fix(codex): recover unsupported live model switches

src/cliproxy/codex-plan-compatibility.ts

@@ -1,4 +1,5 @@
 import { getDefaultAccount } from './account-manager';
+import { getProviderCatalog } from './model-catalog';
 import { fetchCodexQuota } from './quota-fetcher-codex';
 import { getCachedQuota, setCachedQuota } from './quota-response-cache';
 import type { CodexQuotaResult } from './quota-types';
@@ -12,6 +13,9 @@ const FREE_SAFE_FAST_MODEL = 'gpt-5-codex-mini';
 const CODEX_EFFORT_SUFFIX_REGEX = /-(xhigh|high|medium)$/i;
 const CODEX_PAREN_SUFFIX_REGEX = /\((xhigh|high|medium)\)$/i;
 const EXTENDED_CONTEXT_SUFFIX_REGEX = /\[1m\]$/i;
+const KNOWN_CODEX_MODELS = new Set(
+  (getProviderCatalog('codex')?.models ?? []).map((model) => model.id.toLowerCase())
+);
 
 const FREE_PLAN_FALLBACKS = new Map<string, string>([
   ['gpt-5.3-codex', FREE_SAFE_DEFAULT_MODEL],
@@ -19,7 +23,29 @@ const FREE_PLAN_FALLBACKS = new Map<string, string>([
   ['gpt-5.4', FREE_SAFE_DEFAULT_MODEL],
 ]);
 
-function normalizeCodexModelId(model: string): string {
+export interface CodexRuntimeFallbackModelMap {
+  defaultModel?: string;
+  opusModel?: string;
+  sonnetModel?: string;
+  haikuModel?: string;
+}
+
+export interface CodexUnsupportedModelError {
+  message: string | null;
+  code: 'model_not_supported';
+  param: string | null;
+  type: string | null;
+}
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return typeof value === 'object' && value !== null && !Array.isArray(value);
+}
+
+function isKnownCodexModel(model: string): boolean {
+  return KNOWN_CODEX_MODELS.has(model);
+}
+
+export function normalizeCodexModelId(model: string): string {
   return model
     .trim()
     .replace(EXTENDED_CONTEXT_SUFFIX_REGEX, '')
@@ -37,6 +63,74 @@ export function getFreePlanFallbackCodexModel(model: string): string | null {
   return FREE_PLAN_FALLBACKS.get(normalizeCodexModelId(model)) ?? null;
 }
 
+export function parseCodexUnsupportedModelError(
+  statusCode: number | undefined,
+  responseBody: string
+): CodexUnsupportedModelError | null {
+  if (statusCode !== 400 || !responseBody.trim()) {
+    return null;
+  }
+
+  try {
+    const parsed = JSON.parse(responseBody);
+    if (
+      !isRecord(parsed) ||
+      !isRecord(parsed.error) ||
+      parsed.error.code !== 'model_not_supported'
+    ) {
+      return null;
+    }
+
+    return {
+      message: typeof parsed.error.message === 'string' ? parsed.error.message : null,
+      code: 'model_not_supported',
+      param: typeof parsed.error.param === 'string' ? parsed.error.param : null,
+      type: typeof parsed.error.type === 'string' ? parsed.error.type : null,
+    };
+  } catch {
+    return null;
+  }
+}
+
+export function resolveRuntimeCodexFallbackModel(options: {
+  requestedModel: string;
+  modelMap: CodexRuntimeFallbackModelMap;
+  excludeModels?: string[];
+}): string | null {
+  const requestedModel = normalizeCodexModelId(options.requestedModel);
+  if (!requestedModel) {
+    return null;
+  }
+
+  const excludedModels = new Set(
+    (options.excludeModels ?? []).map((model) => normalizeCodexModelId(model)).filter(Boolean)
+  );
+  const candidates = [
+    options.modelMap.defaultModel,
+    getFreePlanFallbackCodexModel(requestedModel),
+    options.modelMap.opusModel,
+    options.modelMap.sonnetModel,
+    options.modelMap.haikuModel,
+    getDefaultCodexModel(),
+  ];
+
+  for (const candidate of candidates) {
+    if (!candidate) continue;
+    const normalizedCandidate = normalizeCodexModelId(candidate);
+    if (
+      !normalizedCandidate ||
+      normalizedCandidate === requestedModel ||
+      excludedModels.has(normalizedCandidate) ||
+      !isKnownCodexModel(normalizedCandidate)
+    ) {
+      continue;
+    }
+    return normalizedCandidate;
+  }
+
+  return null;
+}
+
 export async function reconcileCodexModelForActivePlan(options: {
   settingsPath: string;
   currentModel: string | undefined;

