Add OpenAI Responses inference and discovery

bin/lib/catalog-dev-fixtures.ts

@@ -47,6 +47,10 @@ export const credentialFixtures: Record<string, CredentialFixture> = {
     credentialName: "xAI Direct Key",
     credentialSecret: "xai-fake-key-for-seed-data",
   },
+  "OpenAI Responses": {
+    credentialName: "OpenAI Responses Key",
+    credentialSecret: "sk-responses-fake-key-for-seed-data",
+  },
 };
 
 // Dev pricing as decimal strings, matching the API's string money fields.
@@ -88,6 +92,18 @@ export const priceFixtures: Record<string, Record<string, PriceFixture>> = {
     "grok-4.5": { input: "0.000002", output: "0.000006" },
     "grok-4.6": { input: "0.000002", output: "0.000006" },
   },
+  "OpenAI Responses": {
+    "gpt-6-astra": { input: "0.00001", output: "0.00005" },
+    "gpt-6.1-sol": { input: "0.000002", output: "0.00001" },
+    "gpt-6.1-terra": { input: "0.000001", output: "0.000005" },
+    "gpt-6-luna": { input: "0.0000001", output: "0.0000005" },
+    "gpt-6-sol": { input: "0.000002", output: "0.00001" },
+    "gpt-5.6-sol": { input: "0.000004", output: "0.00002" },
+    "gpt-5.6-terra": { input: "0.000002", output: "0.000012" },
+    "gpt-5.6-luna": { input: "0.0000002", output: "0.0000012" },
+    "gpt-5.5": { input: "0.000005", output: "0.00003" },
+    "gpt-5-mini": { input: "0.00000025", output: "0.000002" },
+  },
 };
 
 // Offerings intentionally seeded without pricing, keyed by provider name to the

bin/lib/discover-registry.test.ts

@@ -20,6 +20,7 @@ const REQUIRED_ENV_BY_PROVIDER: Record<string, readonly string[]> = {
   "google-genai": ["GOOGLE_API_KEY"],
   "opencode-zen": ["OPENCODE_API_KEY", "OPENCODE_BASE_URL"],
   openai: ["OPENAI_API_KEY"],
+  "openai-responses": ["OPENAI_API_KEY"],
   xai: ["XAI_API_KEY"],
 };
 
@@ -31,6 +32,7 @@ describe("PLUGIN_REGISTRY provider contract", () => {
       "google-genai",
       "opencode-zen",
       "openai",
+      "openai-responses",
       "xai",
     ]);
   });
@@ -112,6 +114,20 @@ describe("create factory env wiring", () => {
     expect(() => requireEntry("openai").create({})).toThrow("OPENAI_API_KEY");
   });
 
+  test("openai-responses builds a gpt-5.5 plug-in from OPENAI_API_KEY", () => {
+    const plugin = requireEntry("openai-responses").create({
+      OPENAI_API_KEY: "test-key",
+    });
+    expect(plugin.name).toBe("openai-responses");
+    expect(plugin.models).toContain("gpt-5.5");
+  });
+
+  test("openai-responses raises when OPENAI_API_KEY is absent", () => {
+    expect(() => requireEntry("openai-responses").create({})).toThrow(
+      "OPENAI_API_KEY",
+    );
+  });
+
   test("xai builds a grok-4.5 plug-in from XAI_API_KEY", () => {
     const plugin = requireEntry("xai").create({
       XAI_API_KEY: "test-key",

bin/lib/discover-registry.ts

@@ -20,6 +20,7 @@ import {
 } from "@intx/inference-discovery-google-genai";
 import {
   createOpenAIPlugin,
+  createOpenAIResponsesPlugin,
   createOpencodeZenPlugin,
   createXaiPlugin,
 } from "@intx/inference-discovery-openai";
@@ -79,6 +80,14 @@ function opencodeZenCreate(env: Record<string, string>): ProviderPlugin {
   return createOpencodeZenPlugin({ apiKey, baseUrl });
 }
 
