Pi-Shunt

Cut your pi token usage by ~90% on large file reads and code generation

Port of Spotify's shunt plugin from Claude Code to pi.

Token optimization extension that routes I/O-heavy operations to cheaper worker models – read delegation saves input tokens, write delegation saves output tokens.

---

Quick Start

# Install globally
git clone https://github.com/oguzhaneren/pi-shunt.git ~/.pi/agent/extensions/pi-shunt
cd ~/.pi/agent/extensions/pi-shunt
npm install

# or
pi install npm:@oguzhaneren/pi-shunt

# Start pi – shunt is active immediately
pi

That's it! Shunt automatically intercepts large file reads and routes them to a cheaper worker model.

---

What It Does

• Intercepts read tool calls on files >350 lines
• Intercepts bash cat/head/tail on large files
• Delegates to Gemini Flash (or your chosen worker model)
• Returns structured summaries instead of full file content
• Saves 82-97% of tokens on large file reads
• Delegates boilerplate code generation (tests, stubs, config, docs) to the worker, writing directly to disk

Before (without shunt)

User: "What does Service.java do?"
Pi reads: 4,014 lines → 33,684 tokens

After (with shunt)

User: "What does Service.java do?"
Shunt blocks: "File is 4,014 lines, use shunt_read"
Pi calls: shunt_read --question "What does this service do?" --paths Service.java
Worker analyzes: File → Structured summary (5,737 tokens)
Savings: 82%

---

How It Works

Three layers, from hard enforcement to soft guidance:

1. Hooks – Intercept Read and Bash tool calls, block large file reads
2. Tools – shunt_read delegates file analysis, shunt_write delegates code generation
3. Worker – Fresh pi subprocess with cheaper model analyzes files and generates code

When pi tries to read a file >350 lines (configurable), shunt blocks it and suggests using shunt_read instead. Pi then calls that tool, which spawns a worker pi process to analyze the files and return a structured summary.

Architecture

┌─────────────────────────────────────────────────────────┐
│ Main Pi Session (Frontier Model)                        │
│                                                          │
│  User: "What does Service.java do?"                     │
│  ↓                                                       │
│  Pi attempts: read(Service.java)                        │
│  ↓                                                       │
│  Shunt intercepts: "File too large, use shunt_read"    │
│  ↓                                                       │
│  Pi calls: shunt_read(question, paths)                 │
│  ↓                                                       │
│  ┌─────────────────────────────────────────┐           │
│  │ Worker Pi Subprocess (Gemini Flash)     │           │
│  │                                          │           │
│  │  Receives: <file>...</file> + question  │           │
│  │  Analyzes: Code structure, patterns     │           │
│  │  Returns: Structured summary            │           │
│  └─────────────────────────────────────────┘           │
│  ↓                                                       │
│  Pi receives summary (not full file)                   │
│  ↓                                                       │
│  Pi responds: "This service..."                        │
└─────────────────────────────────────────────────────────┘

Code generation follows the same pattern:

Pi calls: shunt_write(spec, reference, target)
↓
Worker Pi Subprocess (write mode)
  Receives: spec + reference file patterns
  Generates: code matching the reference
  Writes: directly to target file on disk
↓
Pi receives: "Wrote 140 lines to UserService.test.ts [shunt: ... tokens]"
(Generated code never enters the main model's context)

---

Installation

Option 1: Global (all projects)

# Clone to pi extensions directory
git clone https://github.com/oguzhaneren/pi-shunt.git ~/.pi/agent/extensions/pi-shunt

# Install dependencies
cd ~/.pi/agent/extensions/pi-shunt
npm install

Option 2: Project-local

# In your project
git clone https://github.com/oguzhaneren/pi-shunt.git .pi/extensions/pi-shunt
cd .pi/extensions/pi-shunt
npm install

Option 3: Via npm

pi install npm:@oguzhaneren/pi-shunt

---

Configuration

All settings use SHUNT_* environment variables. Pi-shunt resolves them from four sources, highest priority first:

1. Real environment variables (export in your shell profile)
2. .pi/settings.json (project) env block
3. ~/.pi/agent/settings.json (global) env block
4. Built-in defaults

Example ~/.pi/agent/settings.json:

{
  "env": {
    "SHUNT_ENABLED": "true",
    "SHUNT_MIN_LINES": "350",
    "SHUNT_WORKER_MODEL": "gemini-2.5-flash",
    "SHUNT_WORKER_TEMPERATURE": "0.2"
  }
}

Note: pi 0.85.x does not apply the settings.json env block to its own process, so pi-shunt reads those files directly. Restart pi (or /reload) after editing settings.json. If a SHUNT_* var is exported in your shell, the extension sees it via process.env and it overrides settings.json.