src/cliproxy/codex-reasoning-proxy.ts

@@ -1,6 +1,11 @@
 import * as http from 'http';
 import * as https from 'https';
 import { URL } from 'url';
+import {
+  normalizeCodexModelId,
+  parseCodexUnsupportedModelError,
+  resolveRuntimeCodexFallbackModel,
+} from './codex-plan-compatibility';
 import { getModelMaxLevel } from './model-catalog';
 
 export type CodexReasoningEffort = 'medium' | 'high' | 'xhigh';
@@ -29,6 +34,14 @@ export interface CodexReasoningProxyConfig {
   disableEffort?: boolean;
 }
 
+interface ForwardJsonContext {
+  requestPath: string;
+  requestedModel: string | null;
+  attemptedUpstreamModel: string | null;
+  effort: CodexReasoningEffort | null;
+  retryCount: number;
+}
+
 const EXTENDED_CONTEXT_SUFFIX_REGEX = /\[1m\]$/i;
 
 function stripExtendedContextSuffix(model: string): string {
@@ -170,6 +183,7 @@ export class CodexReasoningProxy {
   > &
     Pick<CodexReasoningProxyConfig, 'modelMap' | 'stripPathPrefix'>;
   private readonly modelEffort: Map<string, CodexReasoningEffort>;
+  private readonly sessionFallbackByModel = new Map<string, string>();
   private readonly recent: Array<{
     at: string;
     model: string | null;
@@ -193,6 +207,41 @@ export class CodexReasoningProxy {
     this.modelEffort = buildCodexModelEffortMap(this.config.modelMap, this.config.defaultEffort);
   }
 
+  private getRememberedFallback(model: string | null): string | null {
+    if (!model) return null;
+    return this.sessionFallbackByModel.get(normalizeCodexModelId(model)) ?? null;
+  }
+
+  private rememberFallback(requestedModel: string, fallbackModel: string): void {
+    const normalizedRequestedModel = normalizeCodexModelId(requestedModel);
+    const normalizedFallbackModel = normalizeCodexModelId(fallbackModel);
+    if (!normalizedRequestedModel || !normalizedFallbackModel) return;
+    this.sessionFallbackByModel.set(normalizedRequestedModel, normalizedFallbackModel);
+  }
+
+  private buildForwardBody(
+    body: unknown,
+    upstreamModel: string | null,
+    effort: CodexReasoningEffort | null
+  ): unknown {
+    const withUpstreamModel =
+      upstreamModel && isRecord(body) ? { ...body, model: upstreamModel } : body;
+    if (this.config.disableEffort || !effort) {
+      return withUpstreamModel;
+    }
+    return injectReasoningEffortIntoBody(withUpstreamModel, effort);
+  }
+
+  private sendBufferedResponse(
+    clientRes: http.ServerResponse,
+    statusCode: number,
+    headers: http.IncomingHttpHeaders,
+    responseBody: string
+  ): void {
+    clientRes.writeHead(statusCode, headers);
+    clientRes.end(responseBody);
+  }
+
   /**
    * Treat trailing "-high/-medium/-xhigh" as an effort alias only for known codex models.
    * Prevents stripping legitimate upstream model IDs that happen to end with those tokens.
@@ -365,41 +414,48 @@ export class CodexReasoningProxy {
         ? stripExtendedContextSuffix(originalModel)
         : null;
 
