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。

feat(providers): 完善 Sub2API 接入与新模型代际继承 (#5050)

Chris Zhangd6cd6f3d11dc
diff --git a/apps/desktop/src/main/maker-host/__tests__/catalogDerivedModels.test.ts b/apps/desktop/src/main/maker-host/__tests__/catalogDerivedModels.test.ts
index 8e3566f1d9..7436ef6f96 100644
--- a/apps/desktop/src/main/maker-host/__tests__/catalogDerivedModels.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/catalogDerivedModels.test.ts
@@ -254,9 +254,10 @@ describe('deriveAvailableModels — dynamic-first catalog contract', () => {
efforts: ['low'],
defaultEffort: 'low',
});
+ // The target's exact manufacturer declaration outranks a previous generation.
expect(flatModels.find((m) => m.id === 'grok-4.5')).toMatchObject({
- efforts: [],
- defaultEffort: null,
+ efforts: ['low', 'medium', 'high'],
+ defaultEffort: 'medium',
});
});
@@ -271,7 +272,7 @@ describe('deriveAvailableModels — dynamic-first catalog contract', () => {
pi: {
baseUrl: 'http://127.0.0.1:11434/v1',
wireProtocol: 'openai-responses',
- models: [{ id: 'grok-4.5', name: 'Grok 4.5 without reasoning' }],
+ models: [{ id: 'grok-4.5', name: 'Grok 4.5 without reasoning', reasoning: false }],
},
},
}),
diff --git a/apps/desktop/src/main/maker-host/__tests__/claudeProviderBridge.test.ts b/apps/desktop/src/main/maker-host/__tests__/claudeProviderBridge.test.ts
index bf16d0c790..c2f8911af8 100644
--- a/apps/desktop/src/main/maker-host/__tests__/claudeProviderBridge.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/claudeProviderBridge.test.ts
@@ -12,6 +12,7 @@ async function runBridge(
onRequest: (url: string, init?: RequestInit) => void,
extras: Partial<Parameters<typeof createClaudeProviderBridge>[0]> = {},
effort = 'high',
+ fast = false,
) {
const handler = createClaudeProviderBridge({
url: `https://supplier.example/v1/${protocol === 'openai-chat' ? 'chat/completions' : 'responses'}`,
@@ -29,7 +30,7 @@ async function runBridge(
req.on('end', () => {
void handler.handle({ parsedBody: JSON.parse(Buffer.concat(chunks).toString()),
ctx: { reqId: 1, method: 'POST', url: req.url!, headers: {} }, res,
- prefs: { reasoningEffort: effort },
+ prefs: { reasoningEffort: effort, fast },
}).catch(() => { res.statusCode = 500; res.end(); });
});
});
@@ -102,3 +103,13 @@ describe('Claude Code custom provider translation', () => {
.not.toBe(claudeProviderReasoningNamespace(url));
});
});
+
+
+it.each(['openai-chat', 'openai-responses'] as const)('sends Fast only when enabled and supported on %s', async protocol => {
+ for (const supported of [true, false, undefined]) for (const fast of [true, false]) {
+ let body: Record<string, unknown> | undefined;
+ await runBridge(protocol, true, chatStream, (_url, init) => { body = JSON.parse(String(init?.body)); },
+ { supportsFastMode: supported }, 'high', fast);
+ expect(body?.service_tier).toBe(supported && fast ? 'priority' : undefined);
+ }
+});
diff --git a/apps/desktop/src/main/maker-host/__tests__/codexProxyHost.test.ts b/apps/desktop/src/main/maker-host/__tests__/codexProxyHost.test.ts
index c01e92d616..4eb1034a42 100644
--- a/apps/desktop/src/main/maker-host/__tests__/codexProxyHost.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/codexProxyHost.test.ts
@@ -3325,6 +3325,42 @@ describe('codex proxy host', () => {
expect(host.getCodexProxyEndpoint()).toBe(`${XD_GATEWAY_BASE_URL}/v1`);
});
+ it('keeps Sub2API-discovered Grok effort in the outgoing Responses body', async () => {
+ const host = await freshCodexProxyHost();
+ const { buildUserProvider, parseModelsListResponse, mergeDiscoveredRuntimeModels } = await import('@cindy/model-providers');
+ const { setCustomProviders } = await import('../active-catalog.js');
+ const { setSessionProvider, clearSessionProvider } = await import('../session-provider-store.js');
+ const models = parseModelsListResponse({ data: [{ id: 'grok-4.7',
+ supportsReasoningEffort: true, reasoningEffort: 'high',
+ service_tiers: [{ id: 'priority' }], input_modalities: ['text', 'image'],
+ reasoningEfforts: ['low', 'medium', 'high', 'xhigh'].map(value => ({ value, label: value })),
+ }] })!;
+ const stored = JSON.parse(JSON.stringify(mergeDiscoveredRuntimeModels([], models)));
+ const provider = buildUserProvider({ id: 'sub2api', name: 'Sub2API', runtimes: {
+ codex: { baseUrl: 'https://sub2api.example/custom/v1', wireProtocol: 'openai-responses', models: stored },
+ } });
+ mockState.createAnthropicCompatProxy.mockResolvedValueOnce({
+ url: 'http://127.0.0.1:43210', dispose: vi.fn(async () => undefined),
+ });
+ host.setCodexProxyAuthInjection('env-key');
+ host.registerComposed('sub2api-session', 'sub2api-thread', '');
+ setSessionProvider('sub2api-session', provider.id);
+ setCustomProviders([provider]);
+ try {
+ await host.ensureCodexProxyReady();
+ let body: unknown = { model: 'grok-4.7', input: [], reasoning: { effort: 'xhigh' }, service_tier: 'priority' };
+ for (const transform of mockState.createAnthropicCompatProxy.mock.calls[0][0].transformRequest) {
+ const next = transform(body, { method: 'POST', url: '/responses', headers: { 'thread-id': 'sub2api-thread' } });
+ if (next !== null && next !== undefined) body = next;
+ }
+ expect(body).toHaveProperty('reasoning.effort', 'xhigh');
+ expect(body).toHaveProperty('service_tier', 'priority');
+ } finally {
+ clearSessionProvider('sub2api-session');
+ setCustomProviders([]);
+ }
+ });
+
it('reconciles a saved Responses effort on every catalog refresh without borrowing another route', async () => {
const host = await freshCodexProxyHost();
const { buildUserProvider } = await import('@cindy/model-providers');
diff --git a/apps/desktop/src/main/maker-host/__tests__/genericOAuth.test.ts b/apps/desktop/src/main/maker-host/__tests__/genericOAuth.test.ts
index 13599451aa..0354b59735 100644
--- a/apps/desktop/src/main/maker-host/__tests__/genericOAuth.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/genericOAuth.test.ts
@@ -661,6 +661,18 @@ describe('discoverGenericOAuthModels', () => {
});
describe('deriveModelsDiscoveryUrl', () => {
+ it('keeps the Codex backend path without inserting an extra API version', () => {
+ expect(deriveModelsDiscoveryUrl('https://relay.example/backend-api/codex')).toBe(
+ 'https://relay.example/backend-api/codex/models',
+ );
+ expect(deriveModelsDiscoveryUrl('https://relay.example/team/backend-api/codex/?tenant=a#ignored')).toBe(
+ 'https://relay.example/team/backend-api/codex/models?tenant=a',
+ );
+ expect(deriveModelsDiscoveryUrl('https://relay.example/codex')).toBe(
+ 'https://relay.example/codex/v1/models',
+ );
+ });
+
it('/vN 结尾只追加 /models,其余追加 /v1/models(尾斜杠归一)', () => {
expect(deriveModelsDiscoveryUrl('https://openrouter.ai/api/v1')).toBe(
'https://openrouter.ai/api/v1/models',
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 b67885544a..f28e06006b 100644
--- a/apps/desktop/src/main/maker-host/__tests__/piNativeProviders.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/piNativeProviders.test.ts
@@ -7,7 +7,7 @@ import { existsSync } from 'node:fs';
import fsp from 'node:fs/promises';
import path from 'node:path';
-import { BUNDLED_CATALOG, PROVIDER_MODEL_CATALOG, providerModelRecord, buildUserProvider, providerPresetOAuth, type Catalog } from '@cindy/model-providers';
+import { BUNDLED_CATALOG, PROVIDER_MODEL_CATALOG, providerModelRecord, buildUserProvider, providerPresetOAuth, parseModelsListResponse, mergeDiscoveredRuntimeModels, type Catalog } from '@cindy/model-providers';
// Account discovery persistence is outside this runtime/route fixture.
vi.mock('../model-discovery/xai.js', () => ({
@@ -408,9 +408,9 @@ describe('buildPiNativeProvidersFromConfigs', () => {
it.each([
{ baseUrl: 'https://api.openai.com/v1', wireProtocol: 'openai-responses' as const, expected: true },
- { baseUrl: 'https://private.example/v1', wireProtocol: 'openai-responses' as const, expected: false },
+ { baseUrl: 'https://private.example/v1', wireProtocol: 'openai-responses' as const, expected: true },
{ baseUrl: 'https://api.openai.com/v1', wireProtocol: 'openai-chat' as const, expected: false },
- ])('Astra API metadata requires the matching endpoint and protocol: $baseUrl $wireProtocol', ({ baseUrl, wireProtocol, expected }) => {
+ ])('Astra adapter metadata follows exact identity and protocol, while prices stay route-scoped: $baseUrl $wireProtocol', ({ baseUrl, wireProtocol, expected }) => {
const { providers } = buildPiNativeProvidersFromConfigs([{
id: 'manual-openai', name: 'Manual OpenAI', auth: { method: 'apiKey' },
runtimes: { pi: piRuntime({ baseUrl, wireProtocol, models: [{ id: 'gpt-6-astra', name: 'Astra' }] }) },
@@ -420,8 +420,12 @@ describe('buildPiNativeProvidersFromConfigs', () => {
expect(model).toMatchObject({
contextWindow: 1_050_000, maxTokens: 128_000, reasoning: true,
input: ['text', 'image'], thinkingLevelMap: { off: 'low', max: 'max' },
- cost: { input: 10, output: 50, cacheRead: 1, cacheWrite: 12.5 },
});
+ if (baseUrl === 'https://api.openai.com/v1') {
+ expect(model?.cost).toEqual({ input: 10, output: 50, cacheRead: 1, cacheWrite: 12.5 });
+ } else {
+ expect(model?.cost).toBeUndefined();
+ }
expect(model?.baseUrl).toBeUndefined();
} else {
expect(model?.thinkingLevelMap).toBeUndefined();
@@ -2597,7 +2601,8 @@ describe('buildPiNativeProvidersFromConfigs', () => {
expect(providers[0].models[0]).toEqual({
id,
name,
- contextWindow: undefined,
+ contextWindow: id === 'deepseek-v4-pro' ? 1_000_000 : undefined,
+ ...(id === 'deepseek-v4-pro' ? { maxTokens: 384_000, input: ['text'] } : {}),
...(visual ? { input: ['text', 'image'] } : {}),
reasoning: true,
thinkingLevelMap: {
@@ -2608,8 +2613,13 @@ describe('buildPiNativeProvidersFromConfigs', () => {
xhigh: null,
max: 'max',
},
- // 显式声明能力但无同源元数据,Chat Completions 默认收敛 system role(#3832)。
- compat: { supportsDeveloperRole: false },
+ // Exact manufacturer adapters apply to compatible relays; unknown
+ // models retain the Chat Completions system-role fallback.
