Workflow-driven
Authored workflows define tasks, gates, dependencies, schedules and handoffs rather than embedding one fixed automation path.
Coding workflows / Actor runtime / Built in Hew
Self-hosted coding workflows, built in Hew.
Hewbers brings coding agents, independent review, repository checks and exact-head delivery into one self-hosted workflow. Written in Hew, it uses actors, explicit ownership and durable state to build toward reliable issue-to-merge operation.
HewbersOpen source
Build. Review.
Recover.Hew 0.6.0-rc7-dev.3
Revision de063088
01 / Workflow
From approved issue to merged change.
Hewbers separates selection, implementation, review, verification and merge authority. Each decision is tied to the repository state it covers.
Issue-to-merge lifecycle
Workflow being built- 01
Select and claim
Poll configured backlogs, apply dependency and priority policy, then fence one eligible item to one run.
- 02
Prepare context
Create an isolated worktree and gather the issue, repository instructions, relevant files and workflow inputs.
- 03
Implement
Dispatch the configured coding provider with bounded tools, credentials, time and workspace authority.
- 04
Review independently
A separate reviewer evaluates the exact authored head. Findings return to implementation on the same pull request.
- 05
Verify
Run repository-defined checks, observe provider CI and rebase or repair when the base or result changes.
- 06
Merge and close
A separate merge-review workflow rechecks approval, head, base and CI before merging and closing linked work.
Review findings, failed checks, changed bases and interrupted stages return through bounded repair or recovery paths instead of silently continuing.
Provider-aware
Repository and coding-provider effects are treated as durable external actions whose accepted, failed or uncertain outcomes must be reconciled.
Recovery-first
Claims, journals, exact-head review and intent-before-effect records let a restarted service continue without casually duplicating work.
02 / Execution model
One owner, bounded workers, durable decisions.
The run owner holds lifecycle authority. Selected local workers perform stages; reviewed remote-worker source stays default-off and uncompiled. Provider callbacks and remote results are evidence, not permission to complete or release a run.
Lifecycle authority
Run owner
Pins the workflow and runner inventory, holds the claim or PR lease, evaluates gates and decides completion or release.
Admission → ownership → terminal decision
Bounded execution
Stage workers
Run deterministic commands, coding agents and reviewers in selected local profiles; remote-worker source remains default-off and uncompiled.
Checked input → stage result → surrender
Restartable state
Journal and projections
Record lifecycle events and derive status, trace, recovery and closeout views without rerunning completed effects.
Intent → effect → acknowledgement
Remote stagesReviewed source describes a default-off Kubernetes scratch-worker path with checked identity, scoped credentials and owner-held completion authority. The corrected worker remains uncompiled; worker execution, image behavior, live-cluster, full-chain and live Azure operation are unqualified. Earlier disposable-minikube checks apply to prior-stage scenarios, not this worker.
03 / Implemented
Useful capability, not a demo loop.
Selected local profiles have passed bounded checks; the current remote worker is default-off and uncompiled, not an operating capability.
Continuous backlog service
The supervised service polls backlogs, checks dependencies and selects work under a claim. The Linux service runs under maintainer oversight; its recorded backlog is currently stalled before implementation can begin.
Agentic implementation
Reviewed source supports isolated Git worktree authoring through configured Copilot, Claude Code, Codex or OpenClaw-compatible adapters, with authored-head capture and branch publication in selected profiles. Workflow inputs are pinned to runs. A separate continuation source path can admit one unchanged local-start successor from terminal-owner evidence; it carries no context or execution budget, and successor execution is not established. A finished provider invocation is not successful coding completion: retained partial drafts do not authorize commits. Signing and trusted finalization stay outside the coding stage. Prompt, context and verifier-capture artifacts use restricted private files.
Independent review and repair
Reported selected scenarios exercise independent review of a pinned head and bounded repasses for reviewer, local-check and provider-CI failures on the same pull request. These cases do not establish general PR remediation. Website authorship is separate from protected verification and publication; this content update does not mean those later stages have run.
Exact-head merge review
A distinct workflow selects eligible pull requests, holds a PR-specific lease, refreshes approval, head, base, labels and CI, then merges the reviewed head and closes linked issues.
Authored workflows and scheduling
Selected Goobers DSL 2 workflows have run through normal service commands. Reviewed source describes sequential tasks, bounded read-only branches, joins, repasses and selected schedules; newer command/script and graph profiles remain unqualified for runtime execution. Captured workflow inputs stay distinct from changes admitted for future runs.
Repository providers
GitHub and Gitea paths cover issue and pull-request discovery, labels, comments, branch publication, CI observation, merges and closeout in supported profiles. Controlled cases exercise selected CI polling, pagination and branch publication; this does not establish every live provider profile.
Hewbers preserves Goobers workflow contracts and operational ideas while implementing the runner, ownership and recovery model in Hew.
That makes Hewbers both a practical self-hosted coding system and a substantial concurrent application for Hew. The goal is broad workflow parity through complete operating outcomes, not a line-for-line port.
Meet the Hew language ↗05 / Direction
Reliable delivery, remote execution and a packaged release.
Restore backlog progress and verify newer Hew compatibility, then deepen remote stages, human-feedback repair, Nisshi event history, NATS communication and an install that does not require a checkout.
Read the current development direction →-
Now
Restore reliable local delivery
Resume interrupted backlog work and demonstrate repeated reviewed changes under clear operating authority.
-
Now
Complete remote stage operation
Run checked image stages on disposable Kubernetes workers with scoped identity, recovery and owner-validated completion.
-
Next
Close the human feedback loop
Turn fresh pull-request comments and explicit decisions into bounded repair work.
-
Next
Durable events and inter-agent transport
Use Nisshi for lifecycle events and NATS for authenticated communication between isolated agents.
-
Then
Package the operating system
Install, configure, upgrade, restore and observe Hewbers without a source checkout.
-
Ongoing
Broaden Goobers workflow parity
Run more authored workflows, providers, controls and execution profiles through one Hew runtime.