Environment Variables

Variable	Default	Description
SHUNT_ENABLED	true	Enable/disable extension
SHUNT_MIN_LINES	350	Line count above which reads are blocked
SHUNT_MAX_PAYLOAD_BYTES	400000	Maximum request size
SHUNT_WORKER_MODEL	gemini-2.5-flash	Model to use for analysis
SHUNT_WORKER_TEMPERATURE	0.2	Temperature for worker model
SHUNT_WORKER_INSTRUCTIONS	(default)	Custom system prompt for worker
SHUNT_READ_ENABLED	true	Enable/disable read interception + shunt_read
SHUNT_WRITE_ENABLED	true	Enable/disable shunt_write
SHUNT_WRITE_INSTRUCTIONS	(default)	Custom system prompt for the write worker
SHUNT_CYCLE_KEY	ctrl+alt+s	Keybinding for the cycle shortcut

---

Usage

Automatic (recommended)

Just use pi normally. When it tries to read a large file, shunt intercepts:

You: "Explain how this service works"
Pi: I'll read the service file...
[Shunt blocks read]
Pi: I'll use shunt_read instead
[Worker analyzes file]
Pi: Based on the analysis, this service...

Manual

Call shunt_read directly:

/shunt_read --question "What does this service do?" --paths src/Service.java

Follow-up questions

Each call is independent. Ask again with the same files:

/shunt_read --question "Which methods call the database?" --paths src/Service.java

Re-sending files is free where it matters – they go to the worker, not the frontier model.

Code generation (shunt_write)

For boilerplate that is predictable from existing patterns, shunt_write generates the code with the worker and writes it directly to a file. The generated code never enters the frontier model's context – only a summary comes back.

Pi: I'll write tests for UserService following the existing patterns
Pi calls: shunt_write spec="Write unit tests for UserService..." reference=tests/OrderTest.ts target=tests/UserService.test.ts
Worker generates: code matching OrderTest.ts patterns
Wrote 140 lines to tests/UserService.test.ts   [shunt_write: ... tokens]

Best fits: tests, mocks, fixtures, config stubs, type stubs, docstrings.

Requirements:

• reference is required – context-free generation fits nothing in your project
• For follow-up calls, pass the file the previous call just generated as the reference
• Review the file after writing and make surgical edits for the ~5-20% that needs judgment

Do not use for: editing existing logic or architecture decisions – those need exact content in the frontier model's context.

---

Commands

Command	Effect
/shunt:config	Show current configuration (toggles, cycle key, env vars)
/shunt:toggle	Enable/disable all shunt behavior (hooks + tools)
/shunt:toggle:read	Enable/disable read interception + shunt_read only
/shunt:toggle:write	Enable/disable shunt_write only
/shunt:stats	Show session statistics (reads intercepted, tokens saved, etc.)
/shunt:stats:reset	Reset session statistics

When a tool is disabled, calling it returns a clear "disabled" message instead of delegating to the worker. With the master toggle off, both tools refuse and the hooks let all reads pass through.

All toggles report the resulting state, e.g. /shunt:toggle:read off while write is on reports only write (read off, write on). Possible states: all on, only read, only write, all off.

Cycle shortcut

ctrl+alt+s cycles through all four states: all on → only read → only write → all off. No reload needed – toggles apply to the next call. Rebind via SHUNT_CYCLE_KEY.

---

What Doesn't Get Delegated

Shunt knows when NOT to delegate:

• Targeted reads – read with offset/limit (pi already knows what it needs)
• Small files – Under threshold (delegation overhead > savings)
• Nonexistent files – Let read handle the error
• Piped commands – cat file | grep (targeted read)
• Redirections – cat file > out (not reading into context)

For editing, pi still needs to read the exact content. Shunt allows targeted reads (with offset/limit) to pass through for this reason.

