestPrefsClosed) return Promise.reject(new Error('Pi request preferences are closed'));
- if (fastModels.length === 0) return Promise.resolve();
- const snapshot = JSON.stringify({ fast, models: fastModels });
- if (requestPrefsSnapshot === snapshot) return requestPrefsWriteChain;
- requestPrefsSnapshot = snapshot;
- const write = requestPrefsWriteChain.catch(() => {}).then(async () => {
- try {
- await writeFile(requestPrefsFile, snapshot);
- nativeFastEnabled = fast;
- } catch (error) {
- if (requestPrefsSnapshot === snapshot) requestPrefsSnapshot = undefined;
- throw error;
- }
- });
- requestPrefsWriteChain = write;
- return write;
+ const updateRequestPrefs = (fast = opts.getPriceVariant ? opts.getPriceVariant() === 'priority' : nativeFastEnabled): void => {
+ if (requestPrefsClosed) throw new Error('Pi request preferences are closed');
+ nativeFastEnabled = fast;
};
let activeToolsDisabled = false;
let textOnlyPolicyReady = false;
@@ -3282,7 +3267,6 @@ export class PiAgent extends BaseAgent {
requestPrefsClosed = true;
void rmPath(permissionFile);
void rmPath(subagentRuntimeFile);
- if (fastModels.length > 0) void requestPrefsWriteChain.finally(() => rmPath(requestPrefsFile)).catch(() => {});
};
// 权限档写入串行化 + 代际跳过。并发/连续切档(本地与远程控制端同时切,或用户快速连点)时,
// 无串行的 fs.writeFile 可能让较早的 Full-access 写在较新的 Ask 写之后落盘 —— bridge 每次
@@ -3388,7 +3372,7 @@ export class PiAgent extends BaseAgent {
return run;
};
await writePermissionFile(requestedPermissionSnapshot);
- await writeRequestPrefs();
+ updateRequestPrefs();
// 子代理运行期快照的写入:代际串行,最新意图胜出(理由同权限档 —— 并发/连续 setModel
// 时无串行的 writeFile 可能让较早的模型写在较新的之后落盘,子代理就会读到过期模型)。
@@ -5183,7 +5167,7 @@ export class PiAgent extends BaseAgent {
PI_CODING_AGENT_DIR: configHome,
[PI_BASH_PACKAGE_HOME_ENV]: bashPackageHome,
CINDY_PI_PERMISSION_FILE: permissionFile,
- ...(fastModels.length > 0 ? { CINDY_PI_MODEL_REQUEST_PREFS_FILE: requestPrefsFile } : {}),
+ ...(fastModels.length > 0 ? { CINDY_PI_FAST_MODELS: JSON.stringify(fastModels) } : {}),
CINDY_PI_TURN_TOOL_POLICY: remote ? '' : runtimeInstanceId,
...(allowPiPackageManagement ? { [PI_PACKAGE_MANAGEMENT_ENV]: piPackageManagementToken } : {}),
// 轮 40-w4-t12 HIGH-1:review-only 启动标记 —— 独立于权限文件(文件损坏/
@@ -5280,6 +5264,8 @@ export class PiAgent extends BaseAgent {
resolver: interactionResolver,
permissionMode,
currentModelIds: [mutableModel, mutableWireModel],
+ readNativeFast: (provider, model) => !requestPrefsClosed && nativeFastEnabled &&
+ fastModels.some(candidate => candidate.provider === provider && candidate.id === model),
workspaceRoots: [opts.workingDir],
readRoots: [opts.workingDir, ...mutableExtraDirs, ...mutableWritableDirs],
writableRoots: [opts.workingDir, ...mutableWritableDirs],
@@ -6667,7 +6653,7 @@ export class PiAgent extends BaseAgent {
async send(message: UserMessage, sendOpts?: SendOptions): Promise<void> {
rejectIfCancelled(sendOpts, 'send');
await waitForSessionRpcIdle();
- await writeRequestPrefs();
+ updateRequestPrefs();
rejectIfCancelled(sendOpts, 'send');
if (sendOpts) handle.validateSendOptions?.(sendOpts);
// 本轮策略覆盖:无策略显式清 null,不继承上一轮渠道策略(§7.2.5);内部续跑
@@ -7215,7 +7201,7 @@ export class PiAgent extends BaseAgent {
async setFastMode(enabled: boolean): Promise<void> {
if (reviewMode) return;
- await writeRequestPrefs(enabled);
+ updateRequestPrefs(enabled);
},
async setEffort(effort: Effort): Promise<void> {
@@ -7844,6 +7830,7 @@ export class PiAgent extends BaseAgent {
resolver: InteractionResolver | null;
permissionMode: 'ask' | 'auto' | 'bypassPermissions';
currentModelIds: readonly string[];
+ readNativeFast?: (provider: string, model: string) => boolean;
workspaceRoots: string[];
readRoots: string[];
writableRoots: string[];
@@ -7913,6 +7900,22 @@ export class PiAgent extends BaseAgent {
return;
}
+ // Internal, read-only query over the existing stdio/SSH RPC response channel.
+ // Caller input selects an exact model; it can never supply or update Fast intent.
+ if (method === 'input' && event.title === 'cindy:request-preferences') {
+ const context = getPermissionCtx();
+ let fast = false;
+ try {
+ const requested = JSON.parse(typeof event.placeholder === 'string' ? event.placeholder : '');
+ if (!context.isPermissionContextClosed() && !context.isAccountBoundaryTornDown() &&
+ typeof requested?.provider === 'string' && typeof requested?.model === 'string') {
+ fast = context.readNativeFast?.(requested.provider, requested.model) === true;
+ }
+ } catch { /* Invalid queries never enable a paid tier. */ }
+ proc.send({ type: 'extension_ui_response', id, value: JSON.stringify({ fast }) });
+ return;
+ }
+
if (method === 'input' && event.title === CINDY_SUBAGENT_RUNNER_CONTROL_TITLE) {
const context = getPermissionCtx();
if (context.remote) {
diff --git a/packages/maker-core/src/agents/pi/pi-subagent-runs.ts b/packages/maker-core/src/agents/pi/pi-subagent-runs.ts
index b2f4d6ab08..d19fb92896 100644
--- a/packages/maker-core/src/agents/pi/pi-subagent-runs.ts
+++ b/packages/maker-core/src/agents/pi/pi-subagent-runs.ts
@@ -2800,16 +2800,6 @@ async function resumeClaimedPiSubagentRun(
);
await fs.writeFile(bridgeExtension, bridgeSource, { mode: 0o600, flag: 'wx' });
await writeAtomicJson(permissionFile, launch.permissionSnapshot);
- // Resume requires a live parent. Missing preferences mean no current Fast
- // capability, not permission to resurrect the prior run's capability snapshot.
- const liveRequestPrefsFile = launch.env.CINDY_PI_MODEL_REQUEST_PREFS_FILE;
- delete config.requestPrefsFile;
- if (liveRequestPrefsFile) {
- const requestPrefs = await fs.readFile(liveRequestPrefsFile);
- const requestPrefsFile = path.join(runDir, 'request-prefs.json');
- await fs.writeFile(requestPrefsFile, requestPrefs, { mode: 0o600, flag: 'wx' });
- config.requestPrefsFile = requestPrefsFile;
- }
await fs.writeFile(runnerFile, runnerSource, { mode: 0o600, flag: 'wx' });
await Promise.all([
fs.chmod(runDir, 0o700).catch(() => undefined),
diff --git a/packages/model-providers/src/__tests__/kimiCodeDefaults.test.ts b/packages/model-providers/src/__tests__/kimiCodeDefaults.test.ts
index 808b3f1d5e..251971493d 100644
--- a/packages/model-providers/src/__tests__/kimiCodeDefaults.test.ts
+++ b/packages/model-providers/src/__tests__/kimiCodeDefaults.test.ts
@@ -59,7 +59,7 @@ describe("Kimi K2.8 connection defaults", () => {
defaultEffort: "max",
});
expect(model.cost, agent).toBeUndefined();
- expect(model.referencePrices, agent).toBeUndefined();
+ expect(model, agent).not.toHaveProperty("referencePrices");
}
const native = providerCatalogForPi().providers["kimi-coding"].find(
(m) => m.id === "kimi-for-coding",
diff --git a/packages/model-providers/src/__tests__/mimoV26Regression.test.ts b/packages/model-providers/src/__tests__/mimoV26Regression.test.ts
index 7913a3f049..8859056d50 100644
--- a/packages/model-providers/src/__tests__/mimoV26Regression.test.ts
+++ b/packages/model-providers/src/__tests__/mimoV26Regression.test.ts
@@ -30,6 +30,7 @@ const IMAGE_MODELS = new Set(["mimo-v2.6-pro", "mimo-v2.6-flash"]);
/** 裸 model id → Registry 公共条目(alias 唯一匹配)。 */
function baseModelFor(bareId: string) {
+ if (!registry.ok) throw new Error(registry.error);
const base = findBaseModel(registry.value, bareId);
expect(base, `Registry 缺少 MiMo 公共条目: ${bareId}`).toBeDefined();
return base!;
@@ -75,6 +76,7 @@ describe.each(MIMO_PRESET_IDS)("MiMo 预设连接投影: %s", (presetId) => {
expect(model.id.startsWith("mimo-v2.6"), model.id).toBe(true);
for (const model of models) {
if (!model.id.startsWith("mimo-v2.6")) continue;
+ if (!registry.ok) throw new Error(registry.error);
const resolved = resolveModelMetadata(
registry.value,
presetId,
diff --git a/packages/model-providers/src/__tests__/modelDiscoveryChannels.test.ts b/packages/model-providers/src/__tests__/modelDiscoveryChannels.test.ts
index ee96aae9b0..a7f309bb88 100644
--- a/packages/model-providers/src/__tests__/modelDiscoveryChannels.test.ts
+++ b/packages/model-providers/src/__tests__/modelDiscoveryChannels.test.ts
@@ -5,6 +5,30 @@ import { buildUserProvider } from '../user-provider.js';
import { BUNDLED_CATALOG } from '../catalog.js';
describe('shared provider discovery', () => {
+ it.each([
+ { contextWindow: 128000, contextWindowMax: 64000 },
+ { context_window: 128000, contextWindowMax: 64000 },
+ { contextWindow: 128000, max_context_window: 64000 },
+ { model_info: { max_input_tokens: 128000 }, contextWindowMax: 64000 },
+ { contextWindow: 128000, contextWindowMax: 64000, max_context_window: 32000 },
+ ])('ignores a maximum smaller than the normalized working window: %j', fields => {
+ const models = parseModelsListResponse({ data: [{ id: 'private-model', ...fields }] })!;
+ expect(models[0].discoveredMetadata?.contextWindow).toBe(128000);
+ expect(models[0].discoveredMetadata).not.toHaveProperty('contextWindowMax');
+ const provider = buildUserProvider({ id: 'relay', name: 'Relay', runtimes: {
+ 'claude-code': { baseUrl: 'https://relay.example/v1', models },
+ } });
+ expect(provider.models['claude-code']?.[0]).toMatchObject({ contextWindow: 128000 });
+ expect(provider.models['claude-code']?.[0].contextWindowMax).not.toBe(64000);
+ });
+
+ it.each(['contextWindowMax', 'max_context_window'])('retains a valid %s with or without a working window', key => {
+ for (const fields of [{}, { contextWindow: 128000 }]) {
+ const [model] = parseModelsListResponse({ data: [{ id: 'private-model', ...fields, [key]: 256000 }] })!;
+ expect(model.discoveredMetadata?.contextWindowMax).toBe(256000);
+ }
+ });
+
it('retains the working window when a refresh only reports maximum capacity', () => {
const original = parseModelsListResponse({ data: [{ id: 'private-model',
