Summary

Gateway crashes with unhandled promise rejection when OpenAI embeddings API returns 429 (quota exceeded). The configured openai-codex OAuth accounts are not used as fallback.

Error

[clawdbot] Unhandled promise rejection: Error: openai embeddings failed: 429 {
    "error": {
        "message": "You exceeded your current quota, please check your plan and billing details.",
        "type": "insufficient_quota",
        "code": "insufficient_quota"
    }
}
  at Object.embed [as embedBatch] (file:///Users/oliver/Developer/clawdbot/src/memory/embeddings.ts:1:1562)
  at MemoryIndexManager.indexFile (file:///Users/oliver/Developer/clawdbot/src/memory/manager.ts:25:3078)
  at MemoryIndexManager.runSync (file:///Users/oliver/Developer/clawdbot/src/memory/manager.ts:25:2016)
error: script "clawdbot" exited with code 1

Environment

• Clawdbot version: 2026.1.15
• macOS (arm64)
• Auth profiles: openai-codex:oliver, openai-codex:sylvia (both working for chat)
• OPENAI_API_KEY env var set but account has no quota

Config

"memorySearch": {
  "provider": "openai",
  "fallback": "openai",
  "model": "text-embedding-3-small",
  "sync": { "watch": true }
}

Root Cause

1. Schema restriction: memorySearch.provider only accepts "openai" | "local" — not "openai-codex"
2. No real fallback: Both provider and fallback are "openai", so failure just fails again
3. Credential resolution: resolveApiKeyForProvider({ provider: "openai" }) finds OPENAI_API_KEY env var (which has no quota) instead of using openai-codex OAuth tokens
4. Crash on 429: Unhandled rejection crashes the gateway instead of graceful degradation

Attempted Fix

Setting "provider": "openai-codex" results in:

Invalid config: - agents.defaults.memorySearch.provider: Invalid input

Suggestion: Support Any OpenAI-Compatible Provider

Instead of hardcoding "openai" | "local", allow any provider that resolves to an OpenAI-compatible API:

// Current (restrictive):
provider: "openai" | "local";

// Suggested (flexible):
provider: "openai" | "openai-codex" | "local" | string;

Or better, reuse the existing provider resolution logic:

// In createOpenAiEmbeddingProvider():
const { apiKey } = await resolveApiKeyForProvider({
  provider: options.provider,  // Pass through as-is
  cfg: options.config,
  agentDir: options.agentDir,
});

This would allow:

• "openai" → uses OPENAI_API_KEY or openai auth profiles
• "openai-codex" → uses ChatGPT OAuth tokens (if embeddings are included in scope)
• "openrouter" → uses OpenRouter (supports embeddings)
• Any custom OpenAI-compatible provider

Additional Issues

1. Crash vs graceful degradation: 429 should disable memory search, not crash the gateway
2. Fallback is broken: "fallback": "openai" when provider is already "openai" is a no-op
3. No warning on startup: If embeddings will fail, warn during config validation

Workaround

Use local embeddings:

"memorySearch": {
  "provider": "local",
  "fallback": "none"
}

Labels

bug, crash, memory, embeddings