-      // When effort is disabled (thinking mode: off), strip model suffix but don't inject reasoning
-      if (this.config.disableEffort) {
-        const suffixParsed = this.parseEffortAlias(normalizedRequestModel);
-        const upstreamModel = suffixParsed?.upstreamModel ?? normalizedRequestModel;
-        const forwarded =
-          upstreamModel && isRecord(parsed) ? { ...parsed, model: upstreamModel } : parsed;
-
-        this.log(`[disabled] model=${originalModel ?? 'null'} -> passthrough (no reasoning)`);
-        await this.forwardJson(req, res, fullUpstreamUrl, forwarded);
-        return;
-      }
-
       // Support "model aliases" like `gpt-5.2-codex-xhigh` by translating to:
       // - upstream model: `gpt-5.2-codex`
       // - reasoning.effort: `xhigh`
       //
       // This allows tier→effort mapping without inventing upstream model IDs.
       const suffixParsed = this.parseEffortAlias(normalizedRequestModel);
-      const upstreamModel = suffixParsed?.upstreamModel ?? normalizedRequestModel;
-      const effort =
+      const requestedUpstreamModel = suffixParsed?.upstreamModel ?? normalizedRequestModel;
+      const rememberedFallback = this.getRememberedFallback(requestedUpstreamModel);
+      const upstreamModel = rememberedFallback ?? requestedUpstreamModel;
+      const requestedEffort =
         suffixParsed?.effort ??
         getEffortForModel(normalizedRequestModel, this.modelEffort, this.config.defaultEffort);
+      const effort =
+        !this.config.disableEffort && upstreamModel
+          ? capEffortAtModelMax(upstreamModel, requestedEffort)
+          : !this.config.disableEffort
+            ? requestedEffort
+            : null;
+      const rewritten = this.buildForwardBody(parsed, upstreamModel, effort);
 
-      const withUpstreamModel =
-        upstreamModel && isRecord(parsed) ? { ...parsed, model: upstreamModel } : parsed;
-      const rewritten = injectReasoningEffortIntoBody(withUpstreamModel, effort);
+      if (effort) {
+        this.record(originalModel, upstreamModel, effort, requestPath);
+        this.trace(
+          `[${new Date().toISOString()}] model=${originalModel ?? 'null'} upstreamModel=${
+            upstreamModel ?? 'null'
+          } effort=${effort} path=${requestPath}`
+        );
+      } else {
+        this.log(`[disabled] model=${originalModel ?? 'null'} -> passthrough (no reasoning)`);
+      }
 
-      this.record(originalModel, upstreamModel, effort, requestPath);
-      this.trace(
-        `[${new Date().toISOString()}] model=${originalModel ?? 'null'} upstreamModel=${
-          upstreamModel ?? 'null'
-        } effort=${effort} path=${requestPath}`
-      );
+      if (rememberedFallback && rememberedFallback !== requestedUpstreamModel) {
+        this.log(`Using remembered fallback ${requestedUpstreamModel} -> ${rememberedFallback}`);
+      }
 