---

Examples

Single large file

/shunt_read --question "What does SpotifyUri.java do?" --paths src/main/java/SpotifyUri.java

• Without shunt: 33,684 tokens
• With shunt: 5,737 tokens
• 82% savings

Source + test pair

/shunt_read --question "How do these work together?" --paths src/PromotionRuleRepository.java tests/PromotionRuleRepositoryTest.java

• Without shunt: 75,990 tokens
• With shunt: 4,148 tokens
• 94% savings

Multi-file analysis

/shunt_read --question "How is permission checking implemented across these handlers?" --paths src/handlers/*.java

• Without shunt: 16,221 tokens
• With shunt: 821 tokens
• 94% savings

---

Benchmarks

This repo's test fixtures (worker: qwen3.8-125b)

Scenario	Without Shunt	With Shunt	Savings
Single large file (602 lines)	12,006 tokens	379 tokens	96.8%
Multi-file cross-read (3 files)	12,687 tokens	384 tokens	97.0%
Source + test pair	681 tokens	424 tokens	38-70%*
Code generation	681 tokens	404 tokens	~41%*

Large file reads consistently save 96-97%. The small-pair numbers swing run to run because worker answer verbosity dominates on tiny corpora.

Run your own: npm run test:benchmark

---

Limitations

Can't delegate editing

Worker summaries don't include reliable line numbers. If pi needs to make edits, it has to read the specific section directly. Shunt allows targeted reads (with offset/limit) for this reason.

Can't delegate reasoning

Worker models find surface patterns but can miss subtle bugs. Frontier models are better at architecture decisions, debugging, and safety-critical analysis.

Latency

Each delegation spawns a pi subprocess: 10-30 seconds typical. Acceptable for large reads, counterproductive for small ones (hence the threshold).

Payload size limits

Request must fit in memory. Default max: 400KB. Split large batches if needed.

---

Development

Project Structure

pi-shunt/
├── extensions/              # Pi extension source code
│   ├── index.ts           # Main entry point (hooks, toggles, cycle)
│   ├── lib/               # Core logic (config, worker, stats, utils)
│   ├── tools/             # shunt_read + shunt_write tools
│   └── README.md          # Extension documentation
├── tests/                  # Unit tests, integration tests, benchmarks
│   ├── unit.test.ts       # Pure logic tests (no pi needed)
│   ├── integration.test.ts # Worker delegation tests (needs pi)
│   ├── benchmark.ts       # Token savings measurements
│   ├── fixtures/          # Test files (602-line websocket-handler, etc.)
│   └── README.md          # Test suite documentation
├── skills/                 # Pi skills (shunt-read, shunt-write)
├── shunt/                 # Original Claude Code plugin (reference)
└── README.md              # This file

Local Development

# Install dependencies
npm install

# Typecheck + tests
npm run typecheck
npm test

# Test locally
pi -e ./extensions/index.ts

# Or symlink
ln -s $(pwd)/extensions ~/.pi/agent/extensions/pi-shunt

Running Tests

# All tests (unit + integration, needs pi in PATH)
npm test

# Unit tests only (no pi needed)
npm run test:unit

# Integration tests (needs pi)
npm run test:integration

# Token savings benchmarks
npm run test:benchmark

See tests/README.md for details.

---

Troubleshooting

"Failed to spawn pi"

Ensure pi is in your PATH:

which pi

If not found, install it:

npm install -g @earendil-works/pi-coding-agent

Worker returns empty/wrong answers

Check worker model availability:

pi --model gemini-2.5-flash --message "test"

Try a different worker model:

{
  "env": {
    "SHUNT_WORKER_MODEL": "gemini-2.0-flash"
  }
}

Thinking-only responses: reasoning models sometimes answer with a thinking block and no text (the model has "decided to inspect the repo" but shunt workers run with --no-tools). The error now reports this explicitly, e.g. Worker returned no text. Worker only reasoned: "...". Write-mode instructions include "you have no tools, always output code" to prevent it; a custom worker model may need a similar nudge – set SHUNT_WRITE_INSTRUCTIONS / SHUNT_WORKER_INSTRUCTIONS.

File not found errors

Paths are resolved from pi's working directory. Use absolute paths or relative to cwd:

/shunt_read --question "..." --paths ./src/Service.java

---

Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Add tests for new functionality
4. Submit a pull request

See CONTRIBUTING.md for guidelines.

---

License

MIT

---

Credits

• Original shunt concept: Spotify
• Blog post: Portal by Spotify: Cut My Claude Code Token Usage by 90%

---

Links

• Pi Documentation
• Pi Extensions Guide
• GitHub Repository
• Report Issues