#600 release: Anthropic prompt cache control (#599)

Summary

• Ports open-agents' addCacheControl helper (split into isAnthropicModel, addCacheControlToTools, addCacheControlToMessages per SRP)
• Wires into runAgentStep: tools wrapped once + prepareStep callback wraps messages per step
• Anthropic now caches system prompt + tools across turns; subsequent turns re-read most of the prefix at ~10% cost

Test plan

• 19 unit tests; full suite 3138/3138 pass
• E2E on preview verified: turn 2 cost drops from baseline $0.005959 → $0.00167 (72% cheaper); cacheReadTokens=5020 on turn 2 (was 0)
• Cache shape matches open-agents production (4.5× drop on Haiku vs 12× on Opus — ratio reflects model, shape matches)

🤖 Generated with Claude Code

---

Summary by cubic

Adds Anthropic prompt caching to the chat workflow so the system prompt and tools are reused across turns. Marks the last tool and last message with cacheControl: { type: "ephemeral" }, cutting follow-up turn input cost (~72% in E2E).

• New Features
  • Added isAnthropicModel, addCacheControlToTools, and addCacheControlToMessages.
  • runAgentStep: wraps tools once and uses prepareStep to tag the final message per step (Anthropic only).
  • No changes for non-Anthropic models; 19 unit tests cover helpers and wiring.