Caveman 3.1.0 released
A major release replacing intensity levels with three rethought skills and landing a large fix stack across the codebase.
What changed in 3.1.0
- Three new skills replace the six intensity levels: /caveman, /ultracave for grammar stripping, and /megacave for classical Chinese, each with self-contained rules and structure.
- Session migrations happen automatically on read; legacy aliases /caveman lite|full|ultra|wenyan continue to work.
- Mode IDs changed to caveman, ultracave, megacave with legacy values mapping on every read path; SessionStart hook now injects the active skill.
- Fix stack addresses proxy session correlation, JSON handling, CLI hooks, Codex TOML, streaming and ceiling behavior, and statusline nudge paths.
- Maintenance updates across multiple systems including Python encoding, dependency pins, and plugin manifests.
Check before you update
- Intensity levels no longer exist; use /caveman, /ultracave, or /megacave skills instead.
- Mode IDs are now caveman, ultracave, megacave; stored sessions will migrate automatically on read.
- defaultMode hook behavior changed; Codex SessionStart now honors defaultMode setting.
Summarised by AgentGid from JuliusBrussee's official release notes on GitHub, as reported there; read the original before relying on it. Price, status and release dates come from our daily data.
Caveman at a glance
| What it is | Skill, local proxy and SDK that reduce the tokens coding agents write and read: the skill makes replies terse and the proxy compresses logs, JSON, test output and web pages before they reach the model. It works with Cla… |
|---|---|
| Latest version | 4.1.0, Nov 8, 2026. This page is about an earlier release. |
| Release pace | 5 stable releases in the last 30 days; 3.1.0 came the same day as the previous one |
| Price | Free + $300/mo. Caveman agent and SDK are free and open source (Apache-2.0). Caveman Cloud Free tier is $0/month for one person with usage allowances (100K events, 600 agent minutes, 25K routing … Official pricing ↗ (checked Oct 14, 2026) |
| Gid Score | 31 #41 in Agent tooling for cost, context and memory |