+ compat: id === 'deepseek-v4-pro' ? {
+ supportsDeveloperRole: false, supportsStore: false,
+ maxTokensField: 'max_tokens', requiresReasoningContentOnAssistantMessages: true,
+ thinkingFormat: 'deepseek',
+ } : { supportsDeveloperRole: false },
});
});
});
@@ -2835,3 +2845,37 @@ describe('Anthropic compat per-link pruning (#4982)', () => {
expect(resolvePiCindyGatewayModelSpec('xd', 'moonshot/kimi-k3')?.compat).toMatchObject({ thinkingFormat: 'openai' });
});
});
+
+
+it('carries Sub2API capacity, images, efforts and Fast from discovery into native Pi launch', () => {
+ const models = mergeDiscoveredRuntimeModels([], parseModelsListResponse({ models: [{
+ slug: 'private-model', context_window: 272000, max_context_window: 1050000,
+ input_modalities: ['text', 'image'], service_tiers: [{ id: 'priority' }],
+ supported_reasoning_levels: [{ effort: 'high' }, { effort: 'max' }],
+ }] })!);
+ const config = { id: 'sub2api-test', name: 'Sub2API', runtimes: { pi: {
+ baseUrl: 'https://relay.example/v1', wireProtocol: 'openai-responses' as const, models,
+ } } };
+ const provider = buildUserProvider(config);
+ const result = buildPiNativeProvidersFromConfigs([config], () => 'fixture-key', undefined, undefined,
+ { ...BUNDLED_CATALOG, providers: [provider] });
+ expect(result.providers[0]?.models[0]).toMatchObject({ id: 'private-model',
+ contextWindow: 272000, input: ['text', 'image'], supportsFastMode: true, reasoning: true,
+ thinkingLevelMap: { high: 'high', max: 'max' } });
+});
+
+
+it('materializes a future GPT generation with inherited parameters in native Pi', () => {
+ const config = { id: 'future-sub2api', name: 'Sub2API', runtimes: { pi: {
+ baseUrl: 'https://relay.example/v1', wireProtocol: 'openai-responses' as const,
+ models: mergeDiscoveredRuntimeModels([], parseModelsListResponse({ data: [{ id: 'gpt-9-sol' }] })!),
+ } } };
+ const provider = buildUserProvider(config, { modelRegistry: BUNDLED_CATALOG.modelRegistry });
+ const result = buildPiNativeProvidersFromConfigs([config], () => 'fixture-key', undefined, undefined,
+ { ...BUNDLED_CATALOG, providers: [provider] });
+ expect(result.providers[0]).toMatchObject({ id: 'future-sub2api', baseUrl: 'https://relay.example/v1', api: 'openai-responses' });
+ expect(result.providers[0]?.models[0]).toMatchObject({ id: 'gpt-9-sol',
+ contextWindow: 1050000, maxTokens: 128000, input: ['text', 'image'], reasoning: true,
+ thinkingLevelMap: { low: 'low', medium: 'medium', high: 'high' } });
+ expect(result.providers[0]?.models[0]?.cost).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 599d733d3b..e19ca6145e 100644
--- a/apps/desktop/src/main/maker-host/__tests__/piProviderTransport.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/piProviderTransport.test.ts
@@ -1,6 +1,6 @@
import { describe, expect, it, vi } from 'vitest';
import * as openaiCompletions from '@earendil-works/pi-ai/api/openai-completions';
-import { PROVIDER_MODEL_CATALOG, BUNDLED_CATALOG, buildUserProvider } from '@cindy/model-providers';
+import { PROVIDER_MODEL_CATALOG, BUNDLED_CATALOG, buildUserProvider, parseModelsListResponse, mergeDiscoveredRuntimeModels } from '@cindy/model-providers';
import { createPiProviderFetch, hostCredentialEndpointAllowed, invocationModelRecord, nativeBridgeApiKey, NATIVE_ADAPTER_ERROR_BODY_LIMIT, readBoundedResponseText } from '../pi-provider-transport.js';
vi.mock('@earendil-works/pi-ai/api/openai-completions', async (importOriginal) => ({
@@ -13,6 +13,54 @@ const reply = [
].map(value => `data: ${JSON.stringify(value)}\n\n`).join('') + 'data: [DONE]\n\n';
describe('Pi-owned transport for Cindy harnesses', () => {
+ it.each([true, false, undefined])('sends discovered Sub2API effort and declared Fast (%s) through the native Chat adapter', async supportsFastMode => {
+ const upstream = 'https://sub2api.example/custom/v1';
+ const models = parseModelsListResponse({ data: [{ id: 'grok-4.7',
+ supportsReasoningEffort: true, reasoningEffort: 'high',
+ context_window: 272000, max_context_window: 1050000, supports_fast_mode: supportsFastMode,
+ reasoningEfforts: ['low', 'medium', 'high', 'xhigh'].map(value => ({ value, label: value })),
+ }] })!;
+ const stored = JSON.parse(JSON.stringify(mergeDiscoveredRuntimeModels([], models)));
+ const provider = buildUserProvider({ id: 'sub2api', name: 'Sub2API', runtimes: {
+ codex: { baseUrl: upstream, wireProtocol: 'openai-chat', models: stored },
+ } });
+ const row = invocationModelRecord(provider.models.codex![0], upstream, 'openai-completions')!;
+ expect(row.contextWindow).toBe(1050000);
+ let sent: Record<string, unknown> | undefined;
+ const send = createPiProviderFetch({ row, providerId: provider.id, apiKey: 'fixture-key', fetchImpl: async (url, init) => {
+ expect(String(url)).toBe(`${upstream}/chat/completions`);
+ sent = JSON.parse(String(init?.body));
+ return new Response(reply, { headers: { 'content-type': 'text/event-stream' } });
+ } });
+ const response = await send('https://unused.invalid', { body: JSON.stringify({
+ model: row.id, input: 'hello', reasoning: { effort: 'xhigh' }, service_tier: 'priority', stream: true,
+ }) });
+ expect(await response.text()).toContain('response.completed');
+ expect(sent).toMatchObject({ model: 'grok-4.7', reasoning_effort: 'xhigh' });
+ expect(sent?.service_tier).toBe(supportsFastMode ? 'priority' : undefined);
+ });
+
+ it('sends inherited reasoning for a future model through the Responses serializer', async () => {
+ const upstream = 'https://relay.example/v1';
+ const provider = buildUserProvider({ id: 'future-relay', name: 'Relay', runtimes: {
+ codex: { baseUrl: upstream, wireProtocol: 'openai-responses', models:
+ mergeDiscoveredRuntimeModels([], parseModelsListResponse({ data: [{ id: 'gpt-9-sol' }] })!) },
+ } });
+ const row = invocationModelRecord(provider.models.codex![0], upstream, 'openai-responses')!;
+ let sent: Record<string, unknown> | undefined;
+ const send = createPiProviderFetch({ row, providerId: provider.id, apiKey: 'fixture-key', fetchImpl: async (url, init) => {
+ expect(String(url)).toBe(`${upstream}/responses`);
+ sent = JSON.parse(String(init?.body));
+ return new Response('data: {"type":"response.completed","response":{"id":"r","status":"completed","output":[],"usage":{"input_tokens":0,"output_tokens":0}}}\n\n',
+ { headers: { 'content-type': 'text/event-stream' } });
+ } });
+ await (await send('https://unused.invalid', { body: JSON.stringify({
+ model: row.id, input: 'hello', reasoning: { effort: 'high' }, stream: true,
+ }) })).text();
+ expect(sent).toMatchObject({ model: 'gpt-9-sol', reasoning: { effort: 'high' } });
+ expect(row.cost).toBeUndefined();
+ });
+
it('reconciles saved max when capabilities narrow, disappear and return', async () => {
const request = { model: 'changing-model', input: 'hello', reasoning: { effort: 'max' }, stream: true };
for (const efforts of [['high', 'max'], ['high'], [], ['high', 'max']] as const) {
diff --git a/apps/desktop/src/main/maker-host/__tests__/providerModelFetch.test.ts b/apps/desktop/src/main/maker-host/__tests__/providerModelFetch.test.ts
index c82bdc34cc..0842049b8b 100644
--- a/apps/desktop/src/main/maker-host/__tests__/providerModelFetch.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/providerModelFetch.test.ts
@@ -58,6 +58,21 @@ describe('import discovery limits', () => {
});
describe('buildModelsFetchRequest', () => {
+ it('discovers Sub2API Codex capabilities at the backend path', async () => {
+ const result = await fetchProviderModels(spec({ agent: 'codex',
+ baseUrl: 'https://sub2api.example/backend-api/codex', wireProtocol: 'openai-responses',
+ }), async (url, init) => {
+ expect(url).toBe('https://sub2api.example/backend-api/codex/models');
+ expect(init?.headers).toMatchObject({ authorization: 'Bearer sk-test' });
+ return fakeResponse(200, JSON.stringify({ models: [{ slug: 'gpt-6-sol',
+ supported_reasoning_levels: [{ effort: 'high' }, { effort: 'max' }], default_reasoning_level: 'high',
+ }] }));
+ });
+ expect(result).toMatchObject({ ok: true, models: [{ id: 'gpt-6-sol',
+ discoveredMetadata: { efforts: ['high', 'max'], defaultEffort: 'high' },
+ }] });
+ });
+
it.each(['claude-code', 'codex', 'pi'] as const)('uses the complete OpenRouter catalog for %s without Anthropic ID rewriting', async (agent) => {
const request = buildModelsFetchRequest(spec({
agent,
@@ -441,3 +456,23 @@ it('does not rewrite private Google-compatible discovery or an explicit catalog'
expect(request.init.headers).toMatchObject({ 'x-goog-api-key': 'sk-test' });
}
});
+
+it.each([404, 405])('falls back from an unavailable manifest (%i) on the same authenticated endpoint', async status => {
+ const fetcher = vi.fn()
+ .mockResolvedValueOnce(fakeResponse(status, '{}'))
+ .mockResolvedValueOnce(fakeResponse(200, '{"data":[{"id":"grok-4.7","reasoningEfforts":["low","high"]}]}'));
+ const result = await fetchProviderModels(spec({ baseUrl: 'https://relay.example/proxy/v1',
+ modelsUrl: 'https://relay.example/proxy/v1/models?client_version=0.147.0&tenant=test',
+ }), fetcher);
+ expect(result).toMatchObject({ ok: true, models: [{ id: 'grok-4.7', discoveredMetadata: { efforts: ['low', 'high'] } }] });
+ expect(fetcher).toHaveBeenCalledTimes(2);
+ expect(fetcher.mock.calls[1][0]).toBe('https://relay.example/proxy/v1/models?tenant=test');
+ expect(fetcher.mock.calls[1][1].headers).toEqual(fetcher.mock.calls[0][1].headers);
+});
+
+it.each([401, 403, 429, 503])('does not disguise manifest authorization or availability failures (%i)', async status => {
+ const fetcher = vi.fn(async () => fakeResponse(status, '{}'));
+ const result = await fetchProviderModels(spec({ modelsUrl: 'https://relay.example/v1/models?client_version=0.147.0' }), fetcher);
+ expect(result.ok).toBe(false);
+ expect(fetcher).toHaveBeenCalledOnce();
+});
diff --git a/apps/desktop/src/main/maker-host/anthropic-compat-proxy-host.ts b/apps/desktop/src/main/maker-host/anthropic-compat-proxy-host.ts
index 47fcd0a53a..7caa1ca403 100644
--- a/apps/desktop/src/main/maker-host/anthropic-compat-proxy-host.ts
+++ b/apps/desktop/src/main/maker-host/anthropic-compat-proxy-host.ts
@@ -157,6 +157,7 @@ function attachClaudeProviderBridge(route: RoutingDecision, providerId: string,
const handler = createClaudeProviderBridge({
url: `${base.replace(/\/+$/, '')}/${requestPath.replace(/^\/+/, '')}`,
protocol, headers: route.headerOverride ?? {}, efforts: model.efforts,
+ supportsFastMode: model.supportsFastMode === true,
providerId: provider.id,
...(row && (protocol === 'openai-chat' || model.api) ? { model: row, nativeUpstream: base } : {}),
capabilities: {
diff --git a/apps/desktop/src/main/maker-host/claude-provider-bridge.ts b/apps/desktop/src/main/maker-host/claude-provider-bridge.ts
index b669e66bdc..36524ea50e 100644
--- a/apps/desktop/src/main/maker-host/claude-provider-bridge.ts
+++ b/apps/desktop/src/main/maker-host/claude-provider-bridge.ts
@@ -18,13 +18,16 @@ export function createClaudeProviderBridge(options: {
protocol: 'openai-chat' | 'openai-responses';
headers: Readonly<Record<string, string>>;
efforts: readonly Effort[];
+ supportsFastMode?: boolean;
capabilities?: ChatBridgeCapabilities;
model?: ProviderModelRecord;
providerId?: string;
nativeUpstream?: string;
fetchImpl: typeof fetch;
}): ResponsesBridgeHandler {
- const nativeFetch = options.model ? createPiProviderFetch({ row: options.model,
+ const nativeFetch = options.model ? createPiProviderFetch({ row: {
+ ...options.model, supportsFastMode: options.supportsFastMode ?? options.model.supportsFastMode,
+ },
providerId: options.providerId ?? 'custom',
upstream: options.nativeUpstream,
apiKey: nativeBridgeApiKey(options.headers),
@@ -42,7 +45,10 @@ export function createClaudeProviderBridge(options: {
}
if (nativeFetch) return nativeFetch(_url, init);
if (options.protocol === 'openai-responses') return options.fetchImpl(options.url, init);
- const translated = translateResponsesRequestWithContext(responses, { capabilities: options.capabilities });
+ const translated = translateResponsesRequestWithContext(responses, { capabilities: {
+ ...options.capabilities,
+ ...(options.supportsFastMode ? { passthroughFields: [...(options.capabilities?.passthroughFields ?? []), 'service_tier'] } : {}),
+ } });
const upstream = await options.fetchImpl(options.url, { ...init, body: JSON.stringify(normalizeProviderRequest(translated.request, { harness: 'claude-code', protocol: 'openai-chat', upstreamBase: options.url, model: responses.model }, { reasoningEffortAlreadyMapped: true })) });
if (!upstream.ok || !upstream.body) return upstream;
const translator = new ChatSseTranslator(responses.model, { toolContext: translated.toolContext });
@@ -107,6 +113,7 @@ export function createClaudeProviderBridge(options: {
maxOutputTokensSupported: true,
supportsReasoning: () => options.efforts.length > 0,
supportedReasoningEfforts: () => options.efforts,
+ ...(options.supportsFastMode ? { fastServiceTier: 'priority' } : {}),
}],
fetchImpl: upstreamFetch,
});
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 36d42a8a29..f23d56ab56 100644
--- a/apps/desktop/src/main/maker-host/codex-proxy-host.ts
+++ b/apps/desktop/src/main/maker-host/codex-proxy-host.ts
@@ -1192,7 +1192,9 @@ function createChatBridgeDecision(
|| (isOfficialAnthropicUpstream(standard.upstream) && standard.contextWindow >= 1_000_000))) {
appendCommaSeparatedHeaderToken(nativeHeaders, 'anthropic-beta', 'context-1m-2025-08-07');
}
- const nativeFetch = createPiProviderFetch({ row: standard, providerId,
+ const nativeFetch = createPiProviderFetch({ row: { ...standard,
+ supportsFastMode: selected?.supportsFastMode ?? standard.supportsFastMode,
+ }, providerId,
// Keep the catalog adapter while honoring the host assigned to this account.
upstream: buildRouteDecision(route.routing, null, 'codex', route.apiKey, route.oauthToken)?.upstreamOverride ?? route.routing.upstream,
apiKey: nativeBridgeApiKey(headers),
diff --git a/apps/desktop/src/main/maker-host/generic-oauth.ts b/apps/desktop/src/main/maker-host/generic-oauth.ts
index 77b23eed72..de38a0ac74 100644
--- a/apps/desktop/src/main/maker-host/generic-oauth.ts
+++ b/apps/desktop/src/main/maker-host/generic-oauth.ts
@@ -827,7 +827,9 @@ export function deriveModelsDiscoveryUrl(baseUrl: string): string {
url.hash = '';
let pathname = url.pathname;
while (pathname.length > 1 && pathname.endsWith('/')) pathname = pathname.slice(0, -1);
- url.pathname = /\/v\d+$/i.test(pathname)
+ // Codex-compatible relays (including Sub2API) expose /models directly under
+ // /backend-api/codex; inserting /v1 makes discovery fail while inference works.
+ url.pathname = /(?:\/v\d+|\/backend-api\/codex)$/i.test(pathname)
? `${pathname}/models`
: `${pathname === '/' ? '' : pathname}/v1/models`;
return url.toString();
diff --git a/apps/desktop/src/main/maker-host/pi-host.ts b/apps/desktop/src/main/maker-host/pi-host.ts
index 6eff3a48ae..6d6be04d7f 100644
--- a/apps/desktop/src/main/maker-host/pi-host.ts
+++ b/apps/desktop/src/main/maker-host/pi-host.ts
@@ -54,7 +54,7 @@ import {
runtimeCustomProviderId,
storedCustomProviderId,
} from '@cindy/model-providers';
-import { providerCatalogForPi, providerModelRecord } from '@cindy/model-providers';
+import { providerCatalogForPi, providerModelRecord, providerModelGenerationRecord } from '@cindy/model-providers';
import type {
Catalog,
ModelCost,
@@ -1279,7 +1279,9 @@ export function buildPiNativeProvidersFromConfigs(
const row = providerModelRecord(model.id, model.route?.baseUrl ?? rt.baseUrl,
model.api ?? model.piApi ?? (model.route ? model.route.wireProtocol : rt.wireProtocol),
!model.api && !model.piApi && !model.route)
- ?? ((model.api ?? model.piApi) ? providerPresetModelRecord(rt.catalogPresetId, model.id, model.api ?? model.piApi) : undefined);
+ ?? ((model.api ?? model.piApi) ? providerPresetModelRecord(rt.catalogPresetId, model.id, model.api ?? model.piApi) : undefined)
+ ?? providerModelGenerationRecord(model.id, model.route?.baseUrl ?? rt.baseUrl,
+ model.api ?? model.piApi ?? (model.route ? model.route.wireProtocol : rt.wireProtocol), rt.catalogPresetId);
if (!row || !PI_NATIVE_APIS.has(row.execution.pi.api as PiNativeApi)) return undefined;
return {
id: row.id, name: row.name, baseUrl: row.upstream,
@@ -1401,6 +1403,7 @@ export function buildPiNativeProvidersFromConfigs(
m.route && explicitRouteApi === modelApi ? explicitRoute?.baseUrl : undefined;
const spec = {
id: m.id,
+ ...(resolved?.supportsFastMode !== undefined ? { supportsFastMode: resolved.supportsFastMode } : {}),
...(m.api || m.piApi || modelApi !== providerApi ? { api: modelApi } : {}),
...(authMethod === 'oauth'
? { baseUrl: oauthProxyEndpoint! }
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 3d95ca7951..503de25350 100644
--- a/apps/desktop/src/main/maker-host/pi-provider-transport.ts
+++ b/apps/desktop/src/main/maker-host/pi-provider-transport.ts
@@ -12,12 +12,13 @@ import * as googleVertex from '@earendil-works/pi-ai/api/google-vertex';
import * as azureOpenaiResponses from '@earendil-works/pi-ai/api/azure-openai-responses';
import * as bedrockConverseStream from '@earendil-works/pi-ai/api/bedrock-converse-stream';
import * as mistralConversations from '@earendil-works/pi-ai/api/mistral-conversations';
-import { PI_REASONING_EFFORTS, PROVIDER_MODEL_CATALOG, providerEndpointBindings, providerModelRecord, providerModelAdapterId, providerPresetModelRecord, type CatalogModel, type ProviderModelRecord, type PiModelApi } from '@cindy/model-providers';
+import { PI_REASONING_EFFORTS, PROVIDER_MODEL_CATALOG, providerEndpointBindings, providerModelRecord, providerModelGenerationRecord, providerModelAdapterId, providerPresetModelRecord, type CatalogModel, type ProviderModelRecord, type PiModelApi } from '@cindy/model-providers';
export function invocationModelRecord(model: CatalogModel, upstream: string, api?: PiModelApi): ProviderModelRecord | undefined {
const selected = model.api ?? api;
const known = providerModelRecord(model.id, upstream, selected)
- ?? (selected ? providerPresetModelRecord(model.catalogPresetId, model.id, selected) : undefined);
+ ?? (selected ? providerPresetModelRecord(model.catalogPresetId, model.id, selected) : undefined)
+ ?? providerModelGenerationRecord(model.id, upstream, selected, model.catalogPresetId);
if (!selected && !known) return undefined;
return {
...(known ?? {}), id: model.id, name: model.name, upstream,
@@ -25,6 +26,7 @@ export function invocationModelRecord(model: CatalogModel, upstream: string, api
maxOutput: model.maxOutput ?? known?.maxOutput,
modalities: model.modalities ?? known?.modalities ?? { input: ['text'], output: ['text'] },
supportsImageInput: model.supportsImageInput ?? known?.supportsImageInput ?? false,
+ supportsFastMode: model.supportsFastMode,
reasoning: model.efforts.length > 0, efforts: model.efforts, defaultEffort: model.defaultEffort,
execution: { pi: { ...known?.execution.pi, api: selected ?? known!.execution.pi.api,
thinkingLevelMap: { ...known?.execution.pi.thinkingLevelMap, ...(model.reasoningRequired ? { off: null } : {}) },
@@ -270,6 +272,10 @@ export function createPiProviderFetch(options: PiProviderTransportOptions): type
...(!['google-generative-ai', 'google-vertex', 'bedrock-converse-stream'].includes(model.api)
? { fetch: diagnosticFetch } : {}),
signal, maxRetries: 0,
+ ...(options.row.supportsFastMode === true && request.service_tier === 'priority' &&
+ ['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,
maxTokens: typeof request.max_output_tokens === 'number' ? Math.min(request.max_output_tokens, model.maxTokens) : model.maxTokens,
});
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 ec4a424433..0da3f42b66 100644
--- a/apps/desktop/src/main/maker-host/provider-model-fetch.ts
+++ b/apps/desktop/src/main/maker-host/provider-model-fetch.ts
@@ -271,6 +271,14 @@ export async function fetchProviderModels(
return { ok: false, code: cls.code, detail: cls.detail };
}
if (!res.ok) {
+ // A configured Codex manifest may be unavailable on older compatible gateways.
