JuliusBrussee/caveman v4.0.0

Caveman 4.0.0

on GitHub

Caveman 4.0.0 makes npx @caveman-ai/cli the one install, adds Auto routing through Caveman Cloud on top of the local proxy, and ships Caveman Blocks as a module.

One command to install

• npx @caveman-ai/cli opens a one-screen setup. It finds your coding agents and their logins (every Claude Code profile), shows what it will change, applies on Enter, then offers to start the agent. Without a terminal it applies only with --yes.
• install.sh and install.ps1 end in the same setup, pinned to the CLI this release ships.
• caveman on|off switches modules from a registry. The module index (modules.json) is listed in the signed checksum manifest of the binary release, and the CLI refuses an index that manifest does not list. A module lockfile records what is installed.
• New commands: caveman stop, a status grid, an offline caveman doctor, caveman providers add|remove|login|local and caveman providers cloud on|off.
• One config home: ~/.caveman/cloud.json. Older config migrates once, without races, and the telemetry opt-out carries over.
• Sign-in uses the shared device flow. It refuses plain-HTTP Clouds other than this machine.

Auto routing (signed in to Caveman Cloud)

• caveman-auto shows up in the Claude Code and OpenCode model pickers with a 1M context window. Routing applies only to Auto, and that includes Claude Pro/Max logins and ChatGPT logins (Codex, OpenCode).
• For each routable request the proxy asks Cloud's route stage for a model and effort (contracts route-ask-v1) and picks from the providers you set up. If the key is refused or a billing limit is hit, the request goes through unrouted and caveman status says why.
• Translators between Messages, Chat Completions and Responses cover all nine directions, so Claude Code can run on Responses hosts such as the ChatGPT plan.
• A signed-in runtime sends runtime/v1 event batches to Cloud and respects the telemetry opt-out.

Cache

• The proxy keeps the Anthropic prompt cache warm while a session is paused. Warm timing comes from measured return times, at most two warms run at once, and caveman stats lists every warm.
• A session stays on its model across Cloud pauses and transient failures, so a model switch doesn't rewrite the cached prefix.

Caveman Blocks

• The binary release mirrors caveman-ai/blocks v0.1.0-rc.4 (pinned by the digest of its checksum list). Each Blocks binary is republished under the same signed manifest as the other binaries. Blocks usage counts fall under the same telemetry notice and opt-out as Caveman's.

Catalog

• Adds Sonnet 5.5, Haiku 5.5, GPT-6.1 Sol and GPT-6 Luna.

Versions

• @caveman-ai/cli goes from 2.1.0 to 4.0.0 to match the product version. Binaries move from bin-v2.1.0 to bin-v4.0.0.
• The npx github:JuliusBrussee/caveman installer depends on CLI ^4.0.0, so -- --uninstall runs a CLI that understands what 4.0 wrote.

Companion releases

• Signed binaries, module index and container: bin-v4.0.0
• npm CLI: @caveman-ai/cli@4.0.0

Full changes: v3.2.0...v4.0.0

Release smoke: https://github.com/JuliusBrussee/caveman/actions/runs/38431102855