feat: Phase 1 & 2 implementation plan — T1-T10, T12

T1 (openai-to-claude.ts): response_format injection for json_schema/json_object
T2 (base.ts): intra-URL retry for 429 errors (2x, 2s delay)
T3 (gemini-cli.ts): CLI fingerprint headers (User-Agent, X-Goog-Api-Client)
T5 (modelFamilyFallback.ts + chatCore.ts): intra-family model fallback on 400/404
T9 (pricing.ts): deepseek-3.1, deepseek-3.2, qwen3-coder-next pricing
T10 (scoring.ts + modePacks.ts): tierPriority as 7th scoring factor (Ultra>Pro>Free)
T12 (cliRuntime.ts): isSafePath() guard for CLI_*_BIN env var paths

open-sse/executors/base.ts

@@ -158,6 +158,9 @@ export class BaseExecutor {
     return status === HTTP_STATUS.RATE_LIMITED && urlIndex + 1 < this.getFallbackCount();
   }
 
+  // Intra-URL retry config: retry same URL before falling back to next node
+  static readonly RETRY_CONFIG = { maxAttempts: 2, delayMs: 2000 };
+
   // Override in subclass for provider-specific refresh
   async refreshCredentials(credentials: ProviderCredentials, log: ExecutorLog | null) {
     void credentials;
@@ -179,6 +182,8 @@ export class BaseExecutor {
     const fallbackCount = this.getFallbackCount();
     let lastError: unknown = null;
     let lastStatus = 0;
+    // Track per-URL intra-retry attempts to avoid infinite loops
+    const retryAttemptsByUrl: Record<number, number> = {};
 
     for (let urlIndex = 0; urlIndex < fallbackCount; urlIndex++) {
       const url = this.buildUrl(model, stream, urlIndex, credentials);
@@ -236,6 +241,22 @@ export class BaseExecutor {
 
         const response = await fetch(url, fetchOptions);
 
+        // Intra-URL retry: if 429 and we haven't exhausted per-URL retries, wait and retry the same URL
+        if (
+          response.status === HTTP_STATUS.RATE_LIMITED &&
+          (retryAttemptsByUrl[urlIndex] ?? 0) < BaseExecutor.RETRY_CONFIG.maxAttempts
+        ) {
+          retryAttemptsByUrl[urlIndex] = (retryAttemptsByUrl[urlIndex] ?? 0) + 1;
+          const attempt = retryAttemptsByUrl[urlIndex];
+          log?.debug?.(
+            "RETRY",
+            `429 intra-retry ${attempt}/${BaseExecutor.RETRY_CONFIG.maxAttempts} on ${url} — waiting ${BaseExecutor.RETRY_CONFIG.delayMs}ms`
+          );
+          await new Promise((resolve) => setTimeout(resolve, BaseExecutor.RETRY_CONFIG.delayMs));
+          urlIndex--; // re-run this urlIndex on the next loop iteration
+          continue;
+        }
+
         if (this.shouldRetry(response.status, urlIndex)) {
           log?.debug?.("RETRY", `${response.status} on ${url}, trying fallback ${urlIndex + 1}`);
           lastStatus = response.status;

open-sse/executors/gemini-cli.ts

@@ -2,6 +2,8 @@ import { BaseExecutor } from "./base.ts";
 import { PROVIDERS, OAUTH_ENDPOINTS } from "../config/constants.ts";
 
 export class GeminiCLIExecutor extends BaseExecutor {
+  private _currentModel: string = "";
+
   constructor() {
     super("gemini-cli", PROVIDERS["gemini-cli"]);
   }
@@ -15,11 +17,17 @@ export class GeminiCLIExecutor extends BaseExecutor {
     return {
       "Content-Type": "application/json",
       Authorization: `Bearer ${credentials.accessToken}`,
+      // Fingerprint headers matching native GeminiCLI client (prevents upstream rejection)
+      "User-Agent": `GeminiCLI/0.31.0/${this._currentModel || "unknown"} (linux; x64)`,
+      "X-Goog-Api-Client": "google-genai-sdk/1.41.0 gl-node/v22.19.0",
       ...(stream && { Accept: "text/event-stream" }),
     };
   }
 
   transformRequest(model, body, stream, credentials) {
+    // Capture model so buildHeaders (called after transformRequest) can include it in User-Agent
+    this._currentModel = model || "";
+
     const allowBodyProjectOverride = process.env.OMNIROUTE_ALLOW_BODY_PROJECT_OVERRIDE === "1";
 
