OpenCoworkAI · Lxr-max · Sep 14, 2026
diff --git a/.changeset/provider-model-discovery-modes.md b/.changeset/provider-model-discovery-modes.md
@@ -0,0 +1,7 @@
+---
+"@open-codesign/desktop": patch
+"@open-codesign/i18n": patch
+"@open-codesign/shared": patch
+---
+
+Formalize provider `modelDiscoveryMode` (`models`, `static-hint`, `manual`, `infer-only`) so listing, Settings pickers, imports, and diagnostics no longer treat GET `/models` as universal.
diff --git a/apps/desktop/src/main/connection-ipc.test.ts b/apps/desktop/src/main/connection-ipc.test.ts
@@ -14,11 +14,13 @@ vi.mock('./tls-override', () => ({
 }));
 
 import { createHash } from 'node:crypto';
+import { resolveListForProviderPlan } from '@open-codesign/shared';
 import {
   _clearModelsCache,
   buildAuthHeaders,
   buildAuthHeadersForWire,
   CONNECTION_FETCH_TIMEOUT_MS,
+  type ConnectionTestResponse,
   classifyHttpError,
   classifyNetworkTarget,
   extractIds,
@@ -27,6 +29,7 @@ import {
   getCacheKey,
   handleConfigV1TestEndpoint,
   handleOllamaV1Probe,
+  type ModelsListResponse,
   normalizeBaseUrl,
   normalizeOllamaBaseUrl,
   runProviderTest,
@@ -39,8 +42,6 @@ import { withTlsBypass } from './tls-override';
 // network responses without hitting the network.
 // ---------------------------------------------------------------------------
 
-import type { ConnectionTestResponse, ModelsListResponse } from './connection-ipc';
-
 // ---------------------------------------------------------------------------
 // connection:v1:test test helper
 // ---------------------------------------------------------------------------
@@ -555,6 +556,31 @@ describe('models:v1:list-for-provider input validation', () => {
     const result = validateListForProviderInput('claude-code-anthropic');
     expect(result).toBeNull();
   });
+
+  it('respects declared discovery modes instead of always fetching /models', () => {
+    expect(
+      resolveListForProviderPlan('chatgpt-codex', {
+        wire: 'openai-codex-responses',
+        defaultModel: 'gpt-5.5',
+        modelsHint: ['gpt-5.5', 'gpt-5.4'],
+        requiresApiKey: false,
+      }).action,
+    ).toBe('return');
+    expect(
+      resolveListForProviderPlan('glm', {
+        wire: 'openai-chat',
+        defaultModel: 'glm-4.6',
+        capabilities: { modelDiscoveryMode: 'infer-only' },
+      }),
+    ).toMatchObject({ action: 'return', source: 'local', models: ['glm-4.6'] });
+    expect(
+      resolveListForProviderPlan('openai', {
+        wire: 'openai-chat',
+        defaultModel: 'gpt-4o',
+        capabilities: { modelDiscoveryMode: 'models' },
+      }).action,
+    ).toBe('fetch-remote');
+  });
 });
 
 // ---------------------------------------------------------------------------
@@ -1188,6 +1214,88 @@ describe('runProviderTest degrade-probe (issue #179)', () => {
   });
 });
 