context_window: 272000, max_context_window: 1050000 }] })!;
@@ -22,40 +46,49 @@ describe('shared provider discovery', () => {
}
});
- it('shrinks a saved discovered working window when a max-only refresh reduces capacity', () => {
+ it.each([
+ [{ context_window: 272000 }, { max_context_window: 64000 }],
+ [{ context_window: 272000, max_context_window: 1050000 }, { contextWindowMax: 64000 }],
+ [{ context_window: 32000, max_context_window: 64000 }, { context_window: 272000 }],
+ ])('reconciles working and maximum windows after a sparse refresh: %j + %j', (before, refresh) => {
const original = mergeDiscoveredRuntimeModels([], parseModelsListResponse({ data: [
- { id: 'private-model', context_window: 128000, max_context_window: 256000 },
+ { id: 'private-model', ...before },
] })!);
- const refreshed = mergeDiscoveredRuntimeModels(original, parseModelsListResponse({ data: [
- { id: 'private-model', max_context_window: 64000 },
+ const models = mergeDiscoveredRuntimeModels(original, parseModelsListResponse({ data: [
+ { id: 'private-model', ...refresh },
] })!);
- expect(original[0].discoveredMetadata?.contextWindow).toBe(128000);
- expect(refreshed[0].discoveredMetadata).toMatchObject({ contextWindow: 64000, contextWindowMax: 64000 });
+ expect(original[0].discoveredMetadata?.contextWindow).toBe(before.context_window);
+ expect(models[0].discoveredMetadata?.contextWindow).toBe(272000);
+ expect(models[0].discoveredMetadata).not.toHaveProperty('contextWindowMax');
for (const agent of ['claude-code', 'codex', 'pi'] as const) {
const provider = buildUserProvider({ id: 'relay', name: 'Relay', runtimes: {
- [agent]: { baseUrl: 'https://relay.example/v1', models: refreshed },
+ [agent]: { baseUrl: 'https://relay.example/v1', models },
} });
- expect(provider.models[agent]?.[0]).toMatchObject({ contextWindow: 64000, contextWindowMax: 64000 });
+ expect(provider.models[agent]?.[0].contextWindow).toBe(272000);
+ expect(provider.models[agent]?.[0].contextWindowMax).toBeUndefined();
}
});
- it('uses max-only discovery as the first working window, ahead of inherited defaults', () => {
- const discovered = parseModelsListResponse({ data: [{ id: 'gpt-9-sol', max_context_window: 32000 }] })!;
- for (const models of [discovered, mergeDiscoveredRuntimeModels([], discovered)]) {
- for (const agent of ['claude-code', 'codex', 'pi'] as const) {
- const provider = buildUserProvider({ id: 'relay', name: 'Relay', runtimes: {
- [agent]: { baseUrl: 'https://relay.example/v1', wireProtocol: 'openai-responses', models },
- } }, { modelRegistry: BUNDLED_CATALOG.modelRegistry });
- expect(provider.models[agent]?.[0]).toMatchObject({
- contextWindow: 32000, contextWindowMax: 32000, contextWindowVerified: true,
- });
- }
+ it('uses maximum-only discovery as an unverified working fallback without saving it as a report', () => {
+ const discovered = parseModelsListResponse({ data: [{ id: 'private-model', max_context_window: 64000 }] })!;
+ const models = mergeDiscoveredRuntimeModels([], discovered);
+ expect(models[0].discoveredMetadata?.contextWindow).toBeUndefined();
+ for (const agent of ['claude-code', 'codex', 'pi'] as const) {
+ const provider = buildUserProvider({ id: 'relay', name: 'Relay', runtimes: {
+ [agent]: { baseUrl: 'https://relay.example/v1', models },
+ } });
+ expect(provider.models[agent]?.[0]).toMatchObject({
+ contextWindow: 64000, contextWindowMax: 64000, contextWindowVerified: false,
+ });
+ expect(provider.models[agent]?.[0].userModelConfig?.discoveredMetadata?.contextWindow).toBeUndefined();
}
- const saved = mergeDiscoveredRuntimeModels([], discovered);
- const refreshed = mergeDiscoveredRuntimeModels(saved, parseModelsListResponse({ data: [
- { id: 'gpt-9-sol', max_context_window: 64000 },
+ const refreshed = mergeDiscoveredRuntimeModels(models, parseModelsListResponse({ data: [
+ { id: 'private-model', context_window: 32000 },
] })!);
- expect(refreshed[0].discoveredMetadata).toMatchObject({ contextWindow: 32000, contextWindowMax: 64000 });
+ const provider = buildUserProvider({ id: 'relay', name: 'Relay', runtimes: {
+ pi: { baseUrl: 'https://relay.example/v1', models: refreshed },
+ } });
+ expect(provider.models.pi?.[0]).toMatchObject({ contextWindow: 32000, contextWindowMax: 64000, contextWindowVerified: true });
});
it('imports Vercel token prices, output capacity, image inputs and declared effort levels', () => {
diff --git a/packages/model-providers/src/__tests__/modelGeneration.test.ts b/packages/model-providers/src/__tests__/modelGeneration.test.ts
index c4ee67f145..965c2c8357 100644
--- a/packages/model-providers/src/__tests__/modelGeneration.test.ts
+++ b/packages/model-providers/src/__tests__/modelGeneration.test.ts
@@ -3,7 +3,7 @@ import { previousModelGenerations } from '../modelGeneration.js';
import { buildUserProvider } from '../user-provider.js';
import { isChatEligible } from '../classification.js';
import { BUNDLED_CATALOG } from '../builtin.js';
-import { providerModelGenerationRecord, providerModelRecord } from '../providerModelCatalog.js';
+import { providerModelGenerationRecord, providerModelRecord, providerModelAdapterId } from '../providerModelCatalog.js';
import { parseModelsListResponse } from '../modelDiscovery.js';
import { mergeDiscoveredRuntimeModels } from '../modelMetadataLayers.js';
import type { ProviderRuntimeModelConfig, ProviderWireProtocol } from '../types.js';
@@ -34,6 +34,20 @@ describe('new model generation defaults', () => {
}
});
+ it('uses a target maximum-only report before inherited working windows without verifying it', () => {
+ const provider = build([
+ { id: 'private-6-sol', name: 'Old', discoveredMetadata: { contextWindow: 272000 } },
+ ...discovered('private-7-sol', { max_context_window: 64000 }),
+ ...discovered('gpt-9-sol', { max_context_window: 32000 }),
+ ]);
+ for (const agent of ['codex', 'pi', 'claude-code'] as const) {
+ expect(provider.models[agent]![1]).toMatchObject({ contextWindow: 64000,
+ contextWindowMax: 64000, contextWindowVerified: false });
+ expect(provider.models[agent]![2]).toMatchObject({ contextWindow: 32000,
+ contextWindowMax: 32000, contextWindowVerified: false });
+ }
+ });
+
it('takes the closest same-connection generation, preserves explicit disables, and never changes routing', () => {
const provider = build([
{ id: 'private-6-sol', name: 'Six', discoveredMetadata: { efforts: ['low', 'high'], defaultEffort: 'high', contextWindow: 64000, supportsFastMode: true } },
@@ -138,6 +152,24 @@ describe('new model generation defaults', () => {
expect(previousModelGenerations('qwen4-30b', ['qwen3-235b', 'qwen3-30b'], id => id)).toEqual(['qwen3-30b']);
});
+ it.each([
+ ['private-7b', 'private-70b'],
+ ['gpt-oss-20b', 'gpt-oss-120b'],
+ ['private-0.5b', 'private-1.5b'],
+ ['private-350m', 'private-700m'],
+ ['mixtral-8x7b', 'mixtral-8x22b'],
+ ])('does not treat parameter sizes %s and %s as generations', (small, large) => {
+ expect(previousModelGenerations(large, [small], id => id)).toEqual([]);
+ const target = build([
+ { id: small, name: small, discoveredMetadata: { contextWindow: 64000,
+ efforts: ['high'], supportsImageInput: true, supportsFastMode: true } },
+ { id: large, name: large },
+ ]).models.pi![1]!;
+ expect(target.efforts).toEqual([]);
+ expect(target.supportsFastMode).not.toBe(true);
+ expect(target.supportsImageInput).not.toBe(true);
+ });
+
it('normalizes long trailing slashes without treating internal slashes as the same endpoint', () => {
const endpoint = 'https://api.openai.com/v1';
const slashes = '/'.repeat(100_000);
@@ -216,3 +248,29 @@ describe('new model generation defaults', () => {
expect(providerModelGenerationRecord('gpt-9-sol', 'https://relay.example/v1', 'anthropic-messages')).toBeUndefined();
});
});
+
+
+it('preserves canonical API metadata, including explicit negative capabilities', () => {
+ const metadata = { nativeApi: 'openai-responses', contextWindow: 1000000, maxOutputTokens: 32000,
+ supportsFastMode: false, supportsToolCalls: false, supportsImageInput: false,
+ efforts: ['low', 'high'], defaultEffort: 'low', reasoningRequired: true };
+ const result = parseModelsListResponse({ data: [{ id: 'gpt-7-sol', ...metadata }] });
+ expect(result?.[0]?.discoveredMetadata).toMatchObject(metadata);
+ for (const agent of ['claude-code', 'codex', 'pi'] as const) {
+ const provider = build(mergeDiscoveredRuntimeModels([], result!));
+ expect(provider.models[agent]?.[0]).toMatchObject({ maxOutput: 32000,
+ supportsFastMode: false, supportsToolCalls: false, supportsImageInput: false,
+ efforts: ['low', 'high'] });
+ }
+});
+
+
+it('resolves inherited adapters only within a matching connection or explicit preset', () => {
+ const inherited = providerModelGenerationRecord('claude-opus-9', 'https://account.example/v1', 'anthropic-messages', 'cloudflare-ai-gateway')!;
+ expect(inherited).toBeDefined();
+ expect(providerModelAdapterId(inherited, 'cloudflare-ai-gateway')).toBe('cloudflare-ai-gateway');
+ expect(providerModelAdapterId(inherited)).toBeUndefined();
+ expect(providerModelAdapterId({ ...inherited, execution: { pi: { api: 'google-generative-ai' } } }, 'cloudflare-ai-gateway')).toBeUndefined();
+ expect(inherited.execution.pi.headers).toBeUndefined();
+ expect(inherited.cost).toBeUndefined();
+});
diff --git a/packages/model-providers/src/__tests__/modelMetadataLayers.test.ts b/packages/model-providers/src/__tests__/modelMetadataLayers.test.ts
index 206e3c63ab..58ea71c195 100644
--- a/packages/model-providers/src/__tests__/modelMetadataLayers.test.ts
+++ b/packages/model-providers/src/__tests__/modelMetadataLayers.test.ts
@@ -2,7 +2,11 @@ import { describe, expect, it } from "vitest";
