Bounded Codex to agy delegation, independently checked

codex-agy-worker is an open-source Agent Skill and command-line pipeline for delegating bounded coding tasks from Codex to Google’s Antigravity CLI (agy). It treats every worker report as an untrusted claim, independently checks Git scope, and runs only driver-owned verification commands before accepting a candidate.

The worker can propose changes. Only the evidence gate can accept them, and exit 0 means only that the configured scope and verification checks passed. It does not prove general correctness or security, and it does not commit, push, merge, or release anything.

New to evidence-gated delegation? Start with how to verify AI coding-agent changes without trusting the worker report, then run the offline proof below.

What it is for

Backfilling tests across many files.
Repeating the same bounded edit mechanically.
Read-only repository inventories that are independently spot-checked.
Diff reviews used as a second opinion, never as acceptance evidence.

Each edit job runs in one branch-backed Git worktree. The driver captures an immutable base commit, declares allowed paths, and supplies the verification commands before dispatch. The gate rejects undeclared files, outside-policy changes, malformed worker output, missing edits, failed verification, and verifier-created mutations.

Small runtime, explicit boundaries

The project uses Bash, Python 3, and git. It has no Node runtime dependency, MCP daemon, or background polling service. Model-tier recommendations are visible and advisory: they include rationale, driver-owned evidence, and relative cost impact but never change the selected model automatically.

Dispatches use the external Antigravity CLI and can send approved task text and repository content to Google/Gemini. Read the privacy disclosure before use.

See the evidence boundary in under a minute

From a reviewed clone, run the repository-only ./proof-demo.sh. It builds two private synthetic repositories, confirms one exact edit passes the maintained gate and one plausible scope mismatch is rejected, then cleans up. It is offline and does not run agy or alter the checkout. Its three-line output is starter evidence for those fixed cases only: it is not human review, candidate acceptance, general correctness, a security certification, a benchmark, or production proof.

Integrations and forks can also run the public v1 gate conformance kit. Its eleven versioned synthetic fixtures test exact gate exits, including ignored files, verifier mutation, mutable bases, and human-required routing. Passing is fixture compatibility only, not security certification or real-job proof; the supplied gate runs with the current user’s privileges and must be reviewed first.

For provider-independent release comparisons, the offline Benchmark v1 preregisters fixed synthetic tasks and caller-selected variants, then records only canonical gate receipts and completeness facts. It has no live provider mode, score, ranking, winner, route, or recommendation.

The Persona Evidence Registry v1 separately validates fixed shipped persona sources, mode restrictions, and exact public contract bindings. Its statuses are evidence levels, not trust labels, routing inputs, or acceptance authority.

The data-only Workload Profiles v1 lists fixed non-executable mode/persona/path-shape skeletons. Profiles never select a model or tier, load a repository, carry a command or authorization, dispatch work, or affect acceptance; the caller must still provide every executable and scope-bearing input.

Install and explore

Use the GitHub repository as the source of truth. Review the cloned commit—or the exact release tag selected from GitHub Releases—before installing the standalone Codex skill:

git clone https://github.com/cagdasyurekli/codex-agy-worker.git
cd codex-agy-worker
./install.sh

Checked-in repository files do not alter GitHub About fields, topics, homepage metadata, or social-preview settings. Those remain deliberate repository-owner actions.

Read the full documentation, follow the verification tutorial, review the planned roadmap, or inspect the source and offline tests. Roadmap entries are proposals, not current product capabilities.