Recovering lost code

2026-02-19 · Source: Simon Willison's Weblog · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning · Depth: Intermediate, quick

Summary

A developer experienced a "parallel agent psychosis" event, losing a feature they had worked on the previous day, unable to locate it across various development environments like branches, worktrees, cloud instances, or checkouts. The lost code was eventually discovered to have been developed in a temporary directory (`/tmp`) and subsequently lost due to a computer crash and reboot. Fortunately, the code was recoverable from `~/.claude/projects/` session logs, with Claude Code capable of extracting and restoring the missing feature.

Key takeaway

For developers who frequently prototype or experiment, you should be aware that work in temporary directories is vulnerable to system reboots. If you lose code, check your `~/.claude/projects/` session logs, as Claude Code can often reconstruct lost features from these records, saving significant rework.

Key insights

Lost code can be recovered from session logs, even after system crashes.

Principles

• Temporary directories are volatile.
• Session logs can preserve work.

In practice

• Check `~/.claude/projects/` for lost code.
• Utilize Claude Code for recovery.

Topics

• Code Recovery
• Claude Code
• Session Logging
• Development Workflow

Best for: Machine Learning Engineer, NLP Engineer, Software Engineer, AI Engineer

Related on AIssential

• Your Code Remembers Where it Broke — Code execution can resume from a failure point after a fix, saving development and recovery time.
• Claude Code Leaked 512,000 Lines of Source. Here’s What the Community Found Inside. — A single `.npmignore` oversight exposed 512,000 lines of Claude Code's source, highlighting automation gaps.
• Using Claude Code: session management and 1M context — Effective context management in Claude Code is crucial for optimizing performance and mitigating "context rot."
• Write a script to pull latest code from all repos daily—and let Claude Code write it — Automating Git pulls across multiple repositories streamlines development workflows and ensures codebase currency.
• Breaking Down the .claude Folder — The `.claude` folder manages local state for Claude-powered tools, ensuring consistent and repeatable model interactions.
• The Secret Behind Claude Code’s Retrieval: Why Live Search Fits Better than RAG — Claude Code's retrieval uses live search with `ripgrep` and LSP, not persistent RAG, to explore dynamic codebases like an engineer.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.