omnigateway 0.7.1 → 0.7.3

package/README.md

@@ -21,9 +21,10 @@ omni start
 
 ## What it does
 
-- **Speaks both dialects.** `POST /v1/messages` (Anthropic) and
-  `POST /v1/chat/completions` (OpenAI), including streaming, translated to
-  whichever provider actually serves the request.
+- **Speaks three dialects.** `POST /v1/messages` (Anthropic),
+  `POST /v1/chat/completions` (OpenAI) and `POST /v1/responses` (OpenAI
+  Responses, which is the only API Codex CLI speaks), including streaming,
+  translated to whichever provider actually serves the request.
 - **Filters bulky tool history, optionally.** Built-in RTK filters shorten
   eligible large or repetitive shell and recognizable command output before
   provider dispatch, preserve errors and non-tool-result content, and default off.
@@ -289,6 +290,7 @@ Everything above is also available in the browser at
 | --- | --- | --- |
 | `POST` | `/v1/messages` | Anthropic Messages API |
 | `POST` | `/v1/chat/completions` | OpenAI Chat Completions API |
+| `POST` | `/v1/responses` | OpenAI Responses API (stateless: no stored responses, no `previous_response_id`) |
 | `POST` | `/v1/messages/count_tokens` | Anthropic-compatible local token estimation (authenticated) |
 | `GET` | `/v1/models` | Listing in both dialects, filtered by your key's allowlist |
 | `GET` | `/health` | Unauthenticated liveness check |
@@ -315,7 +317,7 @@ Most tools that accept a custom base URL work unchanged: set it to
 
 Every response carries the limit headers of the surface you asked on, so an SDK
 backs off using the code it already ships — the Anthropic dialect on
-`/v1/messages`, the OpenAI one on `/v1/chat/completions`:
+`/v1/messages`, the OpenAI one on `/v1/chat/completions` and `/v1/responses`:
 
 ```http
 anthropic-ratelimit-requests-limit: 2000        x-ratelimit-limit-requests: 2000

package/bin/omni.js

@@ -1753,6 +1753,10 @@ class GatewayError extends Error {
 }
 // packages/ir/src/logger.ts
 var LEVELS = { debug: 10, info: 20, warn: 30, error: 40 };
+function safeToken(value) {
+  const text = typeof value === "string" ? value : String(value);
+  return /^[A-Za-z0-9_.:/+-]{1,64}$/.test(text) ? text : "(unprintable)";
+}
 var ANSI2 = /\u001B\[[0-9;]*m/g;
 var HEAD = /^(\S+) (DEBUG|INFO|WARN|ERROR) {1,2}(.*)$/;
 var FIELDS = / {2}(?=[a-zA-Z]+=)/;
@@ -17873,6 +17877,9 @@ function codecAdapter(id, capabilities2, codec2, origins) {
   };
 }
 
+// packages/providers/src/body.ts
+import { createHash } from "crypto";
+
 // packages/providers/src/headers.ts
 var SAFE = /^[\x20-\x7E]{1,200}$/;
 function env(name, fallback) {
@@ -17928,14 +17935,14 @@ function mergeHeaders(base, overrides) {
 
 // packages/providers/src/anthropic/profile.ts
 var host = stainlessHost(process.platform, process.arch);
-var ANTHROPIC_CLI_VERSION = env("OMNI_ANTHROPIC_CLI_VERSION", "2.1.219");
+var ANTHROPIC_CLI_VERSION = env("OMNI_ANTHROPIC_CLI_VERSION", "2.1.258");
 var anthropicProfile = {
   headers: [
     ["User-Agent", env("OMNI_UA_ANTHROPIC", `claude-cli/${ANTHROPIC_CLI_VERSION} (external, cli)`)],
     ["x-app", "cli"],
     ["anthropic-dangerous-direct-browser-access", "true"],
     ["X-Stainless-Lang", "js"],
-    ["X-Stainless-Package-Version", env("OMNI_ANTHROPIC_STAINLESS_PACKAGE_VERSION", "0.94.0")],
+    ["X-Stainless-Package-Version", env("OMNI_ANTHROPIC_STAINLESS_PACKAGE_VERSION", "0.112.1")],
     ["X-Stainless-OS", envOrNull("OMNI_ANTHROPIC_STAINLESS_OS") ?? host.os],
     ["X-Stainless-Arch", envOrNull("OMNI_ANTHROPIC_STAINLESS_ARCH") ?? host.arch],
     ["X-Stainless-Runtime", "node"],
@@ -18177,19 +18184,18 @@ function orderFields(obj, order) {
   }
   return out;
 }
-var BUILD_REVISION = envOr("OMNI_ANTHROPIC_BUILD_REVISION", "250");
-var CLI_VERSION = envOr("OMNI_ANTHROPIC_CLI_VERSION", "2.1.219");
 var CCH_PLACEHOLDER = "00000";
 var CCH_SEED = 0x6e52736ac806831en;
 var CCH_MASK = 0xfffffn;
+var CCV_SALT = "59cf53e54c78";
 var BILLING_PREFIX = "x-anthropic-billing-header:";
 var AGENT_PREAMBLE = "You are a Claude agent, built on Anthropic's Claude Agent SDK.";
-function envOr(name, fallback) {