     // Default: prefer OAuth-stored projectId. Incoming body.project can be stale

open-sse/handlers/chatCore.ts

@@ -40,6 +40,7 @@ import {
 } from "@/lib/semanticCache";
 import { getIdempotencyKey, checkIdempotency, saveIdempotency } from "@/lib/idempotencyLayer";
 import { createProgressTransform, wantsProgress } from "../utils/progressTracker.ts";
+import { isModelUnavailableError, getNextFamilyFallback } from "../services/modelFamilyFallback.ts";
 
 /**
  * Core chat handler - shared between SSE and Worker
@@ -248,6 +249,10 @@ export async function handleChatCore({
   // Track pending request
   trackPendingRequest(model, provider, connectionId, true);
 
+  // T5: track which models we've tried for intra-family fallback
+  const triedModels = new Set<string>([model]);
+  let currentModel = model;
+
   // Log start
   appendRequestLog({ model, provider, connectionId, status: "PENDING" }).catch(() => {});
 
@@ -421,7 +426,53 @@ export async function handleChatCore({
     // Update rate limiter from error response headers
     updateFromHeaders(provider, connectionId, providerResponse.headers, statusCode, model);
 
-    return createErrorResult(statusCode, errMsg, retryAfterMs);
+    // ── T5: Intra-family model fallback ──────────────────────────────────────
+    // Before returning a model-unavailable error upstream, try sibling models
+    // from the same family. This keeps the request alive on the same account
+    // instead of failing the entire combo.
+    if (isModelUnavailableError(statusCode, message)) {
+      const nextModel = getNextFamilyFallback(currentModel, triedModels);
+      if (nextModel) {
+        triedModels.add(nextModel);
+        currentModel = nextModel;
+        translatedBody.model = nextModel;
+        log?.info?.("MODEL_FALLBACK", `${model} unavailable (${statusCode}) → trying ${nextModel}`);
+        // Re-execute with the fallback model
+        try {
+          const fallbackResult = await withRateLimit(provider, connectionId, nextModel, () =>
+            executor.execute({
+              model: nextModel,
+              body: translatedBody,
+              stream,
+              credentials,
+              signal: streamController.signal,
+              log,
+              extendedContext,
+            })
+          );
+          if (fallbackResult.response.ok) {
+            providerResponse = fallbackResult.response;
+            providerUrl = fallbackResult.url;
+            providerHeaders = fallbackResult.headers;
+            finalBody = fallbackResult.transformedBody;
+            // Continue processing with the fallback response — skip error return
+            log?.info?.("MODEL_FALLBACK", `Serving ${nextModel} as fallback for ${model}`);
+            // Jump to streaming/non-streaming handling below
+            // We fall through by NOT returning here
+          } else {
+            // Fallback also failed — return original error
+            return createErrorResult(statusCode, errMsg, retryAfterMs);
+          }
+        } catch {
+          return createErrorResult(statusCode, errMsg, retryAfterMs);
+        }
+      } else {
+        return createErrorResult(statusCode, errMsg, retryAfterMs);
+      }
+    } else {
+      return createErrorResult(statusCode, errMsg, retryAfterMs);
+    }
+    // ── End T5 ───────────────────────────────────────────────────────────────
   }
 
   // Non-streaming response

open-sse/services/autoCombo/index.ts

@@ -3,6 +3,7 @@
  */
 export {
   calculateScore,
+  calculateTierScore,
   scorePool,
   validateWeights,
   DEFAULT_WEIGHTS,

open-sse/services/autoCombo/modePacks.ts

@@ -11,37 +11,45 @@
 import type { ScoringWeights } from "./scoring";
 
 export const MODE_PACKS: Record<string, ScoringWeights> = {
+  // Prioritize latency → health. tierPriority replaces 0.05 from stability.
   "ship-fast": {
     quota: 0.15,
     health: 0.3,
     costInv: 0.05,
     latencyInv: 0.35,
     taskFit: 0.1,
-    stability: 0.05,
+    stability: 0.0,
+    tierPriority: 0.05,
   },
+  // Prioritize cost. tierPriority replaces 0.05 from stability.
   "cost-saver": {
     quota: 0.15,
     health: 0.2,
     costInv: 0.4,
     latencyInv: 0.05,
     taskFit: 0.1,
-    stability: 0.1,
+    stability: 0.05,
+    tierPriority: 0.05,
   },
+  // Prioritize task fitness. tierPriority replaces 0.05 from latencyInv.
   "quality-first": {
     quota: 0.1,
     health: 0.2,
     costInv: 0.05,
-    latencyInv: 0.1,
+    latencyInv: 0.05,
     taskFit: 0.4,
     stability: 0.15,
+    tierPriority: 0.05,
   },
+  // Prioritize quota availability. tierPriority replaces 0.05 from taskFit.
   "offline-friendly": {
     quota: 0.4,
     health: 0.3,
     costInv: 0.1,
     latencyInv: 0.05,
-    taskFit: 0.05,
+    taskFit: 0.0,
     stability: 0.1,
+    tierPriority: 0.05,
   },
 };
 

