feat: add three-layer agent loop protection

[Jonnytoshen]

Summary

Adds three layers of protection to the agent loop to prevent runaway execution, API failures, and budget overruns.

Changes

New files

• src/loop-detection.ts — Tool call history tracker with a 30-entry sliding window and three detectors:
  • generic_repeat: same tool + same args repeated N times
  • ping_pong: two different calls alternating
  • global_circuit_breaker: any call repeated without result change
  • Three-level response: warning (5×) → inject nudge message, critical (8×) → halt, breaker (10×) → force stop
• src/retry.ts — Step-level retry utilities: calculateDelay (exponential backoff), isRetryable, sleep
• src/mock-model.ts — Offline mock model for testing all three protection layers without a live API key

Modified files

• src/agent-loop.ts
  • Integrated loop detection (detect / recordCall / recordResult)
  • Step-level retry loop (up to MAX_RETRIES = 3) with exponential backoff
  • Token budget tracking via BudgetState (caller-owned, accumulated across steps)
  • MAX_STEPS bumped 10 → 15 to accommodate retry headroom
  • Switched type to LanguageModelV3 from @ai-sdk/provider
• src/index.ts
  • Pass budget to agentLoop; auto-select mock model when DASHSCOPE_API_KEY is absent
  • Rename askQuestion → ask
  • Add in-app test hints for the three protection scenarios
• package.json — Add @ai-sdk/provider; bump @ai-sdk/openai and ai

Testing

Start the agent without an API key to use the mock model:

pnpm start

Then try the prompts suggested on startup:

• 测试死循环 — triggers loop detection
• 测试重试 — triggers retry logic
• 测试预算 — triggers token budget cutoff