Token Economy HUD: cached tokens counter accumulates across requests

[bonk-moltbot]

What

The cached tokens counter in the Token Economy HUD shows 1M+ tokens, which is physically impossible in a single context window.

Expected behavior

Show per-request cache statistics — how many tokens were cache hits vs cache misses in the current (or most recent) API request.

Actual behavior

Counter appears to sum cache_read_input_tokens cumulatively across all API requests in the session, producing impossibly large numbers.

Related

• Token Economy HUD — rate limits, cache efficiency, and cost tracking #400 — Token Economy HUD (implemented this counter)
• Slice 2: Last-message caching + cache metrics event #397 — Cache metrics event
• Anthropic prompt caching — reduce input token cost ~4x #395 — Prompt caching epic

References

Full exploration: thoughts/shared/notes/2026-04-03/qa-brainstorm-event-loop-mailbox.md (Bug 2: Cached tokens counter shows 1M+)