@bman654/clodex 2.15.0 → 2.17.0

package/README.md

@@ -22,8 +22,8 @@ Contributions are welcome — see [CONTRIBUTING.md](./CONTRIBUTING.md) for how t
 npm install -g @bman654/clodex          # 1. install the CLI (Node 22+)
 clodex providers auth openai   # 2. sign in with your ChatGPT/Codex plan (device-code OAuth)
 clodex models                  # 3. pick favorite models and aliases
-clodex models --alias sol=clodex:openai-oauth:gpt-5.6-sol
-clodex models --alias luna=clodex:openai-oauth:gpt-5.6-luna
+clodex models --alias sol=clodex:openai-oauth:gpt-6-sol
+clodex models --alias luna=clodex:openai-oauth:gpt-6-luna
 clodex models --alias terra=clodex:openai-oauth:gpt-5.6-terra
 clodex patch                   # 4. (optional) patch Claude Code so those models are first-class
 clodex claude                  # 5. launch Claude Code on an OpenAI model
@@ -80,7 +80,7 @@ Clodex avoids this. In proxy mode it uses an HTTP proxy to intercept requests bo
 
 Both `clodex claude` and `clodex server` support two bridge modes. A mode flag applies to **that run only**; to change a command's default, add `--save-mode` (e.g. `clodex claude --endpoint --save-mode`). With no flag and nothing saved, both commands default to **proxy** mode, which works with your existing Claude auth.
 
-- **`--proxy`** (the default): a selective man-in-the-middle proxy for `api.anthropic.com`. Claude Code keeps its normal Anthropic login — Anthropic models work untouched — while models named `clodex:<provider-id>:<model-id>` (or their saved aliases) route to the selected configured provider. Switch with `/model clodex:openai-oauth:gpt-5.6-sol` or `/model sol` after patching.
+- **`--proxy`** (the default): a selective man-in-the-middle proxy for `api.anthropic.com`. Claude Code keeps its normal Anthropic login — Anthropic models work untouched — while models named `clodex:<provider-id>:<model-id>` (or their saved aliases) route to the selected configured provider. Switch with `/model clodex:openai-oauth:gpt-6-sol` or `/model sol` after patching.
 - **`--endpoint`**: clodex runs a local Anthropic-format gateway and launches Claude Code with `ANTHROPIC_BASE_URL` pointed at it. All traffic goes through the gateway. With favorites saved, the gateway is multi-route and Claude Code's `/model` menu lists your starting model plus favorites for live switching.
 
 > [!TIP]
@@ -238,7 +238,7 @@ Manage favorite models (max 20) and short aliases. Favorites feed the endpoint-m
 | --- | --- |
 | *(none)* | Interactive manager: search all providers or browse one at a time |
 | `--list` | Print the exact `clodex:<provider-id>:<model-id>` names (and aliases) without opening the manager |
-| `--alias <name=target>` | Save a short name for a favorite, e.g. `--alias sol=clodex:openai-oauth:gpt-5.6-sol` (the `clodex:` prefix is optional in the target) |
+| `--alias <name=target>` | Save a short name for a favorite, e.g. `--alias sol=clodex:openai-oauth:gpt-6-sol` (the `clodex:` prefix is optional in the target) |
 | `--unalias <name>` | Remove a saved short name |
 | `--context <model=stop>` | Choose how much of a model's context window to use: `standard`, `max`, `default` to clear, or a token count such as `500k`. Applies to this run unless `--save` is given |
 | `--save` | With `--context`: store the stop as that model's default |
@@ -276,6 +276,16 @@ Two things worth knowing about the numbers:
   number first only costs usable context. A provider that declares a share of its own
   is still honoured; clodex just never invents one. Use `--context` if you want a
   smaller window than the provider offers.