import {
resolveModelMetadata,
expandedRegistryEntries,
+ applyModelMetadata,
+ catalogModelMetadata,
+ pickModelMetadata,
} from "../modelMetadataLayers.js";
+import type { CatalogModel } from '../types.js';
import { parseModelRegistry } from "../modelAccessValidator.js";
import type { ModelRegistry } from "../modelAccessBean.js";
@@ -42,6 +46,43 @@ const registry: ModelRegistry = {
],
};
+it('keeps inherited capacity unverified through resolution and repeated catalog projection', () => {
+ const r: ModelRegistry = { schemaVersion: 4, updatedAt: registry.updatedAt, models: [],
+ baseModels: [{ id: 'private-6', aliases: [], defaults: { contextWindow: 64000 } }] };
+ const model: CatalogModel = { id: 'private-7', name: 'New', contextWindow: 200000,
+ contextWindowVerified: true, efforts: [], defaultEffort: null };
+ const metadata = resolveModelMetadata(r, 'supplier', model.id);
+ const inherited = applyModelMetadata(model, metadata);
+ expect(inherited).toMatchObject({ contextWindow: 64000, contextWindowVerified: false });
+ expect(JSON.parse(JSON.stringify(metadata))).toEqual(pickModelMetadata(metadata));
+ const live = catalogModelMetadata(inherited);
+ expect(applyModelMetadata(inherited, resolveModelMetadata(r, 'supplier', model.id, live)))
+ .toMatchObject({ contextWindow: 64000, contextWindowVerified: false });
+ for (const [reported, user] of [[{ contextWindow: 64000 }, undefined], [live, { contextWindow: 64000 }]] as const) {
+ expect(applyModelMetadata(inherited, resolveModelMetadata(r, 'supplier', model.id, reported, user)))
+ .toMatchObject({ contextWindow: 64000, contextWindowVerified: true });
+ }
+ r.models = [{ id: 'new-model', name: 'New', routes: [{ providerId: 'supplier', modelId: model.id,
+ agents: ['codex'], forceOverrides: { contextWindow: 32000 }, overrideReason: 'Verified capacity' }] }];
+ expect(applyModelMetadata(inherited, resolveModelMetadata(r, 'supplier', model.id, live)))
+ .toMatchObject({ contextWindow: 32000, contextWindowVerified: true });
+});
+
+it('drops an inherited maximum below the current model working window through repeated projection', () => {
+ const r: ModelRegistry = { schemaVersion: 4, updatedAt: registry.updatedAt, models: [],
+ baseModels: [{ id: 'private-6', aliases: [], defaults: { contextWindow: 128000, contextWindowMax: 128000 } }] };
+ const base: CatalogModel = { id: 'private-7', name: 'New', contextWindow: 200000, contextWindowMax: 128000,
+ efforts: [], defaultEffort: null };
+ const resolved = resolveModelMetadata(r, 'supplier', base.id, { contextWindow: 256000 });
+ expect(resolved.contextWindow).toBe(256000);
+ expect(resolved.contextWindowMax).toBeUndefined();
+ const projected = applyModelMetadata(base, resolved);
+ expect(projected).toMatchObject({ contextWindow: 256000, contextWindowVerified: true });
+ const again = applyModelMetadata(projected, resolveModelMetadata(r, 'supplier', base.id, catalogModelMetadata(projected)));
+ expect(again.contextWindowMax).toBeUndefined();
+ expect(again.contextWindow).toBe(256000);
+});
+
it("honors entry names below route, live, force and user names", () => {
const r = structuredClone(registry);
r.models[0].name = "Entry";
diff --git a/packages/model-providers/src/__tests__/modelNativeApi.test.ts b/packages/model-providers/src/__tests__/modelNativeApi.test.ts
index cca17d6f6d..e9db50a4db 100644
--- a/packages/model-providers/src/__tests__/modelNativeApi.test.ts
+++ b/packages/model-providers/src/__tests__/modelNativeApi.test.ts
@@ -45,7 +45,7 @@ describe('native declarations for verified catalog identities', () => {
});
it('does not let a retired sibling suppress the active catalog identity', () => {
const r = registry();
- r.baseModels = [{ id: 'gemini-99', name: 'Gemini 99', aliases: [] }];
+ r.baseModels = [{ id: 'gemini-99', defaults: { name: 'Gemini 99' }, aliases: [] }];
r.models = [
{ id: 'google/gemini-99', name: 'Gemini', modelRef: 'gemini-99', nativeApi: 'google-generative-ai', routes: [] },
{ id: 'google/gemini-99-preview', name: 'Gemini preview', modelRef: 'gemini-99', status: 'retired', nativeApi: 'google-generative-ai', routes: [] },
diff --git a/packages/model-providers/src/__tests__/piCatalogSync.test.ts b/packages/model-providers/src/__tests__/piCatalogSync.test.ts
index 60dae2c499..e85f5ffaa3 100644
--- a/packages/model-providers/src/__tests__/piCatalogSync.test.ts
+++ b/packages/model-providers/src/__tests__/piCatalogSync.test.ts
@@ -18,9 +18,10 @@ const piCatalog = providerCatalogForPi();
const root = fileURLToPath(new URL("../../../../", import.meta.url));
describe("Pi catalog sync upstream precedence", () => {
- it.each(["input", "online"] as const)(
- "reads native public API Astra metadata through the %s sync path",
- (source) => {
+ it.each((["input", "online"] as const).flatMap(source =>
+ (["complete", "failed", "omitted"] as const).map(state => ({ source, state }))))(
+ "preserves catalog metadata through the $source sync path ($state)",
+ ({ source, state }) => {
const temporary = mkdtempSync(
path.join(tmpdir(), "cindy-pi-catalog-sync #-"),
);
@@ -59,7 +60,21 @@ describe("Pi catalog sync upstream precedence", () => {
compat: { supportsStore: false },
upstreamField: "preserve-native-metadata",
};
- const input: Record<string, unknown[]> = { ...piCatalog.providers, openai: [native] };
+ const input: Record<string, unknown> = { ...piCatalog.providers, openai: [native],
+ 'new-provider': [{ ...native, id: 'future-model', provider: 'new-provider' }],
+ 'partial-provider': [null, { ...native, id: 'partial-new', provider: 'partial-provider' }],
+ };
+ const snapshotPath = path.join(catalogDirectory, 'provider-models.json');
+ const oldSnapshot = JSON.parse(readFileSync(snapshotPath, 'utf8'));
+ const oldRow = oldSnapshot.providers.openai[0];
+ oldSnapshot.providers['new-provider'] = [{ ...oldRow, id: 'retired-model' }];
+ oldSnapshot.providers['partial-provider'] = [{ ...oldRow, id: 'partial-old' }];
+ oldSnapshot.providers['offline-provider'] = [{ ...oldRow, id: 'offline-old' }];
+ writeFileSync(snapshotPath, JSON.stringify(oldSnapshot));
+ if (state !== "complete") for (const provider of ["xai", "openai", "openai-codex"]) {
+ if (state === "failed") input[provider] = null;
+ else delete input[provider];
+ }
const inputPath = path.join(temporary, "input.json");
writeFileSync(inputPath, JSON.stringify(input));
const args: string[] = [];
@@ -72,7 +87,10 @@ describe("Pi catalog sync upstream precedence", () => {
const providers = JSON.parse(readFileSync(new URL('./input.json', import.meta.url), 'utf8'));
globalThis.fetch = async (url) => {
const provider = decodeURIComponent(new URL(url).pathname.split('/').at(-1));
+ if (provider === 'providers') return new Response(JSON.stringify([...Object.keys(providers), 'offline-provider']));
+ if (provider === 'offline-provider') throw new Error('Offline source');
if (!(provider in providers)) throw new Error('Unexpected provider: ' + provider);
+ if (providers[provider] === null) return new Response("Unavailable", { status: 503 });
return new Response(JSON.stringify(providers[provider]));
};
`,
@@ -92,7 +110,14 @@ describe("Pi catalog sync upstream precedence", () => {
"utf8",
),
);
- expect(result.providers.openai).toMatchObject([{ id: native.id, name: "Native Astra", contextWindow: 1_060_000, execution: { pi: { compat: { supportsStore: false } } } }]);
+ expect(result.providers['new-provider'][0]).toMatchObject({ id: 'future-model', contextWindow: 1_060_000 });
+ expect(result.providers['new-provider']).toHaveLength(1);
+ expect(result.providers['partial-provider'].map((row: { id: string }) => row.id)).toEqual(['partial-new', 'partial-old']);
+ expect(result.providers['offline-provider'].map((row: { id: string }) => row.id)).toEqual(['offline-old']);
+ if (state === "complete") expect(result.providers.openai.find((row: { id: string }) => row.id === native.id)).toMatchObject({ id: native.id, name: "Native Astra", contextWindow: 1_060_000, execution: { pi: { compat: { supportsStore: false } } } });
+ if (state !== "complete") for (const provider of ["xai", "openai", "openai-codex"]) {
+ expect(result.providers[provider]).toEqual(expect.arrayContaining(oldSnapshot.providers[provider]));
+ }
expect(result.providers["openai-codex"].map((m: { id: string }) => m.id).sort()).toEqual(piCatalog.providers["openai-codex"]!.map(m => m.id).sort());
} finally {
rmSync(temporary, { recursive: true, force: true });
diff --git a/packages/model-providers/src/__tests__/piNativeCatalog.test.ts b/packages/model-providers/src/__tests__/piNativeCatalog.test.ts
new file mode 100644
index 0000000000..5c2cab0327
--- /dev/null
+++ b/packages/model-providers/src/__tests__/piNativeCatalog.test.ts
@@ -0,0 +1,38 @@
+import { afterEach, expect, it, vi } from 'vitest';
+import generated from '../../catalog/provider-models.json';
+import { toCindyCatalog } from '../../../../tools/pi/catalog-format.mjs';
+
+afterEach(() => {
+ vi.doUnmock('../../catalog/provider-models.json');
+ vi.resetModules();
+});
+
+it.each([
+ { provider: 'xai', builtin: 'xai', prefix: '', previous: 'grok-4.8', future: 'grok-4.9', api: 'openai-completions' },
+ { provider: 'openai-codex', builtin: 'openai', prefix: 'chatgpt/', previous: 'gpt-6-sol', future: 'gpt-7-sol', api: 'openai-codex-responses' },
+])('initializes the offline catalog with sparse $provider models', async fixture => {
+ const base = { provider: fixture.provider, api: fixture.api, baseUrl: 'https://source.example/v1' };
+ const converted = toCindyCatalog({ [fixture.provider]: [
+ { ...base, id: fixture.previous, contextWindow: 345000 },
+ // A parallel route cannot supply this connection's fallback window.
