Tenkai Daily — April 13, 2026

Open Source Releases

• block/goose v1.30.0 — Terminal UI gets a refresh with tab-expandable tool calls and better message rendering. Also adds sub recipe management in the desktop UI and passes toolInfo to MCP Apps via hostContext.
• JuliusBrussee/caveman — Claude Code skill that cuts token usage by 65% by forcing the model to chat like a caveman. You get smaller bills; results may vary.
• cline/cline v3.78.0 — Ships a ‘Spend Limit Reached’ error UI and fixes read_file line range display in the chat view. Also updates docs.
• claude-code v2.1.101 — Adds a /team-onboarding command for auto-generating ramp-up guides. Also adds OS CA certificate store trust for enterprise TLS proxies with a bundled fallback.

Research Worth Reading

• OpenKedge: Governing Agentic Mutation with Execution-Bound Safety and Evidence Chains — Proposes OpenKedge to control AI agent mutations as a governed process, using execution-bound safety and evidence chains to manage probabilistic risks.
• From Business Events to Auditable Decisions: Ontology-Governed Graph Simulation for Enterprise AI — Grounds LLM-based agent decisions in business event simulation to produce auditable outcomes with traceable decision trails.
• Sustained Impact of Agentic Personalisation in Marketing: A Longitudinal Case Study — Examines how long agentic personalization in marketing CRM holds up and why human-in-the-loop oversight still matters.
• RAMP: Hybrid DRL for Online Learning of Numeric Action Models — Introduces a hybrid DRL approach to learn numeric action models from observations, skipping the need for offline expert traces.
• SPPO: Sequence-Level PPO for Long-Horizon Reasoning Tasks — Addresses temporal credit instability in long-horizon reasoning with sequence-level PPO for LLMs using verifiable rewards.
• StaRPO: Stability-Augmented Reinforcement Policy Optimization — Augments RL with stability considerations to capture the internal logical structure of reasoning, not just final-answer correctness.

AI Dev Tools

• affaan-m/everything-claude-code — Agent harness performance optimization for Claude Code, covering skills, memory, security, and research-first workflows.
• rtk-ai/rtk — CLI proxy that shaves 60–90% off LLM token usage for dev commands. One Rust binary, no dependencies.
• oraios/serena — MCP toolkit for coding with semantic retrieval and editing. Thinks of itself as an IDE for agent-based workflows 🤖.

Today’s Synthesis

Use rtk-ai/rtk to cut token burn on Claude Code workflows, pairing it with affaan-m/everything-claude-code to tune skills, memory, and security for measurable cost control 🤖. Route prompts through the proxy to shave 60–90% of token usage on routine refactors and tests, while the harness framework structures how agents access tools and retain context, reducing wasted cycles. Layer in JuliusBrussee/caveman to force models into tighter, more predictable reasoning patterns that further trim token use without relying on complex fine-tuning. Combine these with insights from SPPO: Sequence-Level PPO for Long-Horizon Reasoning Tasks to design verifiable reward signals that keep long-running agentic sessions on track. This setup gives engineers a concrete stack: proxy for traffic control, harness for governance, and compact prompting strategies to sustain performance. The result is lower bills, clearer audit trails, and fewer surprises when agent workflows scale, making disciplined agent economics a default rather than an edge case.