+- **A window nobody published is not a ceiling.** Some servers list their models
+  without saying how much context each one takes. Clodex reports the 200,000 Claude
+  Code assumes anyway, but that number is a guess, not a measurement — so
+  `clodex models --context <model>=1m --save` raises it instead of being clamped back
+  down to the guess. Clamping still applies in full wherever a window *is* published,
+  by the provider or by clodex's own catalog. If the model is already baked into a
+  patched binary, re-run `clodex patch` so it picks the new window up. Set this model up
+  before this version? Run `clodex providers refresh-models <provider>` once — an older
+  clodex stored the guess as if the server had published it, and until you refresh it is
+  indistinguishable from a real limit and still clamps.
 - **The account ceiling moves.** It is server-side and per-account, and it has
   changed by more than 2x within a single day in the past. `max` reads whatever the
   catalog reports now and clamps to it, so a stale ceiling shrinks the stop rather
@@ -292,7 +302,7 @@ Two things worth knowing about the numbers:
 | `remove <id>` | Remove a provider by id |
 | `refresh-models [id]` | Update cached model lists |
 
-Providers supported: `openai` (API key, platform.openai.com), `openai-oauth` (ChatGPT/Codex plan), and `opencode-go` (OpenCode Go API key). OpenCode Go exposes its Anthropic Messages and Chat Completions models; Responses-only entries are intentionally excluded. See [OpenCode Go provider](docs/opencode-go.md).
+Providers supported: `openai` (API key, platform.openai.com), `openai-oauth` (ChatGPT/Codex plan), and `opencode-go` (OpenCode Go API key). OpenCode Go exposes its Anthropic Messages, Chat Completions, and Responses models; entries whose transport has not been verified against the live endpoint are left out. See [OpenCode Go provider](docs/opencode-go.md).
 
 A **custom OpenAI-compatible server** is any OpenAI-style API, given by its base URL (the part before `/chat/completions`), such as `https://openrouter.ai/api/v1`. `providers add` asks for a name, the base URL and an API key (leave it empty for a local server without auth), lists the server's models to check the connection, and saves the provider as `custom-<name>`. Its models then appear in `clodex models`; once saved as favorites they are addressed as `clodex:custom-<name>:<model>` and can be given a short alias with `clodex models --alias`. The base URL is checked before anything is saved: plain `http://` is accepted only after you confirm it, and only for loopback or private-network addresses, and an `https://` URL whose host is, or resolves to, a loopback (`127.0.0.0/8`, `::1`), private, link-local, unique-local, carrier-grade-NAT or well-known cloud-metadata address is refused. A model's id and name are stored with surrounding whitespace trimmed, and a model whose trimmed id or name still contains a control character is skipped, whether the list is fetched by `providers add` or by `providers refresh-models`, and error text from the server is shown with control characters replaced by spaces. Only the OpenAI-compatible kind is offered in the menu.
 
@@ -357,7 +367,7 @@ clodex --version    # version
   this pre-dispatch intent, not proof of wire serialization. If the provider
   SDK reports that it omitted the tier for a model, clodex warns once and the
   backend default remains in use.
-- **Outbound proxy:** when `HTTP_PROXY`/`HTTPS_PROXY` (and optionally `NO_PROXY`) are set in clodex's environment, all clodex-originated network calls honor them — OAuth sign-in and token refresh, model-list and models.dev refreshes, upstream OpenAI API calls, and the ChatGPT/Codex OAuth WebSocket transport (tunneled via HTTP CONNECT).
+- **Outbound proxy:** when `HTTP_PROXY`/`HTTPS_PROXY` (and optionally `NO_PROXY`) are set in clodex's environment, all clodex-originated network calls honor them — OAuth sign-in and token refresh, model-list and models.dev refreshes, upstream OpenAI API calls, and the ChatGPT/Codex OAuth WebSocket transport (tunneled via HTTP CONNECT). In proxy mode this also covers the web traffic Claude Code itself makes through the clodex proxy, so web fetches and searches work behind a corporate proxy instead of failing on a blocked direct connection.