+function openaiResponsesCreate(env: Record<string, string>): ProviderPlugin {
+  const apiKey = env.OPENAI_API_KEY;
+  if (apiKey === undefined) {
+    throw new Error("OPENAI_API_KEY missing from validated env");
+  }
+  return createOpenAIResponsesPlugin({ apiKey });
+}
+
 function xaiCreate(env: Record<string, string>): ProviderPlugin {
   const apiKey = env.XAI_API_KEY;
   if (apiKey === undefined) {
@@ -108,6 +117,11 @@ export const PLUGIN_REGISTRY: readonly RegisteredPlugin[] = [
     requiredEnv: ["OPENAI_API_KEY"],
     create: openaiCreate,
   },
+  {
+    name: "openai-responses",
+    requiredEnv: ["OPENAI_API_KEY"],
+    create: openaiResponsesCreate,
+  },
   {
     name: "xai",
     requiredEnv: ["XAI_API_KEY"],

packages/inference-catalog/src/catalog.test.ts

@@ -8,6 +8,7 @@ import {
   AnthropicQuirks,
   GoogleGenAIQuirks,
   OpenAIQuirks,
+  ResponsesQuirks,
 } from "@intx/inference/providers";
 import {
   baseURLForCatalogProvider,
@@ -26,6 +27,7 @@ const quirkValidatorByPlugin: Record<CatalogPlugin, Type> = {
   anthropic: AnthropicQuirks,
   openai: OpenAIQuirks,
   "openai-compatible": OpenAIQuirks,
+  "openai-responses": ResponsesQuirks,
   "google-genai": GoogleGenAIQuirks,
 };
 
@@ -71,6 +73,26 @@ describe("first-party OpenAI maxEffort", () => {
   }
 });
 
+describe("OpenAI Responses reasoning", () => {
+  // Same maxEffort contract as above, read from the Responses adapter's
+  // nested key, plus the summary every Responses offering requests.
+  // Exact equality, so a removed or swapped field fails the build the
+  // same as for Direct.
+  const expected: Record<string, unknown> = {
+    "gpt-5.5": { maxEffort: "xhigh", summary: "detailed" },
+    "gpt-5.6-sol": { maxEffort: "max", summary: "detailed" },
+  };
+  const responses = catalogProviders.find((p) => p.name === "OpenAI Responses");
+  if (responses === undefined) throw new Error("OpenAI Responses is missing");
+  for (const offering of responses.offerings) {
+    test(`${offering.model} carries the expected reasoning quirks`, () => {
+      expect(offering.quirks["reasoning"]).toEqual(
+        expected[offering.model] ?? { summary: "detailed" },
+      );
+    });
+  }
+});
+
 describe("catalog offering capabilities", () => {
   for (const provider of catalogProviders) {
     for (const offering of provider.offerings) {

packages/inference-catalog/src/models.ts

@@ -102,4 +102,9 @@ export const catalogModels: CatalogModelSpec[] = [
   { canonicalName: "grok-4.5", displayName: "Grok 4.5" },
   { canonicalName: "grok-4.6", displayName: "Grok 4.6" },
   { canonicalName: "grok-build-0.1", displayName: "Grok Build 0.1" },
+  { canonicalName: "gpt-6-astra", displayName: "GPT-6 Astra" },
+  { canonicalName: "gpt-6.1-sol", displayName: "GPT-6.1 Sol" },
+  { canonicalName: "gpt-6.1-terra", displayName: "GPT-6.1 Terra" },
+  { canonicalName: "gpt-6-luna", displayName: "GPT-6 Luna" },
+  { canonicalName: "gpt-6-sol", displayName: "GPT-6 Sol" },
 ];