+ // Retry only the same endpoint without its manifest selector; never retry auth failures.
+ const ordinaryUrl = new URL(url);
+ if ((res.status === 404 || res.status === 405) && ordinaryUrl.searchParams.has('client_version')) {
+ await res.body?.cancel().catch(() => undefined);
+ ordinaryUrl.searchParams.delete('client_version');
+ return fetchProviderModels({ ...spec, modelsUrl: ordinaryUrl.toString() }, fetchImpl);
+ }
let bodyText = '';
try {
bodyText = spec.responseByteLimit === undefined
diff --git a/apps/desktop/src/main/maker-ipc/__tests__/runtimeSelectionAxes.test.ts b/apps/desktop/src/main/maker-ipc/__tests__/runtimeSelectionAxes.test.ts
index 7352aa4443..cd385520eb 100644
--- a/apps/desktop/src/main/maker-ipc/__tests__/runtimeSelectionAxes.test.ts
+++ b/apps/desktop/src/main/maker-ipc/__tests__/runtimeSelectionAxes.test.ts
@@ -81,11 +81,11 @@ describe('applyRuntimeSelectionAxesWithRecovery', () => {
expect(order).toEqual(['effort', 'fast', 'commit']);
});
- it('applies only the explicitly requested Fast axis', async () => {
+ it.each(['codex', 'pi'] as const)('applies only the explicitly requested Fast axis for %s', async agentKind => {
const setEffort = vi.fn(async () => {});
const setFastMode = vi.fn(async () => {});
await applyRuntimeSelectionAxesWithRecovery({
- session: { agentKind: 'codex', setEffort, setFastMode },
+ session: { agentKind, setEffort, setFastMode },
effort: 'high',
fastMode: true,
applyEffort: false,
diff --git a/apps/desktop/src/main/maker-ipc/register.ts b/apps/desktop/src/main/maker-ipc/register.ts
index de15e722e5..514071bce5 100644
--- a/apps/desktop/src/main/maker-ipc/register.ts
+++ b/apps/desktop/src/main/maker-ipc/register.ts
@@ -17931,19 +17931,7 @@ export function registerMakerIpc(maker: Maker, options: RegisterMakerIpcOptions)
log.debug('set-fast-mode: session not found, no-op', { sessionId });
return remoteResponse;
}
- if (sess.agentKind === 'pi') {
- // Pi 的 ChatGPT 请求不从 pi 请求体携带 Fast,而是由上面的 session store
- // 在 compat-proxy 决策点闭包进 responses bridge prefs。到这里已经即时生效,
- // 无需向 pi RPC 再发一份不存在的 set_fast_mode 控制命令。
- await commitRuntimeAxisAfterPersistence({
- persist: persistFastMode,
- commit: commitFastMode,
- assertCanCommit: assertOwnerCurrent,
- });
- log.debug('set-fast-mode: pi responses bridge state updated', { sessionId, enabled });
- return remoteResponse;
- }
- if (sess.agentKind !== 'codex') {
+ if (sess.agentKind !== 'codex' && sess.agentKind !== 'pi') {
await commitRuntimeAxisAfterPersistence({
persist: persistFastMode,
commit: commitFastMode,
diff --git a/apps/desktop/src/main/maker-ipc/runtimeSelectionAxes.ts b/apps/desktop/src/main/maker-ipc/runtimeSelectionAxes.ts
index c2b3ae1404..4115b33bec 100644
--- a/apps/desktop/src/main/maker-ipc/runtimeSelectionAxes.ts
+++ b/apps/desktop/src/main/maker-ipc/runtimeSelectionAxes.ts
@@ -115,7 +115,7 @@ export async function applyRuntimeSelectionAxesWithRecovery(
await input.session.setEffort(input.effort as Effort);
input.assertCanCommit?.();
}
- if (applyFastMode && input.session.agentKind === 'codex') {
+ if (applyFastMode && (input.session.agentKind === 'codex' || input.session.agentKind === 'pi')) {
await input.session.setFastMode(input.fastMode);
input.assertCanCommit?.();
}
diff --git a/apps/desktop/src/renderer/__tests__/addProviderWizardPresetEntry.test.tsx b/apps/desktop/src/renderer/__tests__/addProviderWizardPresetEntry.test.tsx
index 41bbd489f1..8b6b349fa2 100644
--- a/apps/desktop/src/renderer/__tests__/addProviderWizardPresetEntry.test.tsx
+++ b/apps/desktop/src/renderer/__tests__/addProviderWizardPresetEntry.test.tsx
@@ -1347,3 +1347,32 @@ it('keeps legacy curated recommendations selected and newly discovered models un
const models = vi.mocked(createCustomProvider).mock.calls[0][0].runtimes.pi!.models;
expect(models.filter(model => model.defaultEnabled !== false).map(model => model.id)).toEqual(['recommended']);
});
+
+it('sets up Sub2API from one site address and retains discovered capabilities in all engines', async () => {
+ const preset = BUNDLED_CATALOG.presets!.find(p => p.id === 'sub2api')!;
+ window.electronAPI.maker.listProviderPresets = vi.fn(async () => ({ presets: [preset] }));
+ const models = parseModelsListResponse({ models: [{ slug: 'private-sol', display_name: 'Private Sol',
+ supported_reasoning_levels: [{ effort: 'low' }, { effort: 'high' }], default_reasoning_level: 'high',
+ context_window: 272000, max_context_window: 1050000,
+ input_modalities: ['text', 'image'], service_tiers: [{ id: 'priority' }],
+ }] })!;
+ window.electronAPI.maker.fetchProviderModels = vi.fn(async () => ({ ok: true, models }));
+ vi.mocked(createCustomProvider).mockResolvedValue({ ok: true });
+ renderWizard('sub2api');
+ await screen.findByDisplayValue('Sub2API');
+ fireEvent.change(screen.getAllByDisplayValue('https://{endpoint}/v1')[0], { target: { value: 'https://relay.example/team' } });
+ fireEvent.change(screen.getByPlaceholderText('sk-…'), { target: { value: 'sk-test' } });
+ fireEvent.click(screen.getByText('settings.providers.wizard.next'));
+ fireEvent.click(await screen.findByText('Private Sol'));
+ fireEvent.click(screen.getByText('settings.providers.wizard.finish'));
+ await waitFor(() => expect(createCustomProvider).toHaveBeenCalledOnce());
+ const config = vi.mocked(createCustomProvider).mock.calls[0][0];
+ for (const agent of ['claude-code', 'codex', 'pi'] as const) {
+ const rt = config.runtimes[agent]!;
+ expect(rt.baseUrl).toBe('https://relay.example/team/v1');
+ expect(rt.modelsUrl).toBe('https://relay.example/team/v1/models?client_version=0.147.0');
+ expect(rt.catalogPresetId).toBe('sub2api');
+ expect(rt.models[0].discoveredMetadata).toMatchObject({ efforts: ['low', 'high'],
+ contextWindow: 272000, contextWindowMax: 1050000, supportsFastMode: true, supportsImageInput: true });
+ }
+});
diff --git a/docs/dev-rules/model-catalog-runtime.md b/docs/dev-rules/model-catalog-runtime.md
index e61acbfe47..d1d580609e 100644
--- a/docs/dev-rules/model-catalog-runtime.md
+++ b/docs/dev-rules/model-catalog-runtime.md
@@ -213,3 +213,43 @@ Desktop 本机目录只有在 providers、capabilities 和当前账号的模型
Server 对应入口为 `model-access-server/src/routes/modelCatalog.ts`、`services/catalogSource.ts` 与本仓独立维护的 contracts。
先确认 Server 分支是否具备目标能力;不要把本仓解析器直接视为 Server 已部署的实现。
两仓边界遵守 [协议兼容规则](protocol-compatibility.md)。
+
+### Sub2API 与自定义模型发现
+
+- Sub2API 在供应商预设中提供自托管入口;地址与 API Key 由用户填写,不预填公共站点、
+ 账号或模型清单。三个引擎共用该站点的 Responses 入口,协议兼容开关仍遵循现有用户偏好。
+- `https://{endpoint}/v1` 是整段自托管地址模板,允许 HTTP(S)、端口和部署前缀;绑定后
+ 模型发现 URL 必须落在相同站点。普通站点地址补 `/v1`,已给出的 `/v1` 与
+ `/backend-api/codex` 保留。未完成的模板与内嵌凭证不得发请求或保存。
+- 预设查询 `models?client_version=0.147.0`,取得 Sub2API 的 Codex 格式能力清单;此参数
+ 仅协商清单格式,不代表使用者运行该版本 Codex。404/405 时仅去掉该参数查询普通列表,
+ 不对鉴权、限流或服务异常做静默回退。新预设随客户端 bundled 合并,旧线上目录缺少该
+ 条目也不会遮掉它;不改变已有自定义连接地址。
+- 通用解析器兼容 `supported_reasoning_levels/default_reasoning_level` 与 Grok 的
+ `reasoningEfforts/reasoningEffort/supportsReasoningEffort`;同时读取 `input_modalities`、
+ `service_tiers` 中的 `priority` 和 `max_context_window`。未声明或非法值保持未知,空列表
+ 与显式 false 保持关闭,不根据供应商品牌猜能力。
+- `max_context_window` 仅作为连接实报的 `discoveredMetadata.contextWindowMax` 持久化并投影
+ 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。
+- 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。
+
+
+### 自定义供应商的新代际默认继承
+
+导入、刷新与读取已保存连接均在 `buildUserProvider` 中重算缺项,不把继承值写成实报或用户覆盖。
+同系列、同变体的数字代际按数值比较(含小版本),只从较旧代际继承;Sol / Luna、不同私有命名空间不互借。
+同协议的公共型号资料、同供应商适配表和同连接旧型号提供候选;较近代际优先,同代际连接配置优先。
+连接候选必须保持相同端点、请求路径与协议。预设绑定的账号地址继续使用自己的地址。
+
+继承窗口/输出上限、模态、思考档位/默认值、Fast/工具能力,以及同协议适配器的
+`thinkingLevelMap` / `compat` / `samplingParams`;不继承模型身份、价格、地址、凭证、headers、
+开关或成员资格。型号自身配置、明确实报(包括 false / [] / null)和用户覆盖高于继承层。
+继承默认档不会压过新型号实报的推荐档;既有型号目录默认优先规则保持不变。
+未来型号沿已实现的 API 序列化发送自身 ID。新协议仍需实现,不能靠继承配置获得新传输能力。
+Desktop、Mobile 和远程选择器共同消费执行端目录,不增加 UI 提示或用户步骤。
diff --git a/docs/product-rules/model-metadata-precedence.md b/docs/product-rules/model-metadata-precedence.md
index 49c3c58fd9..40266e4b7d 100644
--- a/docs/product-rules/model-metadata-precedence.md
+++ b/docs/product-rules/model-metadata-precedence.md
@@ -42,7 +42,8 @@ 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。**能力有依据才声明,资料缺失交给供应商默认,不把猜测变成配置。**
+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。
@@ -66,7 +67,7 @@ Pi 明确声明的 `reasoningDefaultEffort` 属于 Harness 默认例外,高于
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。** 新型号先按未知能力处理;目录或接口补齐资料后,
+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。
@@ -80,6 +81,22 @@ Pi 明确声明的 `reasoningDefaultEffort` 属于 Harness 默认例外,高于
[Codex 子 Agent 目录](../../apps/desktop/src/main/maker-host/__tests__/codex-smart-subagent-routing.test.ts)。
这些测试各自证明对应路径,不以目录或 UI 测试通过代替实际请求参数验证。
+### 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。
+
+- 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。
+- 可继承窗口、输出上限、模态、推理档位及默认值、Fast/工具能力和同协议参数映射。
+ 型号自身资料、明确实报(含 `false`、空数组、`null`)和用户覆盖仍优先。
+- 继承结果参与选择器和实际请求适配,但不写成实报或用户 override;不继承身份、价格、地址、
+ 凭证、请求头、开关或成员资格,也不新增上游尚未返回的模型,不凭继承创造新传输协议。
+
+实现与回归边界见[运行时细则](../dev-rules/model-catalog-runtime.md#自定义供应商的新代际默认继承)
+和[代际继承测试](../../packages/model-providers/src/__tests__/modelGeneration.test.ts)。
+
## 字段归属与默认值
| 位置 | 允许表达的内容 | 易混淆的边界 |
diff --git a/packages/maker-core/src/agents/base-agent.ts b/packages/maker-core/src/agents/base-agent.ts
index b5a0585b12..6362afcf4f 100644
--- a/packages/maker-core/src/agents/base-agent.ts
+++ b/packages/maker-core/src/agents/base-agent.ts
@@ -218,6 +218,8 @@ export interface PiNativeModelCost {
/** BYOM:写进 pi models.json 的一个模型(原生 provider 块内)。 */
export interface PiNativeModelSpec {
+ /** Current connection's explicit support for OpenAI priority service tier. */
+ supportsFastMode?: boolean;
/** Cindy/public model id used by provider-aware routing and the UI. */
id: string;
/** PI provider's native model id; omitted when it is identical to id. */
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 6d7f455dd2..434fd87730 100644
--- a/packages/maker-core/src/agents/pi/__tests__/cindyBridgeSource.test.ts
+++ b/packages/maker-core/src/agents/pi/__tests__/cindyBridgeSource.test.ts
@@ -28,6 +28,26 @@ import {
CINDY_PI_BASH_MAX_TIMEOUT_SECONDS,
} from '../cindy-bridge-source.js';
+it('strips Fast control paths from shell environments without mutating the runner environment', () => {
+ const source = CINDY_BRIDGE_EXTENSION_SOURCE;
+ const start = source.indexOf('const SECRET_ENV_NAMES =');
+ const end = source.indexOf('function isolatedBashEnvironment', start);
+ const compiled = ts.transpileModule(source.slice(start, end) + '\nresult = withoutPiSecrets(input);', {
+ compilerOptions: { target: ts.ScriptTarget.ES2022 },
+ }).outputText;
+ const input = { PATH: '/fixture/bin', CINDY_PI_MODEL_REQUEST_PREFS_FILE: '/fixture/prefs.json' };
+ const sandbox = {
+ process: { env: {} }, input, result: undefined,
+ PI_PACKAGE_MANAGEMENT_ENV: 'CINDY_PI_PACKAGE_MANAGEMENT',
+ PI_BASH_PACKAGE_HOME_ENV: 'CINDY_PI_BASH_PACKAGE_HOME',
+ MANAGED_RG_PATH_ENV: 'CINDY_PI_MANAGED_RG_PATH',
+ SUBAGENT_RUN_DIR_ENV: 'CINDY_PI_SUBAGENT_RUN_DIR',
+ };
+ runInNewContext(compiled, sandbox);
+ expect(sandbox.result).toEqual({ PATH: '/fixture/bin' });
+ expect(input.CINDY_PI_MODEL_REQUEST_PREFS_FILE).toBe('/fixture/prefs.json');
+});
+
it('keeps text-only policy local and ordinary tools independent of host UI failures', async () => {
const source = CINDY_BRIDGE_EXTENSION_SOURCE;
const helperStart = source.indexOf('let textOnlyTurnActive = false');
@@ -2307,3 +2327,25 @@ describe('Pi same-turn library native mapping', () => {
expect(await callback!(event)).toBeUndefined();
});
});
+
+
+it('applies native Fast only to the exact declared connection and follows preference changes', () => {
+ const start = CINDY_BRIDGE_EXTENSION_SOURCE.indexOf('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' } });