-      await this.forwardJson(req, res, fullUpstreamUrl, rewritten);
+      await this.forwardJson(req, res, fullUpstreamUrl, rewritten, {
+        requestPath,
+        requestedModel: requestedUpstreamModel,
+        attemptedUpstreamModel: upstreamModel,
+        effort,
+        retryCount: 0,
+      });
     } catch (error) {
       const err = error as Error;
       if (!res.headersSent) {
@@ -487,8 +543,9 @@ export class CodexReasoningProxy {
     originalReq: http.IncomingMessage,
     clientRes: http.ServerResponse,
     upstreamUrl: URL,
-    body: unknown
-  ): Promise<void> {
+    body: unknown,
+    context: ForwardJsonContext
+  ): Promise<number> {
     return new Promise((resolve, reject) => {
       const bodyString = JSON.stringify(body);
       const requestFn = this.getRequestFn(upstreamUrl);
@@ -503,9 +560,73 @@ export class CodexReasoningProxy {
           headers: this.buildForwardHeaders(originalReq.headers, bodyString),
         },
         (upstreamRes) => {
-          clientRes.writeHead(upstreamRes.statusCode || 200, upstreamRes.headers);
-          upstreamRes.pipe(clientRes);
-          upstreamRes.on('end', () => resolve());
+          const statusCode = upstreamRes.statusCode || 200;
+          if (statusCode >= 200 && statusCode < 300) {
+            clientRes.writeHead(statusCode, upstreamRes.headers);
+            upstreamRes.pipe(clientRes);
+            upstreamRes.on('end', () => resolve(statusCode));
+            upstreamRes.on('error', reject);
+            return;
+          }
+
+          const chunks: Buffer[] = [];
+          upstreamRes.on('data', (chunk: Buffer) => chunks.push(chunk));
+          upstreamRes.on('end', async () => {
+            try {
+              const responseBody = Buffer.concat(chunks).toString('utf8');
+              const unsupportedError =
+                context.retryCount === 0
+                  ? parseCodexUnsupportedModelError(statusCode, responseBody)
+                  : null;
+              const fallbackModel =
+                unsupportedError && context.requestedModel
+                  ? resolveRuntimeCodexFallbackModel({
+                      requestedModel: context.requestedModel,
+                      modelMap: this.config.modelMap,
+                      excludeModels: context.attemptedUpstreamModel
+                        ? [context.attemptedUpstreamModel]
+                        : undefined,
+                    })
+                  : null;
+
+              if (unsupportedError && fallbackModel && context.requestedModel) {
+                const retryEffort =
+                  !this.config.disableEffort && context.effort
+                    ? capEffortAtModelMax(fallbackModel, context.effort)
+                    : null;
+                const retryBody = this.buildForwardBody(body, fallbackModel, retryEffort);
+
+                this.log(
+                  `Upstream rejected model "${context.attemptedUpstreamModel}". Retrying ${context.requestPath} with "${fallbackModel}".`
+                );
+
+                const retryStatusCode = await this.forwardJson(
+                  originalReq,
+                  clientRes,
+                  upstreamUrl,
+                  retryBody,
+                  {
+                    ...context,
+                    attemptedUpstreamModel: fallbackModel,
+                    effort: retryEffort,
+                    retryCount: context.retryCount + 1,
+                  }
+                );
+
+                if (retryStatusCode >= 200 && retryStatusCode < 300) {
+                  this.rememberFallback(context.requestedModel, fallbackModel);
+                }
+
+                resolve(retryStatusCode);
+                return;
+              }
+
+              this.sendBufferedResponse(clientRes, statusCode, upstreamRes.headers, responseBody);
+              resolve(statusCode);
+            } catch (error) {
+              reject(error);
+            }
+          });
           upstreamRes.on('error', reject);
         }
       );

tests/unit/cliproxy/codex-plan-compatibility.test.ts

@@ -3,6 +3,8 @@ import { getProviderCatalog, getModelMaxLevel } from '../../../src/cliproxy/mode
 import {
   getDefaultCodexModel,
   getFreePlanFallbackCodexModel,
+  parseCodexUnsupportedModelError,
+  resolveRuntimeCodexFallbackModel,
 } from '../../../src/cliproxy/codex-plan-compatibility';
 
 describe('codex plan compatibility', () => {
@@ -25,6 +27,50 @@ describe('codex plan compatibility', () => {
     expect(getFreePlanFallbackCodexModel('gpt-5.1-codex-mini')).toBeNull();
   });
 
+  it('detects upstream Codex model_not_supported responses', () => {
+    expect(
+      parseCodexUnsupportedModelError(
+        400,
+        JSON.stringify({
+          error: {
+            message: 'The requested model is not supported.',
+            code: 'model_not_supported',
+            param: 'model',
+            type: 'invalid_request_error',
+          },
+        })
+      )
+    ).toEqual({
+      message: 'The requested model is not supported.',
+      code: 'model_not_supported',
+      param: 'model',
+      type: 'invalid_request_error',
+    });
+    expect(
+      parseCodexUnsupportedModelError(500, '{"error":{"code":"model_not_supported"}}')
+    ).toBeNull();
+  });
+
+  it('resolves runtime fallbacks without retrying the rejected model again', () => {
+    expect(
+      resolveRuntimeCodexFallbackModel({
+        requestedModel: 'gpt-5.4',
+        modelMap: { defaultModel: 'gpt-5-codex' },
+      })
+    ).toBe('gpt-5-codex');
+
+    expect(
+      resolveRuntimeCodexFallbackModel({
+        requestedModel: 'gpt-5.4',
+        modelMap: {
+          defaultModel: 'gpt-5.4',
+          haikuModel: 'gpt-5-codex-mini',
+        },
+        excludeModels: ['gpt-5-codex'],
+      })
+    ).toBe('gpt-5-codex-mini');
+  });
+
   it('tracks Codex thinking caps for current safe defaults and paid models', () => {
     expect(getModelMaxLevel('codex', 'gpt-5-codex')).toBe('high');
     expect(getModelMaxLevel('codex', 'gpt-5-codex-mini')).toBe('high');

