SShortSingh.
Back to feed

Claude Code's /rewind Cannot Restore Files Deleted or Modified via Bash

0
·1 views

Anthropic's Claude Code includes a checkpointing feature that saves state before each user prompt, allowing rollback via the /rewind command across up to 100 recent session snapshots. However, official documentation explicitly states that checkpoints do not track files modified by bash commands, meaning deletions, moves, in-place edits, and shell redirects fall entirely outside the restore system. Additional limitations include no support for subagent edits, symlinked or hard-linked files, changes made outside the session, and snapshots that expire roughly 30 days after the last save. Anthropic itself notes that checkpoints are not a replacement for version control, and git does not fully close the gap either, since untracked files and agent-initiated shell commands remain unprotected. Four open-source tools — bashward, git-safepoint, stepback, and others — have emerged specifically to snapshot files before destructive bash commands that Claude Code's native checkpointing cannot cover.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

AI Desktop Tool Helps Surface Unanswered Questions from Meeting Recordings

OGAD (Off Grid AI Desktop) is a locally run AI tool that allows users to transcribe saved meeting recordings and identify questions that were left unanswered or deferred during the discussion. The tool processes audio files in formats such as MP3, WAV, and M4A using downloaded local models for transcription and text analysis, keeping all data off the cloud. Users can then categorize unresolved questions by status — such as deferred, partly answered, or no answer found — to guide more targeted follow-ups. The workflow is part of OGAD's free core app and is available on Mac, Windows, and Linux in beta. The approach is particularly useful for small teams who need to close gaps before a project or delivery deadline.

0
ProgrammingDEV Community ·

How separating queues and workers fixed silent failures in a headless AI server setup

A developer running always-on AI agents discovered their system appeared healthy while webhook jobs silently piled up unprocessed, with Docker reporting containers as live despite workers being effectively frozen. The root cause was mixing webhook intake, cron scheduling, and heavy LLM execution inside a single process — a setup that works locally but breaks under production load. The fix involved splitting ingress from execution using a Redis queue, routing all LLM calls through a single OpenAI-compatible endpoint, and replacing local storage with PostgreSQL and S3-compatible solutions. A critical additional requirement was making all jobs idempotent, since queue systems like Celery can redeliver tasks if a worker crashes before acknowledging completion, risking duplicate side effects such as double charges or repeated emails. Stripe's webhook system, which retries undelivered events for up to three days, was cited as a real-world example of why systems must track whether a given event has already been processed.

0
ProgrammingDEV Community ·

Developer Builds Issue-Conditioned Memory Recall for AI Customer Support Agent

A developer has shared a technical approach to making AI-powered customer support agents retain and apply memory across separate interactions. The system uses a tool called Hindsight to retrieve past support history — such as prior resolutions and known preferences — by anchoring recall to the customer's current issue rather than pulling broad account summaries. A practical example illustrates how a customer's earlier router power-adapter fix could inform a later, unrelated connectivity complaint, allowing the agent to skip redundant troubleshooting steps. The developer outlines three independent quality checks — scope, relevance, and correct use of recalled history — arguing that collapsing these into a single pass risks masking distinct failure modes. The post concludes that retrieval strategy is as much a product decision as a modelling one, with logging and evaluation recommended for production deployments.

0
ProgrammingDEV Community ·

SupportMind AI Agent Brings Persistent Memory to Customer Support Interactions

A developer built SupportMind, an AI-powered customer support agent, as part of the HackHyderabad 3.0 hackathon. The tool is designed to retain context from previous customer interactions, allowing it to recall past issues and successful solutions without requiring customers to repeat themselves. SupportMind maintains individual memory profiles for each customer and a shared playbook of proven fixes to ensure consistent responses. The project was built using Python, Streamlit, and Groq, with an LLM-based memory layer powered by Hindsight AI. The developer has open-sourced the project on GitHub and is seeking community feedback.

Claude Code's /rewind Cannot Restore Files Deleted or Modified via Bash · ShortSingh