02 / Projects
AI Development
An operational system that guides Jira work through governed Codex workflows, from readiness checks to independent evaluation and human hand-off.
AI inside a real delivery process
AI Development is not a chat box that emits code. It is an operational management system for AI-assisted software delivery. A Jira issue can be assigned to a configured AI account, checked for readiness, developed in an isolated environment, tested, independently evaluated, and returned for human acceptance.
From ticket to verifiable change
A signed Jira webhook first starts a correction window so accidental assignments can be undone before work begins. Management then confirms the live assignee and prevents concurrent work in the same repository. Codex evaluates the description, acceptance criteria, comments, and approved Confluence context. If information is missing, the workflow asks focused questions and returns the issue to its owner.
For a ready task, a Bash harness on the AI server prepares a reusable repository workspace, starts DDEV, imports approved mediated environment data when needed, and records the baseline. Codex develops on a feature branch. Tests and structured work summaries flow back to management. A separate evaluation phase reviews the diff against the intended merge branch; only accepted changes can be integrated. Code rejection, merge conflicts, missing information, and infrastructure failures have distinct bounded retry paths.
Trust boundaries instead of shared secrets
The TYPO3 management application is the trusted control plane. It stores encrypted configuration, ticket state, and logs; validates webhooks; starts phases over SSH; and mediates access to Jira, production data, and logs. The AI server is considered untrusted and holds no direct Jira, Confluence, production, or management-database credentials. Short-lived bearer and HMAC contracts expose only the callback required by the current run.
Protected branches remain unavailable to the agent. It may publish feature branches while management owns approved integration. Ansible Vault stores credentials, sensitive context is redacted, and production data reaches local environments only through configured management boundaries with environment and VPN controls.
Observable operation
The backend shows active, waiting, and processed tickets with phases, durations, outcomes, and costs. Every run has an opaque identifier so late callbacks cannot overwrite a newer attempt. Stale detection accounts for working hours, and operators can cancel active work deliberately. Failed Jira tasks can be investigated from the backend with sanitized context.
The system combines a TYPO3 13.4 management extension, Bash harness and Codex CLI on the AI server, DDEV/Docker workspaces, Jira and Atlassian MCP, SSH, Ansible, and structured HTTP callbacks. Unit, functional, and harness tests plus a completed production ticket lifecycle form the current evidence base.
AI Development operates as a production pilot. Its purpose is to make AI work visible, bounded, and repeatable—including the points where a person must decide.