Skip to content
Agents tracked: 378 Downloads (7d): 244M down 5.7% GitHub stars: 8.4M VS Code installs: 151M Releases (7d): 407 Agent status: 1 with issues Updated Oct 10, 2026
AgentGid
✦ Ask

Caveman changelog: what's new each month

Every stable Caveman release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.

October 2026 · so far

5 releases: 2.0.2 → 3.2.0

October brought major updates across multiple versions of Caveman, with improvements to runtime security, compression handling, language support, and session management. The month focused on stability, dependency updates, and expanded proxy routing capabilities.

Highlights

  • Signed runtime companions now available for macOS, Linux, and Windows with Apache 2.0 licensing.
  • Added tree-sitter support for TypeScript, JavaScript, Python, Rust, Java, C, and C++ across platform binaries.
  • Improved rate-limit handling with Retry-After pass-through and conversation-aware raw retry routing.
  • Enhanced compression logic to maintain stability across conversations, including raw bytes decisions.
  • Extended proxy routing to support ChatGPT OAuth traffic and bare /v1/models requests by protocol.

New

  • Tree-sitter language support for seven programming languages in the six-platform binary matrix.
  • Native voice skills included with Claude and Codex installations.
  • Compression support for JUnit, pytest, and Jest test reports.
  • Session mode inheritance for subagents from their parent session.
  • Project-scoped session lookup with stats recorded at SessionEnd.
  • Support for filesystems without hard links via CCR.
  • Headless mode for claude -p and Agent SDK sessions with voice disabled by default.

Improved

  • Compression decisions now remain stable across conversations, including raw bytes handling.
  • Rate-limit responses are passed through with Retry-After headers intact.
  • Raw retries pinned to their conversation lineage for better request tracking.
  • Session telemetry now records caveman-caused changes in local requests.
  • Tool-result compression now requires a recoverable handle.
  • Headless claude -p and Agent SDK sessions start with voice off.
  • Python encoding fixes and merge scope checking applied across releases.

Fixed

  • Python encoding handling corrected.
  • Dead proxy and routing code removed.
  • GDScript elision implemented.
  • C++ parse gate added.
  • Grok Build installation issue resolved.
  • TypeScript 7 and @types/node 26 compatibility issue reverted and stabilized.

September 2026

13 releases: 2.5.0 → 1.0.0

September brought major stability improvements, security hardening, and the arrival of version 3.0.0 with new learning and analytics features. The month also saw a significant licensing change across the platform to Apache-2.0.

Action needed
  • License changed from MIT to Apache-2.0 across the entire platform.
  • Package renamed from @caveman/contracts to @caveman-ai/contracts.
  • Every schema ID changed to new format.
  • Middleware now requires SDK 1.2.0 or later.

Highlights

  • Version 3.0.0 introduced Autopilot for background learning, Memory Doctor for config validation, and week-over-week trends from your own sessions.
  • Security improved: untrusted listeners on proxy ports are now blocked, and provider keys no longer route through foreign processes.
  • Firefox extension now available alongside existing Chrome support, with proper manifest and addons-linter validation.
  • Compression and stats now more reliable: compress validates candidates before touching live files, and multi-block responses are counted correctly.
  • Entire platform relicensed to Apache-2.0 with signed runtime companions for macOS, Linux, and Windows.

New

  • Autopilot background learning re-checks session history at low priority every 6 hours without blocking sessions.
  • Memory Doctor validates MEMORY.md files, detects missing imports, and finds duplicate rules across files.
  • Week-over-week session trends using median data, with "not enough data" shown for fewer than 5 sessions.
  • Built-in model price list covering current Claude and GPT lineups, independent of working directory.
  • Firefox port of the Caveman Mode extension with AMO-clean manifest.
  • Per-model custom provider routing with compat-mount support.
  • Middleware protocol 1.1 with optional version negotiation and new error handling messages.
  • Adapters fail open: if Caveman encounters problems, requests pass through untouched with a warning.

