Pi Auto-Resume Extension

A Pi coding agent extension that automatically resumes when responses are interrupted by token limits, rate limits, or provider plan exhaustion.

The Problem

When working with AI coding assistants, your workflow gets interrupted by:

1. Token limits - Response truncated mid-generation because it hit max_tokens
2. Rate limits - HTTP 429 errors when you exceed API rate limits (Claude Pro, OpenAI, etc.)
3. Plan limits - Quota/billing exhaustion on subscription plans (Claude Max, OpenCode, etc.)
4. Incomplete tool calls - Output cut off mid-tool-call, leaving broken JSON

These interruptions break your flow and require manual intervention to continue.

The Solution

This extension handles all these cases automatically:

• Token truncation → Sends continuation prompt to finish the response
• Rate limits (429) → Waits for retry-after period, then retries
• Plan/quota errors → Notifies user to upgrade or switch provider
• Incomplete tool calls → Prompts model to complete the tool call

Features

Feature	Description
Token limit detection	Detects stopReason: "length" (max_tokens hit)
Rate limit handling	Detects HTTP 429 and rate limit error messages
Plan limit detection	Detects billing/quota exhaustion errors
Incomplete tool call detection	Detects cut-off tool calls with empty arguments
Exponential backoff	Rate limit retries use exponential backoff (1m, 2m, 4m...)
Retry-after support	Respects Retry-After headers from providers
Configurable limits	Max resumes, delays, and retry counts are configurable
System prompt guidance	Tells the model how to handle continuation gracefully
Manual control	/auto-resume command for status, enable/disable, reset

Installation

One-liner install (easiest)

Linux/macOS:

curl -fsSL https://raw.githubusercontent.com/kasaiarashi/pi-auto-resume/master/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/kasaiarashi/pi-auto-resume/master/install.ps1 | iex

Windows (batch): Download and run install.bat from Releases

Pi package install

pi install git:github.com/kasaiarashi/pi-auto-resume

Manual install

cp auto-resume-on-token-limit.ts ~/.pi/agent/extensions/

Pi will automatically load it on next startup.

Configuration

Add to ~/.pi/agent/settings.json (global) or .pi/settings.json (project):

{
  "autoResume": {
    "enabled": true,
    "maxResumes": 5,
    "delayMs": 1000,
    "continuePrompt": "Continue from where you left off. Do not repeat what you've already written.",
    "rateLimit": {
      "enabled": true,
      "maxRetries": 3,
      "baseDelayMs": 60000,
      "maxDelayMs": 3600000
    }
  }
}

Settings Reference

Setting	Default	Description
enabled	true	Enable/disable auto-resume
maxResumes	5	Max consecutive token limit continuations
delayMs	1000	Delay (ms) before sending continuation message
continuePrompt	(see above)	Prompt for truncated responses
rateLimit.enabled	true	Enable rate limit handling
rateLimit.maxRetries	3	Max rate limit retry attempts
rateLimit.baseDelayMs	60000	Base delay (1 min) for exponential backoff
rateLimit.maxDelayMs	3600000	Maximum delay (1 hour)

Commands

Command	Description
/auto-resume or /auto-resume status	Show current status
/auto-resume on or /auto-resume enable	Enable auto-resume
/auto-resume off or /auto-resume disable	Disable auto-resume
/auto-resume reset	Reset all counters

How It Works

Token Limit Handling

1. Detection: Listens for message_end with stopReason === "length"
2. Safety checks: Won't exceed maxResumes, checks agent is idle
3. Delay: Waits delayMs for turn to settle
4. Continuation: Sends follow-up message via sendUserMessage()
5. System prompt: Adds guidance telling model not to repeat content

Rate Limit Handling

1. Detection: HTTP 429 responses or rate limit error messages
2. Wait: Respects Retry-After header or calculates exponential backoff
3. Retry: Sends continue message after delay
4. Limits: Stops after maxRetries attempts

Plan Limit Handling

1. Detection: Error messages matching billing/quota patterns
2. Notification: Alerts user to upgrade plan or switch provider
3. No auto-retry: These require user action

Detected Error Patterns

Rate Limits (auto-retry)

• Anthropic: "rate limit", "too many requests", "overloaded", "capacity"
• OpenAI: "exceeded the rate", "exceeded your quota"
• OpenCode: "plan limit", "usage limit"
• Generic: "retry after", "slow down", "temporarily unavailable"

Plan Limits (user action required)

• "billing", "payment", "subscription"
• "quota exhausted", "plan exhausted"
• "upgrade plan", "out of credits"

Context Overflow (Pi's built-in compaction)

• "context", "too long", "prompt is too long"
• "maximum context", "request too large"

What About Context Overflow?

Pi already has built-in auto-compaction for context window overflow. This extension defers to Pi's built-in handling for those cases and focuses on:

1. Output truncation (max_tokens)
2. Provider rate limits (HTTP 429)
3. Plan/quota exhaustion (billing errors)
4. Incomplete tool calls (truncated JSON)

License

MIT