+describe('runProviderTest discovery modes (issue #210)', () => {
+  beforeEach(() => {
+    vi.useRealTimers();
+  });
+
+  it('infer-only skips GET /models and treats inference success as compatible', async () => {
+    const { calls, restore } = installFakeFetch((url) => {
+      if (url.endsWith('/models')) return { status: 404 };
+      if (url.endsWith('/chat/completions')) return { status: 200, body: { id: 'ok' } };
+      return { status: 500 };
+    });
+    try {
+      const res = await runProviderTest({
+        provider: 'glm',
+        wire: 'openai-chat',
+        apiKey: 'sk-test',
+        baseUrl: 'https://open.bigmodel.cn/api/paas/v4',
+        modelDiscoveryMode: 'infer-only',
+      });
+      expect(res.ok).toBe(true);
+      if (res.ok) {
+        expect(res.probeMethod).toBe('inference');
+        expect(res.compatibility).toBe('compatible');
+      }
+      expect(calls).toHaveLength(1);
+      expect(calls[0]?.url).toMatch(/\/chat\/completions$/);
+      expect(calls.some((c) => c.url.endsWith('/models'))).toBe(false);
+    } finally {
+      restore();
+    }
+  });
+
+  it('manual skips GET /models', async () => {
+    const { calls, restore } = installFakeFetch((url) => {
+      if (url.endsWith('/responses')) return { status: 200, body: { id: 'ok' } };
+      return { status: 404 };
+    });
+    try {
+      const res = await runProviderTest({
+        provider: 'custom',
+        wire: 'openai-responses',
+        apiKey: 'sk-test',
+        baseUrl: 'https://proxy.example.com/v1',
+        modelDiscoveryMode: 'manual',
+      });
+      expect(res.ok).toBe(true);
+      if (res.ok) {
+        expect(res.probeMethod).toBe('inference');
+        expect(res.compatibility).toBe('compatible');
+      }
+      expect(calls.some((c) => c.url.endsWith('/models'))).toBe(false);
+    } finally {
+      restore();
+    }
+  });
+
+  it('models mode still degrade-probes /models 404 as before', async () => {
+    const { calls, restore } = installFakeFetch((url) => {
+      if (url.endsWith('/models')) return { status: 404 };
+      if (url.endsWith('/chat/completions')) return { status: 200, body: { id: 'ok' } };
+      return { status: 500 };
+    });
+    try {
+      const res = await runProviderTest({
+        provider: 'glm',
+        wire: 'openai-chat',
+        apiKey: 'sk-test',
+        baseUrl: 'https://open.bigmodel.cn/api/paas/v4',
+        modelDiscoveryMode: 'models',
+      });
+      expect(res.ok).toBe(true);
+      if (res.ok) {
+        expect(res.probeMethod).toBe('chat_completion_degraded');
+        expect(res.compatibility).toBe('degraded');
+      }
+      expect(calls[0]?.url).toMatch(/\/models$/);
+    } finally {
+      restore();
+    }
+  });
+});
+
 describe('config:v1:test-endpoint response parsing', () => {
   beforeEach(() => {
     vi.useRealTimers();
@@ -1212,6 +1320,25 @@ describe('config:v1:test-endpoint response parsing', () => {
     }
   });
 
+  it('treats GET /models 404 as listing-unavailable rather than a hard failure', async () => {
+    const { restore } = installFakeFetch(() => ({ status: 404 }));
+    try {
+      await expect(
+        handleConfigV1TestEndpoint({
+          wire: 'openai-chat',
+          baseUrl: 'https://open.bigmodel.cn/api/paas/v4',
+          apiKey: 'sk-test',
+        }),
+      ).resolves.toEqual({
+        ok: false,
+        error: 'not-a-model-endpoint',
+        message: 'HTTP 404',
+      });
+    } finally {
+      restore();
+    }
+  });
+
   it('classifies private and metadata network targets', () => {
     expect(classifyNetworkTarget('https://provider.example/v1')).toBe('public');
     expect(classifyNetworkTarget('http://localhost:8317')).toBe('loopback');