+ { ...base, id: fixture.previous.replace(/6|4\.8/, fixture.provider === 'xai' ? '4.8.1' : '6.9'),
+ baseUrl: 'https://other.example/v1', contextWindow: 9000 },
+ { ...base, id: fixture.future },
+ { ...base, id: 'private-model' },
+ ] }, 'fixture');
+ vi.resetModules();
+ vi.doMock('../../catalog/provider-models.json', () => ({ default: {
+ ...generated, providers: { ...generated.providers, ...converted.providers },
+ } }));
+ const { BUILTIN_PROVIDERS } = await import('../builtin.js');
+ const models = BUILTIN_PROVIDERS.find(provider => provider.id === fixture.builtin)!.models.pi!;
+ const future = models.find(model => model.id === fixture.prefix + fixture.future)!;
+ expect(future).toMatchObject({ contextWindow: 345000, contextWindowVerified: false });
+ expect(future.discoveredMetadata).not.toHaveProperty('contextWindow');
+ const unknown = models.find(model => model.id === fixture.prefix + 'private-model')!;
+ expect(unknown.contextWindow).toBeGreaterThan(0);
+ expect(unknown.contextWindowVerified).toBe(false);
+ expect(unknown.discoveredMetadata).not.toHaveProperty('contextWindow');
+ expect(models.find(model => model.id === fixture.prefix + fixture.previous))
+ .toMatchObject({ contextWindow: 345000, contextWindowVerified: true });
+});
diff --git a/packages/model-providers/src/__tests__/providerCatalogImport.test.ts b/packages/model-providers/src/__tests__/providerCatalogImport.test.ts
index 5e582fcc65..ade7155854 100644
--- a/packages/model-providers/src/__tests__/providerCatalogImport.test.ts
+++ b/packages/model-providers/src/__tests__/providerCatalogImport.test.ts
@@ -79,3 +79,85 @@ describe("Pi source import", () => {
expect(row.execution.pi.headers).toEqual({ "NVCF-POLL-SECONDS": "3600" });
});
});
+
+
+it('retains partial upstream successes and previous rows without manufacturing negative capabilities', () => {
+ const previous = toCindyCatalog({ 'new-vendor': [model] }, 'before');
+ const errors: unknown[] = [];
+ const result = toCindyCatalog({ 'new-vendor': [
+ { ...model, id: 'next-model', reasoning: undefined, input: undefined, contextWindow: undefined,
+ supportsFastMode: false, supportsToolCalls: true },
+ { ...model, api: '' },
+ ] }, 'after', { previous, onError: (error: unknown) => errors.push(error) });
+ expect(errors).toHaveLength(1);
+ expect(result.providers['new-vendor']).toContainEqual(previous.providers['new-vendor'][0]);
+ const next = result.providers['new-vendor'].find(row => row.id === 'next-model');
+ expect(next).toMatchObject({ supportsFastMode: false, supportsToolCalls: true });
+ for (const key of ['reasoning', 'efforts', 'defaultEffort', 'contextWindow', 'supportsImageInput'])
+ expect(next).not.toHaveProperty(key);
+});
+
+
+it('preserves an explicit unspecified default effort', () => {
+ expect(toCindyCatalog({ 'new-vendor': [{ ...model, defaultEffort: null }] }, 'now')
+ .providers['new-vendor'][0]).toMatchObject({ efforts: ['high'], defaultEffort: null });
+});
+
+it('removes retired models only after a complete successful provider snapshot', () => {
+ const previous = toCindyCatalog({ 'new-vendor': [model, { ...model, id: 'retired-model' }] }, 'before');
+ const next = toCindyCatalog({ 'new-vendor': [model] }, 'after', { previous });
+ expect(next.providers['new-vendor'].map(row => row.id)).toEqual([model.id]);
+ expect(toCindyCatalog({ 'new-vendor': [] }, 'empty', { previous }).providers['new-vendor']).toEqual([]);
+ expect(toCindyCatalog({ 'new-vendor': [] }, 'incomplete', { previous, incompleteProviders: ['new-vendor'] })
+ .providers['new-vendor']).toEqual(previous.providers['new-vendor']);
+});
+
+it('retains missing adapter fields on the same connection and accepts explicit replacements', () => {
+ const previous = toCindyCatalog({ 'new-vendor': [{ ...model, samplingParams: { temperature: 0.2 } }] }, 'before');
+ const sparse = { ...model, thinkingLevelMap: undefined, compat: undefined, samplingParams: undefined };
+ const kept = toCindyCatalog({ 'new-vendor': [sparse] }, 'after', { previous }).providers['new-vendor'][0];
+ expect(kept.execution.pi).toEqual(previous.providers['new-vendor'][0].execution.pi);
+ expect(kept.efforts).toEqual(['high']);
+ const cleared = toCindyCatalog({ 'new-vendor': [{ ...sparse, thinkingLevelMap: {}, compat: {}, samplingParams: {} }] },
+ 'clear', { previous }).providers['new-vendor'][0];
+ expect(cleared.execution.pi).toMatchObject({ thinkingLevelMap: {}, compat: {}, samplingParams: {} });
+ for (const change of [{ api: 'anthropic-messages' }, { baseUrl: 'https://other.example/v1' }]) {
+ const moved = toCindyCatalog({ 'new-vendor': [{ ...sparse, ...change }] }, 'moved', { previous }).providers['new-vendor'][0];
+ expect(moved.execution.pi).not.toHaveProperty('thinkingLevelMap');
+ expect(moved.execution.pi).not.toHaveProperty('compat');
+ }
+});
+
+it.each([
+ { api: 'anthropic-messages', baseUrl: model.baseUrl },
+ { api: model.api, baseUrl: 'https://other.example/v1' },
+ { api: model.api, baseUrl: undefined },
+])('does not carry channel metadata across connection changes: %j', (connection) => {
+ const previous = toCindyCatalog({ 'new-vendor': [{
+ ...model,
+ nativeApi: 'openai-responses',
+ supportsFastMode: true,
+ supportsToolCalls: true,
+ reasoningRequired: true,
+ samplingParams: { temperature: 0.2 },
+ headers: { 'User-Agent': 'old-channel' },
+ }] }, 'before');
+ const sparse = { id: model.id, provider: model.provider, api: model.api, baseUrl: model.baseUrl };
+ const retained = toCindyCatalog({ 'new-vendor': [sparse] }, 'same', { previous }).providers['new-vendor'][0];
+ expect(retained).toEqual(previous.providers['new-vendor'][0]);
+
+ const moved = toCindyCatalog({ 'new-vendor': [{ ...sparse, ...connection }] }, 'moved', { previous })
+ .providers['new-vendor'];
+ // Keep the discovered model usable, without presenting another channel's
+ // price, native identity or capabilities as this connection's own metadata.
+ expect(moved).toEqual([{
+ id: model.id,
+ name: model.id,
+ upstream: connection.baseUrl ?? '',
+ execution: { pi: { api: connection.api } },
+ }]);
+ const reported = toCindyCatalog({ 'new-vendor': [{ ...sparse, ...connection,
+ cost: { input: 5 }, nativeApi: 'anthropic-messages', supportsFastMode: false,
+ }] }, 'reported', { previous }).providers['new-vendor'][0];
+ expect(reported).toMatchObject({ cost: { input: 5 }, nativeApi: 'anthropic-messages', supportsFastMode: false });
+});
diff --git a/packages/model-providers/src/__tests__/providerModelCatalog.test.ts b/packages/model-providers/src/__tests__/providerModelCatalog.test.ts
index be3ebcf67f..2d87b11dc0 100644
--- a/packages/model-providers/src/__tests__/providerModelCatalog.test.ts
+++ b/packages/model-providers/src/__tests__/providerModelCatalog.test.ts
@@ -135,7 +135,7 @@ it('keeps generated defaults consistent with Cindy model preferences', async ()
// Explicit official defaults: K2.8 max and Grok 4.7 high.