packages/inference-catalog/src/providers.ts

@@ -4,12 +4,16 @@ import {
   OPENAI_FIRSTPARTY_QUIRKS,
   OPENAI_FIRSTPARTY_XHIGH_QUIRKS,
   OPENAI_REASONING_QUIRKS,
+  OPENAI_RESPONSES_MAX_QUIRKS,
+  OPENAI_RESPONSES_REASONING_QUIRKS,
+  OPENAI_RESPONSES_XHIGH_QUIRKS,
 } from "./quirks";
 
 export type CatalogPlugin =
   | "anthropic"
   | "openai"
   | "openai-compatible"
+  | "openai-responses"
   | "google-genai";
 
 export type DiscoverySource = {
@@ -109,6 +113,33 @@ const OPENAI_REASONING_NO_VISION_CAPABILITIES: Capability[] = [
   "structured-output-streaming",
 ];
 
+// The Responses-wire sets proven by the openai-responses matrix rows:
+// gpt-5-mini carries the full builder surface; every other offering carries
+// the representative streaming, buffered, tool follow-up, and reasoning
+// subset. The guard test re-bakes each offering from its provenance tuple.
+const OPENAI_RESPONSES_FULL_OFFERING_CAPABILITIES: Capability[] = [
+  "plain-text",
+  "plain-text-streaming",
+  "function-calling",
+  "function-calling-multi-turn",
+  "function-calling-multi-turn-streaming",
+  "vision-input",
+  "document-input",
+  "code-execution",
+  "code-execution-streaming",
+  "reasoning-content",
+  "reasoning-content-streaming",
+  "grounding",
+  "grounding-streaming",
+];
+
+const OPENAI_RESPONSES_REPRESENTATIVE_OFFERING_CAPABILITIES: Capability[] = [
+  "plain-text",
+  "plain-text-streaming",
+  "function-calling-multi-turn",
+  "reasoning-content",
+];
+
 // Plain-text + reasoning only (no vision, no structured output): glm-5, glm-5.1,
 // deepseek-v4-pro, deepseek-v4-flash.
 const OPENAI_REASONING_TEXT_CAPABILITIES: Capability[] = [
@@ -510,6 +541,102 @@ export const catalogProviders: CatalogProviderSpec[] = [
       },
     ],
   },
+  {
+    name: "OpenAI Responses",
+    plugin: "openai-responses",
+    baseURL: "https://api.openai.com/v1",
+    offerings: [
+      {
+        model: "gpt-6-astra",
+        priority: 0,
+        discoverySource: { provider: "openai-responses", model: "gpt-6-astra" },
+        curatedCapabilities: [],
+        capabilities: OPENAI_RESPONSES_REPRESENTATIVE_OFFERING_CAPABILITIES,
+        quirks: OPENAI_RESPONSES_REASONING_QUIRKS,
+      },
+      {
+        model: "gpt-6.1-sol",
+        priority: 5,
+        discoverySource: { provider: "openai-responses", model: "gpt-6.1-sol" },
+        curatedCapabilities: [],
+        capabilities: OPENAI_RESPONSES_REPRESENTATIVE_OFFERING_CAPABILITIES,
+        quirks: OPENAI_RESPONSES_REASONING_QUIRKS,
+      },
+      {
+        model: "gpt-6.1-terra",
+        priority: 10,
+        discoverySource: {
+          provider: "openai-responses",
+          model: "gpt-6.1-terra",
+        },
+        curatedCapabilities: [],
+        capabilities: OPENAI_RESPONSES_REPRESENTATIVE_OFFERING_CAPABILITIES,
+        quirks: OPENAI_RESPONSES_REASONING_QUIRKS,
+      },
+      {
+        model: "gpt-6-luna",
+        priority: 15,
+        discoverySource: { provider: "openai-responses", model: "gpt-6-luna" },
+        curatedCapabilities: [],
+        capabilities: OPENAI_RESPONSES_REPRESENTATIVE_OFFERING_CAPABILITIES,
+        quirks: OPENAI_RESPONSES_REASONING_QUIRKS,
+      },
+      {
+        model: "gpt-6-sol",
+        priority: 20,
+        discoverySource: { provider: "openai-responses", model: "gpt-6-sol" },
+        curatedCapabilities: [],
+        capabilities: OPENAI_RESPONSES_REPRESENTATIVE_OFFERING_CAPABILITIES,
+        quirks: OPENAI_RESPONSES_REASONING_QUIRKS,
+      },
+      {
+        model: "gpt-5.6-sol",
+        priority: 25,
+        discoverySource: { provider: "openai-responses", model: "gpt-5.6-sol" },
+        curatedCapabilities: [],
+        capabilities: OPENAI_RESPONSES_REPRESENTATIVE_OFFERING_CAPABILITIES,
+        quirks: OPENAI_RESPONSES_MAX_QUIRKS,
+      },
+      {
+        model: "gpt-5.6-terra",
+        priority: 30,
+        discoverySource: {
+          provider: "openai-responses",
+          model: "gpt-5.6-terra",
+        },
+        curatedCapabilities: [],
+        capabilities: OPENAI_RESPONSES_REPRESENTATIVE_OFFERING_CAPABILITIES,
+        quirks: OPENAI_RESPONSES_REASONING_QUIRKS,
+      },
+      {
+        model: "gpt-5.6-luna",
+        priority: 35,
+        discoverySource: {
+          provider: "openai-responses",
+          model: "gpt-5.6-luna",
+        },
+        curatedCapabilities: [],
+        capabilities: OPENAI_RESPONSES_REPRESENTATIVE_OFFERING_CAPABILITIES,
+        quirks: OPENAI_RESPONSES_REASONING_QUIRKS,
+      },
+      {
+        model: "gpt-5.5",
+        priority: 40,
+        discoverySource: { provider: "openai-responses", model: "gpt-5.5" },
+        curatedCapabilities: [],
+        capabilities: OPENAI_RESPONSES_REPRESENTATIVE_OFFERING_CAPABILITIES,
+        quirks: OPENAI_RESPONSES_XHIGH_QUIRKS,
+      },
+      {
+        model: "gpt-5-mini",
+        priority: 45,
+        discoverySource: { provider: "openai-responses", model: "gpt-5-mini" },
+        curatedCapabilities: [],
+        capabilities: OPENAI_RESPONSES_FULL_OFFERING_CAPABILITIES,
+        quirks: OPENAI_RESPONSES_REASONING_QUIRKS,
+      },
+    ],
+  },
   {
     name: "Gemini Direct",
     plugin: "google-genai",