open-sse/services/autoCombo/scoring.ts

@@ -17,6 +17,7 @@ export interface ScoringFactors {
   latencyInv: number;
   taskFit: number;
   stability: number;
+  tierPriority: number; // T10: Ultra > Pro > Free account tier boost
 }
 
 export interface ScoringWeights {
@@ -26,15 +27,18 @@ export interface ScoringWeights {
   latencyInv: number;
   taskFit: number;
   stability: number;
+  tierPriority: number; // T10
 }
 
+// T10: Rebalanced — stability 0.10→0.05, tierPriority 0.05 added. Sum = 1.0.
 export const DEFAULT_WEIGHTS: ScoringWeights = {
   quota: 0.2,
   health: 0.25,
   costInv: 0.2,
   latencyInv: 0.15,
   taskFit: 0.1,
-  stability: 0.1,
+  stability: 0.05,
+  tierPriority: 0.05,
 };
 
 export interface ProviderCandidate {
@@ -47,6 +51,10 @@ export interface ProviderCandidate {
   p95LatencyMs: number;
   latencyStdDev: number;
   errorRate: number;
+  /** T10: Optional account tier for priority boosting (Ultra > Pro > Free) */
+  accountTier?: "ultra" | "pro" | "standard" | "free";
+  /** T10: Optional quota reset interval in seconds (shorter = higher priority when same quota) */
+  quotaResetIntervalSecs?: number;
 }
 
 export interface ScoredProvider {
@@ -70,6 +78,38 @@ export function calculateScore(factors: ScoringFactors, weights: ScoringWeights)
   );
 }
 
+/**
+ * T10: Convert account tier string to a normalized score [0..1].
+ * Ultra = 1.0 (most quota, fastest reset)
+ * Pro   = 0.67
+ * Standard = 0.33
+ * Free  = 0.0
+ * Accounts with faster reset cycles (shorter quotaResetIntervalSecs) also get
+ * a small adjustment: monthly accounts are penalized vs. daily accounts.
+ */
+export function calculateTierScore(
+  tier: string | undefined,
+  quotaResetIntervalSecs: number | undefined
+): number {
+  const BASE_TIER_SCORES: Record<string, number> = {
+    ultra: 1.0,
+    pro: 0.67,
+    standard: 0.33,
+    free: 0.0,
+  };
+  const baseScore = BASE_TIER_SCORES[tier?.toLowerCase() ?? ""] ?? 0.33; // unknown defaults to standard
+
+  // Bonus for faster reset intervals (daily quota > weekly > monthly)
+  // maxInterval ~ 30 days (2_592_000s). Normalize: [0..1] where 0=monthly, 1=per-minute
+  const resetBonus =
+    quotaResetIntervalSecs != null && quotaResetIntervalSecs > 0
+      ? Math.max(0, 1 - quotaResetIntervalSecs / 2_592_000)
+      : 0;
+
+  // Blend: 80% tier level, 20% reset frequency
+  return Math.min(1, baseScore * 0.8 + resetBonus * 0.2);
+}
+
 /**
  * Calculate individual factors for a provider within its pool.
  */
@@ -96,6 +136,7 @@ export function calculateFactors(
     latencyInv: 1 - candidate.p95LatencyMs / maxLatency,
     taskFit: getTaskFitness(candidate.model, taskType),
     stability: 1 - candidate.latencyStdDev / maxStdDev,
+    tierPriority: calculateTierScore(candidate.accountTier, candidate.quotaResetIntervalSecs),
   };
 }
 