 - **Provider timeouts:** `CLODEX_UPSTREAM_IDLE_TIMEOUT_MS` controls how long an
   SDK-backed translated stream may produce no event (default `120000`; range
   `10000`–`3600000` ms). `CLODEX_UPSTREAM_TOTAL_TIMEOUT_MS` limits each call

package/dist/cli.js

@@ -346,7 +346,7 @@ import { join } from "path";
 // package.json
 var package_default = {
   name: "@bman654/clodex",
-  version: "2.15.0",
+  version: "2.17.0",
   publishConfig: {
     access: "public"
   },
@@ -434,7 +434,7 @@ var package_default = {
 
 // src/constants.ts
 var CODEX_RESPONSES_LITE_WS_URL = "wss://chatgpt.com/backend-api/codex/responses";
-var CODEX_RESPONSES_LITE_VERSION = "0.153.3";
+var CODEX_RESPONSES_LITE_VERSION = "0.156.0";
 var CODEX_RESPONSES_WEBSOCKETS_BETA = "responses_websockets=2026-02-06";
 var TEST_TIMEOUT_MS = 1e4;
 var CONFLICTING_ENV_VARS = [
@@ -631,6 +631,7 @@ var HEURISTIC_RULES = [
   [/claude/i, 2e5],
   [/deepseek-v4|deepseek-r1|deepseek-reasoner/i, 1e6],
   [/deepseek/i, 64e3],
+  [/gpt-6-(?:astra|sol|luna)/i, 105e4],
   [/gpt-5|gpt-4\.1|o3-|o4-/i, 1e6],
   [/gpt-4o|gpt-4-turbo|gpt-4/i, 128e3],
   [/gpt-oss/i, 131072],
@@ -694,20 +695,22 @@ function getCacheIndex() {
   }
   return cacheIndex;
 }
-function contextWindowFromHeuristics(modelId) {
-  const cached = heuristicCache.get(modelId);
-  if (cached !== void 0) return cached;
+function matchHeuristicWindow(modelId) {
+  if (heuristicCache.has(modelId)) return heuristicCache.get(modelId);
   for (const [pattern, size] of HEURISTIC_RULES) {
     if (pattern.test(modelId)) {
       heuristicCache.set(modelId, size);
       return size;
     }
   }
-  heuristicCache.set(modelId, DEFAULT_CONTEXT_WINDOW);
-  return DEFAULT_CONTEXT_WINDOW;
+  heuristicCache.set(modelId, void 0);
+  return void 0;
+}
+function lookupKnownContextWindow(modelId) {
+  return getCacheIndex().get(modelId) ?? matchHeuristicWindow(modelId);
 }
 function lookupContextWindow(modelId) {
-  return getCacheIndex().get(modelId) ?? contextWindowFromHeuristics(modelId);
+  return lookupKnownContextWindow(modelId) ?? DEFAULT_CONTEXT_WINDOW;
 }
 function resolveContextWindow(modelId, explicit) {
   if (typeof explicit === "number" && explicit > 0) return explicit;
@@ -3160,17 +3163,19 @@ function parseContextStop(value) {
   return parsed;
 }
 function resolveContextStop(limits, stop = "standard") {
-  const standardRaw = positiveInteger(limits.contextWindow) ?? 0;
-  const ceiling = positiveInteger(limits.maxContextWindow) ?? standardRaw;
-  const requested = stop === "standard" ? standardRaw : stop === "max" ? ceiling : positiveInteger(stop) ?? standardRaw;
-  const raw = Math.min(requested, ceiling);
+  const declaredWindow = positiveInteger(limits.contextWindow);
+  const declaredCeiling = positiveInteger(limits.maxContextWindow);
+  const standardRaw = declaredWindow ?? DEFAULT_CONTEXT_WINDOW;
+  const ceiling = declaredCeiling ?? declaredWindow;
+  const requested = stop === "standard" ? standardRaw : stop === "max" ? declaredCeiling ?? standardRaw : positiveInteger(stop) ?? standardRaw;
+  const raw = ceiling === void 0 ? requested : Math.min(requested, ceiling);
   const effective = effectiveContextWindow(raw, limits.effectiveContextPercent);
