When coding with AI agents, we need the ability to roll back mistakes - both the conversation context AND the filesystem changes need to revert together.
Current state: We can manually revert files (git), and we can revert conversation context, but these are separate operations that can leave the agent's understanding out of sync with actual files.
Git and AI agent conversations face the same fundamental challenge: managing branching threads of work that must be merged back together.
The key insight: Think of git as if it only had commit messages, no file changes. This is exactly what AI conversation management needs - a way to branch, merge, and rebase pure conversation history. The vocabulary and mental models from git can illuminate context management for AI agents.
Effective AI agent performance depends on maintaining high signal-to-noise ratio in the context window. Even with large contexts, noise degrades performance.
Claude Code currently offers only "Compact" - a nuclear option that rebuilds the entire conversation from scratch. This is like having only git rebase -i --root when you often just need git rebase -i HEAD~5.
This design makes Claude Code more responsive by allowing the main agent to continue working while background tasks run. Currently, when Claude spawns multiple long-running tasks, everything freezes - the agent can't continue working on other things, and you can't interact with it. With this change, Claude can process results as they arrive, continue working on independent tasks, and remain available for user interaction.
To understand the problem and solution, you need to know how Claude Code works today.
Or: The Empty Error String That Led Me Down a Rabbit Hole of Broken Networks
TL;DR: httpx AsyncClient was failing with empty error messages. After hours of debugging, I discovered a hotel WiFi firewall that accepts IPv6 TCP connections but kills them during TLS handshake - an edge case so specific it defeats Happy Eyeballs and every other protection mechanism.
It started with the most frustrating error message I've ever seen:
{"detail":"An error occurred while requesting TikAPI: "}
| #!/usr/bin/env python3 | |
| """ | |
| Demonstrate how unusual TikAPI's IPv6 behavior is compared to other services. | |
| """ | |
| import asyncio | |
| import socket | |
| import ssl | |
| import time |
| #!/usr/bin/env python3 | |
| """ | |
| Check why httpx AsyncClient works on some machines but not others | |
| This will help identify environment differences | |
| """ | |
| import socket | |
| import subprocess | |
| import platform | |
| import asyncio |
| #!/usr/bin/env python3 | |
| """ | |
| Phase 1: Network-Level Deep Dive - DNS Resolution Testing | |
| Tests DNS resolution behavior and IPv4/IPv6 connectivity | |
| """ | |
| import asyncio | |
| import httpx | |
| import socket | |
| import ssl |
| #!/usr/bin/env python3 | |
| """ | |
| Phase 1: Network-Level Deep Dive - Packet Capture Analysis | |
| Captures and compares network behavior between sync and async httpx clients | |
| """ | |
| import asyncio | |
| import httpx | |
| import socket | |
| import ssl |
| #!/usr/bin/env python3 | |
| import asyncio | |
| import aiohttp | |
| import json | |
| async def test_tikapi_aiohttp(): | |
| """Test TikAPI using aiohttp instead of httpx.""" | |
| print("🧪 Testing TikAPI with aiohttp...") |