plumber

---

Your AI agent doesn't fix bugs. It patches them. Conditionals on top of conditionals, error handling for things that shouldn't happen, complexity added to fight existing complexity. Every ticket gets a bandage. The pipe keeps leaking.

Plumber fixes the pipe.

Benchmarks

5 tasks. Same model. Same prompt. Vanilla vs Plumber.

Task	LOC Δ	Conditions Δ	Token Δ
Request handler	-18%	-25%	~+1%
Retry with backoff	—	—	~+1%
Config loader	—	—	+1.1%
Multi-format export	+27%	-25%	+1.2%
Logging decorator	+30%	—	+1.2%
Average	+12%	-18%	~+1%

Plumber wrote more lines on average — not fewer. The real metric is conditions: -18% fewer branches across all tasks. Token cost is negligible (~1% overhead).

The biggest wins aren't in the table: sentinel variables eliminated, catch-all exceptions replaced with specific ones, hardcoded lists replaced with formulas, booleans parsed safely. See benchmarks/ for full case studies.

Before / after

The null crash

You report a null crash. Your agent adds three null checks.

def get_user_name(user):
    if user is None:
        return "Unknown"
    if not hasattr(user, 'name'):
        return "Unknown"
    if user.name is None:
        return "Unknown"
    return user.name

Plumber finds the unauthenticated route that should never have passed null here and kills it at the source.

# plumber: fixed the route guard — null shouldn't reach this function
def get_user_name(user):
    return user.name

---

The copy-paste guard

The same check ends up in every function because nobody fixed the caller.

def send_email(user):
    if not user.email_verified:
        return
    ...

def send_notification(user):
    if not user.email_verified:
        return
    ...

def send_sms(user):
    if not user.email_verified:
        return
    ...

Plumber puts one guard where it belongs and removes the rest.

# plumber: @require_verified_email on the route — unverified users never reach these
def send_email(user): ...
def send_notification(user): ...
def send_sms(user): ...

---

The phantom try/except

Your agent wraps everything in try/except just in case.

def get_config(key):
    try:
        return config[key]
    except KeyError:
        return None
    except TypeError:
        return None
    except AttributeError:
        return None

Plumber uses the one line the stdlib already provides.

def get_config(key):
    return config.get(key)

---

The isinstance chain

Your agent makes a function accept "anything" instead of fixing what calls it.

def process(data):
    if isinstance(data, str):
        data = json.loads(data)
    elif isinstance(data, bytes):
        data = json.loads(data.decode())
    elif isinstance(data, list):
        data = {"items": data}
    return transform(data)

Plumber normalizes at the boundary and keeps the function clean.

# plumber: deserialize at the API layer — process() only ever sees a dict
def process(data: dict):
    return transform(data)

How it works

Before touching any code, the agent runs this silently:

1. What is the actual problem?         (not the symptom)
2. Is this patch covering a design flaw?   → fix the design
3. Adding complexity to fight complexity?  → redesign
4. Handling cases that shouldn't exist?    → delete them
5. What can be removed?                → strip it
6. Simple or just compact?             short ≠ simple
7. Write the minimum. Refactor. Repeat.

Redesigns are marked plumber: so they read as intent, not accident.

Install

Claude Code

/plugin marketplace add akshatnerella/plumber
/plugin install plumber@plumber

For Codex, Cursor, Windsurf, Cline, Kiro, Copilot, OpenCode, and more — see INSTALLATION.md.

Commands

Command
/plumber [lite | full | ultra | off]	Set intensity
/plumber-diagnose	Show root cause breakdown before acting
/plumber-review	Flag patchwork in the current diff
/plumber-audit	Scan the whole repo for accumulated patches

/plumber ultra for when the codebase has wronged you personally.

---

MIT