This release adds 3 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Topics
+14 more
Summary
AI summaryPrompt caching, reasoning‑tag scrubbing, and tool‑loop circuit breaking are added to improve efficiency and reliability.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Feature | Low |
Introduces tool-loop circuit breaker warning at 3 failures and blocking at 5 consecutive command failures. Introduces tool-loop circuit breaker warning at 3 failures and blocking at 5 consecutive command failures. Source: llm_adapter@2026-05-29 Confidence: high |
— |
| Feature | Low |
Adds prompt caching for `agent_execute` and `@claude-flow/providers`. Adds prompt caching for `agent_execute` and `@claude-flow/providers`. Source: llm_adapter@2026-05-29 Confidence: low |
— |
| Feature | Low |
Strips reasoning-tag blocks (`think`, `reasoning`) before DISTILL processing. Strips reasoning-tag blocks (`think`, `reasoning`) before DISTILL processing. Source: llm_adapter@2026-05-29 Confidence: low |
— |
| Feature | Low |
Adds prompt caching for `agent_execute` with ~90% token discount and 5‑min TTL. Adds prompt caching for `agent_execute` with ~90% token discount and 5‑min TTL. Source: granite4.1:30b@2026-05-29-audit Confidence: low |
— |
| Feature | Low |
Enables configurable prompt caching in `@claude-flow/providers` (gated by `config.promptCache`). Enables configurable prompt caching in `@claude-flow/providers` (gated by `config.promptCache`). Source: granite4.1:30b@2026-05-29-audit Confidence: low |
— |
| Feature | Low |
Strips `think`, `reasoning`, and similar blocks from trajectory actions/results before DISTILL. Strips `think`, `reasoning`, and similar blocks from trajectory actions/results before DISTILL. Source: granite4.1:30b@2026-05-29-audit Confidence: low |
— |
Full changelog
Ruflo v3.10.10 — Hermes-Agent Tier-1 adoptions
The 3 small, high-confidence, in-repo patterns worth borrowing from NousResearch/Hermes-Agent, from a deep-research capability map (the rest were SKIP or ADR-gated; the map also produced 5 corrections to ADR-113/#1907).
- Prompt caching —
agent_execute's direct Anthropic call now marks the system prompt as an ephemeral cache breakpoint (~90% discount on cached input tokens, 5-min TTL). Same strategy applied to@claude-flow/providers(gated byconfig.promptCache, default on). - Reasoning-tag scrub — trajectory action/result are stripped of
<think>/<thinking>/<reasoning>/<REASONING_SCRATCHPAD>blocks before DISTILL, so extended-thinking tokens don't contaminate the pattern embeddings the learning loop relies on. Boundary-gated. - Tool-loop circuit breaker —
pre-commandwarns at 3 / blocks at 5 consecutive failures of the same command (with a recovery hint), fed bypost-command. Orthogonal to the injection-focused security guardrail; advisory by default.
Honest note: @claude-flow/providers is not consumed by the cli, so the cli-effective prompt-caching fix is in agent-execute-core.ts; the providers change ships on its own cadence.
CI 29/29 green; all three packages at 3.10.10 (latest/alpha/v3alpha lockstep).
🤖 Generated with RuFlo
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
Related context
Related tools
Beta — feedback welcome: [email protected]