   const boundary = positiveInteger(limits.pricingBoundary);
   return {
     stop,
     raw,
     effective,
-    ...requested > ceiling ? { clampedFrom: requested } : {},
+    ...ceiling !== void 0 && requested > ceiling ? { clampedFrom: requested } : {},
     crossesPricingBoundary: boundary !== void 0 && effective > boundary
   };
 }
@@ -5208,6 +5213,52 @@ var opencode_go_models_default = [
     upstreamModelId: "minimax-m3",
     family: "minimax"
   },
+  {
+    id: "muse-spark-1.2-contributor",
+    name: "Muse Spark 1.2 Contributor",
+    contextWindow: 1048576,
+    cost: {
+      input: 0.1,
+      output: 0.2,
+      cache_read: 2e-3
+    },
+    modelFormat: "openai",
+    npm: "@ai-sdk/openai",
+    apiUrl: "https://opencode.ai/zen/go/v1",
+    reasoning: true,
+    modalities: [
+      "text",
+      "image"
+    ],
+    compatibility: {
+      supportsReasoningEffort: false
+    },
+    upstreamModelId: "muse-spark-1.2-contributor",
+    family: "muse"
+  },
+  {
+    id: "muse-spark-1.3-contributor",
+    name: "Muse Spark 1.3 Contributor",
+    contextWindow: 1048576,
+    cost: {
+      input: 0.1,
+      output: 0.2,
+      cache_read: 2e-3
+    },
+    modelFormat: "openai",
+    npm: "@ai-sdk/openai",
+    apiUrl: "https://opencode.ai/zen/go/v1",
+    reasoning: true,
+    modalities: [
+      "text",
+      "image"
+    ],
+    compatibility: {
+      supportsReasoningEffort: false
+    },
+    upstreamModelId: "muse-spark-1.3-contributor",
+    family: "muse"
+  },
   {
     id: "qwen3.6-plus",
     name: "Qwen3.6 Plus",
@@ -5471,6 +5522,10 @@ function openCodeGoPinnedApiUrl(npm) {
   if (npm === "@ai-sdk/anthropic") return OPENCODE_GO_ANTHROPIC_BASE_URL;
   return null;
 }
+function openCodeGoPinnedModelApiUrl(npm) {
+  if (npm === "@ai-sdk/openai") return OPENCODE_GO_COMPLETIONS_BASE_URL;
+  return openCodeGoPinnedApiUrl(npm);
+}
 function isProviderConfiguredForTemplate(provider, templateId) {
   if (provider.id === templateId) return true;
   return templateId === OPENCODE_GO_PROVIDER_ID && isRetainedOpenCodeGoProvider(provider);
@@ -5925,21 +5980,22 @@ function deriveBrand(family) {
 }
 
 // src/data/openai-oauth-models.ts
-var GPT_5_6_PRICING_BOUNDARY = 272e3;
-var GPT_5_6_PRICING_NOTE = "Above it, OpenAI prices the full request at 2x input and 1.5x output.";
+var GPT_HIGH_CONTEXT_PRICING_BOUNDARY = 272e3;
+var GPT_HIGH_CONTEXT_PRICING_NOTE = "Above it, OpenAI prices the full request at 2x input and 1.5x output.";
 var CHATGPT_CODEX_UNSUPPORTED_MODELS = /* @__PURE__ */ new Set([
   "gpt-5.5-fast"
   // confirmed: rejected by chatgpt.com/backend-api/codex
 ]);
 var OPENAI_OAUTH_MODEL_SEEDS = [
-  // GPT-6 family. The window and ceiling are what the live Codex catalog returned on
-  // 2026-09-04 and are deliberately NOT the published API numbers: the model card
-  // lists a 1,050,000 context window, but the Codex client is served a smaller one,
-  // and this path is Codex-only. Output limit and the pricing band come from the
-  // card (https://developers.openai.com/api/docs/models/gpt-6-astra), which states
-  // "Prompts with more than 272K input tokens are priced at 2x input and cache rates
-  // and 1.5x output for the full request" — the same boundary the GPT-5.6 family has.