packages/inference-catalog/src/quirks.ts

@@ -28,6 +28,25 @@ export const OPENAI_FIRSTPARTY_MAX_QUIRKS: Record<string, unknown> = {
   maxEffort: "max",
 };
 
+// Responses-wire counterparts of the maxEffort quirks above: the Responses
+// adapter reads the value from `reasoning.maxEffort`, not the top level, and
+// rejects the Chat keys. The values are carried from the Chat entries for
+// the same models — no Responses-side capture has yet sent `xhigh` or `max`
+// on this wire (INTR-720 tracks per-capture rig options to prove them).
+// Every Responses offering also requests a `detailed` reasoning summary:
+// the fixtures prove the server returns summary text when asked, and the
+// adapter only forwards `summary` when this quirk is set — without it,
+// `reasoning-content` calls would yield empty thinking blocks.
+export const OPENAI_RESPONSES_REASONING_QUIRKS: Record<string, unknown> = {
+  reasoning: { summary: "detailed" },
+};
+export const OPENAI_RESPONSES_XHIGH_QUIRKS: Record<string, unknown> = {
+  reasoning: { maxEffort: "xhigh", summary: "detailed" },
+};
+export const OPENAI_RESPONSES_MAX_QUIRKS: Record<string, unknown> = {
+  reasoning: { maxEffort: "max", summary: "detailed" },
+};
+
 // The reasoning-bearing opencode-zen relay offerings — the kimi models plus
 // qwen, deepseek, glm, and mimo — advertise reasoning_content on the wire. The
 // OpenAI adapter no longer forces reasoning content by default, so