-  const raw = Bun.env[name];
-  return typeof raw === "string" && /^[\x20-\x7E]{1,200}$/.test(raw) ? raw : fallback;
+function ccVersionSuffix(firstUserText) {
+  const picked = [4, 7, 20].map((i) => firstUserText[i] ?? "0").join("");
+  return createHash("sha256").update(`${CCV_SALT}${picked}${ANTHROPIC_CLI_VERSION}`).digest("hex").slice(0, 3);
 }
-function billingBlock() {
-  return `${BILLING_PREFIX} cc_version=${CLI_VERSION}.${BUILD_REVISION}; ` + `cc_entrypoint=cli; cch=${CCH_PLACEHOLDER};`;
+function billingBlock(firstUserText) {
+  return `${BILLING_PREFIX} cc_version=${ANTHROPIC_CLI_VERSION}.${ccVersionSuffix(firstUserText)}; ` + `cc_entrypoint=cli; cch=${CCH_PLACEHOLDER};`;
 }
 var BANNED_SUBSTRINGS = [
   "github.com/anomalyco/opencode",
@@ -18207,7 +18213,7 @@ var REWRITES = [
     "Environment context you are running in:"
   ]
 ];
-function applyAnthropicSystem(system) {
+function applyAnthropicSystem(system, firstUserText) {
   const kept = [];
   for (const block of system) {
     const text = block.text.split(/\n{2,}/).filter((p) => !BANNED_SUBSTRINGS.some((b) => p.includes(b))).filter((p) => !IDENTITY_PREFIXES.some((i) => p.trimStart().startsWith(i))).filter((p) => !p.includes(BILLING_PREFIX)).filter((p) => p.trim() !== AGENT_PREAMBLE).join(`
@@ -18224,17 +18230,26 @@ function applyAnthropicSystem(system) {
       });
     }
   }
-  return [{ type: "text", text: billingBlock() }, { type: "text", text: AGENT_PREAMBLE }, ...kept];
+  return [
+    { type: "text", text: billingBlock(firstUserText) },
+    { type: "text", text: AGENT_PREAMBLE },
+    ...kept
+  ];
 }
 function computeCch(body) {
   const digest = Bun.hash.xxHash64(Buffer.from(body, "utf8"), CCH_SEED);
   return (digest & CCH_MASK).toString(16).padStart(5, "0");
 }
+var BILLING_ANCHOR = `"system":[{"type":"text","text":"${BILLING_PREFIX}`;
 function signAnthropicBody(json2) {
   const needle = `cch=${CCH_PLACEHOLDER};`;
-  if (!json2.includes(needle))
+  const anchor = json2.indexOf(BILLING_ANCHOR);
+  if (anchor < 0)
     return json2;
-  return json2.replace(needle, `cch=${computeCch(json2)};`);
+  const at = json2.indexOf(needle, anchor);
+  if (at < 0)
+    return json2;
+  return `${json2.slice(0, at)}cch=${computeCch(json2)};${json2.slice(at + needle.length)}`;
 }
 
 // packages/providers/src/catalog.ts
@@ -19197,9 +19212,11 @@ var anthropicCodec = {
       cloak,
       autoCache: input.autoCacheEnabled === true
     });
+    const firstUser = input.request.messages.find((m) => m.role === "user");
+    const firstText = firstUser?.content.find((b) => b.type === "text");
     const withSystem = {
       ...body,
-      system: applyAnthropicSystem(body.system ?? []),
+      system: applyAnthropicSystem(body.system ?? [], firstText?.text ?? ""),
       stream: true
     };
     const protocol = [
@@ -20642,10 +20659,10 @@ function toCustomResponsesWire(req, model) {
     note2("custom:context-1m-dropped");
   for (const message of req.messages) {
     const parts = [];
-    const inlined = message.role === "system";
-    if (inlined)
-      note2("custom:system-turn-inlined");
-    const role = inlined ? "user" : message.role;
+    const asDeveloper = message.role === "system";
+    if (asDeveloper)
+      note2("custom:system-turn-as-developer");
+    const role = asDeveloper ? "developer" : message.role;
     const flush = () => {
       if (parts.length === 0)
         return;
@@ -20657,9 +20674,7 @@ function toCustomResponsesWire(req, model) {
         case "text":
           parts.push({
             type: role === "assistant" ? "output_text" : "input_text",
-            text: inlined ? `<system-reminder>
-${block.text}
-</system-reminder>` : block.text
+            text: block.text
           });
           break;
         case "image":
@@ -20737,7 +20752,7 @@ function metadata(data, fail) {
 function endpointUrl(origin, basePath, protocol) {
   const suffix = protocol === "chat_completions" ? "chat/completions" : "responses";
   const base = `${origin}${basePath}`.replace(/\/+$/, "");
-  return base.endsWith("/v1") ? `${base}/${suffix}` : `${base}/v1/${suffix}`;
