A new tool called Replay has hit Hacker News with a mission to audit silent prompt cache misses in AI agent transcripts. The project addresses a critical pain point for developers running AI agents: bills spiking without a single warning or error in the logs. According to the creator, this occurs when a prompt cache breaks, causing providers to silently re-bill the entire conversation history at cold-write prices instead of the expected cached read prices.
The Silent Billing Trap
The core issue identified by Replay is the lack of visibility into cache performance for popular AI coding assistants. Whether a team is using Claude Code, Cursor, or Aider, the underlying provider can fail to serve a cached prompt without triggering any standard log entries. This results in unexpected cost surges that are difficult to trace back to a specific interaction or model behavior, leaving developers guessing at the root cause of their increased API spend.
How Replay Works
Replay functions by analyzing AI agent transcripts to detect these specific cache miss events. By comparing expected cache states with actual billing data, the tool highlights where the cache failed to engage. This allows engineers to pinpoint exactly which prompts or conversation turns resulted in cold-write pricing, providing the necessary data to optimize prompt structures and reduce unnecessary API costs.
Key Takeaways
- Silent cache misses can cause significant, unexplained spikes in AI API bills.
- Major tools like Claude Code, Cursor, and Aider are affected by this lack of visibility.
- Replay provides a mechanism to audit transcripts and identify cold-write pricing events.
- Standard logs often fail to capture cache performance issues, requiring specialized auditing tools.
The Bottom Line
Relying on provider logs alone is insufficient for cost control; teams must actively audit transcripts to catch silent cache misses that drain budgets without warning.