const expected = model.id === 'kimi-for-coding' && model.upstream === 'https://api.kimi.com/coding'
? 'max' : model.id === 'grok-4.7' && model.upstream === 'https://api.x.ai/v1'
- ? 'high' : defaultEffortForCapabilities(model.efforts);
+ ? 'high' : defaultEffortForCapabilities(model.efforts ?? []);
expect(model.defaultEffort, model.id).toBe(expected);
}
});
diff --git a/packages/model-providers/src/__tests__/providerNativeDefaults.test.ts b/packages/model-providers/src/__tests__/providerNativeDefaults.test.ts
index b00d6e3147..3fd14f70ae 100644
--- a/packages/model-providers/src/__tests__/providerNativeDefaults.test.ts
+++ b/packages/model-providers/src/__tests__/providerNativeDefaults.test.ts
@@ -79,11 +79,11 @@ describe('supplier import follows Gateway native model defaults', () => {
expect(refreshed.models['claude-code']![0].defaultEnabled).toBe(false);
});
- it('does not infer native declarations from hand-written IDs on unrelated endpoints', () => {
+ it('uses declared manufacturer families on unrelated endpoints without enabling compatibility', () => {
const provider = buildUserProvider({ id: 'proxy', name: 'Proxy', runtimes: {
codex: { baseUrl: 'https://proxy.example/v1', models: [{ id: 'google/gemini-99', name: 'Gemini', api: 'openai-responses' }] },
} }, options);
- expect(provider.models.codex![0].nativeApi).toBeUndefined();
+ expect(provider.models.codex![0].nativeApi).toBe('google-generative-ai');
expect(provider.models.codex![0].defaultEnabled).toBe(false);
});
diff --git a/packages/model-providers/src/__tests__/user-provider.test.ts b/packages/model-providers/src/__tests__/user-provider.test.ts
index b3104f1b80..e9cae5e835 100644
--- a/packages/model-providers/src/__tests__/user-provider.test.ts
+++ b/packages/model-providers/src/__tests__/user-provider.test.ts
@@ -1753,6 +1753,34 @@ describe('imported model native engine defaults', () => {
expect(agents.map(agent => provider.models[agent]?.[0]?.defaultEnabled)).toEqual(expected);
});
+ it.each([
+ [null, 'openai-responses', null, [false, false, true]],
+ ['anthropic-messages', 'openai-responses', 'anthropic-messages', [true, false, true]],
+ ['openai-responses', 'anthropic-messages', 'openai-responses', [false, true, true]],
+ [undefined, 'anthropic-messages', 'anthropic-messages', [true, false, true]],
+ [undefined, undefined, 'openai-responses', [false, true, true]],
+ ] as const)('applies current native declaration %s ahead of live %s and bundled fallback',
+ (declaration, live, expectedApi, expectedEnabled) => {
+ const agents = ['claude-code', 'codex', 'pi'] as const;
+ const config: CustomProviderConfig = {
+ id: 'native-default-test', name: 'Test',
+ runtimes: Object.fromEntries(agents.map(agent => [agent, {
+ baseUrl: 'https://example.com/v1',
+ wireProtocol: agent === 'claude-code' ? 'anthropic-messages' : 'openai-responses',
+ models: [{ id: 'gpt-6', name: 'GPT 6', discoveredMetadata: { nativeApi: live } }],
+ }])),
+ };
+ const provider = buildUserProvider(config, { modelRegistry: {
+ schemaVersion: 5, updatedAt: '2026-09-24T00:00:00Z',
+ models: [{ id: 'gpt-6', name: 'GPT 6', nativeApi: declaration, routes: [{
+ providerId: config.id, modelId: 'gpt-6', agents: ['claude-code', 'codex'],
+ }] }],
+ } });
+ expect(agents.map(agent => provider.models[agent]?.[0]?.nativeApi)).toEqual(agents.map(() => expectedApi));
+ expect(agents.map(agent => provider.models[agent]?.[0]?.defaultEnabled)).toEqual(expectedEnabled);
+ expect(config.runtimes?.codex?.models[0].discoveredMetadata?.nativeApi).toBe(live);
+ });
+
it('preserves explicit configuration defaults', () => {
const provider = buildUserProvider({ id: 'explicit-defaults', name: 'Test', runtimes: {
pi: { baseUrl: 'https://example.com/v1', models: [{ id: 'test', name: 'Test', api: 'anthropic-messages', defaultEnabled: false }] },
diff --git a/packages/model-providers/src/index.ts b/packages/model-providers/src/index.ts
index 06cec933c2..8754ce4ba3 100644
--- a/packages/model-providers/src/index.ts
+++ b/packages/model-providers/src/index.ts
@@ -111,6 +111,7 @@ export {
decideModelRegistrySnapshot,
findModelRegistryRoute,
resolveModelNativeApi,
+ resolveCatalogModelNativeApi,
resolveModelReferencePrice,
resolveBaseModelReferencePrice,
} from "./modelRegistry.js";
@@ -296,6 +297,7 @@ export {
registryEntryDefaults,
expandedRegistryEntries,
pickModelMetadata,
+ MODEL_METADATA_FIELDS,
validModelMetadata,
mergeModelMetadata,
} from "./modelMetadataLayers.js";
diff --git a/packages/model-providers/src/modelDiscovery.ts b/packages/model-providers/src/modelDiscovery.ts
index 9422510212..26669ea22f 100644
--- a/packages/model-providers/src/modelDiscovery.ts
+++ b/packages/model-providers/src/modelDiscovery.ts
@@ -147,7 +147,7 @@ export function parseModelsListResponse(
(Array.isArray(architecture?.input_modalities) && Array.isArray(architecture?.output_modalities)
? { input: architecture.input_modalities, output: architecture.output_modalities } : undefined);
const inputModalities = (modalities as { input?: unknown } | undefined)?.input ?? record.input_modalities;
- const contextWindowMax = pickModelMetadata({ contextWindowMax: record.max_context_window }).contextWindowMax;
+ const contextWindowMax = pickModelMetadata({ contextWindowMax: record.max_context_window ?? record.contextWindowMax }).contextWindowMax;
const serviceTiers = record.service_tiers;
const isVercel = sourceUrl !== undefined && (() => {
try { const url = new URL(sourceUrl); return url.origin === 'https://ai-gateway.vercel.sh'
@@ -156,7 +156,8 @@ export function parseModelsListResponse(
// Vercel marks image/video/etc. as type, but Cindy chat import only executes language models.
// Keep them out of the picker instead of saving a mode that later disappears from every list.
if (isVercel && record.type !== undefined && record.type !== 'language') continue;
- const discoveredMetadata = pickModelMetadata({
+ const discoveredMetadata = { ...pickModelMetadata(record), ...pickModelMetadata({
+ nativeApi: record.nativeApi !== undefined ? record.nativeApi : record.native_api,
...([rec?.display_name, rec?.name, google?.displayName].some(
(value) => typeof value === "string" && value.trim().length > 0,
)
@@ -169,9 +170,7 @@ export function parseModelsListResponse(
group: record.group,
contextWindow:
typeof rawWindow === "number" ? Math.floor(rawWindow) : info.max_input_tokens,
- contextWindowMax: contextWindowMax !== undefined &&
- (typeof rawWindow !== 'number' || contextWindowMax >= Math.floor(rawWindow))
- ? contextWindowMax : undefined,
+ contextWindowMax,
maxOutputTokens:
record.max_output_tokens ?? info.max_output_tokens ?? google?.outputTokenLimit ??
record.maxOutputTokens ??
@@ -209,7 +208,14 @@ export function parseModelsListResponse(
(Array.isArray(inputModalities) && inputModalities.every(value => typeof value === 'string')
? inputModalities.includes("image")
: undefined),
- });
+ }) };
+ // Check the normalized pair so every spelling (including model_info) obeys
+ // the same capacity constraint, without discarding the valid working window.
+ if (discoveredMetadata.contextWindowMax !== undefined &&
+ discoveredMetadata.contextWindow !== undefined &&
+ discoveredMetadata.contextWindowMax < discoveredMetadata.contextWindow) {
+ delete discoveredMetadata.contextWindowMax;
+ }
// OpenRouter and Vercel document USD per token. Never apply these units to arbitrary proxies.
const prices =
sourceUrl && (isOpenRouterModelsUrl(sourceUrl) || (isVercel && record.type === 'language'))
diff --git a/packages/model-providers/src/modelGeneration.ts b/packages/model-providers/src/modelGeneration.ts
index 8f86957e00..ccbaac658b 100644
--- a/packages/model-providers/src/modelGeneration.ts
+++ b/packages/model-providers/src/modelGeneration.ts
@@ -6,6 +6,9 @@ import { pickModelMetadata, type ModelMetadata } from './modelMetadataLayers.js'
*/
function generation(id: string) {
const normalized = id.toLowerCase().replace(/^(?:openai|anthropic|google|x-ai|xai|deepseek|qwen)\//, '');
+ // A first numeric token carrying a parameter-size unit is not a version.
+ // Later sizes (qwen4-30b) remain in the variant handled by the match below.
+ if (/^[a-z][a-z/-]*?\d+(?:\.\d+)?(?:x\d+(?:\.\d+)?)?[kmbt](?=[-_/.:]|$)/.test(normalized)) return undefined;
const match = /^([a-z][a-z/-]*?)(\d{1,3}(?:[.-]\d{1,3}(?=[.-]|$))*)([^0-9].*|$)/.exec(normalized);
if (!match) return undefined;
return { family: match[1] + '|' + match[3], version: match[2]!.split(/[.-]/).map(Number) };
@@ -29,6 +32,6 @@ export function previousModelGenerations<T>(id: string, candidates: readonly T[]
}
/** Inherit capabilities, not identity, documentation, price, routing or membership. */
export function generationCapabilities(value: ModelMetadata | undefined): ModelMetadata {
- const { name: _name, description: _description, officialDocs: _docs, group: _group, mode: _mode, ...capabilities } = pickModelMetadata(value);
+ const { name: _name, nativeApi: _nativeApi, description: _description, officialDocs: _docs, group: _group, mode: _mode, ...capabilities } = pickModelMetadata(value);
return capabilities;
}
diff --git a/packages/model-providers/src/modelMetadataLayers.ts b/packages/model-providers/src/modelMetadataLayers.ts
index 44b4e8a367..651408191e 100644
--- a/packages/model-providers/src/modelMetadataLayers.ts
+++ b/packages/model-providers/src/modelMetadataLayers.ts
@@ -1,3 +1,4 @@
+import { generationCapabilities, previousModelGenerations } from './modelGeneration.js';
import type { CatalogModel } from "./types.js";
import type {
ModelRegistry,
@@ -5,10 +6,18 @@ import type {
ModelRegistryRoute,
ModelEffort,
ModelReferencePriceGroup,
+ ModelAccessWireProtocol,
} from "./modelAccessBean.js";
+// Projection-only provenance; a symbol keeps it out of the public metadata schema
+// and JSON storage. CatalogModel carries the serializable verification flag.
+const inheritedContextWindow = Symbol('inheritedContextWindow');
+type ResolvedModelMetadata = ModelMetadata & { [inheritedContextWindow]?: true };
+
/** Data only. Membership, credentials, routing and billed prices never inherit. */
export interface ModelMetadata {
+ /** Manufacturer language, independent of the connection's execution protocol. */
+ nativeApi?: ModelAccessWireProtocol | null;
mode?: string;
modalities?: { input: string[]; output: string[] };
officialDocs?: string;
@@ -34,6 +43,7 @@ export interface BaseModel {
referencePriceGroups?: ModelReferencePriceGroup[];
}
export const MODEL_METADATA_FIELDS = [
+ "nativeApi",
"mode",
"modalities",
"officialDocs",
@@ -64,6 +74,8 @@ export function validModelMetadata(value: unknown): value is ModelMetadata {
return Object.entries(value).every(([key, v]) => {
if (!(MODEL_METADATA_FIELDS as readonly string[]).includes(key))
return false;
+ if (key === 'nativeApi') return v === null ||
+ ['anthropic-messages', 'openai-responses', 'openai-completions', 'google-generative-ai'].includes(v as string);
if (key === "mode")
return typeof v === "string" && /^[a-z][a-z0-9_]{0,63}$/.test(v);
if (key === "modalities") {
@@ -173,13 +185,13 @@ export function resolveModelMetadata(
registry: ModelRegistry | undefined,
providerId: string,
modelId: string,
- live?: ModelMetadata,
+ live?: ResolvedModelMetadata,
user?: ModelMetadata,
agent?: string,
providerDefaults?: ModelMetadata,
declaredDefaultEffort?: ModelMetadata["defaultEffort"],
generationDefaults?: ModelMetadata,
-): ModelMetadata {
+): ResolvedModelMetadata {
const ids = [modelId];
if (providerId === "openai" && modelId.startsWith("chatgpt/"))
ids.push(modelId.slice(8));
@@ -215,10 +227,25 @@ export function resolveModelMetadata(
findBaseModel(registry, modelId)?.defaults,
providerDefaults,
);
- const result = mergeModelMetadata(
- generationDefaults,
+ // Public family capabilities also cover subscription discovery. Exact/live data and
+ // server corrections below still win; inheritance never adds account membership.
+ const familyDefaults = generationDefaults ?? mergeModelMetadata(
+ ...previousModelGenerations(ids.at(-1)!, registry?.baseModels?.flatMap(model =>
+ [model.id, ...model.aliases].map(id => ({ id, defaults: model.defaults }))) ?? [], model => model.id)
+ .map(model => generationCapabilities(model.defaults)),
+ ...previousModelGenerations(modelId,
+ registry?.models.flatMap(entry => entry.routes
+ .filter(route => route.providerId === providerId && (!agent || agent === 'pi' || route.agents.includes(agent as never)))
+ .map(route => ({ entry, route }))) ?? [], candidate => candidate.route.modelId)
+ .map(({ entry, route }) => generationCapabilities(registryEntryDefaults(registry!, entry, route, agent))),
+ );
+ const inheritedLiveWindow = live?.[inheritedContextWindow] === true;
+ const currentLive = inheritedLiveWindow ? { ...live, contextWindow: undefined } : live;
+ const result: ResolvedModelMetadata = mergeModelMetadata(
+ familyDefaults,
+ inheritedLiveWindow ? { contextWindow: live?.contextWindow } : undefined,
defaults,
- live,
+ currentLive,
// A Harness's suggested default is not a model capability. Keep the shared
// model intent (including explicit route/Harness exceptions), then adapt it
// to the live effort membership below. Explicit force/user settings still win.
