NOW you can let AI agents make code changes without losing engineering control.
RunWield runs entirely on your machine. Use any model subscription or API key. No vendor lock-in.
curl -fsSL https://raw.githubusercontent.com/RunWield/runwield/main/install.sh | bashFrom chat and hope, to plan and prove.
Finally, a coding harness optimized for more than just getting an agent spitting out code as fast as possible.
RunWield — Plan-by-Default HarnessProject context loaded. What would you like to build?
Add CSV export to the orders page. Let people export just the orders they’re looking at.
AGENTS.md · project context · orders page · export utilities
Export filtered orders → opening Plan Review
Plan ReviewExport filtered orders
Objective
Let admins download a CSV of all orders matching the current filters. Exclude customer email addresses.
Approach
- Reuse the existing admin guard and orders query.
- Include every matching order, across all pages.
- Add an Export CSV button to the orders toolbar.
Verification
Test admin access, active filters, pagination, CSV escaping, and empty results. Run project CI and AI review.
export.ts · csv.ts · OrdersPage.tsx · csv.test.ts
Plan ReviewExport filtered orders
Objective
Let admins download a CSV of all orders matching the current filters. Exclude customer email addresses.
Approach
- Reuse the existing admin guard and orders query.
- Include every matching order, across all pages.
- Add an Export CSV button to the orders toolbar.
Verification
Test admin access, active filters, pagination, CSV escaping, and empty results. Run project CI and AI review.
export.ts · csv.ts · OrdersPage.tsx · csv.test.ts
Plan ReviewExport filtered orders
Objective
Let admins download a CSV of all orders matching the current filters. Exclude customer email addresses.
Approach
- Reuse the existing admin guard and orders query.
- Include every matching order, across all pages.
- Add an Export CSV button to the orders toolbar.
- Disable export when no orders match. Explain why beside the button.
Verification
Test admin access, active filters, pagination, CSV escaping, and empty results. Run project CI and AI review.
export.ts · csv.ts · OrdersPage.tsx · csv.test.ts
Export filtered orders · Autonomous execution
orders-csv-export · isolated from your working checkout
src/orders/export.ts + export filtered orders src/orders/csv.ts + serialize CSV safely src/orders/OrdersPage.tsx + Export CSV button
Admin access · filters · pagination · empty results
Changes and tests committed in the worktree. Running project CI.
bun run ci
Type check · lint · authorization · filter coverage
csv.test.ts › escapes quotes in order notes Expected: "A ""quoted"" note" Received: "A "quoted" note"
bun run ci
All project checks passed. Starting independent AI review.
CSV fields beginning with = can run as formulas in spreadsheet apps. Quoting alone does not neutralize them.
All checks passed after repair.
Finding resolved. No remaining blocking findings.
Your turn to inspect the change.
Code Review 1 export function csvCell(value: string) { 2 − return `"${value}"`; 2 + const safe = /^[=+@\-\t\r\n]/.test(value) 3 + ? `'${value}` : value; 4 + return `"${safe.replaceAll('"', '""')}"`; 5 } 40 <ExportButton 41 + disabled={matchingOrders === 0} 42 + filters={activeFilters} 43 /> 44 + {matchingOrders === 0 && ( 45 + <p>No matching orders to export</p> 46 + )}Reviewed: permissions, filter query, CSV safety, and empty-state behavior.
A shipped change.
A better starting point.
Filtered CSV export is on main. The next Plan can build on what this run learned.
- Decisions
- Admins only. All matching orders. No emails.
- Validation
- CI, AI review, and code review passed.
- Lesson
- CSV quoting and formula safety need separate tests.
Describe the change. RunWield brings in the Planner before any code is changed.
Key features
Any model. Subscription or API key.
Use the models you want with the subscriptions or API keys you already have. Switch providers as your needs change. RunWield uses pi-ai, the same library that powers the Pi Coding Agent.
All local. Yours to control.
RunWield runs on your machine. Your Plans, project context, memory, and Work Records live alongside your work.
Review the Plan before the code.
Shape the approach, add context, and choose how the agent should work before execution begins.
Validation built into delivery.
Run project checks, get an independent AI review, and inspect the diff. Findings feed directly into repair.
Every change informs the next.
Carry decisions and lessons forward through memory and Work Records, giving the next Plan a better starting point.
Make the harness fit your workflow.
Work from your terminal, browser, IDE, or chat. Extend RunWield with skills, MCP servers, and prompt templates.
Engineers who want more from agents than faster code generation.
- 01 Experienced developers working in real, existing codebases
- 02 Teams learning where agents should collaborate and where they can run
- 03 Anyone turning ambiguous product or engineering work into validated change
Build faster. Keep the understanding.
Steve Krouse’s “Vibe code is legacy code” captures the concern that led me to build RunWield: it’s easy to accumulate code faster than you understand it.
I built RunWield around reviewing the approach, checking the result, and carrying decisions and lessons into the next change. The aim is for your understanding to grow alongside your codebase.
Planning gets better when everyone can point at the same thing.
A Plan is not ceremony. It is the shared working surface where intent becomes specific enough for agents to act without guessing.
The agent said it's done. RunWield delivers proof.
- 01Mechanical validation
Tests, types, lint, formatting, and project-specific checks run automatically. Failures go back through automatic repair.
- 02Automated LLM reviewer
An independent reviewer compares the implementation with the Plan. Material findings trigger another repair pass.
- 03Human review, guided
Inspect the diff, follow review cues, annotate exact lines, and approve only when the evidence supports the result.
Help shape a better way to build with agents.
We're looking for experienced developers using AI on real codebases.