Ai

Caveman v1.7.0 Adds Usage Receipts, Smarter Installation, and Leaner AI Agent Handoffs

Caveman v1.7.0 Adds Usage Receipts, Smarter Installation, and Leaner AI Agent Handoffs

Caveman v1.7.0 is positioned as the project’s biggest update since v1.0, and this release is clearly centered on measurable efficiency improvements for AI coding workflows. The update introduces direct session-based usage receipts, a smarter installer that detects and configures more than 30 agents, slimmer subagent handoff behavior, MCP middleware for reducing tool-description overhead, and an important macOS installer fix that previously broke detection for compound-spec providers.

What Changed

The headline addition in v1.7.0 is /caveman-stats, a new command that reads active Claude Code session JSONL files directly to calculate token output, cache-read activity, and estimated savings. Instead of relying on rough model-side estimates, the feature uses model-id prefix matching for pricing logic, making cost and token reporting more reliable across model variants and point releases.

The release also adds a default statusline savings badge, giving users a persistent view of cumulative token savings after the first stats run. That makes efficiency tracking much more visible in day-to-day usage rather than burying it in occasional reports.

Another major change is the upgraded installer. According to the release notes, it now detects more than 30 agents and runs each one’s native install flow. That is a meaningful usability improvement for teams working across multiple agent environments, especially where setup consistency matters.

Version 1.7.0 also introduces three caveman-mode subagents designed to emit around 60% fewer handoff tokens than standard alternatives. In parallel, the new MCP middleware shrinks tool descriptions in flight, further reducing prompt and orchestration overhead during agent-tool interactions.

Finally, the release includes a critical macOS installer fix. The bug had reportedly been silently breaking detection for every compound-spec provider, including Cursor, Windsurf, Continue, and many others. That fix is likely one of the most practically important parts of this version for affected users.

Why It Matters

This update matters because it shifts Caveman from optimization by intuition to optimization by measurement. With direct receipt-style token tracking, teams can now quantify the impact of workflow compression, cached context usage, and memory reduction in a more concrete way.

The installer improvements also lower friction for adoption. In agent-heavy development environments, installation complexity can slow experimentation and create inconsistent setups. Native install detection across dozens of agents reduces that operational burden.

The leaner subagent handoffs and MCP-shrinking middleware are especially relevant for AI engineering teams trying to control context bloat and inference costs. As multi-agent and tool-driven development workflows grow more common, token overhead becomes a real performance and budget concern. Caveman v1.7.0 directly targets that problem.

Meanwhile, the macOS detection fix improves trust in the setup process. Silent failures in provider detection can waste significant engineering time, so resolving a bug that affected a broad class of providers strengthens the release beyond just new features.

Official Source: https://github.com/JuliusBrussee/caveman/releases/tag/v1.7.0