diff --git a/apps/desktop/src/main/connection-ipc.ts b/apps/desktop/src/main/connection-ipc.ts
@@ -4,11 +4,15 @@ import {
   BUILTIN_PROVIDERS,
   CodesignError,
   canonicalBaseUrl,
+  connectionTestProbesModelsEndpoint,
   type DiagnosticCategory,
   ERROR_CODES,
   ensureVersionedBase,
   isSupportedOnboardingProvider,
   type ProviderEntry,
+  type ProviderModelDiscoveryMode,
+  resolveListForProviderPlan,
+  resolveProviderCapabilities,
   type SupportedOnboardingProvider,
   stripInferenceEndpointSuffix,
   type WireApi,
@@ -72,6 +76,8 @@ export interface ConnectionTestResult {
   ok: true;
   /**
    * `models` when the standard GET /models probe succeeded.
+   * `inference` when the provider's discovery mode does not expect /models
+   * and the real invoke path answered.
    * `chat_completion_degraded` when /models 404'd but POST /chat/completions
    * proved the openai-chat wire is alive (e.g. Zhipu GLM — no public /models).
    * `responses_degraded` when /models 404'd but POST /responses proved the
@@ -81,6 +87,7 @@ export interface ConnectionTestResult {
    */
   probeMethod?:
     | 'models'
+    | 'inference'
     | 'chat_completion_degraded'
     | 'responses_degraded'
     | 'anthropic_messages_degraded';
@@ -100,12 +107,18 @@ export interface ConnectionTestError {
 export type ConnectionTestResponse = ConnectionTestResult | ConnectionTestError;
 
 export type ModelsListResponse =
-  | { ok: true; models: string[] }
+  | {
+      ok: true;
+      models: string[];
+      discoveryMode?: ProviderModelDiscoveryMode;
+      source?: 'remote' | 'static-hint' | 'local';
+    }
   | {
       ok: false;
       code: 'IPC_BAD_INPUT' | 'NETWORK' | 'HTTP' | 'PARSE';
       message: string;
       hint: string;
+      discoveryMode?: ProviderModelDiscoveryMode;
     };
 
 function parseConnectionTestPayload(raw: unknown): ConnectionTestPayloadV1 {
@@ -492,6 +505,9 @@ export interface ActiveProviderCredentials {
   builtin?: boolean;
   /** Opt-in TLS verification bypass; only honored when `builtin === false`. */
   tlsRejectUnauthorized?: boolean;
+  /** Listing strategy for this provider. Defaults to `models` when omitted so
+   *  existing tests and older callers keep probing GET /models. */
+  modelDiscoveryMode?: ProviderModelDiscoveryMode;
 }
 
 function resolveCredentialsForProvider(
@@ -539,6 +555,7 @@ function resolveCredentialsForProvider(
     apiKey,
     baseUrl: entry.baseUrl,
     builtin: entry.builtin === true,
+    modelDiscoveryMode: resolveProviderCapabilities(providerId, entry).modelDiscoveryMode,
     ...(entry.httpHeaders !== undefined ? { httpHeaders: entry.httpHeaders } : {}),
     ...(entry.tlsRejectUnauthorized !== undefined
       ? { tlsRejectUnauthorized: entry.tlsRejectUnauthorized }
@@ -607,10 +624,15 @@ export async function runProviderTest(
     return testChatGPTCodexOAuth();
   }
 
+  const discoveryMode = creds.modelDiscoveryMode ?? 'models';
+  const bypass = creds.builtin !== true && creds.tlsRejectUnauthorized === true;
+  if (!connectionTestProbesModelsEndpoint(discoveryMode)) {
+    return withTlsBypass(bypass, () => probeDeclaredInference(creds));
+  }
+
   // Bypass is the per-provider opt-in, force-gated so a tampered config can
   // never weaken TLS for built-in providers. Wrapping the whole body covers
   // both the GET /models probe and the inner POST inside tryDegradeProbe.
-  const bypass = creds.builtin !== true && creds.tlsRejectUnauthorized === true;
   return withTlsBypass(bypass, async () => {
     const { url, normalizedBaseUrl } = buildEndpointForWire(creds.wire, creds.baseUrl);
     const headers = buildAuthHeadersForWire(
@@ -666,6 +688,55 @@ export async function runProviderTest(
   });
 }
 
