Agentic AI Delivery · ADR-Governed · Release-Telemetry Backed
Claude Code + Codex · ADR-driven delivery · Release telemetry you can curl
This approach uses repository instructions, architecture decisions and explicit verification steps around AI-assisted changes. Build scripts and release records show what was checked for a particular revision; they do not guarantee correctness or regulatory compliance.
Agentic workflows need bounded tasks, human review and tested release gates. The repository documents this approach in scripts/build_release.sh, docs/cicd/pipeline.md and the architecture records. Source configuration is not evidence that every production deployment ran every check.
Three standard shapes, each delivered against named artifacts. No open-ended retainers without a named deliverable.
The referenced repository records describe the method. They do not establish independent validation, client outcomes or universal release compliance.
Related pages explain the approach and its scope.
How subagents, hooks, and auto-approvals combine into a repeatable delivery chain. Built around the CLAUDE.md and auto-approvals docs in this repo.
Where Codex fits and where it does not. Choice heuristics from real delivery use — not benchmarks and not leaderboards.
Research and applied documents that anchor the systems-first and verification-first claims this page depends on.
Bring your repository, your gate expectations, and your regulatory context. The first conversation defines what we inspect, what we report, and who owns what after the handoff.