+  // GPT-6 family. The window, ceiling and Responses-Lite flags are what the live
+  // Codex catalog returned (Astra on 2026-09-04, Sol and Luna on 2026-09-22) and are
+  // deliberately NOT the published API numbers: the model cards list a 1,050,000
+  // context window, but the Codex client is served a smaller one, and this path is
+  // Codex-only. All three cards publish a 128,000 output limit and the same 272K
+  // pricing boundary as the GPT-5.6 family.
   { id: "gpt-6-astra", name: "GPT-6 Astra", contextWindow: 272e3, maxContextWindow: 872e3, maxOutputTokens: 128e3, reasoning: true, useResponsesLite: true, preferWebSockets: true },
+  { id: "gpt-6-sol", name: "GPT-6 Sol", contextWindow: 272e3, maxContextWindow: 872e3, maxOutputTokens: 128e3, reasoning: true, useResponsesLite: true, preferWebSockets: true },
+  { id: "gpt-6-luna", name: "GPT-6 Luna", contextWindow: 272e3, maxContextWindow: 872e3, maxOutputTokens: 128e3, reasoning: true, useResponsesLite: true, preferWebSockets: true },
   // "An alias for our flagship general-purpose models, with safeguards calibrated
   // for defensive cybersecurity work" — access is gated on a separate opt-in
   // program, so most installs will never see this id in their catalog.
@@ -5957,10 +6013,10 @@ var OPENAI_OAUTH_MODEL_SEEDS = [
   { id: "gpt-5", name: "GPT-5", contextWindow: 272e3, reasoning: true },
   // o-series reasoning (Plus+)
   { id: "o4-mini", name: "o4 Mini", reasoning: true },
-  { id: "o3", name: "o3", reasoning: true },
+  { id: "o3", name: "o3", contextWindow: 2e5, reasoning: true },
   { id: "o3-mini", name: "o3 Mini", reasoning: true },
-  { id: "o1", name: "o1", reasoning: true },
-  { id: "o1-mini", name: "o1 Mini", reasoning: true }
+  { id: "o1", name: "o1", contextWindow: 2e5, reasoning: true },
+  { id: "o1-mini", name: "o1 Mini", contextWindow: 128e3, reasoning: true }
 ];
 function hasPricingBoundary(id) {
   const version = /^gpt-(\d+)(?:\.(\d+))?(?:-|$)/i.exec(id);
@@ -5974,8 +6030,8 @@ function hasPricingBoundary(id) {
 function openAiPricingMetadata(id) {
   if (!hasPricingBoundary(id)) return {};
   return {
-    pricingBoundary: GPT_5_6_PRICING_BOUNDARY,
-    pricingBoundaryNote: GPT_5_6_PRICING_NOTE
+    pricingBoundary: GPT_HIGH_CONTEXT_PRICING_BOUNDARY,
+    pricingBoundaryNote: GPT_HIGH_CONTEXT_PRICING_NOTE
   };
 }
 var LEGACY_IMPOSED_CONTEXT_PERCENT = 95;
@@ -6008,7 +6064,7 @@ function buildOpenAiOAuthModels() {
       upstreamModelId: seed.id,
       family: prefix,
       brand: deriveBrand(prefix),
-      contextWindow: resolveContextWindow(seed.id, seed.contextWindow),
+      contextWindow: seed.contextWindow ?? lookupKnownContextWindow(seed.id),
       maxContextWindow: seed.maxContextWindow,
       effectiveContextPercent: seed.effectiveContextPercent,
       pricingBoundary: seed.pricingBoundary ?? pricing.pricingBoundary,
@@ -6524,7 +6580,9 @@ function parseModelList(body, npm, allowBareArray) {
     const freeStatus = classifyFreeStatus({
       model: { cost, isFree: row.isFree }
     });
-    const contextWindow = row.context_length ?? row.contextWindow ?? row.context_window ?? row.max_model_len ?? resolveContextWindow(id);
+    const contextWindow = row.context_length ?? row.contextWindow ?? row.context_window ?? row.max_model_len ?? // Left undefined when nothing is known: persisting the invented 200,000
+    // would make it this model's permanent ceiling.
