docs: verify prompt caching optimization is already implemented (#851)

docs/issue-851-verification.md

@@ -0,0 +1,55 @@
+# Issue #851 Verification
+
+## Status: ✅ ALREADY IMPLEMENTED
+
+Issue #851 is a research/audit issue whose own conclusion is that prompt caching is already extensively implemented in hermes-agent and that the remaining work is operational, not a repo-side code change.
+
+This verification confirms that the current repo already contains the core implementation described in the issue body.
+
+## Acceptance Criteria Check
+
+1. ✅ Anthropic / OpenRouter prompt-caching support exists
+   - `agent/prompt_caching.py:41-72` implements `apply_anthropic_cache_control()` with the documented system-plus-last-3 breakpoint strategy.
+   - `run_agent.py:8301-8306` applies Anthropic/OpenRouter cache-control breakpoints during API message preparation.
+
+2. ✅ OpenAI/Codex prompt-cache key support exists
+   - `run_agent.py:6199-6213` sets `prompt_cache_key = self.session_id` on the responses path for non-GitHub responses.
+   - `run_agent.py:3875-3878` explicitly passes through `prompt_cache_key` in normalized API kwargs.
+
+3. ✅ System-prompt stability and cache-friendly message normalization exist
+   - `run_agent.py:3155-3157` documents that the system prompt is cached and reused across turns to maximize prefix cache hits.
+   - `run_agent.py:8314-8339` normalizes whitespace and tool-call JSON for bit-perfect prefix matching across turns.
+
+4. ✅ Cache hit/miss logging infrastructure exists
+   - `run_agent.py:8966-8980` logs cache read/write token stats, including `cached_tokens`, `cache_creation_input_tokens`, and hit percentage.
+
+## Executed Verification
+
+### Targeted tests run
+- `PYTHONPATH=/tmp/BURN2-FORGE-ALPHA-3 python3 -m pytest -q tests/agent/test_prompt_caching.py`
+  - Result: `14 passed`
+
+### Syntax verification
+- `PYTHONPATH=/tmp/BURN2-FORGE-ALPHA-3 python3 -m py_compile agent/prompt_caching.py run_agent.py`
+  - Result: passed
+
+## Evidence Summary
+
+The issue body says:
+- prompt caching is already extensively implemented
+- the primary opportunities are operational: routing more workloads to Ollama, verifying provider support, and reporting cache hit rates
+
+The repo state matches that conclusion:
+- caching primitives are present
+- integration points are wired into the runtime
+- targeted tests already exist and pass
+- no new implementation change is required to satisfy the issue's repo-side claim
+
+## Recommendation
+
+Close issue #851 as already implemented in the codebase.
+
+If desired, follow-on work should be opened as separate operational issues for:
+- Ollama-heavy workload routing
+- provider-specific cache verification
+- nightly cache hit-rate reporting