Hewbers / Setup and use
Build, configure and operate Hewbers.
Hewbers is currently a source-built service. The public concepts below are stable enough for evaluation and contribution; exact instance files and deployment paths remain repository-specific.
01 / Guide
What you need
Use Linux, Git, a matching Hew compiler and standard library, and the repository's Make targets. A continuously supervised installation also needs systemd.
Choose at least one repository connection and one coding-provider adapter. Authenticate provider CLIs in the agent home selected for that instance.
- GitHub or Gitea repository access
- GitHub Copilot CLI, Claude Code, Codex or an OpenClaw-compatible harness
- Repository-defined build and test commands
- A private state directory separate from the source checkout
02 / Guide
Build and inspect
The Makefile is the supported build and test interface. It stages the service programs and operator tooling into bin/ without installing or activating a service.
make all
make test
bin/hewbers version
bin/hewbers check
bin/hewbers status
bin/hewbers runs03 / Guide
Configuration concepts
An instance pins its state directory, repository connections, workflow inventory, runner inventory, provider commands and operating limits. Existing runs retain the choices with which they were admitted.
Connections define backlog and repository access. Workflows define tasks and gates. Runners advertise the operating system, capabilities and restrictions they can satisfy. A task starts only when one runner satisfies its complete requirement set.
04 / Guide
Run and observe
For foreground evaluation, build the local runtime artifacts and start the service with an instance configuration. The daemon keeps a state-directory lock and owns claims, workers and recovery.
Use the CLI for health and run listings. Use workflow status and trace views for the recorded task sequence, gate decisions and terminal outcome.
make local-runtime-artifacts
bin/hewbers --config /path/to/instance.json serve
bin/hewbers --json runs 2005 / Guide
Design workflows around authority
Keep implementation, review and merge authority separate. Bind every review and check to an exact head. Treat repository writes and remote submissions as effects that may be accepted even when their acknowledgement is lost.
Use explicit gates for repository policy, local verification and provider CI. Give workers only the inputs and credentials required for their stage; completion and claim release remain owner decisions.
06 / Guide
Plan for interruption
Hewbers records intent before external effects, then records the accepted result. After restart it reads the provider or durable result before deciding whether an action may be repeated.
Unknown effects stay uncertain and retain their fence. A clean restart does not invent failure, success or permission to repeat a merge, comment, label change, remote submission or model invocation.
Start with the workflow, not the daemon.
Define who may select work, what each stage receives, which exact result a gate covers and how recovery distinguishes failure from an unknown external effect.