open-sse/services/modelFamilyFallback.ts

@@ -0,0 +1,157 @@
+/**
+ * Model Family Fallback — Phase 2 Feature (T5)
+ *
+ * Implements two-phase model resolution:
+ *   Phase 1 (static, pre-request): already done by model.ts alias resolution.
+ *   Phase 2 (dynamic, post-error): when a provider returns a model-not-available
+ *   error (400 with specific message or 404), we try sibling models within the
+ *   same "family" before giving up.
+ *
+ * Inspired by Antigravity Manager's account-aware dynamic model remapping
+ * (commit 6cea566, Mar 8 2026).
+ */
+
+// ── Model Family Definitions ─────────────────────────────────────────────────
+
+/**
+ * Ordered candidate lists per model family.
+ * First entry is the most preferred; fallback proceeds in order.
+ */
+const MODEL_FAMILIES: Record<string, string[]> = {
+  // Gemini 3 / 3.1 Pro family — ordered by preference
+  "gemini-3-pro": [
+    "gemini-3.1-pro-preview",
+    "gemini-3-pro-preview",
+    "gemini-3.1-pro-high",
+    "gemini-3-pro-high",
+    "gemini-3.1-pro-low",
+    "gemini-3-pro-low",
+  ],
+  "gemini-3.1-pro": [
+    "gemini-3.1-pro-preview",
+    "gemini-3-pro-preview",
+    "gemini-3.1-pro-high",
+    "gemini-3-pro-high",
+    "gemini-3.1-pro-low",
+    "gemini-3-pro-low",
+  ],
+  "gemini-3-pro-preview": [
+    "gemini-3.1-pro-preview",
+    "gemini-3-pro-high",
+    "gemini-3.1-pro-high",
+    "gemini-3-pro-low",
+    "gemini-3.1-pro-low",
+  ],
+  "gemini-3.1-pro-preview": [
+    "gemini-3-pro-preview",
+    "gemini-3.1-pro-high",
+    "gemini-3-pro-high",
+    "gemini-3.1-pro-low",
+    "gemini-3-pro-low",
+  ],
+  "gemini-3-pro-high": [
+    "gemini-3.1-pro-high",
+    "gemini-3-pro-preview",
+    "gemini-3.1-pro-preview",
+    "gemini-3-pro-low",
+    "gemini-3.1-pro-low",
+  ],
+  "gemini-3.1-pro-high": [
+    "gemini-3-pro-high",
+    "gemini-3.1-pro-preview",
+    "gemini-3-pro-preview",
+    "gemini-3.1-pro-low",
+    "gemini-3-pro-low",
+  ],
+
+  // Gemini 2.5 Pro family
+  "gemini-2.5-pro": ["gemini-2.5-pro-preview-06-05", "gemini-2.5-pro-exp-03-25"],
+  "gemini-2.5-pro-preview-06-05": ["gemini-2.5-pro", "gemini-2.5-pro-exp-03-25"],
+
+  // Claude Opus family
+  "claude-opus-4-6": ["claude-opus-4-6-thinking", "claude-opus-4-5-20251101", "claude-sonnet-4-6"],
+  "claude-opus-4-6-thinking": ["claude-opus-4-6", "claude-opus-4-5-20251101"],
+
+  // Claude Sonnet family
+  "claude-sonnet-4-6": ["claude-sonnet-4-5-20250929", "claude-sonnet-4-20250514"],
+  "claude-sonnet-4-5-20250929": ["claude-sonnet-4-6", "claude-sonnet-4-20250514"],
+
+  // GPT-5 family
+  "gpt-5": ["gpt-5-mini", "gpt-4o"],
+  "gpt-5.1": ["gpt-5.1-mini", "gpt-5", "gpt-4o"],
+};
+
+// ── Error Detection ──────────────────────────────────────────────────────────
+
+/**
+ * Error message fragments that indicate the requested model is unavailable
+ * for the current account/provider, as opposed to a transient error.
+ */
+const MODEL_UNAVAILABLE_FRAGMENTS = [
+  "model not found",
+  "model_not_found",
+  "model not available",
+  "model is not available",
+  "no such model",
+  "unsupported model",
+  "unknown model",
+  "this model does not exist",
+  "invalid model",
+  "model not supported",
+  "does not support",
+  "not enabled for",
+  "access to model",
+];
+
+/**
+ * Returns true if the HTTP status + error message indicates the model
+ * itself is not available, not a transient server error.
+ */
+export function isModelUnavailableError(status: number, errorMessage: string): boolean {
+  if (status === 404) return true;
+  if (status !== 400 && status !== 403) return false;
+
+  const msg = errorMessage.toLowerCase();
+  return MODEL_UNAVAILABLE_FRAGMENTS.some((fragment) => msg.includes(fragment));
+}
+
+// ── Fallback Resolution ──────────────────────────────────────────────────────
+
+/**
+ * Get the next fallback model from the same family.
+ *
+ * @param currentModel  The model that just failed
+ * @param triedModels   Set of model IDs already tried (to avoid cycles)
+ * @returns             Next model to try, or null if family exhausted
+ */
+export function getNextFamilyFallback(
+  currentModel: string,
+  triedModels: Set<string>
+): string | null {
+  const family = MODEL_FAMILIES[currentModel];
+  if (!family) return null;
+
+  for (const candidate of family) {
+    if (!triedModels.has(candidate)) {
+      return candidate;
+    }
+  }
+
+  return null; // family exhausted
+}
+
+/**
+ * Check if a model belongs to any registered family.
+ */
+export function isInModelFamily(model: string): boolean {
+  return model in MODEL_FAMILIES;
+}
+
+/**
+ * Get all members of a model's family (including itself).
+ */
+export function getModelFamily(model: string): string[] {
+  const family = MODEL_FAMILIES[model];
+  if (!family) return [model];
+  return [model, ...family];
+}