+async function probeDeclaredInference(
+  creds: ActiveProviderCredentials,
+): Promise<ConnectionTestResponse> {
+  if (
+    creds.wire !== 'openai-chat' &&
+    creds.wire !== 'openai-responses' &&
+    creds.wire !== 'anthropic'
+  ) {
+    return {
+      ok: false,
+      code: 'IPC_BAD_INPUT',
+      message: `Discovery mode does not probe /models for wire "${creds.wire}"`,
+      hint: 'Use a listed model id or switch this provider to a listing-capable endpoint',
+      compatibility: 'incompatible',
+      reasonCategory: 'unknown',
+    };
+  }
+  const { normalizedBaseUrl } = buildEndpointForWire(creds.wire, creds.baseUrl);
+  const headers = buildAuthHeadersForWire(
+    creds.wire,
+    creds.apiKey,
+    creds.httpHeaders,
+    creds.baseUrl,
+  );
+  const probe = await probeInferenceEndpoint(creds.wire, normalizedBaseUrl, headers);
+  if (probe.kind === 'pass') {
+    return { ok: true, probeMethod: 'inference', compatibility: 'compatible' };
+  }
+  if (probe.kind === 'network') {
+    return {
+      ok: false,
+      code: 'NETWORK',
+      message: probe.message,
+      hint: 'Cannot reach provider inference endpoint',
+      compatibility: 'incompatible',
+      reasonCategory: 'unknown',
+    };
+  }
+  const { code, hint } = classifyHttpError(probe.status);
+  return {
+    ok: false,
+    code,
+    message: `HTTP ${probe.status}`,
+    hint,
+    compatibility: 'incompatible',
+    reasonCategory: connectionCategoryForStatus(probe.status, normalizedBaseUrl),
+  };
+}
+
 async function tryDegradeProbe(
   wire: 'openai-chat' | 'openai-responses' | 'anthropic',
   normalizedBaseUrl: string,
@@ -977,19 +1048,24 @@ async function handleModelsV1ListForProvider(raw: unknown): Promise<ModelsListRe
   if ('ok' in resolved) return resolved;
   const { providerId, entry } = resolved;
 
-  // Providers that expose a static hint (e.g. chatgpt-codex, whose /models
-  // endpoint requires OAuth bearer + ChatGPT-Account-Id headers that this
-  // keyless discovery path cannot supply) short-circuit with modelsHint.
-  if (entry.modelsHint !== undefined && entry.modelsHint.length > 0) {
-    return { ok: true, models: entry.modelsHint };
+  const plan = resolveListForProviderPlan(providerId, entry);
+  if (plan.action === 'return') {
+    return {
+      ok: true,
+      models: plan.models,
+      discoveryMode: plan.discoveryMode,
+      source: plan.source,
+    };
   }
 
   const keyResult = resolveApiKeyForListing(providerId, entry);
   if ('ok' in keyResult) return keyResult;
   const { apiKey } = keyResult;
 
   const cached = getCachedModels(providerId, entry.baseUrl, apiKey);
-  if (cached !== null) return { ok: true, models: cached };
+  if (cached !== null) {
+    return { ok: true, models: cached, discoveryMode: 'models', source: 'remote' };
+  }
 
   const { url } = buildEndpointForWire(entry.wire, entry.baseUrl);
   const headers = buildAuthHeadersForWire(entry.wire, apiKey, entry.httpHeaders, entry.baseUrl);
@@ -1001,8 +1077,11 @@ async function handleModelsV1ListForProvider(raw: unknown): Promise<ModelsListRe
       hint: 'Check provider /models endpoint compatibility',
     }),
   );
-  if (result.ok) setCachedModels(providerId, entry.baseUrl, apiKey, result.models);
-  return result;
+  if (result.ok) {
+    setCachedModels(providerId, entry.baseUrl, apiKey, result.models);
+    return { ...result, discoveryMode: 'models', source: 'remote' };
+  }
+  return { ...result, discoveryMode: 'models' };
 }
 
 async function fetchModelListResponse(