5 posts
Left to taste, agents write two styles, collide on a file and deploy on the developer's word. sdlc names each agent on a board, gives each rule one owner and one gate, and ends development at qa-handoff.md: G10 will not deploy until QA's report passes, dated after the shipped commit's G8 run.
A quiet edit to an approved file leaves the design and plan built on it drifting. In sdlc its hash changes, so its gates go STALE along the dependency graph, with three ways out: reaffirm the diff, re-run, or a named person's override.
An AI review you cannot audit is a nod. A first sdlc run shows one that is not: it rejects 'fast' and 'usable', gets file paths only, and writes a log with the SHA-256 of every file it judged and relied on.
Your AI coding agent says done after grading its own work, and you find what nobody decided two weeks later. In sdlc the agent that writes a stage never approves it: a different subagent writes the review log, and the next command refuses to start without its PASS.
sdlc and stlc meet at one file out and one report back: G9 writes qa-handoff.md with the REQ and AC ids, stlc returns a QA report whose overall verdict is computed from its gate lines, and G10 refuses to deploy without a passing report dated after the shipped commit's G8 run.
