bootstrap-monorepo

name	bootstrap-monorepo
description	Autonomous polyglot monorepo bootstrap meta-prompt on the moon + proto + Bun stack (Nx-convergent). TRIGGERS - new monorepo, new repository, polyglot setup, scaffold repo, moon proto bootstrap, monorepo from scratch.
allowed-tools	Read
disable-model-invocation	false

Bootstrap Polyglot Monorepo (moon + proto + Bun, Nx-convergent)

Canonical reference lives in THIS skill:

→ See: references/bootstrap-monorepo.md → Cross-language boundaries: references/cross-language-interop.md — boundary decision ladder + verified tool status (research-verified 2026-06)

Self-Evolving Skill: This skill improves through use. If instructions are wrong, parameters drifted, or a workaround was needed — fix this file immediately, don't defer. Only update for real, reproducible issues.

When to Use This Skill

Use this skill when:

• Starting ANY new repository — single-language or polyglot (the structure scales down)
• Setting up a moon-orchestrated, proto-pinned, Bun/TypeScript-wired monorepo
• Wiring engines (Rust/Python/Go) behind contracts under a TS control plane
• You want the repo Nx-convergent from day one (later nx init is mechanical, not a restructure)

Stack

Tool	Responsibility
proto	Toolchain versions (bun, python, rust, node, …) pinned in repo-local .prototools
moon	Project graph + task orchestration + caching + affected detection (moon ci)
Bun	TS runtime for every script, CLI, glue tool, and test; root workspaces
uv / cargo	Python / Rust engines invoked natively from moon script: tasks

TypeScript is the control plane; other languages are engines behind language-neutral contracts (JSON Schema 2020-12 / proto) with drift gates and parity tests.

Language-Selection Default (greenfield tiebreaker)

Choose which language owns the code before choosing how to cross between them. This is a tiebreaker, not a mandate — a SOTA-native ecosystem and existing repo convention override it.

• TypeScript by default for the control plane: app logic, orchestration, CLIs, APIs, config, validation, agent/workflow automation, and anything future agents must refactor. (TS is structurally typed by default — the most agent-navigable surface.)
• An engine language only where its ecosystem is SOTA-native: Python for ML/data/science/quant or Python-only libraries; Rust/Go for hot kernels and systems work (Go > Rust as the tiebreaker).
• Never let an engine language leak across the repo. A Python-only library does NOT make the project Python — wrap it behind ONE typed boundary (ladder below) and keep TS as the public interface. Validate at the boundary: TS types are compile-time only and Python hints are not runtime-enforced.

Crossing a Language Boundary (decision ladder)

When work must cross languages, climb from cheapest to most coupled — stop at the first rung that satisfies the need (full doctrine + verified tool status: references/cross-language-interop.md):

1. Process boundary — CLI + JSON/NDJSON on stdio (moon script: task). Default.
2. Schema-typed RPC — protobuf + buf / ConnectRPC, when a long-lived service or stream exists.
3. Data-plane — Apache Arrow IPC/Flight/ADBC, the moment two languages exchange tables.
4. In-process FFI — PyO3/maturin, napi-rs, Bun FFI, when call frequency makes 1–3 too slow.
5. One WASM component core — WIT + Component Model, ONLY if the kernel needs no threads (WASI threading still unshipped as of mid-2026).

Pattern B gate amendment: after the 8-float-op test selects Pattern B, ask "does the kernel need threads?" → if yes, native core (cdylib/PyO3), not WASM.

TypeScript ↔ Python (the most common pair) verified picks, June 2026: rung 1 = Bun.spawn/child_process running Python via uv run --python 3.14 (NOT bun:ffi — C-ABI only); rung 2 = FastAPI + @hey-api/openapi-ts (HTTP) or buf-generated protobuf-es + protocolbuffers/python stubs (RPC); schema SSoT = Pydantic v2 → JSON Schema 2020-12 → TS, with a CI drift gate. Full table in the reference.

Quick Commands

# After bootstrap:
moon ci                          # affected quality pipeline
moon run <project>:check         # one project's full gate (lint+test+drift)
moon query projects              # machine-readable project graph (agents read this)
moon query tasks                 # machine-readable task surface
proto use                        # install all .prototools pins on a fresh machine

Legacy Path

Pre-2026-06 repos on Pants + mise: the old reference remains at ../mise-tasks/references/bootstrap-monorepo.md. Migrate per-repo (parity-first, cut tasks over one at a time), never big-bang.

Related Skills

• itp:semantic-release - Release automation (local-first; Actions only for release/CodeQL/Dependabot/deploy)
• itp:mise-tasks / itp:mise-configuration - legacy mise-era orchestration (still valid for unmigrated repos)

---

Troubleshooting

Issue	Cause	Solution
proto::detect::failed for moon	tool installed but no version pinned	proto pin --to global moon <version> (or add to .prototools)
moon not found over ssh	shims not on non-interactive PATH	export PROTO_HOME + shims PATH in ~/.zshenv (NOT only .zshrc)
Python tests "No module named pytest"	uv run prunes dev extras	uv run --extra dev -p <version> pytest <path> from repo root
PyO3 crate cargo test link errors (_PyBool_Type)	tests reference #[pyfunction] under extension-module	keep logic in pure-Rust core fns; tests call the core, wrapper stays thin
Task runs in wrong cwd	script assumes repo root	options: { runFromWorkspaceRoot: true } in the task
Guard task wrongly cached	moon caches by default	options: { cache: false } on guards/parity/network tasks
Commit "passed" but didn't land	a hook auto-fixed files and aborted	re-stage and retry; ALWAYS verify git log --oneline -1 after commit

Post-Execution Reflection

After this skill completes, check before closing:

1. Did the command succeed? — If not, fix the instruction or error table that caused the failure.
2. Did parameters or output change? — If the underlying tool's interface drifted, update Usage examples and Parameters table to match.
3. Was a workaround needed? — If you had to improvise (different flags, extra steps), update this SKILL.md so the next invocation doesn't need the same workaround.

Only update if the issue is real and reproducible — not speculative.