Intelligent routing layer for pi (coding agent) with multi-level failover and observability.

English | 简体中文

• Features
• Quick Start
• Commands
• How It Works
• Configuration Reference
• Performance
• Uninstall
• Architecture
• Further Reading
• License

• Auto Router mode — select router/auto and let pi-router handle all routing automatically
• Channel failover for the same model across different providers
• Model fallback with context transfer
• Smart routing by latency, capability, cost, or manual order
• Circuit breaker and cooldown protection
• Persistent sticky routing — remembers the last successful channel across restarts
• Decision logging, latency tracking, and health monitoring
• Footer status showing the active provider channel
• Interactive menus and tab completion for all commands
• Interactive configuration wizard
• Fast startup with lazy loading and caching

# From npm (after publication)
pi install npm:pi-router

# Or from GitHub
pi install git:github.com/jiangjilin/pi-router

# Or from local directory (for development)
pi install /path/to/pi-router

See INSTALL.md for detailed installation options.

Recommended: run the interactive wizard.

/router config wizard

The wizard walks through:

1. Routing strategy (channelFirst / custom)
2. Sort strategy (latency / capabilityFirst / cost / manual)
3. Auto-sync (enable / disable)
4. Health probe (10 minutes / disabled)
5. Sticky mode (enable / disable)
6. Optional channel order adjustment

Channel Classification: The wizard automatically classifies channels into three categories:

• 🔵 OAuth (Official) - Official API endpoints from AI providers (e.g., api.anthropic.com, api.deepseek.com)
• 🟡 Aggregator (Third-party) - Third-party aggregation services
• 🟢 Free (Local) - Local deployments and free services

Classification is based on:

• auth.json OAuth markers (type: "oauth")
• Official domain whitelist (40+ providers including Anthropic, OpenAI, Google, DeepSeek, Qwen, GLM, Kimi, and more)
• Local URL detection (localhost, 127.0.0.1)

Configuration is stored at:

~/.pi/agent/pi-router.json

Advanced users can edit the file directly. A companion file is also generated:

~/.pi/agent/pi-router.README.md

A reference example is available at examples/router.config.json.

In pi, select the Auto Router model:

/model router/auto

This routes all requests through pi-router automatically, using your configured strategy and model chain.

You can also select a specific model with routing:

/model router/your-model-id

pi-router will then:

• try channels in order (based on strategy)
• fail over on errors
• apply circuit breaker and cooldown rules
• optionally fall back to another model
• record health and latency information
• display the active channel in the footer (e.g., via anthropic)
• remember the last successful route for next time (sticky mode)

Run /router without arguments to open an interactive menu, or use tab completion.

/router config wizard    # Interactive configuration wizard
/router config order     # Adjust existing model/channel order only
/router config show      # Show current configuration
/router config reset     # Reset to default configuration

Shortcuts:

/router config w         # = wizard
/router config o         # = order
/router config s         # = show
/router config r         # = reset

/router status           # Show config summary
/router list             # List configured router models
/router explain          # Show failures, latency, health, circuits
/router decisions        # Show recent routing decisions
/router probes           # Show background health probe results
/router pricing          # Show per-channel pricing breakdown

/router sticky           # Show current sticky routing records
/router sticky clear     # Clear all sticky records (re-route from beginning)
/router sticky clear <m> # Clear sticky record for a specific model

/router sync             # Check models.json changes
/router sync accept      # Apply detected changes
/router diff             # Preview config differences

When you select router/auto, pi-router manages the full model chain:

User Request: router/auto
    ↓
Check sticky record → found "model-X@channel-Y"?
    ↓ yes                    ↓ no
Try sticky first         Follow strategy order
    ↓ fail                   ↓
Clear sticky, fall back to strategy order
    ↓
channelFirst: Model-A[ch1,ch2,ch3] → Model-B[ch1,ch2] → ...
custom:       explicit model@channel order from customOrder
    ↓
On success: update sticky record, stream response
    ↓
Footer shows: via <channel-name>

Example:

Try channels: Provider-A -> Provider-B -> Provider-C

Typical flow:

User Request: router/example-model
    ↓
Router intercepts
    ↓
Try Provider-A -> Provider-B -> Provider-C
- check cooldown
- check circuit breaker
- forward to real provider
- on error: record failure, try next
    ↓
If all channels fail, try fallback model
    ↓
Stream events back to user

Sticky mode remembers the last successful route and tries it first on next request:

Request 1: Provider-A succeeds → sticky = Provider-A
Request 2: Provider-A tried first → succeeds
Request 3: Provider-A fails → clear sticky, try Provider-B → succeeds → sticky = Provider-B
Request 4: Provider-B tried first
...
(Persists across pi restarts)

Use /router sticky clear to reset and re-route from the beginning.