+    lookupKnownContextWindow(id);
     models.push({
       id,
       name: normalizeGoogleDisplayName(row.name, id),
@@ -6556,7 +6614,7 @@ function materializeTemplateModel(template, model, baseUrl) {
     upstreamModelId: model.upstreamModelId ?? normalizedUpstream,
     family,
     brand: model.brand ?? deriveBrand(family),
-    contextWindow: model.contextWindow ?? resolveContextWindow(id),
+    contextWindow: model.contextWindow ?? lookupKnownContextWindow(id),
     isFree: model.isFree ?? isFreeStatus(freeStatus),
     freeStatus,
     modelFormat: model.modelFormat ?? modelFormatForNpm(npm),
@@ -6758,10 +6816,10 @@ function resolveMaterializedApiUrl(cached, provider, npm) {
   if (!isRetainedOpenCodeGoProvider(provider)) {
     return cached.apiUrl ?? provider.api.url ?? "";
   }
-  return openCodeGoPinnedApiUrl(npm);
+  return openCodeGoPinnedModelApiUrl(npm);
 }
 function projectContextStop(cached, providerId, modelId) {
-  const limits = contextLimitsFrom(cached, resolveContextWindow(modelId));
+  const limits = contextLimitsFrom(cached, lookupKnownContextWindow(modelId));
   const stop = selectContextStop(providerId, modelId);
   const resolved = resolveContextStop(limits, stop);
   return {
@@ -7893,7 +7951,7 @@ function outboundHttpProxyAgent(targetUrl, env = process.env) {
     }
     return new HttpsProxyAgent(parsedProxy, { keepAlive: true });
   } catch (err) {
-    console.error(
+    emitParentNotice(
       `clodex: HTTP(S)_PROXY cannot be used for a CONNECT tunnel; using a direct connection (${err instanceof Error ? err.message : String(err)})`
     );
     return void 0;
@@ -10777,7 +10835,7 @@ async function createLanguageModel(spec) {
       apiKey: "",
       ...spec.headers ? { headers: spec.headers } : {},
       fetch: fetchWithoutCredentialHeaders
-    } : { apiKey, ...spec.headers ? { headers: spec.headers } : {} };
+    } : { apiKey, ...baseURL ? { baseURL } : {}, ...spec.headers ? { headers: spec.headers } : {} };
     const openai = createOpenAI(oauthOptions);
     return useResponsesEndpoint ? openai.responses(modelId) : openai.chat(modelId);
   }
@@ -11024,7 +11082,7 @@ function supportsExtendedCodexEffort(modelId) {
   return /^gpt-daybreak(?:-|$)/i.test(modelId);
 }
 function supportsNoneEffort(modelId) {
-  return /^gpt-5\.6(?:-|$)/i.test(modelId) || /^gpt-daybreak-blue(?:-|$)/i.test(modelId);
+  return /^gpt-5\.6(?:-|$)/i.test(modelId) || /^gpt-6-(?:sol|luna)(?:-|$)/i.test(modelId) || /^gpt-daybreak-blue(?:-|$)/i.test(modelId);
 }
 function isReasoningSummaryUnsupportedModel(modelId) {
   return /codex-spark(?:-|$)/i.test(modelId);
@@ -14109,7 +14167,7 @@ function buildDesiredPatchConfig(options = {}) {
         compatibility: model.compatibility,
         upstreamModelId: upstreamModelId2
       });
-      const limits = contextLimitsFrom(model, resolveContextWindow(model.id));
+      const limits = contextLimitsFrom(model, lookupKnownContextWindow(model.id));
       const stop = resolveContextStop(
         limits,
         options.sessionStops ? selectContextStop(provider.id, model.id, prefs.modelContextModes) : readContextStop(prefs.modelContextModes, provider.id, model.id) ?? "standard"
@@ -14678,7 +14736,9 @@ async function runPatchCommand(opts = {}) {
     return 1;
   }
   for (const id of desired.unknownWindows) {
-    p2.log.warn(`No context window metadata for ${id} \u2014 Claude Code will assume the 200k default.`);
+    p2.log.warn(