+  return /\/v\d+$/i.test(base) ? `${base}/${suffix}` : `${base}/v1/${suffix}`;
 }
 function protocolOf(state, fail) {
   if (state !== null && typeof state === "object") {
@@ -20778,7 +20793,7 @@ var customCodec = {
 // packages/providers/src/custom/index.ts
 var customAdapter = codecAdapter("custom", customDescriptor.capabilities, customCodec);
 // packages/providers/src/grok/codec.ts
-import { createHash as createHash2 } from "crypto";
+import { createHash as createHash3 } from "crypto";
 
 // packages/providers/src/grok/decode.ts
 var ERROR_CODE = {
@@ -20977,7 +20992,7 @@ async function* decodeGrokResponses(messages) {
 }
 
 // packages/providers/src/grok/wire.ts
-import { createHash } from "crypto";
+import { createHash as createHash2 } from "crypto";
 var MAX_TOOLS = 200;
 function encodeToolChoice2(c) {
   switch (c.type) {
@@ -20993,7 +21008,7 @@ function encodeToolChoice2(c) {
 }
 function promptCacheKey(instructions, firstInput) {
   const stable = JSON.stringify({ instructions, firstInput });
-  return createHash("sha256").update(stable).digest("hex").slice(0, 32);
+  return createHash2("sha256").update(stable).digest("hex").slice(0, 32);
 }
 function toGrokWire(req, model) {
   const degradations = [];
@@ -21006,10 +21021,10 @@ function toGrokWire(req, model) {
     note2("grok:context-1m-dropped");
   for (const message of req.messages) {
     const parts = [];
-    const inlined = message.role === "system";
-    if (inlined)
-      note2("grok:system-turn-inlined");
-    const role = inlined ? "user" : message.role;
+    const asDeveloper = message.role === "system";
+    if (asDeveloper)
+      note2("grok:system-turn-as-developer");
+    const role = asDeveloper ? "developer" : message.role;
     const flush = () => {
       if (parts.length === 0)
         return;
@@ -21021,9 +21036,7 @@ function toGrokWire(req, model) {
         case "text":
           parts.push({
             type: role === "assistant" ? "output_text" : "input_text",
-            text: inlined ? `<system-reminder>
-${block.text}
-</system-reminder>` : block.text
+            text: block.text
           });
           break;
         case "image":
@@ -21106,7 +21119,7 @@ var PROXY_URL = "https://cli-chat-proxy.grok.com/v1/responses";
 var API_URL = "https://api.x.ai/v1/responses";
 var CONV_NAMESPACE = "omni-grok-conv";
 function derivedUuid(value) {
-  const h = createHash2("sha256").update(`${CONV_NAMESPACE}:${value}`).digest("hex");
+  const h = createHash3("sha256").update(`${CONV_NAMESPACE}:${value}`).digest("hex");
   const variant = (Number.parseInt(h[16], 16) & 3 | 8).toString(16);
   return [
     h.slice(0, 8),
@@ -21815,6 +21828,7 @@ var ERROR_CODE2 = {
   content_policy_violation: "CONTENT_FILTER"
 };
 var KNOWN_EVENTS3 = new Set([
+  "keepalive",
   "response.created",
   "response.queued",
   "response.in_progress",
@@ -21848,7 +21862,8 @@ function json7(data) {
     return null;
   }
 }
-async function* decodeResponses(messages) {
+async function* decodeResponses(messages, options = {}) {
+  const nativeReasoning = options.nativeReasoning === true;
   const indices = new Map;
   let next = 0;
   const irIndex = (outputIndex, contentIndex = 0) => {
@@ -21893,7 +21908,7 @@ async function* decodeResponses(messages) {
           yield {
             type: "blockStart",
             index: irIndex(d.output_index ?? 0),
-            block: { type: "thinking" }
+            block: nativeReasoning ? { type: "providerNative", provider: "openai", blockType: "reasoning", data: {} } : { type: "thinking" }
           };
         } else if (item.type === "function_call") {
           sawToolCall = true;
@@ -21926,7 +21941,12 @@ async function* decodeResponses(messages) {
         yield {
           type: "blockDelta",
           index: irIndex(d.output_index ?? 0),
-          delta: { type: "thinking", text: String(d.delta ?? "") }
+          delta: nativeReasoning ? {
+            type: "providerNative",
+            provider: "openai",
+            deltaType: "reasoning_summary_text.delta",