+ 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(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 d5dc7a3a10..9215e1aeef 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 }], 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 + (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 } } },
}) + '\\n');
process.stdout.write(JSON.stringify({ type: 'agent_end' }) + '\\n');
process.stdout.write(JSON.stringify({ type: 'agent_settled' }) + '\\n');
@@ -154,8 +154,12 @@ 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,
@@ -195,7 +199,14 @@ 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 input = hostInput(f);
+ 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 result = await registered.execute!(
'tool-foreground',
{ agent: 'worker', task: 'write the fixture', model: 'claude-fable-5' },
@@ -210,6 +221,9 @@ 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 0c30394d48..f4a7e71640 100644
--- a/packages/maker-core/src/agents/pi/__tests__/cindySubagentRunner.test.ts
+++ b/packages/maker-core/src/agents/pi/__tests__/cindySubagentRunner.test.ts
@@ -976,6 +976,9 @@ 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 {
@@ -999,6 +1002,7 @@ 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,
@@ -1019,6 +1023,10 @@ 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')))
@@ -1048,6 +1056,7 @@ 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,
@@ -1067,6 +1076,12 @@ 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-provider-routing.test.ts b/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
index 871af071c9..a73a18e9e5 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
@@ -234,6 +234,32 @@ 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;
+ const agent = new PiAgent({
+ auth: { getState: async () => ({ authenticated: true, authSource: 'api-key' as const }),
+ triggerLogin: async () => ({ authenticated: true }), logout: async () => {}, getAuthEnv: async () => ({}) },
+ runtimeConfig: { endpoint: 'http://127.0.0.1:9' }, binaryPath: path.join(agentHome, 'pi'),
+ logger: noopLogger, resolvePiAgentHome: () => agentHome,
+ resolvePiNativeProviders: async () => ({ providers: [{ id: 'relay', name: 'Relay',
+ baseUrl: 'https://relay.example/v1', api: 'openai-responses' as const,
+ 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' }] });
+ await handle.setFastMode!(false);
+ expect(JSON.parse(readFileSync(file, 'utf8')).fast).toBe(false);
+ fast = false;
+ await handle.send({ type: 'user', content: 'hello' });
+ expect(JSON.parse(readFileSync(file, 'utf8')).fast).toBe(false);
+ 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)));
+ });
+
it("uses providerId as the primary key when duplicate model ids exist", async () => {
const authProviderIds: Array<string | null | undefined> = [];
const apiResolver = vi.fn((providerId: string | null | undefined) =>
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 316a9f84b0..d10c978522 100644
--- a/packages/maker-core/src/agents/pi/cindy-bridge-source.ts
+++ b/packages/maker-core/src/agents/pi/cindy-bridge-source.ts
@@ -112,6 +112,7 @@ const projectPiManagedCommandFailure = ${projectPiManagedCommandFailure.toString
const SECRET_ENV_NAMES = new Set<string>([
'CINDY_PI_SECRET_ENV_NAMES',
'CINDY_PI_PERMISSION_FILE',
+ 'CINDY_PI_MODEL_REQUEST_PREFS_FILE',
'CINDY_PI_TURN_TOOL_POLICY',
PI_PACKAGE_MANAGEMENT_ENV,
PI_BASH_PACKAGE_HOME_ENV,
@@ -3637,13 +3638,31 @@ function astraResponsesPayload(payload, model) {
return out;
}
+function nativeFastPayload(payload, model) {
+ if (!payload || typeof payload !== 'object' || Array.isArray(payload)) return undefined;
+ if (!model || !['openai-responses', 'azure-openai-responses', 'openai-completions'].includes(model.api)) return undefined;
+ 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; }
+}
+
${PI_NATIVE_PROVIDER_ADAPTER_SOURCE}
export default async function cindyBridge(pi: any) {
installTextOnlyTurnPolicy(pi);
await registerCindyNativeProviderAdapters(pi);
if (!currentPermissionState().reviewOnly) registerCindyQuestionTool(pi);
- pi.on('before_provider_request', (event, ctx) => astraResponsesPayload(event.payload, ctx.model));
+ pi.on('before_provider_request', (event, ctx) => {
+ const payload = astraResponsesPayload(event.payload, ctx.model) ?? event.payload;
+ return nativeFastPayload(payload, ctx.model) ?? payload;
+ });
const mcpGateway = new CindyMcpGateway();
// bash 隔离 home 经 resolveBashPackageHome 解析(首次加载读删 + 防篡改 stash,
// 扩展重载(#3070)经双重验证取回,而不是拿到 undefined 让 bash 永久 fail-closed)。
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 6c54c5262d..c3c9c0233d 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
@@ -772,6 +772,7 @@ 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 f5bececf24..2790a59621 100644
--- a/packages/maker-core/src/agents/pi/cindy-subagent-source.ts
+++ b/packages/maker-core/src/agents/pi/cindy-subagent-source.ts
@@ -794,6 +794,7 @@ 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) {
@@ -821,6 +822,7 @@ 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,
@@ -910,6 +912,11 @@ 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 e9493a3031..fee0db5245 100644
--- a/packages/maker-core/src/agents/pi/index.ts
+++ b/packages/maker-core/src/agents/pi/index.ts
@@ -1900,9 +1900,8 @@ export class PiAgent extends BaseAgent {
return {
switchModel: { supported: true },
availableModels: [],
- // Pi 的 ChatGPT 模型经 Desktop responses bridge 调用。Fast 状态由 host 按
- // sessionId 注入 bridge prefs,再映射为 Codex `service_tier: priority`;实际
- // 是否显示开关仍由目录里该 (provider, model, pi) 的 supportsFastMode 门控。
+ // ChatGPT 订阅由 Desktop bridge 注入 Fast;自定义直连由原生请求钩子读取
+ // 本运行时偏好,映射 service_tier: priority。两者均由该连接的能力声明门控。
hasFastMode: true,
effort: { supported: true },
effortLevels: [
@@ -3231,6 +3230,32 @@ 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;
+ 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;
+ };
let activeToolsDisabled = false;
let textOnlyPolicyReady = false;
const textOnlyInput = (text: string): string => {
@@ -3254,8 +3279,10 @@ export class PiAgent extends BaseAgent {
const cleanupRuntimeFiles = (): void => {
if (runtimeFilesCleaned) return;
runtimeFilesCleaned = true;
+ 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
@@ -3361,6 +3388,7 @@ export class PiAgent extends BaseAgent {
return run;
};
await writePermissionFile(requestedPermissionSnapshot);
+ await writeRequestPrefs();
// 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)。
@@ -5118,6 +5146,7 @@ export class PiAgent extends BaseAgent {
// 扩展照旧现读快照,能力不受影响。
CINDY_SUBAGENT_ENV.runtimeFile,
CINDY_SUBAGENT_ENV.ownerId,
+ 'CINDY_PI_MODEL_REQUEST_PREFS_FILE',
// 受管工具路径同属控制面:不得让获批 bash 改写/替换后影响后续自动放行的 grep/find。
...(managedRipgrepPath ? [PI_MANAGED_RG_PATH_ENV] : []),
]));
@@ -5154,6 +5183,7 @@ 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 } : {}),
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/
@@ -6637,6 +6667,7 @@ export class PiAgent extends BaseAgent {
async send(message: UserMessage, sendOpts?: SendOptions): Promise<void> {
rejectIfCancelled(sendOpts, 'send');
await waitForSessionRpcIdle();
+ await writeRequestPrefs();
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
@@ -7182,6 +7213,11 @@ export class PiAgent extends BaseAgent {
return run;
},
+ async setFastMode(enabled: boolean): Promise<void> {
+ if (reviewMode) return;
+ await writeRequestPrefs(enabled);
+ },
+
async setEffort(effort: Effort): Promise<void> {
if (reviewMode) return;
assertStartupEffortAllowed(activeEffortSnapshot, effort);
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 d19fb92896..b2f4d6ab08 100644
--- a/packages/maker-core/src/agents/pi/pi-subagent-runs.ts
+++ b/packages/maker-core/src/agents/pi/pi-subagent-runs.ts
@@ -2800,6 +2800,16 @@ 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/catalog/providers.json b/packages/model-providers/catalog/providers.json
index 79c0520438..f2d79aac14 100644
--- a/packages/model-providers/catalog/providers.json
+++ b/packages/model-providers/catalog/providers.json
@@ -427,6 +427,34 @@
}
],
"presets": [
+ {
+ "id": "sub2api",
+ "name": "Sub2API",
+ "docsUrl": "https://github.com/Wei-Shaw/sub2api",
+ "runtimes": {
+ "claude-code": {
+ "baseUrl": "https://{endpoint}/v1",
+ "baseUrlEditable": true,
+ "wireProtocol": "openai-responses",
+ "modelsUrl": "https://{endpoint}/v1/models?client_version=0.147.0",
+ "models": []
+ },
+ "codex": {
+ "baseUrl": "https://{endpoint}/v1",
+ "baseUrlEditable": true,
+ "wireProtocol": "openai-responses",
+ "modelsUrl": "https://{endpoint}/v1/models?client_version=0.147.0",
+ "models": []
+ },
+ "pi": {
+ "baseUrl": "https://{endpoint}/v1",
+ "baseUrlEditable": true,
+ "wireProtocol": "openai-responses",
+ "modelsUrl": "https://{endpoint}/v1/models?client_version=0.147.0",
+ "models": []
+ }
+ }
+ },
{
"id": "openrouter",
"name": "OpenRouter",
diff --git a/packages/model-providers/src/__tests__/modelDiscoveryChannels.test.ts b/packages/model-providers/src/__tests__/modelDiscoveryChannels.test.ts
index 68150d4f1c..ee96aae9b0 100644
--- a/packages/model-providers/src/__tests__/modelDiscoveryChannels.test.ts
+++ b/packages/model-providers/src/__tests__/modelDiscoveryChannels.test.ts
@@ -2,8 +2,62 @@ import { describe, expect, it } from 'vitest';
import { parseModelsListResponse } from '../modelDiscovery.js';
import { mergeDiscoveredRuntimeModels } from '../modelMetadataLayers.js';
import { buildUserProvider } from '../user-provider.js';
+import { BUNDLED_CATALOG } from '../catalog.js';
describe('shared provider discovery', () => {
+ 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 }] })!;
+ const refresh = parseModelsListResponse({ data: [{ id: 'private-model',
+ max_context_window: 2000000 }] })!;
+ expect(refresh[0].contextWindow).toBeUndefined();
+ expect(refresh[0].discoveredMetadata?.contextWindow).toBeUndefined();
+ const models = mergeDiscoveredRuntimeModels(original, refresh);
+ expect(models[0].discoveredMetadata).toMatchObject({ contextWindow: 272000, contextWindowMax: 2000000 });
+ 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: 272000, contextWindowMax: 2000000 });
+ }
+ });
+
+ it('shrinks a saved discovered working window when a max-only refresh reduces capacity', () => {
+ const original = mergeDiscoveredRuntimeModels([], parseModelsListResponse({ data: [
+ { id: 'private-model', context_window: 128000, max_context_window: 256000 },
+ ] })!);
+ const refreshed = mergeDiscoveredRuntimeModels(original, parseModelsListResponse({ data: [
+ { id: 'private-model', max_context_window: 64000 },
+ ] })!);
+ expect(original[0].discoveredMetadata?.contextWindow).toBe(128000);
+ expect(refreshed[0].discoveredMetadata).toMatchObject({ contextWindow: 64000, contextWindowMax: 64000 });
+ 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 },
+ } });
+ expect(provider.models[agent]?.[0]).toMatchObject({ contextWindow: 64000, contextWindowMax: 64000 });
+ }
+ });
+
+ 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,
+ });
+ }
+ }
+ const saved = mergeDiscoveredRuntimeModels([], discovered);
+ const refreshed = mergeDiscoveredRuntimeModels(saved, parseModelsListResponse({ data: [
+ { id: 'gpt-9-sol', max_context_window: 64000 },
+ ] })!);
+ expect(refreshed[0].discoveredMetadata).toMatchObject({ contextWindow: 32000, contextWindowMax: 64000 });
+ });
+
it('imports Vercel token prices, output capacity, image inputs and declared effort levels', () => {
const models = parseModelsListResponse({ data: [{ id: 'vendor/new', name: 'New', type: 'language',
context_window: 128000, max_tokens: 32000,
@@ -85,3 +139,130 @@ it('does not offer thinking when a complete parameter list explicitly excludes i
expect(models?.[1]?.discoveredMetadata?.efforts).toBeUndefined();
expect(models?.[2]?.discoveredMetadata?.efforts).toBeUndefined();
});
+
+
+describe('Sub2API Codex manifest reasoning', () => {
+ // Sub2API configuredCodexModelDescriptor uses slug, supported_reasoning_levels
+ // ({ effort, description } objects) and default_reasoning_level.
