🤖 fix: cap GPT-5.6 context over Codex OAuth

src/common/constants/codexOAuth.test.ts

@@ -1,5 +1,6 @@
 import { describe, expect, it } from "bun:test";
 import {
+  getCodexOauthContextWindowOverride,
   isCodexOauthAllowedModel,
   isCodexOauthAllowedModelId,
   isCodexOauthRequiredModel,
@@ -21,11 +22,20 @@ describe("codexOAuth model gating", () => {
 
   it("allows the GPT-5.6 family through Codex OAuth without requiring it", () => {
     // Includes the bare alias: it is a servable model id (OpenAI routes it to Sol).
-    for (const model of ["gpt-5.6", "gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna"]) {
+    const contextLimits = {
+      "gpt-5.6": 272_000,
+      "gpt-5.6-sol": 272_000,
+      "gpt-5.6-terra": 128_000,
+      "gpt-5.6-luna": 128_000,
+    } as const;
+
+    for (const [model, contextLimit] of Object.entries(contextLimits)) {
       expect(isCodexOauthAllowedModelId(model)).toBe(true);
       expect(isCodexOauthAllowedModelId(`openai:${model}`)).toBe(true);
       expect(isCodexOauthRequiredModelId(model)).toBe(false);
       expect(isCodexOauthRequiredModelId(`openai:${model}`)).toBe(false);
+      expect(getCodexOauthContextWindowOverride(model)).toBe(contextLimit);
+      expect(getCodexOauthContextWindowOverride(`openai:${model}`)).toBe(contextLimit);
     }
   });
 

src/common/constants/codexOAuth.ts

@@ -131,9 +131,14 @@ export const CODEX_OAUTH_REQUIRED_MODELS = new Set<string>([
  * still use the public OpenAI limits.
  */
 const CODEX_OAUTH_CONTEXT_WINDOW_OVERRIDES: Record<string, number> = {
-  // User-reported routing limit: GPT-5.5's public API window is 1.05M, but the
-  // ChatGPT/Codex OAuth backend rejects prompts near that size and must compact at ~270K.
+  // The public API exposes a 1.05M window for these models, but the ChatGPT/Codex
+  // OAuth routing layer has smaller tier-specific limits. Keep the auth-route caps
+  // separate from model metadata so API-key requests retain the full public window.
   "gpt-5.5": 272_000,
+  "gpt-5.6": 272_000,
+  "gpt-5.6-sol": 272_000,
+  "gpt-5.6-terra": 128_000,
+  "gpt-5.6-luna": 128_000,
 };
 
 function normalizeCodexOauthModelId(modelId: string): string {

src/common/utils/compaction/contextLimit.test.ts

@@ -77,9 +77,6 @@ describe("getEffectiveContextLimit", () => {
     expect(defaultOauthLimit).toBe(272_000);
   });
 
-  // Pinned to the literal "gpt-5.5" (not KNOWN_MODELS.GPT.id): the GPT alias now
-  // tracks gpt-5.6-sol, which has no Codex OAuth context override, while the cap
-  // under test belongs to the retired-but-still-priced gpt-5.5 model string.
   test("inherits the Codex OAuth context cap from a mapped OpenAI model", () => {
     const limit = getEffectiveContextLimit(
       "openai:team-gpt",
@@ -93,13 +90,22 @@ describe("getEffectiveContextLimit", () => {
     expect(limit).toBe(272_000);
   });
 
-  test("does not cap GPT-5.6 Sol under Codex OAuth (no context-window override published)", () => {
-    const oauthOnlyLimit = getEffectiveContextLimit(
-      "openai:gpt-5.6-sol",
-      false,
-      providersWithOpenAI({ codexOauthSet: true })
-    );
-    expect(oauthOnlyLimit).toBe(1_050_000);
+  test("caps the GPT-5.6 family at each tier's Codex OAuth context window", () => {
+    const contextLimits = {
+      "gpt-5.6": 272_000,
+      "gpt-5.6-sol": 272_000,
+      "gpt-5.6-terra": 128_000,
+      "gpt-5.6-luna": 128_000,
+    } as const;
+
+    for (const [model, contextLimit] of Object.entries(contextLimits)) {
+      const oauthOnlyLimit = getEffectiveContextLimit(
+        `openai:${model}`,
+        false,
+        providersWithOpenAI({ codexOauthSet: true })
+      );
+      expect(oauthOnlyLimit).toBe(contextLimit);
+    }
   });
 
   test("does not apply the GPT-5.5 OAuth cap to gateway-routed models", () => {

src/common/utils/tokens/tokenMeterUtils.test.ts

@@ -77,8 +77,6 @@ describe("calculateTokenMeterData", () => {
   });
 
   test("uses the Codex OAuth cap for GPT-5.5 token meter percentages", () => {
-    // Pin to the literal model string: gpt-5.5's 272K Codex OAuth cap remains in
-    // production even though KNOWN_MODELS.GPT now points at gpt-5.6-sol.
     const result = calculateTokenMeterData(SAMPLE_USAGE, "openai:gpt-5.5", false, false, {
       openai: {
         apiKeySet: false,
@@ -92,7 +90,7 @@ describe("calculateTokenMeterData", () => {
     expect(result.totalPercentage).toBeCloseTo((11_000 / 272_000) * 100);
   });
 
-  test("does not cap GPT-5.6 Sol under Codex OAuth (no context-window override published)", () => {
+  test("uses the Codex OAuth cap for GPT-5.6 token meter percentages", () => {
     const result = calculateTokenMeterData(SAMPLE_USAGE, "openai:gpt-5.6-sol", false, false, {
       openai: {
         apiKeySet: false,
@@ -102,9 +100,8 @@ describe("calculateTokenMeterData", () => {
       },
     });
 
-    // GA model page: 1.05M-token context window for every GPT-5.6 tier.
-    expect(result.maxTokens).toBe(1_050_000);
-    expect(result.totalPercentage).toBeCloseTo((11_000 / 1_050_000) * 100);
+    expect(result.maxTokens).toBe(272_000);
+    expect(result.totalPercentage).toBeCloseTo((11_000 / 272_000) * 100);
   });
 
   test("uses Claude Sonnet 4.6's native 1M context even when the beta toggle is off", () => {