open-sse/translator/request/openai-to-claude.ts

@@ -24,6 +24,7 @@ type ClaudeTool = {
   description: string;
   input_schema: Record<string, unknown>;
   cache_control?: { type: string; ttl?: string };
+  defer_loading?: boolean;
 };
 
 // Convert OpenAI request to Claude format
@@ -193,6 +194,23 @@ export function openaiToClaudeRequest(model, body, stream) {
     result.tool_choice = convertOpenAIToolChoice(body.tool_choice);
   }
 
+  // response_format: inject JSON structured output instruction into system prompt.
+  // Claude doesn't natively support response_format, so we insert a system-level instruction.
+  // NOTE: systemParts are consumed later (after this block) — they're accumulated here.
+  if (body.response_format) {
+    const fmt = body.response_format;
+    if (fmt.type === "json_schema" && fmt.json_schema?.schema) {
+      const schemaJson = JSON.stringify(fmt.json_schema.schema, null, 2);
+      systemParts.push(
+        `You must respond with valid JSON that strictly follows this JSON schema:\n\`\`\`json\n${schemaJson}\n\`\`\`\nRespond ONLY with the JSON object, no other text.`
+      );
+    } else if (fmt.type === "json_object") {
+      systemParts.push(
+        "You must respond with valid JSON. Respond ONLY with a JSON object, no other text."
+      );
+    }
+  }
+
   // Thinking configuration
   if (body.thinking) {
     result.thinking = {

src/shared/constants/pricing.ts

@@ -138,6 +138,14 @@ export const DEFAULT_PRICING = {
       reasoning: 6.0,
       cache_creation: 1.0,
     },
+    // Next-generation Qwen Coder tier (added Mar 2026 from decolua/9router catalog)
+    "qwen3-coder-next": {
+      input: 2.0,
+      output: 8.0,
+      cached: 1.0,
+      reasoning: 12.0,
+      cache_creation: 2.0,
+    },
     "qwen3-coder-flash": {
       input: 0.5,
       output: 2.0,
@@ -198,6 +206,21 @@ export const DEFAULT_PRICING = {
       reasoning: 4.5,
       cache_creation: 0.75,
     },
+    // Short-form aliases used by decolua/9router catalog (Mar 2026)
+    "deepseek-3.1": {
+      input: 0.27,
+      output: 1.1,
+      cached: 0.07,
+      reasoning: 2.2,
+      cache_creation: 0.27,
+    },
+    "deepseek-3.2": {
+      input: 0.27,
+      output: 1.1,
+      cached: 0.07,
+      reasoning: 2.2,
+      cache_creation: 0.27,
+    },
     "minimax-m2": {
       input: 0.5,
       output: 2.0,

src/shared/services/cliRuntime.ts

@@ -198,7 +198,26 @@ const resolveToolCommands = (toolId: string): string[] => {
   return tool.defaultCommand ? [tool.defaultCommand] : [];
 };
 
+/**
+ * T12: Validate a CLI executable path to prevent shell injection.
+ * Enforces: absolute path, no dangerous shell metacharacters, must exist and be a file.
+ * Inspired by Antigravity Manager commit 96732c2 (Mar 11, 2026).
+ */
+const DANGEROUS_PATH_CHARS = ["&", "|", ";", "<", ">", "(", ")", "`", "$", "^", "%", "!"];
+
+const isSafePath = (execPath: string): boolean => {
+  if (!execPath || !path.isAbsolute(execPath)) return false;
+  if (DANGEROUS_PATH_CHARS.some((c) => execPath.includes(c))) return false;
+  // Allow path.sep and path.delimiter — no further character filtering needed
+  return true;
+};
+
 const checkExplicitPath = async (commandPath: string) => {
+  // Reject paths that look like injection attempts
+  if (!isSafePath(commandPath)) {
+    return { installed: false, commandPath: null, reason: "unsafe_path" };
+  }
+
   try {
     await fs.access(commandPath, fs.constants.F_OK);
   } catch {