+ it.each(['models', 'data'] as const)('imports declared efforts from the %s envelope into all engines', (envelope) => {
+ for (const id of ['gpt-6-sol', 'gpt-6-luna', 'team-coding-model']) {
+ const models = parseModelsListResponse({ [envelope]: [{
+ ...(envelope === 'models' ? { slug: id } : { id }),
+ display_name: id,
+ supported_reasoning_levels: [{ effort: 'high', description: 'High' }, { effort: 'max', description: 'Maximum' }],
+ default_reasoning_level: 'high',
+ }] })!;
+ expect(models[0].discoveredMetadata).toMatchObject({ efforts: ['high', 'max'], defaultEffort: 'high' });
+ for (const agent of ['claude-code', 'codex', 'pi'] as const) {
+ for (const modelRegistry of [undefined, BUNDLED_CATALOG.modelRegistry]) {
+ const provider = buildUserProvider({ id: 'custom-sub2api', name: 'Sub2API', runtimes: {
+ [agent]: { baseUrl: 'https://relay.example/v1', models: mergeDiscoveredRuntimeModels([], models) },
+ } }, { modelRegistry });
+ expect(provider.models[agent]?.[0]).toMatchObject({ id, efforts: ['high', 'max'], defaultEffort: 'high' });
+ }
+ }
+ }
+ });
+
+ it('keeps absent, invalid, empty and none-only declarations distinct', () => {
+ const entries = [
+ {},
+ { supported_reasoning_levels: null },
+ { supported_reasoning_levels: [{ description: 'missing effort' }] },
+ { supported_reasoning_levels: [] },
+ { supported_reasoning_levels: [{ effort: 'none' }], default_reasoning_level: 'none' },
+ { supported_reasoning_levels: ['none', 'low', 'high'], default_reasoning_level: 'high' },
+ { supported_reasoning_levels: [{ effort: 'low' }, { effort: 'high' }], default_reasoning_level: null },
+ ];
+ const models = parseModelsListResponse({ models: entries.map((entry, i) => ({ slug: `model-${i}`, ...entry })) })!;
+ expect(models.map(model => model.discoveredMetadata?.efforts))
+ .toEqual([undefined, undefined, undefined, [], [], ['low', 'high'], ['low', 'high']]);
+ expect(models[4].discoveredMetadata?.defaultEffort).toBeNull();
+ expect(models[5].discoveredMetadata?.defaultEffort).toBe('high');
+ expect(models[6].discoveredMetadata?.defaultEffort).toBeNull();
+ });
+
+ it('preserves existing reasoning fields when both formats are present', () => {
+ const [model] = parseModelsListResponse({ data: [{ id: 'gpt-6-sol',
+ reasoning: { supportedEfforts: [], defaultEffort: null },
+ supported_reasoning_levels: [{ effort: 'high' }], default_reasoning_level: 'high',
+ }] })!;
+ expect(model.discoveredMetadata).toMatchObject({ efforts: [], defaultEffort: null });
+ });
+});
+
+describe('Sub2API Grok model discovery', () => {
+ it.each(['grok-4.6', 'grok-4.7', 'team-grok'])('imports and refreshes %s without rewriting user choices', id => {
+ const payload = { data: [{ id, supportsReasoningEffort: true, reasoningEffort: 'high',
+ reasoningEfforts: [
+ { value: 'low', label: 'Low' }, { value: 'medium', label: 'Medium' },
+ { value: 'high', label: 'High', default: true }, { value: 'xhigh', label: 'xHigh' },
+ ],
+ }] };
+ const discovered = parseModelsListResponse(payload)!;
+ expect(discovered[0].discoveredMetadata).toMatchObject({
+ efforts: ['low', 'medium', 'high', 'xhigh'], defaultEffort: 'high',
+ });
+ const models = mergeDiscoveredRuntimeModels([{ id, name: 'My model', discoveredMetadata: {} }], discovered);
+ // Refresh must repair old name-only imports and survive configuration serialization.
+ for (const agent of ['claude-code', 'codex', 'pi'] as const) {
+ const config = { id: 'sub2api', name: 'Relay', runtimes: {
+ [agent]: { baseUrl: 'https://relay.example/v1', models },
+ } };
+ const provider = buildUserProvider(JSON.parse(JSON.stringify(config)));
+ expect(provider.models[agent]?.[0]).toMatchObject({ id,
+ efforts: ['low', 'medium', 'high', 'xhigh'], defaultEffort: 'high',
+ });
+ }
+ const overridden = mergeDiscoveredRuntimeModels([{ id, name: id, reasoning: false }], discovered);
+ expect(buildUserProvider({ id: 'sub2api', name: 'Relay', runtimes: {
+ codex: { baseUrl: 'https://relay.example/v1', models: overridden },
+ } }).models.codex?.[0]).toMatchObject({ efforts: [], defaultEffort: null });
+ });
+
+ it('does not invent levels from the support flag and respects explicit disabling', () => {
+ const models = parseModelsListResponse({ data: [
+ { id: 'unknown', supportsReasoningEffort: true, reasoningEffort: 'high' },
+ { id: 'disabled', supportsReasoningEffort: false, reasoningEfforts: [{ value: 'high' }] },
+ { id: 'empty', supportsReasoningEffort: true, reasoningEfforts: [] },
+ { id: 'invalid', reasoningEfforts: [{ label: 'High' }] },
+ { id: 'strings', reasoningEfforts: ['low', 'high'], reasoningEffort: null },
+ { id: 'canonical', reasoning: { supportedEfforts: ['low'], defaultEffort: null },
+ reasoningEfforts: [{ value: 'high' }], reasoningEffort: 'high' },
+ ] })!;
+ expect(models.map(model => model.discoveredMetadata?.efforts))
+ .toEqual([undefined, [], [], undefined, ['low', 'high'], ['low']]);
+ expect(models[4].discoveredMetadata?.defaultEffort).toBeNull();
+ expect(models[5].discoveredMetadata?.defaultEffort).toBeNull();
+ });
+});
+
+describe('Sub2API manifest capabilities', () => {
+ it('preserves vision, Fast and capacity separately through save and refresh for every engine', () => {
+ const discovered = parseModelsListResponse({ models: [{ slug: 'private-model',
+ context_window: 272000, max_context_window: 1050000,
+ input_modalities: ['text', 'image'], service_tiers: [{ id: 'priority', name: 'Fast' }],
+ }] })!;
+ const saved = JSON.parse(JSON.stringify(mergeDiscoveredRuntimeModels([], discovered)));
+ const refreshed = mergeDiscoveredRuntimeModels(saved, parseModelsListResponse({ data: [{ id: 'private-model' }] })!);
+ 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 },
+ } });
+ expect(provider.models[agent]?.[0]).toMatchObject({ contextWindow: 272000,
+ contextWindowMax: 1050000, supportsFastMode: true, supportsImageInput: true });
+ }
+ });
+ it('distinguishes omitted, malformed, disabled and explicit canonical capability fields', () => {
+ const values = [ {}, { input_modalities: [null], service_tiers: [{}], max_context_window: -1 },
+ { input_modalities: [], service_tiers: [] },
+ { supports_image_input: false, supports_fast_mode: false,
+ input_modalities: ['image'], service_tiers: [{ id: 'priority' }] },
+ { context_window: 100000, max_context_window: 32000 },
+ ];
+ const models = parseModelsListResponse({ models: values.map((value, i) => ({ slug: `model-${i}`, ...value })) })!;
+ expect(models.map(m => m.discoveredMetadata?.supportsImageInput)).toEqual([undefined, undefined, false, false, undefined]);
+ expect(models.map(m => m.discoveredMetadata?.supportsFastMode)).toEqual([undefined, undefined, false, false, undefined]);
+ expect(models.every(m => m.discoveredMetadata?.contextWindowMax === undefined)).toBe(true);
+ });
+});
diff --git a/packages/model-providers/src/__tests__/modelGeneration.test.ts b/packages/model-providers/src/__tests__/modelGeneration.test.ts
new file mode 100644
index 0000000000..c4ee67f145
--- /dev/null
+++ b/packages/model-providers/src/__tests__/modelGeneration.test.ts
@@ -0,0 +1,218 @@
+import { describe, expect, it } from 'vitest';
+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 { parseModelsListResponse } from '../modelDiscovery.js';
+import { mergeDiscoveredRuntimeModels } from '../modelMetadataLayers.js';
+import type { ProviderRuntimeModelConfig, ProviderWireProtocol } from '../types.js';
+
+const build = (models: ProviderRuntimeModelConfig[], wireProtocol: ProviderWireProtocol = 'openai-responses') => {
+ const runtime = { baseUrl: 'https://relay.example/v1', wireProtocol, models };
+ return buildUserProvider({ id: 'my-sub2api', name: 'Relay', runtimes: {
+ codex: runtime, pi: runtime, 'claude-code': runtime,
+ } }, { modelRegistry: BUNDLED_CATALOG.modelRegistry });
+};
+const discovered = (id: string, metadata = {}) => mergeDiscoveredRuntimeModels([], parseModelsListResponse({ data: [{ id, ...metadata }] })!);
+
+describe('new model generation defaults', () => {
+ it('imports future GPT generations across all engines without requiring the old model in the account', () => {
+ for (const id of ['gpt-7-sol', 'gpt-8-sol', 'openai/gpt-9-luna']) {
+ const oldId = id.includes('luna') ? 'gpt-6-luna' : 'gpt-6-sol';
+ const previous = build(discovered(oldId)).models.codex![0]!;
+ const saved = discovered(id);
+ const provider = build(saved);
+ for (const agent of ['codex', 'pi', 'claude-code'] as const) {
+ const model = provider.models[agent]![0]!;
+ expect(model).toMatchObject({ id, name: id, efforts: previous.efforts,
+ defaultEffort: previous.defaultEffort, contextWindow: previous.contextWindow,
+ maxOutput: previous.maxOutput, supportsImageInput: true, contextWindowVerified: false });
+ expect(model.cost).toBeUndefined();
+ expect(model.userModelConfig).toEqual(saved[0]);
+ }
+ }
+ });
+
+ 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 } },
+ { id: 'private-7-sol', name: 'Seven', discoveredMetadata: { efforts: [], supportsFastMode: false } },
+ { id: 'private-8-sol', name: 'Eight' },
+ ]);
+ expect(provider.models.codex![2]).toMatchObject({ id: 'private-8-sol', name: 'Eight',
+ efforts: [], defaultEffort: null, contextWindow: 64000, supportsFastMode: false });
+ expect(provider.routing.codex?.upstream).toBe('https://relay.example/v1');
+ });
+
+ it('prefers a newer account declaration over the older public catalog', () => {
+ const models = build([
+ { id: 'gpt-8-sol', name: 'Eight', discoveredMetadata: { efforts: ['high'], defaultEffort: null,
+ contextWindow: 48000, maxOutputTokens: 4000, supportsFastMode: true } },
+ ...discovered('gpt-9-sol'),
+ ]).models.codex!;
+ expect(models[1]).toMatchObject({ efforts: ['high'], defaultEffort: null,
+ contextWindow: 48000, maxOutput: 4000, supportsFastMode: true });
+ });
+
+ it('refresh replaces inherited values, while target user settings survive', () => {
+ const initial = discovered('gpt-9-sol');
+ expect(build(initial).models.pi![0]!.efforts.length).toBeGreaterThan(0);
+ const refreshed = mergeDiscoveredRuntimeModels(initial, parseModelsListResponse({ data: [{ id: 'gpt-9-sol',
+ supported_reasoning_levels: ['low', 'high'], default_reasoning_level: 'high',
+ context_window: 32000, input_modalities: ['text'], service_tiers: [],
+ }] })!);
+ expect(build(refreshed).models.pi![0]).toMatchObject({ efforts: ['low', 'high'], defaultEffort: 'high',
+ contextWindow: 32000, supportsImageInput: false, supportsFastMode: false });
+ refreshed[0]!.contextWindow = 16000;
+ refreshed[0]!.reasoning = false;
+ expect(build(refreshed).models.pi![0]).toMatchObject({ efforts: [], defaultEffort: null, contextWindow: 16000 });
+ });
+
+ it('verifies only the target window across engines, including declarations equal to inherited defaults', () => {
+ const initial: [ProviderRuntimeModelConfig, ProviderRuntimeModelConfig] = [
+ { id: 'private-6-sol', name: 'Old', discoveredMetadata: { contextWindow: 64000 } },
+ { id: 'private-7-sol', name: 'New' },
+ ];
+ for (const target of [
+ initial[1],
+ { ...initial[1], discoveredMetadata: { contextWindow: 64000 } },
+ { ...initial[1], contextWindow: 32000 },
+ ]) {
+ const provider = build([initial[0], target]);
+ for (const agent of ['codex', 'pi', 'claude-code'] as const) {
+ expect(provider.models[agent]![0]).toMatchObject({ contextWindowVerified: true });
+ expect(provider.models[agent]![1]).toMatchObject({
+ contextWindow: target.contextWindow ?? 64000,
+ contextWindowVerified: target !== initial[1],
+ });
+ }
+ }
+ });
+
+ it.each([
+ ['image_generation', 'imageModels'],
+ ['video_generation', 'videoModels'],
+ ['embedding', 'embeddingModels'],
+ ] as const)('does not inherit %s membership, but preserves the target declaration', (mode, field) => {
+ for (const declared of [false, true]) {
+ const provider = build([
+ { id: 'private-6-sol', name: 'Old', discoveredMetadata: { mode, contextWindow: 64000 } },
+ { id: 'private-7-sol', name: 'New', discoveredMetadata: declared ? { mode } : {} },
+ ]);
+ for (const agent of ['codex', 'pi', 'claude-code'] as const) {
+ const target = provider.models[agent]![1]!;
+ expect(target.contextWindow).toBe(64000);
+ expect(target.mode).toBe(declared ? mode : undefined);
+ expect(isChatEligible(target)).toBe(!declared);
+ }
+ expect(provider[field]?.map(model => model.id)).toEqual(
+ declared ? ['private-6-sol', 'private-7-sol'] : ['private-6-sol'],
+ );
+ }
+ });
+
+ it('does not inherit from another variant, protocol, private namespace or endpoint', () => {
+ const target = { id: 'private-9-sol', name: 'New' };
+ for (const source of [
+ { id: 'private-8-luna', name: 'Different variant' },
+ { id: 'team/private-8-sol', name: 'Private namespace' },
+ { id: 'private-8-sol', name: 'Other API', api: 'anthropic-messages' as const },
+ { id: 'private-8-sol', name: 'Other endpoint', route: { baseUrl: 'https://other.example/v1', wireProtocol: 'openai-responses' as const } },
+ ]) {
+ expect(build([{ ...source, reasoning: true, reasoningEfforts: ['high'] }, target]).models.pi![1]!.efforts).toEqual([]);
+ }
+ });
+
+ it('does not switch the adapter of an already known model while filling defaults', () => {
+ const model = build(discovered('claude-sonnet-4-6'), 'anthropic-messages').models.codex![0]!;
+ expect(model.api).toBeUndefined();
+ expect(model.route).toBeUndefined();
+ });
+
+ it('compares numeric generations and keeps model sizes and variants separate', () => {
+ const ids = ['gpt-9-sol', 'gpt-6-sol', 'gpt-10-sol', 'gpt-7-luna'];
+ expect(previousModelGenerations('gpt-11-sol', ids, id => id)).toEqual(['gpt-6-sol', 'gpt-9-sol', 'gpt-10-sol']);
+ expect(previousModelGenerations('grok-4.10', ['grok-4.9', 'grok-4.11'], id => id)).toEqual(['grok-4.9']);