@@ -231,16 +258,28 @@ export function resolveModelMetadata(
matched?.route.forceOverrides,
user,
);
- // A predecessor's capacity cannot constrain the target's larger working
- // window. Discard only inherited capacity, not target/route/user declarations.
- if (generationDefaults?.contextWindowMax !== undefined &&
- result.contextWindow !== undefined && result.contextWindowMax !== undefined &&
- result.contextWindow > result.contextWindowMax &&
- [defaults, live, matched?.route.forceOverrides, user].every(
- (layer) => layer?.contextWindowMax === undefined,
- )) {
+ // Resolve the pair only after all layers: a maximum-only report is a usable
+ // fallback, not a verified working-window report. Never replace a known window.
+ const hasOwnWorkingWindow = [defaults, currentLive, matched?.route.forceOverrides, user]
+ .some(source => source?.contextWindow !== undefined);
+ if (!hasOwnWorkingWindow && currentLive?.contextWindowMax !== undefined) {
+ // This model's reported maximum takes precedence over a predecessor's
+ // working-window fallback, without claiming a verified working window.
+ result.contextWindow = result.contextWindowMax;
+ result[inheritedContextWindow] = true;
+ }
+ if (result.contextWindow === undefined && result.contextWindowMax !== undefined) {
+ result.contextWindow = result.contextWindowMax;
+ result[inheritedContextWindow] = true;
+ }
+ if (result.contextWindow !== undefined && result.contextWindowMax !== undefined &&
+ result.contextWindowMax < result.contextWindow) {
delete result.contextWindowMax;
}
+ if (result.contextWindow !== undefined &&
+ !hasOwnWorkingWindow) {
+ result[inheritedContextWindow] = true;
+ }
if (result.efforts?.length === 0) result.defaultEffort = null;
else if (
result.defaultEffort != null &&
@@ -384,9 +423,11 @@ export function expandedRegistryEntries(
export function catalogModelMetadata(
model: Partial<CatalogModel>,
-): ModelMetadata {
+): ResolvedModelMetadata {
return {
...pickModelMetadata(model),
+ ...(model.contextWindowVerified === false && model.contextWindow !== undefined
+ ? { [inheritedContextWindow]: true as const } : {}),
...(model.maxOutput !== undefined
? { maxOutputTokens: model.maxOutput }
: {}),
@@ -394,17 +435,20 @@ export function catalogModelMetadata(
}
export function applyModelMetadata(
model: CatalogModel,
- metadata: ModelMetadata,
+ metadata: ResolvedModelMetadata,
): CatalogModel {
- const { maxOutputTokens, ...fields } = metadata;
+ const { maxOutputTokens, [inheritedContextWindow]: inheritedWindow, ...fields } = metadata;
const result = {
...model,
...fields,
...(metadata.contextWindow !== undefined
- ? { contextWindowVerified: true }
+ ? { contextWindowVerified: inheritedWindow !== true }
: {}),
...(maxOutputTokens !== undefined ? { maxOutput: maxOutputTokens } : {}),
};
+ if (result.contextWindowMax !== undefined && result.contextWindowMax < result.contextWindow) {
+ delete result.contextWindowMax;
+ }
if (
result.efforts.length === 0 ||
(result.defaultEffort != null &&
@@ -431,18 +475,17 @@ export function mergeDiscoveredRuntimeModels(
for (const model of discovered) {
if (!model.id || !model.name || seen.has(model.id)) continue;
seen.add(model.id);
- const discoveredMetadata = pickModelMetadata(
- model.discoveredMetadata ?? model,
- );
const index = models.findIndex((m) => m.id === model.id);
- // Max-only refreshes preserve smaller working budgets, but a reduced
- // capacity must bound the old discovered window. User overrides stay separate.
- const previousWindow = models[index]?.discoveredMetadata?.contextWindow;
- if (discoveredMetadata.contextWindow === undefined &&
+ const discoveredMetadata = mergeModelMetadata(
+ index >= 0 ? models[index].discoveredMetadata : undefined,
+ pickModelMetadata(model.discoveredMetadata ?? model),
+ );
+ // Sparse refreshes can supply either half of this pair. Validate after merging
+ // with the last snapshot, without shrinking its working window to a bad maximum.
+ if (discoveredMetadata.contextWindow !== undefined &&
discoveredMetadata.contextWindowMax !== undefined &&
- ((previousWindow === undefined && models[index]?.contextWindow === undefined) ||
- (previousWindow !== undefined && previousWindow > discoveredMetadata.contextWindowMax))) {
- discoveredMetadata.contextWindow = discoveredMetadata.contextWindowMax;
+ discoveredMetadata.contextWindowMax < discoveredMetadata.contextWindow) {
+ delete discoveredMetadata.contextWindowMax;
}
if (index < 0)
models.push({
@@ -456,7 +499,7 @@ export function mergeDiscoveredRuntimeModels(
models[index] = {
...models[index],
...(!models[index].discoveredMetadata ? { nameExplicit: true } : {}),
- discoveredMetadata: mergeModelMetadata(models[index].discoveredMetadata, discoveredMetadata),
+ discoveredMetadata,
...(model.discoveredCost ? { discoveredCost: model.discoveredCost } : {}),
};
}
@@ -467,12 +510,9 @@ export function mergeDiscoveredRuntimeModels(
export function runtimeUserModelMetadata(
m: import("./types.js").ProviderRuntimeModelConfig,
): ModelMetadata {
+ const { name: _name, ...metadata } = pickModelMetadata(m);
return pickModelMetadata({
- ...pickModelMetadata({
- mode: m.mode,
- modalities: m.modalities,
- officialDocs: m.officialDocs,
- }),
+ ...metadata,
...(!m.discoveredMetadata || m.nameExplicit ? { name: m.name } : {}),
...(m.contextWindow !== undefined
? { contextWindow: m.contextWindow }
@@ -480,9 +520,8 @@ export function runtimeUserModelMetadata(
...(m.supportsImageInput !== undefined
? { supportsImageInput: m.supportsImageInput }
: {}),
- ...(m.reasoning !== undefined
- ? { efforts: m.reasoning ? (m.reasoningEfforts ?? []) : [] }
- : {}),
+ ...(m.reasoning === false ? { efforts: [] }
+ : m.reasoningEfforts !== undefined ? { efforts: m.reasoningEfforts } : {}),
...(m.reasoningDefaultEffort !== undefined
? { defaultEffort: m.reasoningDefaultEffort }
: {}),
diff --git a/packages/model-providers/src/modelRegistry.ts b/packages/model-providers/src/modelRegistry.ts
index 858b5fe4c9..0ff1b89577 100644
--- a/packages/model-providers/src/modelRegistry.ts
+++ b/packages/model-providers/src/modelRegistry.ts
@@ -63,10 +63,10 @@ export function resolveModelNativeApi(
?.nativeApi;
}
-/** Model identity only, for a model verified in an imported catalog or the Registry.
+/** Model identity only, using exact declarations or server-declared family prefixes.
* Reuse the declarations already used by Gateway; do not turn an execution API
* into a manufacturer declaration or copy a Gateway endpoint/capability override.
- * Callers must not pass arbitrary names from a hand-written connection here.
+ * Unknown/private namespaces stay unknown; connection transport is never evidence.
*/
export function resolveCatalogModelNativeApi(
registry: ModelRegistry | undefined,
diff --git a/packages/model-providers/src/piNativeCatalog.ts b/packages/model-providers/src/piNativeCatalog.ts
index 84a847e3e6..f10a6415b2 100644
--- a/packages/model-providers/src/piNativeCatalog.ts
+++ b/packages/model-providers/src/piNativeCatalog.ts
@@ -5,6 +5,7 @@ import {
import { defaultEffortForCapabilities } from "./effortResolution.js";
import { piSupportedEfforts } from "./piThinkingLevels.mjs";
+import { previousModelGenerations } from "./modelGeneration.js";
import type { ModelMetadata } from "./modelMetadataLayers.js";
import type {
CatalogModel,
@@ -19,7 +20,7 @@ interface PiCatalogRow {
api?: string;
provider: string;
baseUrl?: string;
- contextWindow: number;
+ contextWindow?: number;
maxTokens?: number;
input?: string[];
reasoning?: boolean;
@@ -81,31 +82,37 @@ export function piNativeCatalogModels(
);
}
return rows.map((row, index) => {
- if (
- row.provider !== piProviderId ||
- !Number.isFinite(row.contextWindow) ||
- row.contextWindow <= 0
- ) {
+ if (row.provider !== piProviderId) {
throw new Error(
`[model-providers] invalid Pi catalog row '${piProviderId}/${row.id}'`,
);
}
const efforts = piSupportedEfforts(row);
const piApi = portablePiApi(row.api);
+ const declaredWindow = Number.isSafeInteger(row.contextWindow) && row.contextWindow! > 0
+ ? row.contextWindow : undefined;
+ const previousWindow = declaredWindow === undefined
+ ? previousModelGenerations(row.id, rows.filter(candidate =>
+ candidate.baseUrl === row.baseUrl && candidate.api === row.api &&
+ Number.isSafeInteger(candidate.contextWindow) && candidate.contextWindow! > 0), candidate => candidate.id)
+ .at(-1)?.contextWindow
+ : undefined;
return {
id: `${options.idPrefix ?? ""}${row.id}`,
name: row.name ?? row.id,
...(options.group ? { group: options.group } : {}),
sortOrder: index,
- contextWindow: row.contextWindow,
- contextWindowVerified: true,
+ // Unknown models remain usable. Neither a predecessor window nor the
+ // generic working budget is a verified limit of the newly imported model.
+ contextWindow: declaredWindow ?? previousWindow ?? 200_000,
+ contextWindowVerified: declaredWindow !== undefined,
...(Number.isFinite(row.maxTokens) && row.maxTokens! > 0
? { maxOutput: row.maxTokens }
: {}),
efforts,
discoveredMetadata: {
...(row.name ? { name: row.name } : {}),
- contextWindow: row.contextWindow,
+ ...(declaredWindow !== undefined ? { contextWindow: declaredWindow } : {}),
// Thinking tiers are imported defaults, not account discovery. Keep
// them on the fallback model so shared Registry efforts can replace them.