+            data: { text: String(d.delta ?? "") }
+          } : { type: "thinking", text: String(d.delta ?? "") }
         };
         break;
       case "response.function_call_arguments.delta":
@@ -21942,6 +21962,20 @@ async function* decodeResponses(messages) {
       case "response.output_item.done": {
         const outputIndex = d.output_index ?? 0;
         if (ownsBlock.delete(outputIndex)) {
+          const item = d.item ?? {};
+          if (nativeReasoning && item.type === "reasoning") {
+            yield {
+              type: "blockDelta",
+              index: irIndex(outputIndex),
+              delta: {
+                type: "providerNative",
+                provider: "openai",
+                deltaType: "response.output_item.done",
+                fold: "merge",
+                data: item
+              }
+            };
+          }
           yield { type: "blockEnd", index: irIndex(outputIndex) };
         }
         break;
@@ -22001,7 +22035,7 @@ async function* decodeResponses(messages) {
 }
 
 // packages/providers/src/openai/wire.ts
-import { createHash as createHash3 } from "crypto";
+import { createHash as createHash4 } from "crypto";
 function suppliedKey(req) {
   const vendor = req.vendor?.openai;
   for (const name of ["prompt_cache_key", "session_id"]) {
@@ -22013,7 +22047,24 @@ function suppliedKey(req) {
 }
 function cacheKey(req, instructions, firstInput) {
   const stable = req.conversationId !== undefined && req.conversationId.length > 0 ? JSON.stringify({ conversation: req.conversationId }) : JSON.stringify({ instructions, firstInput: firstInput ?? null });
-  return createHash3("sha256").update(stable).digest("hex").slice(0, 32);
+  return createHash4("sha256").update(stable).digest("hex").slice(0, 32);
+}
+function repairOrphanedCalls(input) {
+  const answered = new Set;
+  for (const item of input) {
+    const record2 = item;
+    if (record2.type === "function_call_output" && typeof record2.call_id === "string") {
+      answered.add(record2.call_id);
+    }
+  }
+  for (let i = input.length - 1;i >= 0; i--) {
+    const record2 = input[i];
+    if (record2.type !== "function_call" || typeof record2.call_id !== "string")
+      continue;
+    if (answered.has(record2.call_id))
+      continue;
+    input.splice(i + 1, 0, { type: "function_call_output", call_id: record2.call_id, output: "" });
+  }
 }
 function encodeToolChoice5(c) {
   switch (c.type) {
@@ -22038,10 +22089,10 @@ function toResponsesWire(req, model, opts = { oauth: false }) {
     note2("openai:context-1m-dropped");
   for (const message of req.messages) {
     const parts = [];
-    const inlined = message.role === "system";
-    if (inlined)
-      note2("openai:system-turn-inlined");
-    const role = inlined ? "user" : message.role;
+    const asDeveloper = message.role === "system";
+    if (asDeveloper)
+      note2("openai:system-turn-as-developer");
+    const role = asDeveloper ? "developer" : message.role;
     const flush = () => {
       if (parts.length === 0)
         return;
@@ -22053,9 +22104,7 @@ function toResponsesWire(req, model, opts = { oauth: false }) {
         case "text":
           parts.push({
             type: role === "assistant" ? "output_text" : "input_text",
-            text: inlined ? `<system-reminder>
-${block.text}
-</system-reminder>` : block.text
+            text: block.text
           });
           break;
         case "image":
@@ -22087,12 +22136,19 @@ ${block.text}
           });
           break;
         case "providerNative":
+          if (block.provider === "openai") {
+            flush();
+            const { id: _id, ...rest } = block.data;
+            input.push({ type: block.blockType, ...rest });
+            break;
+          }
           note2("openai:anthropic-native-block-dropped");
           break;
       }
     }
     flush();
   }
+  repairOrphanedCalls(input);
   const body = { model, input, stream: req.stream, store: false };
   const instructions = systemText(req.system, "openai", note2);
   if (instructions !== undefined && instructions.length > 0)