Failures: 0 -> 1 -> 2 -> 3 -> 4 -> 5 (open)
    ↓
Block requests for a cooldown window
    ↓
Half-open: allow a test request
    ↓
Success -> closed
Failure -> open again

When switching models:

1. summary mode: summarize the conversation
2. sanitize incompatible context fields if needed
3. forward the adapted context to the fallback model

~/.pi/agent/pi-router.json

{
  "strategy": "channelFirst",
  "sortBy": "latency",
  "autoSync": true,
  "sticky": true,
  "healthProbe": {
    "enabled": false
  },
  "models": [
    {
      "id": "example-model",
      "channels": ["Provider-A", "Provider-B", "Provider-C"]
    }
  ]
}

{
  "strategy": "channelFirst",
  "auto": true,
  "sortBy": "latency",
  "autoSync": true,
  "sticky": true,
  "request": {
    "timeoutMs": 120000,
    "maxRetries": 0,
    "maxRetryDelayMs": 1000,
    "maxTokens": 32768
  },
  "footer": {
    "rightAlignRoute": true,
    "statusLine": true
  },
  "contextTransfer": "summary",
  "summaryModel": "optional-summary-model",
  "summaryPrompt": "optional custom prompt",
  "summaryMaxTokens": 2000,
  "failover": {
    "cooldownMs": 60000
  },
  "healthProbe": {
    "enabled": true,
    "intervalMs": 600000,
    "timeoutMs": 10000,
    "probeMessage": "ping"
  }
}

Footer defaults:

• footer.rightAlignRoute defaults to true: while router status is active, pi-router replaces pi's footer so route status can be right-aligned
• set footer.rightAlignRoute to false to keep pi's built-in footer layout
• footer.statusLine defaults to true: when the replacement footer is disabled, pi-router still shows a short built-in status item
• set footer.statusLine to false to suppress that fallback status item

Default behavior:

• summaryModel is not required
• if the current context still fits inside the selected target model's context window, pi-router skips summary generation and forwards the full context
• when a summary is needed and summaryModel is unset, pi-router first uses the target model for summary generation
• if AI summary generation fails, pi-router falls back to a plain text non-AI summary path
• summaryMaxTokens defaults to 2000

Optional dedicated summary AI configuration:

{
  "contextTransfer": "summary",
  "summaryModel": "cheap-summary-model",
  "summaryPrompt": "Summarize the conversation for model handoff.",
  "summaryMaxTokens": 2000
}

summaryModel supports either:

• model-id
• model-id@provider

{
  "id": "example-model",
  "channels": ["Provider-A", "Provider-B"],
  "sticky": true,
  "sortBy": "latency",
  "contextTransfer": "summary",
  "fallbackModels": [
    {
      "id": "fallback-model",
      "channels": ["Provider-A", "Provider-B"]
    }
  ],
  "failover": {
    "cooldownMs": 30000
  }
}

You can edit ~/.pi/agent/pi-router.json directly.

After editing:

/reload

or restart pi.

pi-router is optimized for fast startup with minimal overhead.

• smart file hash caching
• lazy loading of models.json
• deferred health probe startup
• reduced duplicate file I/O

See PERFORMANCE_OPTIMIZATION.md for details.

# npm install
pi remove npm:pi-router

# git install
pi remove git:github.com/jiangjilin/pi-router

# local install
pi remove /path/to/pi-router

Configuration files are not removed automatically.

rm -f ~/.pi/agent/pi-router.json ~/.pi/agent/pi-router.README.md

See ARCHITECTURE.md for detailed design documentation.

• INSTALL.md
• TESTING.md
• ARCHITECTURE.md
• CHANGELOG.md

MIT

Built with the pi coding agent extension API.