+      `No context window metadata for ${id} \u2014 Claude Code will assume the 200k default. Refresh the provider's models, then set a window with \`clodex models --context <model=stop> --save\`, then re-run \`clodex patch\`.`
+    );
   }
   reportRejectedModelAliases(desired.rejectedAliasRejections);
   const localPatches = inspectLocalPatchSource(
@@ -15781,7 +15841,7 @@ async function fetchAnthropicModels(baseUrl, apiKey, extraHeaders) {
           upstreamModelId: id,
           family: id.split("-")[0] ?? id,
           brand: deriveBrand(id),
-          contextWindow: resolveContextWindow(id),
+          contextWindow: lookupKnownContextWindow(id),
           modelFormat: "anthropic",
           npm: "@ai-sdk/anthropic",
           apiUrl: root
@@ -16328,7 +16388,7 @@ function buildDynamicOAuthModel(entry, seedById, codexCatalog) {
     upstreamModelId: id,
     family: prefix,
     brand: deriveBrand(prefix),
-    contextWindow: entry.context_window ?? resolveContextWindow(id),
+    contextWindow: entry.context_window ?? lookupKnownContextWindow(id),
     maxContextWindow: entry.max_context_window,
     // Absent means no reduction. clodex reports the window the provider actually
     // gives; deciding how much of it to leave free is the client's job, and Claude
@@ -20433,6 +20493,14 @@ async function startHttpProxy(options) {
     );
   });
   const sockets = /* @__PURE__ */ new Set();
+  let warnedConnectSelfProxy = false;
+  const connectTunnelAgents = /* @__PURE__ */ new Map();
+  const connectTunnelAgent = (proxyUrl, targetUrl) => {
+    if (!connectTunnelAgents.has(proxyUrl)) {
+      connectTunnelAgents.set(proxyUrl, outboundHttpProxyAgent(targetUrl));
+    }
+    return connectTunnelAgents.get(proxyUrl);
+  };
   mitmServer.on("upgrade", (req, socket, head) => {
     forwardAnthropicUpgrade(
       req,
@@ -20462,6 +20530,65 @@ async function startHttpProxy(options) {
       clientSocket.end("HTTP/1.1 400 Bad Request\r\n\r\n");
       return;
     }
+    const targetUrl = `https://${req.url}`;
+    const outboundProxyUrl = outboundProxyUrlForTarget(targetUrl);
+    const bound = proxyServer.address();
+    const selfTargeting = outboundProxyUrl !== void 0 && bound !== null && typeof bound !== "string" && proxyUrlTargetsListener(outboundProxyUrl, bound.address, bound.port);
+    if (selfTargeting && !warnedConnectSelfProxy) {
+      warnedConnectSelfProxy = true;
+      emitParentNotice(
+        "clodex: HTTP(S)_PROXY points at this proxy; tunneling CONNECT direct"
+      );
+    }
+    const outboundAgent = outboundProxyUrl !== void 0 && !selfTargeting ? connectTunnelAgent(outboundProxyUrl, targetUrl) : void 0;
+    if (outboundAgent) {
+      let upstream2;
+      let proxyConnectStatus;
+      req.once("proxyConnect", (response) => {
+        proxyConnectStatus = response.statusCode;
+      });
+      clientSocket.once("close", () => {
+        if (upstream2 && !upstream2.destroyed) upstream2.destroy();
+      });
+      void outboundAgent.connect(req, {
+        host: target.host,
+        port: target.port,
+        secureEndpoint: false
+      }).then((connected) => {
+        upstream2 = connected;
+        if (proxyConnectStatus !== 200 || clientSocket.destroyed) {
+          connected.destroy();
+          if (!clientSocket.destroyed) {
+            clientSocket.end("HTTP/1.1 502 Bad Gateway\r\n\r\n", () => clientSocket.destroy());
+          }
+          return;
+        }
+        let tunnelEstablished2 = false;
+        sockets.add(connected);