tests/unit/cliproxy/codex-reasoning-proxy-extended-context.test.ts

@@ -230,6 +230,90 @@ describe('CodexReasoningProxy extended-context compatibility', () => {
     expect(capturedBody?.model).toBe('enterprise-internal-high');
   });
 
+  it('retries unsupported live-session models once and remembers the fallback', async () => {
+    const capturedModels: string[] = [];
+    const capturedEfforts: Array<string | undefined> = [];
+
+    const upstream = http.createServer((req, res) => {
+      let rawBody = '';
+      req.setEncoding('utf8');
+      req.on('data', (chunk) => {
+        rawBody += chunk;
+      });
+      req.on('end', () => {
+        const requestBody = rawBody ? (JSON.parse(rawBody) as JsonRecord) : {};
+        const reasoning = requestBody.reasoning as JsonRecord | undefined;
+        const model = String(requestBody.model ?? '');
+        const effort = typeof reasoning?.effort === 'string' ? reasoning.effort : undefined;
+
+        capturedModels.push(model);
+        capturedEfforts.push(effort);
+
+        if (model === 'gpt-5.4') {
+          res.writeHead(400, { 'Content-Type': 'application/json' });
+          res.end(
+            JSON.stringify({
+              error: {
+                message: 'The requested model is not supported.',
+                code: 'model_not_supported',
+                param: 'model',
+                type: 'invalid_request_error',
+              },
+            })
+          );
+          return;
+        }
+
+        res.writeHead(200, { 'Content-Type': 'application/json' });
+        res.end(
+          JSON.stringify({
+            ok: true,
+            model,
+            effort: effort ?? null,
+          })
+        );
+      });
+    });
+    cleanupServers.push(upstream);
+
+    const upstreamPort = await listenOnRandomPort(upstream);
+    const proxy = new CodexReasoningProxy({
+      upstreamBaseUrl: `http://127.0.0.1:${upstreamPort}`,
+      modelMap: {
+        defaultModel: 'gpt-5.4',
+        haikuModel: 'gpt-5-codex-mini',
+      },
+      defaultEffort: 'medium',
+    });
+
+    const proxyPort = await proxy.start();
+    const firstResponse = await postJson(
+      `http://127.0.0.1:${proxyPort}/api/provider/codex/v1/messages`,
+      {
+        model: 'gpt-5.4-xhigh',
+        messages: [],
+      }
+    );
+    const secondResponse = await postJson(
+      `http://127.0.0.1:${proxyPort}/api/provider/codex/v1/messages`,
+      {
+        model: 'gpt-5.4-xhigh',
+        messages: [],
+      }
+    );
+
+    proxy.stop();
+
+    expect(firstResponse.statusCode).toBe(200);
+    expect(secondResponse.statusCode).toBe(200);
+    expect(firstResponse.body.model).toBe('gpt-5-codex');
+    expect(firstResponse.body.effort).toBe('high');
+    expect(secondResponse.body.model).toBe('gpt-5-codex');
+    expect(secondResponse.body.effort).toBe('high');
+    expect(capturedModels).toEqual(['gpt-5.4', 'gpt-5-codex', 'gpt-5-codex']);
+    expect(capturedEfforts).toEqual(['xhigh', 'high', 'high']);
+  });
+
   it('keeps reasoning enabled when CCS_THINKING=high overrides config off', async () => {
     let capturedBody: JsonRecord | null = null;