name: rate-limit-recovery
description: >
Collects recent transcripts and logging information from agent platforms that were rate-limited mid-task.
Supports recovery from Codex, Claude Code, Pi, and Antigravity rate limits by gathering session data,
logs, and partial results to resume interrupted work.
allowed-tools: Bash, Read, Glob, Grep, Write
triggers:
- rate limit recovery
- recover from rate limit
- collect transcripts
- gather logs
- resume interrupted task
- rate limited session
metadata:
short-description: Recover from rate limits by collecting session data and logs

Rate Limit Recovery Skill

Recovers from rate limiting interruptions across multiple agent platforms by collecting recent transcripts, logs, and session data to resume interrupted tasks.

Supported Platforms

• OpenAI Codex: Collects from codex CLI sessions and sandbox logs
• Claude Code: Gathers from Claude Code workspace logs and session files
• Pi: Retrieves from Pi's episodic memory and session archives
• Antigravity: Collects from Antigravity sandbox logs and session data

Features

1. Automatic Platform Detection: Identifies which agent platform was interrupted
2. Session Data Collection: Gathers recent transcripts, logs, and partial results
3. Rate Limit Context: Captures error details and retry timing information
4. Recovery Summary: Provides structured overview of what was collected
5. Resume Guidance: Suggests next steps for continuing interrupted work

Usage

Basic Recovery

./run.sh recover  # Auto-detect platform and collect recent data

Platform-Specific Recovery

./run.sh recover --platform codex --session-id abc123
./run.sh recover --platform claude --workspace /path/to/project
./run.sh recover --platform pi --session-id recent
./run.sh recover --platform antigravity --task-id task456

Advanced Options

# Export to specific format
./run.sh recover --format json --output recovery_report.json

# Custom output location
./run.sh recover --format markdown --output /path/to/custom/report.md

Data Storage

Recovery data is stored in ~/.pi/rate-limit-recovery/ by default. This ensures:

• Consistent location across all projects
• Proper organization of recovery files
• Easy access for future reference

Recovery Data Structure

The skill collects and organizes data into these categories:

Session Context

• Recent conversation history and tool calls
• Partial results and intermediate outputs
• User inputs and agent responses

Error Information

• Rate limit error details (429 responses, quota info)
• Retry timing and backoff information
• Platform-specific error codes and messages

Log Files

• Platform-specific log locations and formats
• Recent activity timestamps and sequences
• Debug and verbose logging when available

System State

• Workspace and file system state at interruption
• Environment variables and configuration
• Running processes and background tasks

Integration with Other Skills

This skill works well with:

• memory: Store recovered session data for future reference
• episodic-archiver: Archive the recovery session for analysis
• task-monitor: Monitor recovery progress and retry attempts
• agent-inbox: Communicate recovery status to other agents

Platform-Specific Details

Codex Recovery

• Collects from ~/.codex/sessions/ and current workspace
• Gathers reasoning effort and model configuration
• Captures sandbox execution logs and tool outputs

Claude Code Recovery

• Retrieves from Claude Code workspace .claude/ directory
• Collects conversation history and context files
• Gathers Claude-specific configuration and settings

Pi Recovery

• Accesses Pi's episodic memory and session archives
• Collects from .pi/sessions/ and memory stores
• Gathers ArangoDB-backed conversation history

Antigravity Recovery

• Collects from Antigravity sandbox logs and session data
• Grows Google Cloud Code Assist integration logs
• Captures multi-model conversation context

Error Handling

The skill handles various failure scenarios:

• Missing or corrupted session files
• Inaccessible log directories
• Platform-specific authentication issues
• Network connectivity problems during recovery

Output Formats

Recovery data can be exported in multiple formats:

• JSON: Structured data for programmatic access
• Markdown: Human-readable report with sections
• Plain Text: Simple chronological log format
• HTML: Rich formatted report with navigation