Orbi
Turns GitHub Issues into independently reviewed code, merged changes, and tagged releases.
Per-dimension scores and reasoning
The evidence shows isolated worktrees, a frozen base SHA, independent review, reviewed-head merge gates, and explicitly scoped CI permissions. External effects and data movement through Issues, PRs, labels, comments, logs, branches, and release tags are described clearly, while prior-run evidence is preserved. Deductions apply because normal operation is fully automated: the ai-ready label supplies initial authorization, but merge and release actions have no per-step confirmation. The supplied files do not expose enough runtime implementation to audit every credential, redaction, and permission path; Actions use major-version tags rather than immutable commits; and explicit recovery for a bad merge or release is limited. Authorship and licensing are attributed, but the stated publisher remains unverified and the responsible legal organization cannot be established.
The README, package metadata, workflows, and test fakes consistently describe the same Issue-backed, isolated-development, review-gated delivery model. The fakes fail loudly on unsupported commands, objects, fields, and qualifiers and preserve useful exit-code and stderr distinctions. Dependency availability is reduced because operation relies on GitHub, gh, uv, Pi, a model provider, user-session scheduling, and the relatively new Python 3.14; readiness checks are documented, but these files cannot show durable availability across ordinary environments. Failure reporting and escalation to ai-blocked are documented thoroughly, supporting full credit for failure messages.
The material identifies bootstrap and external single-repository modes, Linux systemd and macOS launchd operation, and development, review, release, status, session, and diagnostic scenarios. Boundaries are explicit: the Agent stops at a commit, the Runner performs delivery, AI does not push protected branches, and unrecoverable cases require human action. Labels, workflow events, branch identity, and SHA gates make triggers precise. Environment fit is reduced because real-hardware macOS operation is expressly unverified, Ubuntu 24.04 defaults do not satisfy the Python or gh requirements, and Windows support is neither provided nor discussed.
The README has strong navigation across quick start, workflow, features, documentation, development, and licensing. Installation notes are detailed, and the relationship among the Orbi brand, orbi-cli distribution, and orbi command is explicitly stabilized. The complete license and practical commercial-use boundaries are present. Deductions apply because there is no dedicated FAQ content and the principal example is one showcased loop; changelog information is delegated to GitHub Releases without an included history or release-policy text; and maintenance is linked to an author, repository, and Discussions but lacks a defined team, response expectations, security contact, or succession path.
The intended outputs are directly actionable: commits, PRs, review fixes, merge records, tagged releases, Issue comments, and run_id-linked logs form an operational delivery chain, supplemented by queue, session, and diagnostic commands. This offers plausible value over manually coordinating the same tools, but the supplied evidence is chiefly project-authored description and test scaffolding rather than comparative measurements or independent cases, so marginal value is not fully established. Cost-benefit is also reduced because benefits are clear while the deployment and operating costs of Python 3.14, gh, uv, Pi, provider credentials, scheduling, and five-minute ticks are not quantified or systematically compared.
The files make extensive references to concrete Issue numbers, commands, responsibilities, SHAs, PRs, and release mechanics. Product claims are corroborated across the README, package metadata, CI permissions, macOS workflow, and GitHub/Git adapter fakes, earning strong traceability and cross-source corroboration. Fact/inference separation is reduced because comments assert matters such as verification against live APIs, real REST shapes, and complete records without supplying all corresponding tests, outputs, or external record contents here. The README does use cautious wording such as “claims” and discloses the macOS boundary, but not every strong assertion is labeled with an equally clear evidence level.
- After an ai-ready label triggers work, the system can autonomously change repository state, push branches, create and comment on PRs, merge code, and publish tags. Validate branch protection, token scopes, release authority, and emergency-stop behavior in a constrained test repository before deployment.
- This was a static review of only the supplied files. No tests, installer, GitHub API operations, full Issue-to-release journey, or real-hardware macOS deployment were executed.
- The Sustainable Use License is not a conventional permissive open-source license and restricts commercial distribution and selling Orbi as a hosted service or embedded paid product. Review the full terms and obtain authorization before commercial integration.
- Operation depends on Python 3.14, a sufficiently recent GitHub CLI, uv, Pi, a model provider, and a systemd or launchd user session. Confirm target-platform support, credential storage, spending limits, and dependency availability.
- Workflows reference actions/checkout@v5 and actions/setup-python@v6 by major-version tags rather than immutable commits. High-assurance deployments should assess stronger supply-chain pinning.
What does this agent do, and when should you use it?
Orbi is a self-hosted software-delivery system that claims `ai-ready` GitHub Issues and develops each one inside an isolated Git worktree. A Pi development session plans, implements, tests, and verifies the change; the runner then updates the base, pushes the branch, and opens a pull request containing `Fixes #N`. A separate review session examines the PR and fixes its findings, while the merge gate accepts only the reviewed head. Release Issues can freeze a specific SHA and publish a tag, so the resulting artifacts include commits, pull requests, merges, and releases. GitHub Issues, comments, PRs, CI, and releases are the sole state ledger—there is no database, queue, or resident daemon—and scheduled ticks run through systemd on Linux or launchd on macOS. It fits teams prepared to center delivery on GitHub and operate the required Python 3.14, Pi, GitHub CLI, and self-hosted scheduler stack.