+        connected.once("close", () => {
+          sockets.delete(connected);
+          if (tunnelEstablished2 && !clientSocket.destroyed) clientSocket.destroy();
+        });
+        connected.once("error", () => {
+          if (clientSocket.destroyed) return;
+          if (tunnelEstablished2) {
+            clientSocket.destroy();
+            return;
+          }
+          clientSocket.end("HTTP/1.1 502 Bad Gateway\r\n\r\n", () => clientSocket.destroy());
+        });
+        tunnelEstablished2 = true;
+        clientSocket.write("HTTP/1.1 200 Connection Established\r\n\r\n");
+        if (head.length > 0) connected.write(head);
+        clientSocket.pipe(connected);
+        connected.pipe(clientSocket);
+      }).catch(() => {
+        if (!clientSocket.destroyed) {
+          clientSocket.end("HTTP/1.1 502 Bad Gateway\r\n\r\n", () => clientSocket.destroy());
+        }
+      });
+      return;
+    }
     const upstream = net.connect(target.port, target.host);
     let tunnelEstablished = false;
     sockets.add(upstream);
@@ -20527,6 +20654,8 @@ async function startHttpProxy(options) {
       for (const socket of sockets) socket.destroy();
       await new Promise((resolve4) => proxyServer.close(() => resolve4()));
       mitmServer.close();
+      for (const agent of connectTunnelAgents.values()) agent?.destroy();
+      connectTunnelAgents.clear();
       anthropicAgent?.destroy();
       adapter?.close();
     }
@@ -21965,7 +22094,7 @@ ${pc13.bold("Examples:")}
   clodex claude --trace --resume abc-123
   clodex claude --endpoint
   clodex claude --endpoint --save-mode
-  clodex claude --provider openai-oauth --model gpt-5.6-sol
+  clodex claude --provider openai-oauth --model gpt-6-sol
   clodex claude -- --print "hello"
   clodex claude -- --dangerously-skip-permissions`;
 }
@@ -22053,7 +22182,7 @@ Manage favorite models for mid-session switching.
 ${pc13.bold("Usage:")}
   clodex favorites
   clodex models --list
-  clodex models --alias sol=clodex:openai-oauth:gpt-5.6-sol
+  clodex models --alias sol=clodex:openai-oauth:gpt-6-sol
   clodex models --unalias sol
   clodex models --context sol=max --save
   clodex models --json
@@ -22086,13 +22215,13 @@ ${pc13.bold("How it works:")}
   Claude Code binary so they pass model validation and report real context.
 
 ${pc13.bold("Context stops and cost:")}
-  Some providers price large prompts in bands: GPT-5.6 bills the full request at
+  Some providers price large prompts in bands: GPT-5.6 and GPT-6 bill the full request at
   a higher rate above 272,000 input tokens. The standard stop stays under that
   line, so max is opt-in and reports the boundary when you select it.
 
 ${pc13.bold("Examples:")}
   clodex favorites
-  clodex models --alias sol=clodex:openai-oauth:gpt-5.6-sol
+  clodex models --alias sol=clodex:openai-oauth:gpt-6-sol
   clodex models --context sol=max --save
   clodex claude    # switch menu active when favorites are set`;
 }
@@ -22242,7 +22371,7 @@ function applyContextStopAssignments(assignments, save, report, saveHint = "Add
       );
       return 1;
     }
-    const limits = contextLimitsFrom(cached, resolveContextWindow(assignment.modelId));
+    const limits = contextLimitsFrom(cached, lookupKnownContextWindow(assignment.modelId));
     const resolved = resolveContextStop(limits, assignment.stop ?? "standard");
     const target = modelAliasTarget(assignment);
     const label = assignment.label === target ? target : `${assignment.label} (${target})`;