...(row.maxTokens ? { maxOutputTokens: row.maxTokens } : {}),
diff --git a/packages/model-providers/src/providerModelCatalog.ts b/packages/model-providers/src/providerModelCatalog.ts
index 8f5d97806f..861bb61555 100644
--- a/packages/model-providers/src/providerModelCatalog.ts
+++ b/packages/model-providers/src/providerModelCatalog.ts
@@ -4,24 +4,24 @@ import { providerEndpointBindings } from './providerEndpointTemplate.js';
import { declaredModelInterface } from './providerInterfaceRoutes.js';
import interfaceModels from "../catalog/provider-interface-models.json" with { type: "json" };
import generated from "../catalog/provider-models.json" with { type: "json" };
-import type { ModelMetadata } from "./modelMetadataLayers.js";
+import { pickModelMetadata, type ModelMetadata } from "./modelMetadataLayers.js";
import type { ModelCost, PiModelApi, ProviderWireProtocol } from "./types.js";
-export interface ProviderModelRecord {
+export interface ProviderModelRecord extends ModelMetadata {
/** Internal fallback provenance, not an upstream capability declaration. */
inheritedFrom?: string;
id: string;
name: string;
upstream: string;
- contextWindow: number;
+ contextWindow?: number;
maxOutput?: number;
- modalities: { input: string[]; output: string[] };
- supportsImageInput: boolean;
+ modalities?: { input: string[]; output: string[] };
+ supportsImageInput?: boolean;
/** Connection-scoped declaration; absent does not authorize a Fast request. */
supportsFastMode?: boolean;
- reasoning: boolean;
- efforts: NonNullable<ModelMetadata["efforts"]>;
- defaultEffort: ModelMetadata["defaultEffort"];
+ reasoning?: boolean;
+ efforts?: NonNullable<ModelMetadata["efforts"]>;
+ defaultEffort?: ModelMetadata["defaultEffort"];
cost?: ModelCost;
execution: {
pi: {
@@ -134,16 +134,25 @@ export function providerPresetModelRecord(presetId: string | undefined, modelId:
}
/** Preserve the upstream adapter identity, independently of a user's connection UUID. */
-export function providerModelAdapterId(row: ProviderModelRecord): string | undefined {
- const matches = Object.entries(PROVIDER_MODEL_CATALOG.providers).filter(([, rows]) =>
- rows.some(candidate => candidate.id === row.id && (normalize(candidate.upstream) === normalize(row.upstream) || (candidate.upstream.includes('{') && providerEndpointBindings(candidate.upstream, row.upstream) !== null))
- && candidate.execution.pi.api === row.execution.pi.api));
+export function providerModelAdapterId(row: ProviderModelRecord, presetId?: string): string | undefined {
+ const presetProvider = presetId ? sourceProviderForPreset(presetId) : undefined;
+ const matches = Object.entries(PROVIDER_MODEL_CATALOG.providers).filter(([provider, rows]) => {
+ const connectionRows = rows.filter(candidate => candidate.execution.pi.api === row.execution.pi.api &&
+ (normalize(candidate.upstream) === normalize(row.upstream) ||
+ (candidate.upstream.includes('{') && providerEndpointBindings(candidate.upstream, row.upstream) !== null) ||
+ (row.inheritedFrom !== undefined && provider === presetProvider)));
+ return connectionRows.some(candidate => candidate.id === row.id) ||
+ // Only the matched connection/preset lends transport identity. Manufacturer
+ // capability fallback on an unrelated relay must not lend its adapter.
+ (row.inheritedFrom !== undefined && previousModelGenerations(row.id, connectionRows, candidate => candidate.id).length > 0);
+ });
// Duplicated subscription catalogs can share endpoints; never guess between distinct identities.
return matches.length === 1 ? matches[0][0] : undefined;
}
export function providerModelMetadata(row: ProviderModelRecord): ModelMetadata {
return {
+ ...pickModelMetadata(row),
name: row.name,
contextWindow: row.contextWindow,
...(row.maxOutput ? { maxOutputTokens: row.maxOutput } : {}),
@@ -173,7 +182,7 @@ export function providerCatalogForPi() {
baseUrl: row.upstream,
contextWindow: row.contextWindow,
maxTokens: row.maxOutput,
- input: row.modalities.input,
+ input: row.modalities?.input,
reasoning: row.reasoning,
...(row.cost ? { cost: row.cost } : {}),
...row.execution.pi,
diff --git a/packages/model-providers/src/types.ts b/packages/model-providers/src/types.ts
index 52421904f3..67135f1c73 100644
--- a/packages/model-providers/src/types.ts
+++ b/packages/model-providers/src/types.ts
@@ -590,10 +590,7 @@ export interface Provider {
* contextWindow 缺省时由 `buildUserProvider` 使用保守默认;预设可显式携带厂商文档确认的值,
* 并随用户配置持久化,避免已知长上下文模型被错误降级。
*/
-export interface ProviderRuntimeModelConfig extends Pick<
- ModelMetadata,
- "mode" | "modalities" | "officialDocs"
-> {
+export interface ProviderRuntimeModelConfig extends ModelMetadata {
discoveredMetadata?: ModelMetadata;
discoveredCost?: ModelCost;
nameExplicit?: boolean;
diff --git a/packages/model-providers/src/user-provider.ts b/packages/model-providers/src/user-provider.ts
index 682d85be9f..dc317f296b 100644
--- a/packages/model-providers/src/user-provider.ts
+++ b/packages/model-providers/src/user-provider.ts
@@ -338,10 +338,6 @@ function toCatalogModel(
...(supportsFastMode ? { supportsFastMode: true } : {}),
};
const user = runtimeUserModelMetadata(m);
- const live = m.discoveredMetadata?.contextWindow === undefined &&
- m.discoveredMetadata?.contextWindowMax !== undefined
- ? { ...m.discoveredMetadata, contextWindow: m.discoveredMetadata.contextWindowMax }
- : m.discoveredMetadata;
const resolved =
(modelRegistry?.schemaVersion ?? 0) >= 4 ||
m.discoveredMetadata ||
@@ -350,7 +346,7 @@ function toCatalogModel(
modelRegistry ?? undefined,
metadataProviderId,
m.id,
- live,
+ m.discoveredMetadata,
pickModelMetadata(user),
agent,
providerDefaults,
@@ -358,22 +354,7 @@ function toCatalogModel(
generationDefaults,
)
: pickModelMetadata(user);
- const result = applyModelMetadata(model, resolved);
- if (generationDefaults?.contextWindow !== undefined) {
- // A predecessor's window remains a usable default, not verified capacity
- // for this model/route. Only this model's own metadata can establish that.
- const declared = resolveModelMetadata(
- modelRegistry ?? undefined,
- metadataProviderId,
- m.id,
- live,
- pickModelMetadata(user),
- agent,
- providerDefaults,
- );
- if (declared.contextWindow === undefined) result.contextWindowVerified = false;
- }
- return result;
+ return applyModelMetadata(model, resolved);
}
function defaultWireProtocol(agent: AgentKind): ProviderWireProtocol {
@@ -621,30 +602,30 @@ export function buildUserProvider(
const baseModel = findBaseModel(source, m.id);
const routeNativeApi = resolveModelNativeApi(source, config.id, m.id);
return routeNativeApi !== undefined ? routeNativeApi
- : imported || baseModel || (followsPreset && (presetModel || m.discoveredMetadata))
- ? resolveCatalogModelNativeApi(source, baseModel?.id ?? m.id)
- : undefined;
+ : resolveCatalogModelNativeApi(source, baseModel?.id ?? m.id);
};
+ const projected = toCatalogModel(
+ m,
+ followsPreset && sameRoute ? preset!.id : config.id,
+ agent,
+ options.modelRegistry,
+ defaults,
+ nativeCodex ? 'openai' : undefined,
+ generationDefaults,
+ );
const currentDeclaration = resolveDeclaration(registry);
- // Same fallback as Gateway: Server omissions use local native declarations;
- // explicit corrections/unknowns win. Never backfill another route's capabilities.
+ // Apply current Server identity after capability projection, including an
+ // explicit unknown. Only absent declarations may use discovery/local fallback.
const declaration = currentDeclaration !== undefined ? currentDeclaration
+ : projected.nativeApi !== undefined ? projected.nativeApi
: resolveDeclaration(BUNDLED_CATALOG.modelRegistry);
const nativeApi = declaration === null || declaration === 'anthropic-messages'
|| declaration === 'openai-responses' || declaration === 'openai-completions'
|| declaration === 'google-generative-ai' ? declaration : undefined;
return {
- ...(nativeApi !== undefined ? { nativeApi } : {}),
...(imported?.cost ? { cost: imported.cost } : {}),
- ...toCatalogModel(
- m,
- followsPreset && sameRoute ? preset!.id : config.id,
- agent,
- options.modelRegistry,
- defaults,
- nativeCodex ? 'openai' : undefined,
- generationDefaults,
- ),
+ ...projected,
+ ...(nativeApi !== undefined ? { nativeApi } : {}),
// Projection is not a user edit. Save only the original configuration.
userModelConfig: structuredClone(storedModel),
...(m.api ? { api: m.api, ...(agent === 'pi' ? { piApi: m.api } : {}) } : {}),
diff --git a/tools/pi/catalog-format.d.mts b/tools/pi/catalog-format.d.mts
index 232fd57fc0..83e588d285 100644
--- a/tools/pi/catalog-format.d.mts
+++ b/tools/pi/catalog-format.d.mts
@@ -5,11 +5,16 @@ export interface PiImportModel {
api: string;
baseUrl?: string;
name?: string;
- contextWindow: number;
+ contextWindow?: number;
maxTokens?: number;
input?: string[];
+ output?: string[];
+ nativeApi?: string;
+ supportsFastMode?: boolean;
+ supportsToolCalls?: boolean;
+ reasoningRequired?: boolean;
reasoning?: boolean;
- defaultEffort?: string;
+ defaultEffort?: string | null;
thinkingLevelMap?: Record<string, string | null>;
cost?: Record<string, unknown>;
headers?: Record<string, string>;
@@ -17,7 +22,11 @@ export interface PiImportModel {
samplingParams?: Record<string, unknown>;
}
export function toCindyProviderModel(row: PiImportModel): ProviderModelRecord;
-export function toCindyCatalog(providers: Record<string, PiImportModel[]>, generatedAt: string): {
+export function toCindyCatalog(providers: Record<string, PiImportModel[]>, generatedAt: string, options?: {
+ previous?: { providers: Record<string, ProviderModelRecord[]> };
+ onError?: (error: unknown) => void;
+ incompleteProviders?: Iterable<string>;
+}): {
schemaVersion: number;
generatedAt: string;
source: string;
diff --git a/tools/pi/catalog-format.mjs b/tools/pi/catalog-format.mjs
index 795a6f0ccc..6fae590202 100644
--- a/tools/pi/catalog-format.mjs
+++ b/tools/pi/catalog-format.mjs
@@ -25,11 +25,9 @@ export function toCindyProviderModel(row) {
const efforts = piSupportedEfforts(row);
const { id, provider, baseUrl, api } = row;
if (
- !id ||
- !provider ||
- !api ||
- !Number.isSafeInteger(row.contextWindow) ||
- row.contextWindow <= 0
+ typeof id !== 'string' || !id.trim() || id.length > 256 ||
+ typeof provider !== 'string' || !provider.trim() ||
+ typeof api !== 'string' || !api.trim()
) {
throw new Error(`Invalid upstream model: ${provider}/${id}`);
}
@@ -61,25 +59,30 @@ export function toCindyProviderModel(row) {
if (Array.isArray(row.cost?.tiers)) cost.tiers = row.cost.tiers;
return {
id,
- name: row.name ?? id,
+ name: typeof row.name === 'string' && row.name.trim() ? row.name : id,
upstream: baseUrl ?? "",
- contextWindow: row.contextWindow,
+ ...(Number.isSafeInteger(row.contextWindow) && row.contextWindow > 0 ? { contextWindow: row.contextWindow } : {}),
...(row.maxTokens > 0 ? { maxOutput: row.maxTokens } : {}),
- modalities: { input: row.input ?? ["text"], output: ["text"] },
- supportsImageInput: row.input?.includes("image") ?? false,
- reasoning: row.reasoning === true,
- efforts,
+ ...(Array.isArray(row.input) ? {
+ modalities: { input: row.input, output: row.output ?? ["text"] },
+ supportsImageInput: row.input.includes("image"),
+ } : {}),
+ ...(typeof row.reasoning === "boolean" ? { reasoning: row.reasoning, efforts } : {}),
+ ...Object.fromEntries(["supportsFastMode", "supportsToolCalls", "reasoningRequired"]
+ .filter(key => typeof row[key] === "boolean").map(key => [key, row[key]])),
+ ...(['anthropic-messages', 'openai-responses', 'openai-completions', 'google-generative-ai'].includes(row.nativeApi)
+ ? { nativeApi: row.nativeApi } : {}),
// Preserve a supported explicit default; otherwise use Cindy's generic preference.