Orbi claims a GitHub Issue carrying the ai-ready label, records the current origin/main SHA, and creates a feature branch plus an isolated worktree. The Pi session follows plan → implement → test → verify and stops at a commit. orbi.runner then synchronizes the latest base, pushes the branch, and opens a PR whose body includes Fixes #N. An independent review session evaluates that PR and addresses findings within the same session; the merge gate allows only the reviewed head to merge, and AI does not push protected branches. Recoverable failures return to the same PR for further work, while unrecoverable failures label the Issue ai-blocked for human action. Each run ties its branch, worktree, logs, and PR to one run_id; retries create a new run without replacing earlier evidence. A Release Issue can freeze a SHA and publish a tag. The CLI also provides orbi add for dispatch, status for the queue, session for following Pi, install-units for scheduler installation, and doctor for read-only diagnostics.
- A small engineering team already managing work in GitHub Issues wants well-scoped Issues converted into reviewed pull requests automatically.
- A platform team operating a self-hosted delivery pipeline needs each task isolated in its own worktree with traceable branches, logs, PRs, and
run_idvalues. - A maintainer wants an independent session to review generated changes, repair findings, and prevent merging a head that was not reviewed.
- A release team uses GitHub Issues to coordinate versions and wants a release task to freeze an exact SHA and publish a corresponding tag.
- A GitHub-centered project wants automation without maintaining a separate task database or message queue and accepts GitHub as the system of record.
What are this agent's strengths and limitations?
- Issues, comments, PRs, CI, and releases serve as both workflow state and audit trail, avoiding a separate database, queue, or task system.
- Every task receives an isolated worktree, and its branch, logs, and PR share one
run_id; retries preserve prior-run evidence. - The independent review session, reviewed-head merge gate, and protected-branch restriction provide concrete controls beyond the development session's own checks.
- Failure handling is explicit: recoverable cases continue on the PR, unrecoverable cases become
ai-blocked, and command errors remain visible in logs. - Operational commands cover idempotent setup, scheduler-unit installation, queue inspection, session following, and read-only health diagnostics.
- GitHub Issues are the only state store, and no GitLab, Bitbucket, or generic tracker adapter is documented, so non-GitHub teams would need to change their workflow.
- Python 3.14 or newer is mandatory; the Python 3.12 supplied by Ubuntu 24.04 is insufficient and requires uv to provision another interpreter.
- Operation depends on Pi and its provider, GitHub CLI 2.94+, GitHub authentication, Git, and a working user-level scheduler session—not merely the Python package.
- The documented macOS launchd path is explicitly described as not yet verified on real hardware.
- The Sustainable Use License permits internal use and self-hosting, but selling hosted Orbi or embedding it in a paid product requires commercial authorization.
How do you install or deploy this agent?
Prerequisites are Python 3.14 or newer, uv, Pi with its provider configured, GitHub CLI 2.94 or newer, Git, an authenticated GitHub account, and either a Linux systemd user session or a macOS launchd GUI session. Run:
git clone https://github.com/orbi-build/orbi.git && cd orbi
uv tool install --force --reinstall --editable --python python3 .If the system Python is older than 3.14, replace the final argument with --python 3.14 so uv provisions a compatible interpreter. For the CLI alone, use uv tool install orbi-cli; on an older system Python, use uv tool install orbi-cli --python 3.14. Check the environment with uv --version, pi --version, pi --print "reply with the single word: ok", gh auth login, and gh auth status. On Linux also run systemctl --user status; on macOS run launchctl print gui/$(id -u).
How do you use this agent?
Create a configuration and verify one deployment tick:
cp src/orbi/example_config.toml orbi.toml
orbi setup --config orbi.toml
PYTHONPATH=src python3 -m orbi.runner --config orbi.toml
orbi doctor --config orbi.tomlorbi setup checks the existing GitHub CLI authentication, labels, systemd or launchd scheduler units, and checkout; it is documented as idempotent. The direct runner command executes one manual tick, while normal operation is triggered by the timer every five minutes. To make work eligible, label a GitHub Issue ai-ready. Use status to inspect the queue and session to follow the Pi session through the installed orbi CLI.
How does this agent compare with similar options?
Compared with automation systems that copy task state into a separate database, queue, or project-management service, Orbi uses GitHub Issues, comments, PRs, CI, and releases as its ledger. That removes additional state infrastructure but makes GitHub central to the design. Compared with a one-shot coding assistant, Orbi also covers scheduled task claiming, isolated worktrees, PR creation, independent review, merge gating, and tagged release publication.
FAQ
Is internal use or large-scale self-hosting free?
Does Orbi require a database, queue, or long-running daemon?
Can generated code be pushed directly to a protected branch?
What happens when a task fails?
ai-blocked for a human decision.Can it run with Ubuntu 24.04's default Python?
--python 3.14 to the uv installation command lets uv provision a compatible interpreter.