+ expect(previousModelGenerations('claude-sonnet-6', ['claude-sonnet-4-6', 'claude-sonnet-5', 'claude-opus-5'], id => id)).toEqual(['claude-sonnet-4-6', 'claude-sonnet-5']);
+ expect(previousModelGenerations('qwen4-30b', ['qwen3-235b', 'qwen3-30b'], id => id)).toEqual(['qwen3-30b']);
+ });
+
+ 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);
+ const source = providerModelRecord('gpt-5.6-sol', endpoint, 'openai-responses');
+ expect(source).toBeDefined();
+ expect(providerModelRecord('gpt-5.6-sol', ` ${endpoint}${slashes} `, 'openai-responses')).toBe(source);
+ expect(providerModelRecord('gpt-5.6-sol', `${endpoint}${slashes}other`, 'openai-responses')).toBeUndefined();
+ });
+
+ it('uses the exact manufacturer adapter before a predecessor on compatible relays', () => {
+ const exact = providerModelRecord('gpt-5.4', 'https://api.openai.com/v1', 'openai-responses')!;
+ expect(exact).toBeDefined();
+ const relay = providerModelGenerationRecord('gpt-5.4', 'https://relay.example/v1', 'openai-responses')!;
+ const { headers: _headers, ...parameters } = exact.execution.pi;
+ expect(relay.inheritedFrom).toBe('gpt-5.4');
+ expect(relay.execution.pi).toEqual(parameters);
+ expect(relay.upstream).toBe('https://relay.example/v1');
+ expect(relay.cost).toBeUndefined();
+ expect(relay.execution.pi.headers).toBeUndefined();
+ expect(providerModelGenerationRecord('gpt-5.4', 'https://relay.example/v1', 'anthropic-messages')).toBeUndefined();
+ });
+
+ it('retains exact adapter metadata across engines even without Registry entries', () => {
+ const exact = providerModelRecord('grok-4.7', 'https://api.x.ai/v1', 'openai-responses')!;
+ expect(exact).toBeDefined();
+ const runtime = { baseUrl: 'https://relay.example/v1', wireProtocol: 'openai-responses' as const,
+ models: discovered('grok-4.7') };
+ const provider = buildUserProvider({ id: 'relay', name: 'Relay', runtimes: {
+ codex: runtime, pi: runtime, 'claude-code': runtime,
+ } }, { modelRegistry: { schemaVersion: 4, updatedAt: '2026-09-24T00:00:00Z', models: [] } });
+ for (const agent of ['codex', 'pi', 'claude-code'] as const) {
+ expect(provider.models[agent]![0]).toMatchObject({
+ contextWindow: exact.contextWindow, efforts: exact.efforts,
+ supportsImageInput: exact.supportsImageInput, contextWindowVerified: true,
+ });
+ }
+ });
+
+ it('clears inherited image input after text-only discovery without changing predecessor or output modalities', () => {
+ const source: ProviderRuntimeModelConfig = { id: 'private-6-sol', name: 'Old', discoveredMetadata: {
+ modalities: { input: ['text', 'image'], output: ['text', 'image'] }, supportsImageInput: true,
+ } };
+ const provider = build([source, ...discovered('private-7-sol', { input_modalities: ['text'] })]);
+ for (const agent of ['codex', 'pi', 'claude-code'] as const) {
+ expect(provider.models[agent]![1]).toMatchObject({ supportsImageInput: false,
+ modalities: { input: ['text'], output: ['text', 'image'] } });
+ expect(provider.models[agent]![0]).toMatchObject({ supportsImageInput: true,
+ modalities: { input: ['text', 'image'], output: ['text', 'image'] } });
+ }
+ expect(source.discoveredMetadata?.modalities?.input).toEqual(['text', 'image']);
+ });
+
+ it('drops only inherited capacity when the target declares a larger working window', () => {
+ for (const ownMax of [undefined, 256000]) {
+ const provider = build([
+ { id: 'private-6-sol', name: 'Old', discoveredMetadata: { contextWindow: 64000, contextWindowMax: 64000 } },
+ { id: 'private-7-sol', name: 'New', discoveredMetadata: { contextWindow: 128000,
+ ...(ownMax !== undefined ? { contextWindowMax: ownMax } : {}) } },
+ ]);
+ for (const agent of ['codex', 'pi', 'claude-code'] as const) {
+ const model = provider.models[agent]![1]!;
+ expect(model.contextWindow).toBe(128000);
+ expect(model.contextWindowMax).toBe(ownMax);
+ }
+ }
+ });
+
+ it('reuses serializer mappings without copying the predecessor identity, prices, endpoint or headers', () => {
+ const source = providerModelRecord('gpt-5.6-sol', 'https://api.openai.com/v1', 'openai-responses')!;
+ const inherited = providerModelGenerationRecord('gpt-9-sol', 'https://relay.example/v1', 'openai-responses')!;
+ expect(inherited).toMatchObject({ id: 'gpt-9-sol', name: 'gpt-9-sol', upstream: 'https://relay.example/v1', inheritedFrom: 'gpt-5.6-sol' });
+ expect(inherited.execution.pi.thinkingLevelMap).toEqual(source.execution.pi.thinkingLevelMap);
+ expect(inherited.execution.pi.compat).toEqual(source.execution.pi.compat);
+ expect(inherited.cost).toBeUndefined();
+ expect(inherited.execution.pi.headers).toBeUndefined();
+ expect(providerModelGenerationRecord('gpt-9-sol', 'https://relay.example/v1', 'anthropic-messages')).toBeUndefined();
+ });
+});
diff --git a/packages/model-providers/src/__tests__/modelMetadataValidation.test.ts b/packages/model-providers/src/__tests__/modelMetadataValidation.test.ts
index 753ff30f77..cd419e58ba 100644
--- a/packages/model-providers/src/__tests__/modelMetadataValidation.test.ts
+++ b/packages/model-providers/src/__tests__/modelMetadataValidation.test.ts
@@ -1,6 +1,6 @@
import { describe, expect, it } from "vitest";
import { parseModelRegistry } from "../modelAccessValidator.js";
-import { expandedRegistryEntries } from "../modelMetadataLayers.js";
+import { expandedRegistryEntries, validModelMetadata } from "../modelMetadataLayers.js";
import type { ModelRegistry } from "../modelAccessBean.js";
function fixture(): ModelRegistry {
@@ -22,6 +22,23 @@ function fixture(): ModelRegistry {
}
describe("V4 metadata validation boundaries", () => {
+ it.each([4, 5] as const)("rejects connection capacity in every V%s Registry metadata layer", (schemaVersion) => {
+ for (const layer of ["public", "defaults", "forceOverrides"] as const) {
+ const r = fixture();
+ r.schemaVersion = schemaVersion;
+ const metadata = { contextWindow: 272_000, contextWindowMax: 1_050_000 };
+ expect(validModelMetadata(metadata)).toBe(true);
+ if (layer === "public") r.baseModels![0].defaults = metadata;
+ else {
+ r.models[0].routes[0][layer] = metadata;
+ r.models[0].routes[0].overrideReason = "Correct supplier metadata";
+ }
+ expect(parseModelRegistry(r).ok).toBe(false);
+ Reflect.deleteProperty(metadata, "contextWindowMax");
+ expect(parseModelRegistry(r).ok).toBe(true);
+ }
+ });
+
it.each([1, 247, 248, 256])(
"keeps split routes valid with a %i-character ID and reserved ID collisions",
(length) => {
diff --git a/packages/model-providers/src/__tests__/providerEndpointTemplate.test.ts b/packages/model-providers/src/__tests__/providerEndpointTemplate.test.ts
index 9318d8e7bb..048dc576b8 100644
--- a/packages/model-providers/src/__tests__/providerEndpointTemplate.test.ts
+++ b/packages/model-providers/src/__tests__/providerEndpointTemplate.test.ts
@@ -134,3 +134,23 @@ describe('Vertex official endpoint family', () => {
}
});
});
+
+describe('Sub2API self-hosted setup', () => {
+ const preset = BUNDLED_CATALOG.presets!.find(p => p.id === 'sub2api')!;
+ it.each([
+ ['https://relay.example', 'https://relay.example/v1'],
+ ['https://relay.example/proxy/v1/', 'https://relay.example/proxy/v1'],
+ ['http://localhost:8080', 'http://localhost:8080/v1'],
+ ['https://relay.example/backend-api/codex', 'https://relay.example/backend-api/codex'],
+ ])('binds %s without sending discovery to the template host', (endpoint, expected) => {
+ for (const runtime of Object.values(preset.runtimes)) {
+ const bound = bindProviderPresetRuntime(runtime!, endpoint);
+ expect(bound.baseUrl).toBe(expected);
+ expect(bound.modelsUrl).toBe(`${expected}/models?client_version=0.147.0`);
+ expect(bound.models).toEqual([]);
+ }
+ });
+ it.each(['https://{endpoint}/v1', 'https://user:[email protected]/v1', 'file:///tmp/models', 'https://relay.example/v1?key=secret'])('rejects unsafe or unresolved endpoint %s', endpoint => {
+ expect(providerEndpointBindings(preset.runtimes.codex!.baseUrl, endpoint)).toBeNull();
+ });
+});
diff --git a/packages/model-providers/src/__tests__/user-provider.test.ts b/packages/model-providers/src/__tests__/user-provider.test.ts
index 8003ba5e6e..b3104f1b80 100644
--- a/packages/model-providers/src/__tests__/user-provider.test.ts
+++ b/packages/model-providers/src/__tests__/user-provider.test.ts
@@ -306,6 +306,53 @@ describe("buildUserProvider (per-runtime)", () => {
});
});
+ it.each(["claude-code", "codex", "pi"] as const)(
+ "%s inherits GPT-6 Sol/Luna reasoning for an unknown custom supplier",
+ (agent) => {
+ for (const id of ["gpt-6-sol", "gpt-6-luna", "openai/gpt-6-sol", "openai/gpt-6-luna"]) {
+ const config: CustomProviderConfig = {
+ id: "custom-xdtai",
+ name: "XDTAI",
+ runtimes: {
+ [agent]: {
+ baseUrl: "https://custom.example/v1",
+ models: [{ id, name: id, discoveredMetadata: { supportsImageInput: true } }],
+ },
+ },
+ };
+ const before = structuredClone(config);
+ const oldRegistry = structuredClone(BUNDLED_CATALOG.modelRegistry!);
+ oldRegistry.baseModels = oldRegistry.baseModels?.filter(
+ (model) => !["openai/gpt-6-sol", "openai/gpt-6-luna"].includes(model.id),
+ );
+ oldRegistry.models = oldRegistry.models.filter(
+ (model) => !["openai/gpt-6-sol", "openai/gpt-6-luna"].includes(model.modelRef ?? model.id),
+ );
+ // Without current-generation catalog data, compatible Responses routes now
+ // inherit the previous generation; Messages/Chat do not borrow its protocol.
+ expect(buildUserProvider(config, { modelRegistry: oldRegistry }).models[agent]?.[0])
+ .toMatchObject(agent === 'codex'
+ ? { efforts: ['low', 'medium', 'high', 'xhigh', 'max'], defaultEffort: 'medium' }
+ : { efforts: [], defaultEffort: null });
+ const provider = buildUserProvider(config, { modelRegistry: BUNDLED_CATALOG.modelRegistry });
+ expect(provider.models[agent]?.[0]).toMatchObject({
+ id,
+ efforts: ["low", "medium", "high", "xhigh", "max"],
+ defaultEffort: "medium",
+ });
+ expect(provider.routing[agent]?.upstream).toBe("https://custom.example/v1");
+ expect(config).toEqual(before);
+ const stored = config.runtimes[agent]!.models[0];
+ stored.discoveredMetadata = { efforts: ["low", "high"] };
+ expect(buildUserProvider(config, { modelRegistry: BUNDLED_CATALOG.modelRegistry }).models[agent]?.[0])
+ .toMatchObject({ efforts: ["low", "high"], defaultEffort: "low" });
+ stored.reasoning = false;
+ expect(buildUserProvider(config, { modelRegistry: BUNDLED_CATALOG.modelRegistry }).models[agent]?.[0])
+ .toMatchObject({ efforts: [], defaultEffort: null });
+ }
+ },
+ );
+
it.each(["claude-code", "codex", "pi"] as const)(
"%s leaves unknown reasoning unspecified and preserves declared capabilities",
(agent) => {
diff --git a/packages/model-providers/src/index.ts b/packages/model-providers/src/index.ts
index 6b6356304a..06cec933c2 100644
--- a/packages/model-providers/src/index.ts
+++ b/packages/model-providers/src/index.ts
@@ -313,7 +313,7 @@ export { runtimeUserModelMetadata } from "./modelMetadataLayers.js";
export { PROVIDER_MEDIA_FIELDS, providerMediaField, projectProviderMediaModels } from "./providerMediaModels.js";
-export { PROVIDER_MODEL_CATALOG, providerModelRecord, providerModelAdapterId, providerPresetModelRecord, providerModelMetadata, providerCatalogForPi, providerModelsForRoute } from "./providerModelCatalog.js";
+export { PROVIDER_MODEL_CATALOG, providerModelRecord, providerModelGenerationRecord, providerModelAdapterId, providerPresetModelRecord, providerModelMetadata, providerCatalogForPi, providerModelsForRoute } from "./providerModelCatalog.js";
export { parseModelsListResponse, isOpenRouterModelsUrl } from "./modelDiscovery.js";
diff --git a/packages/model-providers/src/modelAccessValidator.ts b/packages/model-providers/src/modelAccessValidator.ts
index 39e90812e0..8ca740393c 100644
--- a/packages/model-providers/src/modelAccessValidator.ts
+++ b/packages/model-providers/src/modelAccessValidator.ts
@@ -1,6 +1,7 @@
import {
validModelMetadata,
expandedRegistryEntries,
+ type ModelMetadata,
} from "./modelMetadataLayers.js";
import { parseLocalModelCatalog } from "./localModelCatalog.js";
import {
@@ -999,6 +1000,11 @@ function referencePriceGroupsError(value: unknown): string | null {
return null;
}
+// Capacity is connection-local metadata, never a public Registry default.
+function validRegistryMetadata(value: unknown): value is ModelMetadata {
+ return validModelMetadata(value) && !("contextWindowMax" in value);
+}
+
function registryRouteError(
value: unknown,
path: string,
@@ -1022,7 +1028,7 @@ function registryRouteError(
);
if (error) return error;
for (const field of ["defaults", "forceOverrides"] as const) {
- if (value[field] !== undefined && !validModelMetadata(value[field]))
+ if (value[field] !== undefined && !validRegistryMetadata(value[field]))
return `${path}.${field} is invalid`;
}
if (
@@ -1361,7 +1367,7 @@ export function parseModelRegistry(
!base.aliases.every(
(a) => typeof a === "string" && a.length > 0 && a.length <= 256,
) ||
- !validModelMetadata(base.defaults)
+ !validRegistryMetadata(base.defaults)
)
return fail("modelRegistry.baseModels contains invalid data");
if (base.referencePriceGroups !== undefined) {
diff --git a/packages/model-providers/src/modelDiscovery.ts b/packages/model-providers/src/modelDiscovery.ts
index 2ce4087d4a..9422510212 100644
--- a/packages/model-providers/src/modelDiscovery.ts
+++ b/packages/model-providers/src/modelDiscovery.ts
@@ -11,6 +11,19 @@ export function isOpenRouterModelsUrl(value: string): boolean {
} catch { return false; }
}
+/** Codex manifests use { effort, description }; Sub2API's Grok list uses { value, label }.
+ * Missing/malformed lists remain unknown; an explicit empty or none-only list stays empty.