Improved

  • Caveman prose now mixes ASD-STE100 clarity register to keep compressed answers unambiguous.
  • Per-turn reinforcement strengthened so mid-conversation competing instructions don't drift the style.
  • Array-shaped tool output is now compressed, and route eligibility is recorded.
  • Proxy native-hook bridge returns on first complete JSON payload instead of waiting for EOF.
  • Cavecrew subagents work with Gemini CLI after dropping Claude-only tools frontmatter.
  • Stats number formatting pinned to en-US for consistency.
  • Claude Code plugin and Gemini extension manifests now declare the release version instead of showing commit sha.
  • Shell commands no longer rewritten in Codex, preserving saved approval rules.

Fixed

  • Provider-key security: wrap now launches agent directly instead of routing keys through unverified foreign processes.
  • Compress no longer self-corrupts its config or stacks duplicate hooks.
  • Stats now reports only what transcripts show; unmeasured savings claims removed from docs.
  • Caveman learn background re-check validates each candidate in staging file before touching live file.
  • Multi-block response counting fixed to count each response once.
  • Claude Code Remote Control now launches direct under wrap when proxied base URL is detected.

August 2026

13 releases: 1.10.0 → 2.4.0

August brought Caveman 2, a major engine release that adds real compression to reduce what models read, alongside numerous stability fixes and a comprehensive reporting system for measuring actual savings.

Highlights

  • Caveman 2 ships with a compression engine that squeezes tool outputs, logs, code, and JSON before the model reads them—achieving 93% reduction on real data.
  • New context-depth and cost reporting shows your session's peak context usage, actual savings measured on your machine, and 30-day cost illustrations.
  • Caveman learn reconcile ties measured spend to provider usage exports for accurate accounting without projecting savings.
  • Fixed critical installation and uninstall issues on Windows and Linux that could corrupt settings or leak half-written files.
  • Full test suite now runs on every push across Node 18/20/22 and Python 3.11.

New

  • Compression engine squeezes input data (logs, JSON, code, tool outputs) before models read it, with byte-safe, fully recoverable output.
  • Context-depth report buckets each session's peak context-window share and flags sessions over 30% or 50%.
  • Controlled-holdout harness runs fixes on and off on the same machine to measure impact beyond the edited file.
  • Spend attribution names the confirmation method used (deterministicremeasure, controlledholdout, counterfactualreplay, or interruptedtimeseries).
  • Privacy-safe export digest for sharing cost reports without shipping prompt text.
  • Twenty-chapter technical manual under docs/technical/.
  • Proxy buffers non-streaming upstream bodies before replying and logs serve-mode warnings to ~/.caveman/proxy.log.
  • Skill forbids caveman phrasing that adds tokens and names defect/ticket text.

Improved

  • Token savings now split by workload—65% for chat-prose; agentic coding shows 8.5% across 86 tasks with no quality change.
  • /caveman-stats reports net savings (gross minus ~1–1.5k input tokens per turn) instead of gross only.
  • Hooks share one parser and survive resume/clear/compact re-fires; mid-session /caveman <level> switches no longer reset on SessionStart re-fire.
  • Installer moved bin/ to cli/ with hardened Windows quoting and clean uninstall path.
  • Installer installs Cursor skills globally and leaves other plugins' hooks/package.json untouched.
  • Validator now recognizes CommonMark indented code blocks (4-space indent).
  • Outer-fence strip no longer merges a document's first and last code blocks into prose.
  • Lost file paths and renamed headings are now errors instead of warnings.

Fixed

  • Compress no longer loses data on Windows; UTF-8 and atomic writes prevent file truncation on crash mid-write.
  • Uninstall no longer deletes hook files while settings.json points at them, fixing repeated errors on session start.
  • Concurrent binary installers no longer publish each other's half-written files by using unique .part filenames.
  • Settings file that cannot be parsed or rewritten now halts hard instead of proceeding silently.
  • Installer temp downloads moved to 0700 mkdtemp to prevent symlink pre-planting attacks.
  • cavecrew model overrides no longer rewrite tracked agents/.md when plugin root is a source checkout.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Caveman: pricing, features and alternatives · All changelogs