- defaultEffort: efforts.includes(row.defaultEffort)
+ ...(typeof row.reasoning === "boolean" ? { defaultEffort: row.defaultEffort === null ? null : efforts.includes(row.defaultEffort)
? row.defaultEffort
: (["medium", "high", "low", "xhigh", "max", "minimal", "ultra"].find(
(effort) => efforts.includes(effort),
- ) ?? null),
+ ) ?? null) } : {}),
...(Object.keys(cost).length ? { cost } : {}),
execution: {
pi: {
api,
- ...(Object.keys(headers).length ? { headers } : {}),
+ ...(row.headers !== undefined ? { headers } : {}),
...(row.thinkingLevelMap
? { thinkingLevelMap: row.thinkingLevelMap }
: {}),
@@ -90,7 +93,8 @@ export function toCindyProviderModel(row) {
};
}
-export function toCindyCatalog(providers, generatedAt) {
+export function toCindyCatalog(providers, generatedAt, { previous, onError, incompleteProviders = [] } = {}) {
+ const incomplete = new Set(incompleteProviders);
return {
schemaVersion: 1,
generatedAt,
@@ -100,9 +104,37 @@ export function toCindyCatalog(providers, generatedAt) {
.sort(([a], [b]) => a.localeCompare(b))
.map(([id, models]) => [
id,
- models
- .map(toCindyProviderModel)
- .sort((a, b) => a.id.localeCompare(b.id)),
+ (() => {
+ const prior = new Map((previous?.providers?.[id] ?? []).map(row => [row.id, row]));
+ const converted = new Map();
+ let complete = !incomplete.has(id);
+ for (const row of models) {
+ try {
+ const old = prior.get(row?.id);
+ const sameConnection = old?.upstream === (row.baseUrl ?? '') && old?.execution.pi.api === row.api;
+ const adapterDefaults = sameConnection ? old.execution.pi : {};
+ // Convert after filling missing adapter fields, so efforts and
+ // the request mapping stay consistent. Explicit {} replaces.
+ const next = toCindyProviderModel({ ...adapterDefaults, ...row,
+ ...Object.fromEntries(['headers', 'thinkingLevelMap', 'compat', 'samplingParams']
+ .filter(key => row[key] === undefined && adapterDefaults[key] !== undefined)
+ .map(key => [key, adapterDefaults[key]])),
+ });
+ // Last-good channel metadata is valid only for the same route.
+ // A moved model remains imported; shared model defaults are
+ // resolved separately rather than copied from the old channel.
+ converted.set(next.id, sameConnection ? { ...old, ...next } : next);
+ } catch (error) {
+ if (!onError) throw error;
+ complete = false;
+ onError(error);
+ }
+ }
+ if (!complete) for (const [modelId, old] of prior) {
+ if (!converted.has(modelId)) converted.set(modelId, old);
+ }
+ return [...converted.values()].sort((a, b) => a.id.localeCompare(b.id));
+ })(),
]),
),
};
diff --git a/tools/pi/sync-model-catalog.mjs b/tools/pi/sync-model-catalog.mjs
index 9a61504630..0dbdb07ec8 100644
--- a/tools/pi/sync-model-catalog.mjs
+++ b/tools/pi/sync-model-catalog.mjs
@@ -24,7 +24,7 @@ import {
applyAstraCatalogAdditions,
applyPinnedAstraCorrections,
} from "./openai-catalog-corrections.mjs";
-function catalogEntries(providerId, value) {
+function catalogEntries(providerId, value, incompleteProviders) {
const entries = Array.isArray(value)
? value
: value && typeof value === "object" && Array.isArray(value.models)
@@ -34,21 +34,19 @@ function catalogEntries(providerId, value) {
: null;
if (!entries)
throw new Error(`Pi catalog '${providerId}' is not an array or model map`);
- return entries.map((entry) => {
- if (!entry || typeof entry !== "object" || typeof entry.id !== "string") {
- throw new Error(`Pi catalog '${providerId}' contains an invalid model`);
+ return entries.filter((entry) => {
+ if (!entry || typeof entry !== "object" || typeof entry.id !== "string" || entry.provider !== providerId) {
+ console.warn(`Skipping invalid model in '${providerId}'`);
+ incompleteProviders.add(providerId);
+ return false;
}
- if (entry.provider !== providerId) {
- throw new Error(
- `Pi catalog '${providerId}' model '${entry.id}' has provider '${entry.provider}'`,
- );
- }
- return entry;
+ return true;
});
}
async function main() {
const providers = {};
+ const incompleteProviders = new Set();
const previous = JSON.parse(await fs.readFile(SNAPSHOT_PATH, "utf8"));
const PROVIDER_IDS = Object.keys(previous.providers)
.filter((id) => id !== ".manifest")
@@ -74,39 +72,55 @@ async function main() {
await fs.readFile(path.resolve(bundlePath), "utf8"),
)
: JSON.parse(await fs.readFile(path.resolve(inputPath), "utf8"));
- for (const providerId of PROVIDER_IDS) {
- if (!(providerId in input))
- throw new Error(`Input catalog lacks provider '${providerId}'`);
- }
for (const providerId of Object.keys(input)) {
if (providerId === ".manifest") continue;
- if (!(providerId in input))
- throw new Error(`Input catalog lacks provider '${providerId}'`);
- providers[providerId] = catalogEntries(providerId, input[providerId]);
+ try { providers[providerId] = catalogEntries(providerId, input[providerId], incompleteProviders); }
+ catch {
+ incompleteProviders.add(providerId);
+ console.warn(`Keeping previous catalog for '${providerId}': invalid source`);
+ }
}
} else {
// Refresh the complete imported catalog, including channels not curated as GUI presets.
- const providerIds = [...new Set(PROVIDER_IDS)].sort();
+ let discovered = [];
+ try {
+ const index = await fetch(PI_CATALOG_BASE, { signal: AbortSignal.timeout(15_000) });
+ if (!index.ok) throw new Error(`HTTP ${index.status}`);
+ const data = await index.json();
+ if (Array.isArray(data)) discovered = data.filter(id => typeof id === 'string' && /^[a-z0-9][a-z0-9-]*$/.test(id));
+ } catch { console.warn('Provider index unavailable; refreshing previously known providers'); }
+ const providerIds = [...new Set([...PROVIDER_IDS, ...discovered])].sort();
for (const providerId of providerIds) {
- const response = await fetch(
- `${PI_CATALOG_BASE}/${encodeURIComponent(providerId)}`,
- {
- headers: {
- accept: "application/json",
- "user-agent": "cindy-pi-catalog-sync",
+ try {
+ const response = await fetch(
+ `${PI_CATALOG_BASE}/${encodeURIComponent(providerId)}`,
+ {
+ signal: AbortSignal.timeout(15_000),
+ headers: {
+ accept: "application/json",
+ "user-agent": "cindy-pi-catalog-sync",
+ },
},
- },
- );
- if (!response.ok)
- throw new Error(
- `Pi catalog '${providerId}' returned HTTP ${response.status}`,
);
- const modified = Date.parse(response.headers.get("last-modified") ?? "");
- if (!Number.isNaN(modified))
- newestModified = Math.max(newestModified, modified);
- providers[providerId] = catalogEntries(providerId, await response.json());
+ if (!response.ok)
+ throw new Error(
+ `Pi catalog '${providerId}' returned HTTP ${response.status}`,
+ );
+ const modified = Date.parse(response.headers.get("last-modified") ?? "");
+ if (!Number.isNaN(modified))
+ newestModified = Math.max(newestModified, modified);
+ providers[providerId] = catalogEntries(providerId, await response.json(), incompleteProviders);
+ } catch {
+ incompleteProviders.add(providerId);
+ console.warn(`Keeping previous catalog for '${providerId}': source unavailable`);
+ }
}
}
+ // An omitted/failed source is not a complete snapshot, even if corrections
+ // below create a provider containing only an additive model.
+ for (const providerId of PROVIDER_IDS) {
+ if (!Object.hasOwn(providers, providerId)) incompleteProviders.add(providerId);
+ }
if (providers.xai) providers.xai = applyKnownXaiCorrections(providers.xai);
applyGrok47CatalogAddition(providers);
applyAstraCatalogAdditions(providers);
@@ -119,12 +133,18 @@ async function main() {
? new Date(generatedAtArg).toISOString()
: new Date(newestModified || Date.now()).toISOString();
const standard = {
- ...toCindyCatalog(providers, generatedAt),
+ ...toCindyCatalog(providers, generatedAt, { previous, incompleteProviders, onError: () => console.warn('Skipping invalid model; keeping last good record') }),
...(sourceVersion ? { sourceVersion } : {}),
};
- // Conversion completes before replacing the last good catalog.
+ standard.providers = { ...previous.providers, ...standard.providers };
+ // A partial upstream outage never removes the last usable model record.
+ // Replace atomically so interruption cannot truncate the last good catalog.
const modelText = `${JSON.stringify(standard, null, 2)}\n`;
- await fs.writeFile(SNAPSHOT_PATH, modelText);
+ const temporary = `${SNAPSHOT_PATH}.${process.pid}.tmp`;
+ try {
+ await fs.writeFile(temporary, modelText, { flag: 'wx' });
+ await fs.rename(temporary, SNAPSHOT_PATH);
+ } finally { await fs.rm(temporary, { force: true }); }
console.log(
`Synced ${Object.keys(providers).length} Pi providers at ${generatedAt}`,
);