+ */
+function declaredReasoningEfforts(value: unknown, key: "effort" | "value") {
+ if (!Array.isArray(value)) return undefined;
+ const levels = value.map((level: unknown) =>
+ typeof level === "string" ? level
+ : level && typeof level === "object" ? (level as Record<string, unknown>)[key]
+ : undefined,
+ );
+ return pickModelMetadata({ efforts: [...new Set(levels.filter(level => level !== "none"))] }).efforts;
+}
+
/**
* 解析 OpenAI / Anthropic「列模型」响应的三种形状(`{data:[{id}]}` / `{models:[{id|slug}]}` /
* 字符串数组)为去重后的 `{id, name, contextWindow?}[]`;无法识别返回 null。显示名优先取
@@ -72,6 +85,7 @@ export function parseModelsListResponse(
context_window?: unknown;
max_context_length?: unknown;
max_input_tokens?: unknown;
+ max_context_window?: unknown;
})
: null;
const name = google && typeof google.displayName === 'string' ? google.displayName :
@@ -124,11 +138,17 @@ export function parseModelsListResponse(
? reasoning.defaultEffort
: reasoning?.default_effort !== undefined
? reasoning.default_effort
- : record.default_effort;
+ : record.default_effort !== undefined
+ ? record.default_effort
+ : record.default_reasoning_level !== undefined
+ ? record.default_reasoning_level
+ : record.reasoningEffort;
const modalities = record.modalities ??
(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;
+ const inputModalities = (modalities as { input?: unknown } | undefined)?.input ?? record.input_modalities;
+ const contextWindowMax = pickModelMetadata({ contextWindowMax: record.max_context_window }).contextWindowMax;
+ const serviceTiers = record.service_tiers;
const isVercel = sourceUrl !== undefined && (() => {
try { const url = new URL(sourceUrl); return url.origin === 'https://ai-gateway.vercel.sh'
&& url.pathname.replace(/\/+$/, '') === '/v1/models'; } catch { return false; }
@@ -149,6 +169,9 @@ 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,
maxOutputTokens:
record.max_output_tokens ?? info.max_output_tokens ?? google?.outputTokenLimit ??
record.maxOutputTokens ??
@@ -161,6 +184,9 @@ export function parseModelsListResponse(
? reasoning.supported_efforts.filter((value) => value !== "none")
: undefined) ??
record.supported_efforts ??
+ declaredReasoningEfforts(record.supported_reasoning_levels, "effort") ??
+ (record.supportsReasoningEffort === false
+ ? [] : declaredReasoningEfforts(record.reasoningEfforts, "value")) ??
(isVercel && Array.isArray(record.reasoning_options)
? record.reasoning_options.find((option: unknown) =>
option && typeof option === 'object' && (option as { type?: unknown }).type === 'effort')?.values
@@ -174,10 +200,13 @@ export function parseModelsListResponse(
: info.supports_function_calling ?? capabilities.trained_for_tool_use,
reasoningRequired: reasoning?.required ?? reasoning?.mandatory,
defaultEffort: rawDefault === "none" ? null : rawDefault,
- supportsFastMode: record.supports_fast_mode ?? record.supportsServiceTier,
+ supportsFastMode: record.supports_fast_mode ?? record.supportsServiceTier ??
+ (Array.isArray(serviceTiers) && serviceTiers.every(tier =>
+ tier && typeof tier === 'object' && typeof tier.id === 'string')
+ ? serviceTiers.some(tier => tier.id === 'priority') : undefined),
supportsImageInput:
record.supports_image_input ?? info.supports_vision ?? capabilities.vision ??
- (Array.isArray(inputModalities)
+ (Array.isArray(inputModalities) && inputModalities.every(value => typeof value === 'string')
? inputModalities.includes("image")
: undefined),
});
diff --git a/packages/model-providers/src/modelGeneration.ts b/packages/model-providers/src/modelGeneration.ts
new file mode 100644
index 0000000000..8f86957e00
--- /dev/null
+++ b/packages/model-providers/src/modelGeneration.ts
@@ -0,0 +1,34 @@
+import { pickModelMetadata, type ModelMetadata } from './modelMetadataLayers.js';
+
+/** Compare numbered generations within one family/variant, never lexicographically.
+ * Vendor namespaces are aliases; private namespaces remain part of the identity.
+ * Dates and size suffixes stay in the variant, rather than becoming generations.
+ */
+function generation(id: string) {
+ const normalized = id.toLowerCase().replace(/^(?:openai|anthropic|google|x-ai|xai|deepseek|qwen)\//, '');
+ 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) };
+}
+function compare(left: number[], right: number[]) {
+ for (let i = 0; i < Math.max(left.length, right.length); i++) {
+ const delta = (left[i] ?? 0) - (right[i] ?? 0);
+ if (delta) return delta;
+ }
+ return 0;
+}
+/** Oldest first so nearer generations override older defaults field by field. */
+export function previousModelGenerations<T>(id: string, candidates: readonly T[], getId: (candidate: T) => string): T[] {
+ const target = generation(id);
+ if (!target) return [];
+ return candidates.flatMap(candidate => {
+ const source = generation(getId(candidate));
+ return source && source.family === target.family && compare(source.version, target.version) < 0
+ ? [{ candidate, version: source.version }] : [];
+ }).sort((a, b) => compare(a.version, b.version)).map(entry => entry.candidate);
+}
+/** 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);
+ return capabilities;
+}
diff --git a/packages/model-providers/src/modelMetadataLayers.ts b/packages/model-providers/src/modelMetadataLayers.ts
index 1e2813234c..44b4e8a367 100644
--- a/packages/model-providers/src/modelMetadataLayers.ts
+++ b/packages/model-providers/src/modelMetadataLayers.ts
@@ -16,6 +16,8 @@ export interface ModelMetadata {
description?: string;
group?: string;
contextWindow?: number;
+ /** Upstream capacity, distinct from the recommended working window. */
+ contextWindowMax?: number;
maxOutputTokens?: number;
efforts?: ModelEffort[];
defaultEffort?: ModelEffort | null;
@@ -39,6 +41,7 @@ export const MODEL_METADATA_FIELDS = [
"description",
"group",
"contextWindow",
+ "contextWindowMax",
"maxOutputTokens",
"efforts",
"defaultEffort",
@@ -95,7 +98,7 @@ export function validModelMetadata(value: unknown): value is ModelMetadata {
typeof v === "string" && v.trim().length > 0 && v.length <= maxLength
);
}
- if (["contextWindow", "maxOutputTokens"].includes(key))
+ if (["contextWindow", "contextWindowMax", "maxOutputTokens"].includes(key))
return typeof v === "number" && Number.isSafeInteger(v) && v > 0;
if (key === "efforts")
return (
@@ -119,7 +122,21 @@ export function pickModelMetadata(value: object | undefined): ModelMetadata {
export function mergeModelMetadata(
...layers: (ModelMetadata | undefined)[]
): ModelMetadata {
- return Object.assign({}, ...layers.map(pickModelMetadata));
+ const result: ModelMetadata = {};
+ for (const layer of layers) {
+ const fields = pickModelMetadata(layer);
+ Object.assign(result, fields);
+ // An explicit image-input denial supersedes lower-priority input modalities.
+ // Do not mutate the source catalog or discard unrelated output capabilities.
+ if (fields.supportsImageInput === false && fields.modalities === undefined &&
+ result.modalities?.input.includes('image')) {
+ result.modalities = {
+ ...result.modalities,
+ input: result.modalities.input.filter((modality) => modality !== 'image'),
+ };
+ }
+ }
+ return result;
}
export function findBaseModel(
registry: ModelRegistry | undefined,
@@ -161,6 +178,7 @@ export function resolveModelMetadata(
agent?: string,
providerDefaults?: ModelMetadata,
declaredDefaultEffort?: ModelMetadata["defaultEffort"],
+ generationDefaults?: ModelMetadata,
): ModelMetadata {
const ids = [modelId];
if (providerId === "openai" && modelId.startsWith("chatgpt/"))
@@ -198,6 +216,7 @@ export function resolveModelMetadata(
providerDefaults,
);
const result = mergeModelMetadata(
+ generationDefaults,
defaults,
live,
// A Harness's suggested default is not a model capability. Keep the shared
@@ -212,6 +231,16 @@ 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,
+ )) {
+ delete result.contextWindowMax;
+ }
if (result.efforts?.length === 0) result.defaultEffort = null;
else if (
result.defaultEffort != null &&
@@ -406,6 +435,15 @@ export function mergeDiscoveredRuntimeModels(
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 &&
+ discoveredMetadata.contextWindowMax !== undefined &&
+ ((previousWindow === undefined && models[index]?.contextWindow === undefined) ||
+ (previousWindow !== undefined && previousWindow > discoveredMetadata.contextWindowMax))) {
+ discoveredMetadata.contextWindow = discoveredMetadata.contextWindowMax;
+ }
if (index < 0)
models.push({
id: model.id,
diff --git a/packages/model-providers/src/providerBranding.ts b/packages/model-providers/src/providerBranding.ts
index 56b2b9eef1..87cec1cfec 100644
--- a/packages/model-providers/src/providerBranding.ts
+++ b/packages/model-providers/src/providerBranding.ts
@@ -10,6 +10,8 @@
// Additional supplier marks: Lobe Icons 5.14.0 (MIT; included in THIRD-PARTY-NOTICES).
export const PROVIDER_LOGO_PATHS = {
+ // Original S monogram for the self-hosted Sub2API connection (not upstream artwork).
+ sub2api: 'M5 3h14v3H8v4h8a3 3 0 0 1 3 3v5a3 3 0 0 1-3 3H5v-3h11v-5H8a3 3 0 0 1-3-3V3z',
antgroup: "M15.984 1.383c-.259.593-.877.914-1.334 1.348-.702.669-.998 1.726-1.218 2.682 3.37.775 5.995 3.532 6.486 7.025.566 4.016-1.85 7.859-5.744 9.134-3.894 1.276-8.153-.38-10.126-3.935A8.339 8.339 0 016.457 6.833a18.703 18.703 0 01-.376-.802l-.174-.408a8.698 8.698 0 01-.618-1.987 7.523 7.523 0 01-.051-.37 3.717 3.717 0 01-.024-.34 2.506 2.506 0 01.132-.872 1.26 1.26 0 01.384-.594.842.842 0 011.165.11c.694.82.169 2.078.142 3.02.009.575.111 1.144.301 1.684l-.103.052a8.404 8.404 0 001.318-.79c.28-.243.45-.742.781-.864.33-.121.682.146.946.297.278.145.587.219.901.214l.462.007c.076.002.152.005.226.011.212.01.422.026.63.05l.196-1.08c.04-.216.082-.43.126-.645l.141-.641c.14-.7.42-1.367.823-1.96.305-.431.75-.746 1.262-.89.877-.217 1.24.629.937 1.348zm-2.244 5.44c-2.841.002-4.898 2.532-4.628 5.64.27 3.11 2.781 5.642 5.613 5.666 2.832.024 4.932-2.512 4.658-5.644-.273-3.133-2.802-5.666-5.643-5.663zm.357.57a4.707 4.707 0 013.304 1.445 5.703 5.703 0 011.643 3.49 4.841 4.841 0 01-1.141 3.705v.015a3.923 3.923 0 01-2.808 1.303 4.664 4.664 0 01-3.367-1.387 5.678 5.678 0 01-1.71-3.518 4.832 4.832 0 011.14-3.753 3.923 3.923 0 012.813-1.3h.126z",
baseten: "M2.316 4.8h14.682v4.8H7.31a.302.302 0 00-.308.3v4.2c0 .171.14.3.308.3h9.688v4.8h-4.686a.302.302 0 00-.308.3v4.2c0 .171.141.3.308.3h4.378a.297.297 0 00.308-.3v-4.5h4.694a.302.302 0 00.308-.3v-4.2c0-.171-.14-.3-.308-.3h-4.694V9.6h4.694A.302.302 0 0022 9.3V5.1c0-.171-.14-.3-.308-.3h-4.694V.3c0-.171-.14-.3-.308-.3H2.316A.31.31 0 002 .3v4.2c0 .171.14.3.316.3z",
cerebras: "M14.121 2.701a9.299 9.299 0 000 18.598V22.7c-5.91 0-10.7-4.791-10.7-10.701S8.21 1.299 14.12 1.299V2.7zm4.752 3.677A7.353 7.353 0 109.42 17.643l-.901 1.074a8.754 8.754 0 01-1.08-12.334 8.755 8.755 0 0112.335-1.08l-.901 1.075zm-2.255.844a5.407 5.407 0 00-5.048 9.563l-.656 1.24a6.81 6.81 0 016.358-12.043l-.654 1.24zM14.12 8.539a3.46 3.46 0 100 6.922v1.402a4.863 4.863 0 010-9.726v1.402z M15.407 10.836a2.24 2.24 0 00-.51-.409 1.084 1.084 0 00-.544-.152c-.255 0-.483.047-.684.14a1.58 1.58 0 00-.84.912c-.074.203-.11.416-.11.631 0 .218.036.43.11.631a1.594 1.594 0 00.84.913c.2.093.43.14.684.14.216 0 .417-.046.602-.135.188-.09.35-.225.475-.392l.928 1.006c-.14.14-.3.261-.482.363a3.367 3.367 0 01-1.083.38c-.17.026-.317.04-.44.04a3.315 3.315 0 01-1.182-.21 2.825 2.825 0 01-.961-.597 2.816 2.816 0 01-.644-.929 2.987 2.987 0 01-.238-1.21c0-.444.08-.847.238-1.21.15-.35.368-.666.643-.929.278-.261.605-.464.962-.596a3.315 3.315 0 011.182-.21c.355 0 .712.068 1.072.204.361.138.685.36.944.649l-.962.97z",
@@ -83,6 +85,7 @@ export type ProviderLogoRouting = Readonly<
/** Bundled regional presets share a brand and never infer it from display names. */
const PROVIDER_LOGO_KIND_BY_ID: Readonly<Record<string, ProviderLogoKind>> = {
+ sub2api: 'sub2api',
"antgroup": "antgroup",
"baseten": "baseten",
"cerebras": "cerebras",
diff --git a/packages/model-providers/src/providerEndpointTemplate.ts b/packages/model-providers/src/providerEndpointTemplate.ts
index beffd1beab..f2fac23e10 100644
--- a/packages/model-providers/src/providerEndpointTemplate.ts
+++ b/packages/model-providers/src/providerEndpointTemplate.ts
@@ -7,6 +7,8 @@ const AZURE_RESOURCE = /^[a-z0-9](?:[a-z0-9-]{0,62}[a-z0-9])$/;
const VERTEX_REGIONAL_TEMPLATE = 'https://{location}-aiplatform.googleapis.com';
const VERTEX_MULTIREGION_TEMPLATE = 'https://aiplatform.{location}.rep.googleapis.com';
const VERTEX_GLOBAL_HOST = 'https://aiplatform.googleapis.com';
+// Self-hosted gateways have no fixed host, scheme, port or deployment prefix.
