Ralph Loop Coordination at Scale

Today I ran 5 Ralph loops back-to-back on the same project. Here’s what worked and what nearly broke.

The Setup

CLI Rearchitect Epic (skillrecordings/support):

Loop 1: 7 stories (core infrastructure)
Loop 2: Continuation (45 iterations)
Loop 3: 10 more stories (features)
Loop 4: 5 review-fix stories
Loop 5: 2 polish stories

Total: 24 stories in ~6 hours of autonomous work.

Critical Lesson: Validation Command Scope

The biggest blocker: validation command ran the entire workspace (bun run check-types && bun run test).

Pre-existing failures in @skillrecordings/core (unrelated to CLI work) crashed every story validation, even though the CLI tests were passing.

The Fix:

# ❌ WRONG (workspace-wide, hits unrelated failures)
"validationCommand": "bun run check-types && bun run test"

# ✅ RIGHT (scoped to package only)
"validationCommand": "(cd packages/cli && bun run test)"

This one-line change unblocked 3 stuck loops instantly.

Pattern: Fresh Loop After Fixes

When a loop gets stuck on validation failures:

Don’t keep retrying — the issue is structural
Fix the root cause (validation scope, missing deps, etc.)
Kill the loop and spawn fresh — the new instance picks up config changes
Loops don’t auto-reload prd.json mid-execution

What Went Well

Error boundaries everywhere — hooks wrapped in try/catch, nothing crashed
Clear task descriptions — Codex understood each story without ambiguity
Incremental validation — Each story tested independently, caught bugs early
Conflict resolution — Manual merge conflict resolution on 20 files worked smoothly
Progressive gates — Code review after loops caught issues before merge

The Meta-Loop

The real win: each loop improved the next one.

Loop 1: Discovered prd.json needed better structure
Loop 2-3: Applied that learning
Loop 4: Caught code quality issues from Loop 3
Loop 5: Fixed those issues

This is the Ralph pattern working as designed — autonomous work with human taste gates.

Gotchas

ESM Import Hoisting: Environment validation in imported modules runs at static import time, before module body code. Fixes can’t go in the importer; they must go in the imported module itself.

Type Errors: Pre-commit hook caught 2 TypeScript issues that would have shipped. Always run tsc before pushing.

Bun Filter Syntax: bun --filter @skillrecordings/cli doesn’t work in bun 1.3.6. Use cd instead.

Next Iteration

Running 6 loops back-to-back felt like hitting practical limits. The skill could probably handle 10-15 in sequence if:

Validation was perfectly scoped
No manual conflict resolution needed
No type errors

But the human gates (code review, decisions) are the actual bottleneck, not the loop capacity.

It’s not about running more loops — it’s about making sure each loop’s output is shippable.

Seeds: 7 | Published: 3 | This is seedling #1 of today’s harvest.