dev-flow — from idea to production

A battle-tested loop for shipping with an AI agent: a feature reaches production in roughly 40 minutes to 2 hours. The spec is plain markdown written via the KB skills, the work happens in an isolated git worktree, and only green reaches prod. The KB (see kb-curate) is the carrier of knowledge — onboarding, hand-off and scaling all start from it, not from the code.

The loop

1. Research — understand from facts, not guesses: read logs, traces, and the code itself. Reproduce before fixing.
2. Tasks — turn the research into a list of tracked tasks with dependencies. Nothing gets lost.
3. Goal — crystallize one clear goal and the "done" criterion from those tasks.
4. Spec — write it as markdown in the KB via kb-curate (ontology: node_type, frontmatter, typed links links.documents: [src/…]). The spec is durable knowledge, not a throwaway ticket — searchable, linkable, graphable.
5. Isolate — work in a dedicated git worktree: main stays untouched, parallel agents don't collide, and rollback is just dropping the worktree.
6. Implement — code to the spec inside the worktree; keep doc↔code linked (implemented_by).
7. Tests — write/adjust unit + E2E for the feature. The test is part of the feature, not an afterthought.
8. Independent review — run an independent model (e.g. Codex CLI, read-only) over the diff for logic and security bugs before rollout. A second model catches what the author's model misses — on a real production codebase this pass caught 191 bugs before they reached prod.
9. Dev-tests — open an MR with the commits into the dev branch; run the full suite there. Red → fix in the worktree, don't merge.
10. Prod-tests — E2E/smoke against the real prod contour, not only mocks or dev. Verify behaviour where users live.
11. Ship — merge dev → main and deploy (build the new image before stopping the old container, then poll the healthcheck to measure real downtime).

Git-flow

git worktree  →  MR + commits  →  dev branch (dev-tests)  →  main (prod-tests + deploy)  →  prod

Principles

• Spec = markdown + skills. No ceremony, but ontological — so it stays searchable, linkable and graphable (kb-search / gitmark map).
• Worktree isolation by default — clean parallelism and clean rollback.
• Tests and independent review are gates, not afterthoughts.
• Verify on prod, not just dev.
• md + git = source of truth; everything derived (search index, graph) is regenerated.

When to apply

Starting any non-trivial feature or fix. For a one-line change you can collapse steps, but keep the gates (review + tests) — they are where the 191 bugs were caught.