+const SELF_HOSTED_TEMPLATE = 'https://{endpoint}/v1';
function stripTrailingSlash(value: string): string {
return value.replace(/\/+$/, '');
@@ -88,6 +90,13 @@ function genericBindings(template: string, endpoint: string): Record<string, str
/** Only declared account/location placeholders bind. Official host aliases stay explicit; no cross-origin guessing. */
export function providerEndpointBindings(template: string, endpoint: string): Record<string, string> | null {
+ if (template === SELF_HOSTED_TEMPLATE) {
+ const url = parseHttpUrl(endpoint);
+ if (!url || /[{}]/.test(endpoint) || url.search || url.hash) return null;
+ const pathname = stripTrailingSlash(url.pathname);
+ url.pathname = /(?:\/v\d+|\/backend-api\/codex)$/i.test(pathname) ? pathname : `${pathname}/v1`;
+ return { endpoint: stripTrailingSlash(url.toString()) };
+ }
if (isVertexFamilyTemplate(template)) {
const location = recognizeVertexLocation(endpoint);
return location ? { location } : null;
@@ -104,6 +113,9 @@ export function bindProviderEndpoint(
bindings: Record<string, string>,
sourceEndpoint?: string,
): string {
+ if (bindings.endpoint && (template === SELF_HOSTED_TEMPLATE || template.startsWith(`${SELF_HOSTED_TEMPLATE}/`))) {
+ return bindings.endpoint + template.slice(SELF_HOSTED_TEMPLATE.length);
+ }
if (isVertexFamilyTemplate(template) && bindings.location) return officialVertexEndpoint(bindings.location);
if (isAzureOpenAIFamilyTemplate(template) && bindings.resource) {
const fromSource = sourceEndpoint ? recognizeAzureBinding(sourceEndpoint) : null;
diff --git a/packages/model-providers/src/providerModelCatalog.ts b/packages/model-providers/src/providerModelCatalog.ts
index dee3e058b9..8f5d97806f 100644
--- a/packages/model-providers/src/providerModelCatalog.ts
+++ b/packages/model-providers/src/providerModelCatalog.ts
@@ -1,3 +1,4 @@
+import { previousModelGenerations } from './modelGeneration.js';
import { sourceProviderForPreset } from './providerPresetIdentity.js';
import { providerEndpointBindings } from './providerEndpointTemplate.js';
import { declaredModelInterface } from './providerInterfaceRoutes.js';
@@ -7,6 +8,8 @@ import type { ModelMetadata } from "./modelMetadataLayers.js";
import type { ModelCost, PiModelApi, ProviderWireProtocol } from "./types.js";
export interface ProviderModelRecord {
+ /** Internal fallback provenance, not an upstream capability declaration. */
+ inheritedFrom?: string;
id: string;
name: string;
upstream: string;
@@ -14,6 +17,8 @@ export interface ProviderModelRecord {
maxOutput?: number;
modalities: { input: string[]; output: string[] };
supportsImageInput: boolean;
+ /** Connection-scoped declaration; absent does not authorize a Fast request. */
+ supportsFastMode?: boolean;
reasoning: boolean;
efforts: NonNullable<ModelMetadata["efforts"]>;
defaultEffort: ModelMetadata["defaultEffort"];
@@ -52,7 +57,12 @@ for (const [provider, declaration] of Object.entries(interfaceModels)) {
const byEndpointAndId = new Map<string, ProviderModelRecord[]>();
const apisByEndpoint = new Map<string, Set<string>>();
-const normalize = (url: string) => url.trim().replace(/\/+$/, "");
+const normalize = (url: string) => {
+ const trimmed = url.trim();
+ let end = trimmed.length;
+ while (end > 0 && trimmed.charCodeAt(end - 1) === 47) end--;
+ return trimmed.slice(0, end);
+};
for (const rows of Object.values(PROVIDER_MODEL_CATALOG.providers)) {
for (const row of rows) {
const endpoint = normalize(row.upstream);
@@ -69,7 +79,11 @@ for (const rows of Object.values(PROVIDER_MODEL_CATALOG.providers)) {
}
}
-/** Exact route identity, shared by every harness. No fuzzy names or cross-proxy borrowing. */
+const generationRows = Object.values(PROVIDER_MODEL_CATALOG.providers).flat();
+const manufacturerGenerationRows = ['openai', 'anthropic', 'google', 'xai', 'deepseek', 'qwen']
+ .flatMap(provider => PROVIDER_MODEL_CATALOG.providers[provider] ?? []);
+
+/** Exact route identity, shared by every harness. */
export function providerModelRecord(
modelId: string,
upstream: string,
@@ -86,6 +100,31 @@ export function providerModelRecord(
return matches?.length === 1 ? matches[0] : undefined;
}
+/** A new generation reuses parameters, but keeps its own connection and identity. */
+export function providerModelGenerationRecord(modelId: string, upstream: string, protocol?: ProviderWireProtocol | PiModelApi, presetId?: string): ProviderModelRecord | undefined {
+ const api = protocol === 'openai-chat' ? 'openai-completions' : protocol;
+ // Transport must already be selected; inheritance never changes a connection's API.
+ if (!api) return undefined;
+ const endpointRows = generationRows
+ .filter(row => normalize(row.upstream) === normalize(upstream) && row.execution.pi.api === api);
+ // Manufacturer adapters are reusable across compatible relays. Never borrow another
+ // relay's headers, endpoint, prices or adapter identity.
+ const manufacturerRows = manufacturerGenerationRows
+ .filter(row => row.execution.pi.api === api);
+ const presetRows = presetId ? (PROVIDER_MODEL_CATALOG.providers[sourceProviderForPreset(presetId)] ?? [])
+ .filter(row => row.execution.pi.api === api) : [];
+ const candidates = [...manufacturerRows, ...presetRows, ...endpointRows];
+ // A relay endpoint mismatch does not make a known model a new generation.
+ // Reuse its own same-protocol adapter before considering older models.
+ const previous = candidates.filter(row => row.id === modelId).at(-1)
+ ?? previousModelGenerations(modelId, candidates, row => row.id).at(-1);
+ if (!previous) return undefined;
+ const { cost: _cost, execution, ...metadata } = previous;
+ const { headers: _headers, ...parameters } = execution.pi;
+ return { ...metadata, id: modelId, name: modelId, upstream, inheritedFrom: previous.id,
+ execution: { pi: structuredClone(parameters) } };
+}
+
/** Account-specific endpoints may still explicitly reference a maintained connection template. */
export function providerPresetModelRecord(presetId: string | undefined, modelId: string, api?: PiModelApi): ProviderModelRecord | undefined {
if (!presetId) return undefined;
@@ -110,6 +149,7 @@ export function providerModelMetadata(row: ProviderModelRecord): ModelMetadata {
...(row.maxOutput ? { maxOutputTokens: row.maxOutput } : {}),
modalities: row.modalities,
supportsImageInput: row.supportsImageInput,
+ ...(row.supportsFastMode !== undefined ? { supportsFastMode: row.supportsFastMode } : {}),
efforts: row.efforts,
defaultEffort: row.defaultEffort,
...(row.execution.pi.thinkingLevelMap?.off === null
diff --git a/packages/model-providers/src/providerSetupLinks.ts b/packages/model-providers/src/providerSetupLinks.ts
index ab5aedf229..19f83355a0 100644
--- a/packages/model-providers/src/providerSetupLinks.ts
+++ b/packages/model-providers/src/providerSetupLinks.ts
@@ -65,7 +65,7 @@ export function providerSetupLink(preset: Pick<ProviderPreset, 'id' | 'docsUrl'
const link = links[sourceProviderForPreset(preset.id)];
if (link) return link;
// Local servers have no provider account. Keep their setup help without asking for a cloud key.
- if (['litellm', 'lmstudio', 'llamacpp', 'vllm'].includes(preset.id) && preset.docsUrl) {
+ if (['litellm', 'lmstudio', 'llamacpp', 'vllm', 'sub2api'].includes(preset.id) && preset.docsUrl) {
return { url: preset.docsUrl, kind: 'setup' };
}
return undefined;
diff --git a/packages/model-providers/src/user-provider.ts b/packages/model-providers/src/user-provider.ts
index 05846afed5..682d85be9f 100644
--- a/packages/model-providers/src/user-provider.ts
+++ b/packages/model-providers/src/user-provider.ts
@@ -1,10 +1,11 @@
+import { previousModelGenerations, generationCapabilities } from './modelGeneration.js';
import { alignModelApiRoute, providerInterfaceModelRoute, hasDeclaredProviderInterface, providerWireProtocolForApi, providerBaseUrlForApi } from './providerInterfaceRoutes.js';
import { nativeModelAgents } from './modelProtocol.js';
import { isMimoTokenPlanProvider } from './mimoPresentation.js';
import { resolveCatalogModelNativeApi, resolveModelNativeApi } from './modelRegistry.js';
import { providerEndpointBindings, bindProviderEndpoint, bindProviderPresetRuntime, canonicalProviderEndpoint } from './providerEndpointTemplate.js';
import { PI_MODEL_APIS } from "./types.js";
-import { providerModelRecord, providerPresetModelRecord, providerModelMetadata } from "./providerModelCatalog.js";
+import { providerModelRecord, providerModelGenerationRecord, providerPresetModelRecord, providerModelMetadata } from "./providerModelCatalog.js";
import { BUNDLED_CATALOG, BUILTIN_PROVIDERS } from './builtin.js';
import { providerMediaField } from "./providerMediaModels.js";
import {
@@ -284,9 +285,10 @@ function toCatalogModel(
modelRegistry: ModelRegistry | null | undefined,
providerDefaults?: ModelMetadata,
metadataProviderId = providerId,
+ generationDefaults?: ModelMetadata,
): CatalogModel {
// 显式 runtime 能力优先:reasoning:true 才导出 efforts;false = 明确无思考档。
- // 未知能力不借用其它模型的档位;后续仍按目录、预设、实报和用户配置逐层继承。
+ // 新代际缺项可沿同系列/变体/协议继承;本型号目录、实报和用户配置优先。
// 空档位只表示不指定推理强度,不代表要求供应商关闭思考。
const efforts: Effort[] =
m.reasoning === true
@@ -336,21 +338,42 @@ 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 ||
- providerDefaults
+ providerDefaults || (generationDefaults && Object.keys(generationDefaults).length > 0)
? resolveModelMetadata(
modelRegistry ?? undefined,
metadataProviderId,
m.id,
- m.discoveredMetadata,
+ live,
pickModelMetadata(user),
agent,
providerDefaults,
+ undefined,
+ generationDefaults,
)
: pickModelMetadata(user);
- return applyModelMetadata(model, resolved);
+ 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;
}
function defaultWireProtocol(agent: AgentKind): ProviderWireProtocol {
@@ -471,6 +494,15 @@ export function buildUserProvider(
// This lends only the endpoint protocol, never another model's window or capabilities.
const presetApis = new Set<ProviderRuntimeModelConfig['api']>(followsPreset ? presetRuntime.models.map(model => model.api ?? model.piApi) : []);
const presetApi = presetApis.size === 1 ? [...presetApis][0] : undefined;
+ const registrySources = [BUNDLED_CATALOG.modelRegistry, options.modelRegistry].filter(
+ (source): source is ModelRegistry => source != null,
+ );
+ const publicGenerationCandidates = registrySources.flatMap(source =>
+ (source.baseModels ?? []).flatMap(base => {
+ const protocol = resolveCatalogModelNativeApi(source, base.id);
+ return [base.id, ...base.aliases].map(id => ({ id, protocol, metadata: base.defaults }));
+ }),
+ );
models[agent] = rt.models.map((storedModel) => {
// Old ID-only imports must pick up newly known per-model interfaces too.
// An explicit model API/path remains a user choice, not a preset default.
@@ -532,14 +564,55 @@ export function buildUserProvider(
? providerPresetModelRecord(rt.catalogPresetId, m.id) : undefined)
?? (followsPreset && sameRoute && m.api && presetModel?.api === m.api
? providerPresetModelRecord(preset?.id, m.id, m.api) : undefined)
+ ?? providerModelGenerationRecord(m.id, m.route?.baseUrl ?? resolvedBaseUrl, m.api ?? m.piApi ?? wire, followsPreset ? preset?.id : undefined)
: undefined;
- const importedApi = imported &&
+ // Existing catalog identities keep their established execution path. A
+ // predecessor supplies parameters, not permission to switch their adapter.
+ const keepExistingApi = imported?.inheritedFrom && registrySources.some(source => findBaseModel(source, m.id));
+ const importedApi = imported && !keepExistingApi &&
PI_MODEL_APIS.some(api => api === imported.execution.pi.api)
? imported.execution.pi.api as NonNullable<ProviderRuntimeModelConfig['piApi']>
: !m.route && presetApi ? presetApi
: wire === 'google-generative-ai' ? 'google-generative-ai' : undefined;
+ // Recompute inherited defaults from declarations, never save them as user edits.
+ // Same-connection declarations win within a generation; public defaults fill gaps.
+ const protocol = m.api ?? m.piApi ?? (wire === 'openai-chat' ? 'openai-completions' : wire);
+ const generationCandidates: Array<{ id: string; metadata: ModelMetadata }> =
+ publicGenerationCandidates.filter(candidate => candidate.protocol === protocol);
+ if (imported?.inheritedFrom) generationCandidates.push({
+ id: imported.inheritedFrom, metadata: providerModelMetadata(imported),
+ });
+ for (const candidate of previousModelGenerations(m.id, [...(followsPreset ? presetRuntime.models : []), ...rt.models], candidate => candidate.id)) {
+ const candidateWire = candidate.route?.wireProtocol ?? rt.wireProtocol ?? defaultWireProtocol(agent);
+ const candidateProtocol = candidate.api ?? candidate.piApi ??
+ (candidateWire === 'openai-chat' ? 'openai-completions' : candidateWire);
+ if (candidateProtocol !== protocol ||
+ withoutTrailingSlashes(candidate.route?.baseUrl ?? resolvedBaseUrl) !== withoutTrailingSlashes(m.route?.baseUrl ?? resolvedBaseUrl) ||
+ (candidate.route?.requestPath ?? rt.requestPath) !== (m.route?.requestPath ?? rt.requestPath)) continue;
+ const candidateRow = providerModelRecord(candidate.id, candidate.route?.baseUrl ?? resolvedBaseUrl, candidateProtocol);
+ generationCandidates.push({ id: candidate.id, metadata: mergeModelMetadata(
+ ...registrySources.map(source => findBaseModel(source, candidate.id)?.defaults),
+ candidateRow ? providerModelMetadata(candidateRow) : undefined,
+ candidate.discoveredMetadata,
+ runtimeUserModelMetadata(candidate),
+ ) });
+ }
+ const generationDefaults = mergeModelMetadata(
+ ...previousModelGenerations(m.id, generationCandidates, candidate => candidate.id)
+ .map(candidate => generationCapabilities(candidate.metadata)),
+ );
+ // An exact manufacturer record reused on a relay is still this model's
+ // own declaration. Only a strictly older ID is a generation fallback.
+ const exactRelayDefaults = imported?.inheritedFrom === m.id && !registrySources.some(source =>
+ findBaseModel(source, m.id) || source?.models.some(entry =>
+ entry.id === m.id || entry.routes.some(route => route.modelId === m.id),
+ ),
+ ) ? mergeModelMetadata(providerModelMetadata(imported), {
+ // Reused adapter data is a fallback, not a configured route default.
+ defaultEffort: m.discoveredMetadata?.defaultEffort,
+ }) : undefined;
const defaults = imported || catalogDefaults || presetDefaults
- ? mergeModelMetadata(imported ? providerModelMetadata(imported) : undefined, catalogDefaults, presetDefaults)
+ ? mergeModelMetadata(imported && !imported.inheritedFrom ? providerModelMetadata(imported) : exactRelayDefaults, catalogDefaults, presetDefaults)
: undefined;
// A verified catalog identity can reuse the manufacturer's declaration.
// Execution protocols and prices still belong to this exact connection.
@@ -570,6 +643,7 @@ export function buildUserProvider(
options.modelRegistry,
defaults,
nativeCodex ? 'openai' : undefined,
+ generationDefaults,
),
// Projection is not a user edit. Save only the original configuration.
userModelConfig: structuredClone(storedModel),