The Loop Architecture
How Dymont Labs frames work, ships, reviews, and learns
One workflow that scales across every project the lab touches — Coexo Memory, the wisdom site, sandbox.wtf, ops. Three modes, six stages inside the heaviest mode, and a clear set of settled defaults.
Three modes
Pick at the start of every meaningful chunk of work.
Sprint
Building or shipping a feature
Frame → Plan → Build → Verify → Review → Learn
Forage
Exploration, frontier scouting, research spikes
Read SOTA → run in disposable sandboxes → write findings to Coexo
Tend
Ops, upgrades, maintenance
Apply skill → verify → capture new patterns as new skills
Sprint mode — six stages
- 01
Frame — SPEC.md
One-pager per feature: goal, constraints, success criteria, blast radius. Lives in the repo, written by Claude, greenlit by Stephan. Trivial fixes skip it; multi-file work gets one. Reference patterns: GitHub Spec-Kit, Kiro IDE.
- 02
Plan
Plan mode for any work where the CTO might pick the wrong abstraction. Stephan greenlights or redirects. Skipped for routine work. The plan is the CTO's to defend; Stephan adjudicates scope, not technique.
- 03
Build — autonomy matched to the work
Interactive for fast feedback (UI, debugging). Bypass-and-run for mechanical work. Ralph loop for long autonomous runs (requires confirmation — see Autonomy Boundaries). Worktree fanout for truly independent subtasks.
- 04
Verify — hooks own this
Every repo of any size gets PostToolUse hooks (typecheck + lint on Edit/Write) and a Stop hook (full test suite must be green before turn ends). SessionStart hook injects project context. Process is the product, wired into the harness.
- 05
Review
/review for every non-trivial commit. /ultrareview for production-bound or hard-to-reverse changes (multi-agent cloud review). /security-review for auth, secrets, external surfaces. Solo dev means we manufacture adversarial review.
- 06
Learn
Session ends → transcript flushes to Coexo via /events/from-transcript. Lessons become procedural skills, decisions become semantic events, next session inherits. The self-improving loop, hard-wired.
Settled defaults
Decisions locked 2026-05-03. The CTO follows these without asking; Stephan can override per-task.
| Spec | SPEC.md per non-trivial feature in Coexo + Foxops |
| Plan | Plan mode when wrong abstraction is plausible; Stephan greenlights scope |
| Ralph loops | Never without explicit confirmation, regardless of stakes |
| Worktree fanout | Within-session technique; no per-launch confirmation |
| Verify | Hooks own typecheck/lint (PostToolUse) + tests (Stop) |
| Review | /review for non-trivial; /ultrareview for production / hard-to-reverse |
| Learn | Transcript → Coexo on session end; lessons promoted to skills |
Autonomy boundaries
Ralph loops require explicit confirmation
Long-running autonomous Claude runs (while true; do claude -p PROMPT.md; done) require Stephan's explicit go-ahead, regardless of stakes. They burn tokens and run for hours; he wants to know one is alive. Stephan can request one with "ralph this" or similar.