GSD workflow tackles AI context rot

This feels less like a prompt pack than a state machine for agentic coding, which is why it is more durable than vibe-based workflows. Moving memory out of chat and into files lets the agent reread, verify, and resume from disk; /gsd:map-codebase and the phase files rehydrate architecture, conventions, and open questions, with PROJECT.md, REQUIREMENTS.md, ROADMAP.md, STATE.md, and .planning/ carrying persistent state. The repo is deliberately runtime-agnostic, so the same workflow layer can travel across Claude Code, OpenCode, Gemini CLI, Codex, Copilot, Cursor, Windsurf, and Antigravity. Fresh subagent contexts, atomic task plans, and 200k-token implementation windows directly attack drift in long sessions, and verify-work plus milestone checks force proof before the next phase starts. The tradeoff is ceremony: great for multi-step builds and messy repos, probably overkill for small edits.