From the deep-research map of NousResearch/Hermes-Agent vs ruflo (3 small,
high-confidence, in-repo adoptions; the rest were SKIP/ADR-gated).
- #8 Prompt caching: agent_execute's direct Anthropic call (agent-execute-core.ts
— the path the cli actually uses) now marks the system prompt as an ephemeral
cache breakpoint (~90% discount on cached input tokens, 5-min TTL). Also applied
the system+trailing-message breakpoint strategy to @claude-flow/providers'
anthropic-provider (separate package; gated by new config.promptCache, default on).
- #14 Reasoning-tag scrub: scrubReasoningBlocks() strips <think>/<thinking>/
<reasoning>/<REASONING_SCRATCHPAD> blocks from trajectory action/result before
DISTILL, so extended-thinking tokens don't contaminate pattern embeddings.
Boundary-gated (leaves prose that merely mentions the tags untouched).
- #6 Tool-loop circuit breaker: new tool-loop-guardrail.ts; post-command records
(command, success), pre-command warns at 3 / blocks at 5 consecutive failures of
the same command with a recovery hint. Orthogonal to the security
tool-output-guardrail (injection). Advisory by default.
Note: @claude-flow/providers is NOT consumed by the cli (separate published
package), so the cli-effective #8 fix is in agent-execute-core.ts; the providers
change ships on its own cadence.
Verification: cli + providers build clean; 29/29 targeted tests pass
(hermes-tier1 6 + bandit 8 + bug-cluster 15). Deep-research map also produced 5
corrections to ADR-113 (#1907).
rUv