packages/inference-discovery-openai/README.md

@@ -10,8 +10,9 @@ is organised in two layers:
 - `deployments/` — concrete deployments built on the protocol
   layer. Each deployment names a provider, lists its models,
   declares its auth and redaction policy, and (where needed)
-  overrides reasoning extraction. Three deployments ship today: the
-  OpenCode Zen relay, first-party OpenAI, and first-party xAI.
+  overrides reasoning extraction. Four deployments ship today: the
+  OpenCode Zen relay, first-party OpenAI (Chat Completions), first-party
+  OpenAI Responses, and first-party xAI.
 
 See [`@intx/inference-discovery`](../inference-discovery/README.md)
 for the runtime, the plug-in contract, and the `discover` CLI.
@@ -88,6 +89,46 @@ limitation.
 | ---------------- | ------------------------------------------------------------ |
 | `OPENAI_API_KEY` | Sent as `Authorization: Bearer <key>`. Redacted in fixtures. |
 
+## OpenAI Responses
+
+The `openai-responses` deployment probes first-party `api.openai.com` on
+the Responses wire (`/v1/responses`), under provider name
+`openai-responses` (sessions land in this package's `sessions/` tree
+alongside the Chat Completions deployments). The base URL is fixed to
+`https://api.openai.com/v1`; the deployment reads only `OPENAI_API_KEY`.
+
+```ts
+import { createOpenAIResponsesPlugin } from "@intx/inference-discovery-openai";
+
+const plugin = createOpenAIResponsesPlugin({
+  apiKey: process.env.OPENAI_API_KEY,
+});
+// Hand off to runCapture from @intx/inference-discovery.
+```
+
+Models: the reasoning-capable first-party models (`gpt-6-astra`,
+`gpt-6.1-sol`, `gpt-6.1-terra`, `gpt-6-luna`, `gpt-6-sol`, `gpt-5.6-sol`,
+`gpt-5.6-terra`, `gpt-5.6-luna`, `gpt-5.5`, `gpt-5`, `gpt-5-mini`,
+`gpt-5-nano`, `gpt-5.1`, `gpt-5.2`, `gpt-5.4`, `gpt-5.4-mini`,
+`gpt-5.4-nano`, `o1`, `o3`, `o3-mini`, `o4-mini`). The gpt-4 family is
+excluded pending verification that those models serve `/responses`.
+
+The Responses body builder emits `plain-text`, `plain-text-streaming`,
+`function-calling`, `function-calling-multi-turn`, `reasoning-content`,
+`reasoning-content-streaming`, `vision-input`, `document-input`,
+`code-execution`, `code-execution-streaming`, `grounding`, and
+`grounding-streaming`. Structured output and URL-form file input are out of
+scope until the runtime adapter supports them and discovery verifies them;
+audio, video, and image-output are not built. Refresh the fixtures with a
+live capture run (`bin/discover --provider openai-responses --all` against
+a funded account).
+
+### Environment
+
+| Variable         | Purpose                                                      |
+| ---------------- | ------------------------------------------------------------ |
+| `OPENAI_API_KEY` | Sent as `Authorization: Bearer <key>`. Redacted in fixtures. |
+
 ## xAI
 
 The `xai` deployment probes first-party `api.x.ai` directly, under provider

packages/inference-discovery-openai/sessions/openai-responses/gpt-5-mini/code-execution-streaming/exchanges/0/request-headers.json

@@ -0,0 +1,4 @@
+{
+  "Content-Type": "application/json",
+  "Authorization": "<REDACTED>"
+}