feat: enable Anthropic prompt caching for system prompts

[samarthpatel24]

Problem:

The Anthropic backend re-sends the full system prompt (~5K+ tokens) on everysingle agent step at full price. On a 100-step Claude Opus run, that's ~$7.50 burned just on the same static system prompt repeating 100 times. The irony — the codebase already tracks cached_input_tokens, but it's always 0 because caching was never wired up.

Fix:

Flips on Anthropic prompt caching by sending the system prompt as a content block with cache_control ephemeral. Step 1 writes to cache (1.25x); steps 2–100 read from it (0.1x). That's roughly ~90% savings on system prompt input cost per run.

Changes:

• anthropic_model.py — System prompt sent as content block array with cache_control
• anthropic_model.py — Metrics helper now handles both string and list-of-blocks system format
• anthropic_model.py — Usage parser now captures cache_creation_input_tokens
• base.py — Added cache_creation_input_tokens to usage metric keys

Not affected:

• OpenAI / OpenRouter — untouched, just report 0 for the new metric key
• API version — 2023-06-01 already supports caching, no bump needed
• Short prompts — if system prompt is under 4096 tokens, API silently skips caching (no error)

Test plan:

• All 7 existing unit tests pass
• Live run: confirm cache_read_input_tokens > 0 in usage metrics after step 1

[samarthpatel24]

@microsoft-github-policy-service agree