Velme

I'm building a programming language in Rust with Claude Code. In 5 days: 4 phases, ~21.6k lines of Rust, 236 acceptance criteria tracked to tests. My first attempt ran everything in one long session. It hit context limits and the code got worse after each /compact. What fixed it: the main session only coordinates. It spawns subagents and reads their short reports, never diffs. The loop for each slice (a slice is about one commit): • Build. A slice-builder subagent gets the slice ID, plan text and the exact spec sections. The model and effort for each slice come from a progress table. • Review. A separate slice-reviewer (Opus, read-only) checks the uncommitted diff. • Fix. Findings go back to the same builder via SendMessage. It still has the context, so fixes are cheap. • Commit. The main session re-runs verification itself. It never trusts "all tests pass" from the builder. Before closing a phase, three reviews run one after another, each fully fixed before the next: architect review (Opus, high) → code review of the whole phase → /simplify. Running them in parallel caused conflicting edits. What the reviews caught after everything already passed tests: • A sandbox cost bug: comparing strings cost the same regardless of length, so generated code could burn huge CPU cheaply. • A cache key that left out some limits, so two different contracts could share cached code. • The artifact store followed symlinks and read unbounded files. • The writer accepted files the reader would refuse, so you could save something you couldn't load. None were caught by the agent that wrote the code. What it cost, in tokens (from my Claude Code transcripts): • ~460M processed; 97% cache reads. • Main session: 39%, even though it writes no code. Every turn re-reads the conversation. • Subagents: 61%. • ~115M per phase. • Opus ~90%, Sonnet ~10%, Haiku 0%. I planned Sonnet builders and Haiku for mechanical work, but kept bumping "hard" slices to Opus. Next phase I'm pushing more to Sonnet to see if review catches the difference. Lessons: • A separate reviewer beats self-review. • Brief agents with rule IDs (INV-4, AC-RDM-08) so they grep instead of reading whole files. • Add one line to every agent prompt: "If it's ambiguous or needs a judgment call, stop and report instead of guessing." • Measure your model mix; the plan and reality drift. What I'm building: Velme (https://github.com/velme-lang/velme), a language where an LLM writes the implementation but a deterministic compiler and sandbox decide whether it's accepted. It's the same idea as the workflow: the LLM never judges its own work. Happy to share the agent files and CLAUDE.md setup in the comments. How do others split work between models?