Roadmap
The next steps target failure diagnostics and decision quality rather than simply increasing worker counts or retry limits.
Private internal project. Features described here are evidence-limited and are not offered as a public execution service.
First: actionable feedback
Correct misleading retry messages and establish which visible-test failures give the worker useful information. Do not pass withheld acceptance answers or diagnostics into model prompts.
Next: bounded experiments
Run authorized same-task early-stop checks, then a representative paired benchmark containing tasks that fail visibly, tasks that fail only withheld checks, and tasks that should not be attempted.
Then: quality and cost
Compare held-out success at clearly declared attempt, time and token budgets. Verify signatures, source integrity and cleanup. Measure local compute, operator review, and external API fees separately.
After evidence: operator workflow
Improve the review experience so scope, provenance, risk, test results, costs and the proposed patch can be inspected before any apply or merge. Approval remains independent of model-generated text.
Not enabled
Unattended grants and deployments, privileged tool inheritance, and multi-node code execution require their own certification and are not current promised capabilities.