@@ -22156,7 +22212,10 @@ var openaiCodec = {
       throw input.fail("AUTH", "openai credential has no token");
     }
     const headers = orderHeaders(mergeHeaders(openaiProfile.headers, protocol), openaiProfile.order);
+    const include = body.include;
+    const nativeReasoning = Array.isArray(include) && include.includes("reasoning.encrypted_content");
     return {
+      decodeState: { nativeReasoning },
       request: {
         url: oauth ? OAUTH_URL2 : API_URL3,
         method: "POST",
@@ -22166,8 +22225,9 @@ var openaiCodec = {
       degradations
     };
   },
-  decode({ body }) {
-    return decodeResponses(parseSse(body));
+  decode({ body, decodeState }) {
+    const state = decodeState;
+    return decodeResponses(parseSse(body), { nativeReasoning: state?.nativeReasoning === true });
   }
 };
 
@@ -22344,8 +22404,10 @@ function parseOrThrow(schema, body2) {
   if (result.success)
     return result.data;
   const issue2 = result.error.issues[0];
-  const path = issue2?.path.join(".") ?? "(root)";
-  throw new GatewayError("BAD_REQUEST", `${path}: ${issue2?.message ?? "invalid request"}`);
+  const segments = issue2?.path ?? [];
+  const path = segments.length === 0 ? "(root)" : segments.map((s) => safeToken(String(s))).join(".");
+  const detail = issue2 === undefined ? "invalid request" : issue2.code === "unrecognized_keys" ? `unrecognized key "${safeToken(issue2.keys[0])}"` : issue2.message;
+  throw new GatewayError("BAD_REQUEST", `${path}: ${detail}`);
 }
 function optionalNumber(value, fallback) {
   if (value === undefined)
@@ -22705,10 +22767,13 @@ function bestPositive(values2) {
   return positive.length > 0 ? Math.min(...positive) : null;
 }
 var EXPECTED_OUTPUT_TOKENS = 1000;
-function requestCost(target, request2) {
+function estimateTokens(request2) {
   const cachedTok = estimateCachedInputTokens(request2);
   const freshTok = Math.max(0, estimateInputTokens(request2) - cachedTok);
   const outTok = Math.min(request2.maxTokens ?? EXPECTED_OUTPUT_TOKENS, EXPECTED_OUTPUT_TOKENS);
+  return { cachedTok, freshTok, outTok };
+}
+function requestCost(target, { cachedTok, freshTok, outTok }) {
   return freshTok * target.costPerMTok.input + cachedTok * cacheReadRate(target.costPerMTok) + outTok * target.costPerMTok.output;
 }
 function healthScore(h) {
@@ -22721,7 +22786,8 @@ function score(pairs, input) {
   const tiers = pairs.map((p) => p.target.tier);
   const minTier = Math.min(...tiers);
   const maxTier = Math.max(...tiers);
-  const costs = pairs.map((p) => requestCost(p.target, input.request));
+  const tokens2 = estimateTokens(input.request);
+  const costs = pairs.map((p) => requestCost(p.target, tokens2));
   const bestCost = bestPositive(costs);
   const bestLatency = bestPositive(pairs.flatMap((p) => {
     const h = snapshot.health.get(healthKey(p.credential.id, p.target.model));
@@ -23403,16 +23469,16 @@ async function createKey(store, input) {
 }
 async function setKeyLimits(store, id, input, now = Date.now()) {
   const body2 = parseOrThrow(keyLimitsSchema, input);
-  const key = (await store.keys.list()).find((entry) => entry.id === id);
-  if (key === undefined)
+  const key = await store.keys.get(id);
+  if (key === null)
     throw new GatewayError("BAD_REQUEST", "no such api key");
   await store.keys.setLimits(id, body2.limits);
   return toSummary(store, { ...key, limits: body2.limits }, now);
 }
 async function setKeyModels(store, id, input, now = Date.now()) {
   const body2 = parseOrThrow(keyModelsSchema, input);
-  const key = (await store.keys.list()).find((entry) => entry.id === id);
-  if (key === undefined)
+  const key = await store.keys.get(id);
+  if (key === null)
     throw new GatewayError("BAD_REQUEST", "no such api key");
   await store.keys.setModelAllowlist(id, body2.modelAllowlist);
   return toSummary(store, { ...key, modelAllowlist: body2.modelAllowlist }, now);
@@ -27062,7 +27128,7 @@ function createServiceDeps(input) {
 }
 
 // apps/cli/src/version.ts
-var VERSION = "0.7.1";
+var VERSION = "0.7.3";
 
 // apps/cli/src/commands/service.ts
 function gatewayArgv(root) {