Claude Code bug: automatic git reset destroys uncommitted changes every 10 minutes

Bug Details
Claude Code performs git fetch origin + git reset --hard origin/main on the user's project repository every 10 minutes via programmatic git operations (no external git binary spawned). This silently destroys all uncommitted changes to tracked files. Untracked files survive, and git worktrees are immune to this behavior.
Environment and Evidence
Affected version: Claude Code 2.1.87 (Homebrew cask, compiled Bun binary)
OS: macOS 15.4 (Darwin 25.3.0, arm64)
Evidence from investigation:
- Git reflog shows 95+ entries at exact 10-minute intervals with entries like
reset: moving to origin/main - Live reproduction confirmed modified tracked files are silently reverted at the 10-minute mark
fswatchcaptured the file operations pattern matching git fetch + reset- Only the Claude Code CLI process (PID 70111,
claude --dangerously-skip-permissions) was found with CWD in the affected repo - No external git processes spawned - operations are programmatic within Claude Code (likely libgit2)
Technical Analysis
Binary analysis of the compiled binary at /opt/homebrew/Caskroom/claude-code/2.1.87/claude revealed:
hg1()function does["fetch","origin"]viat_(C8(), _)without explicit CWD, defaulting toprocess.cwd()io1()function is a git pull wrapper logginggit pull: cwd=${H} ref=${_??"default"}fileHistorystate tracks{snapshots: [], trackedFiles: new Set, snapshotSequence: 0}
What Was Ruled Out
The investigation eliminated: git hooks, Claude Code user hooks, plugin marketplace updater, macOS cloud sync (iCloud, Dropbox, Syncthing, Synology, Google Drive), cron/LaunchAgents, Vite/SvelteKit dev servers, IDEs/editors, Time Machine, and external file watchers.
📖 Read the full source: HN AI Agents
👀 See Also

Claude for Word Add-in Evidence Found in Analytics API
Anthropic's analytics API now returns metrics for Claude for Word alongside existing Excel and PowerPoint add-ins, indicating the Word integration is in development. The API shows zero usage counts for Word, suggesting it's not yet publicly available.

Agentic AI Failure Modes and Developmental Scaffolding
Agentic AI systems fail in production through alignment drift, context loss across handoffs, boundary violations, and coordination collapse. The source proposes a 'developmental scaffolding' approach with five components: coherence monitoring, coordination repair, consent and boundary awareness, relational continuity, and adaptive governance.

70% of devs say AI code has more vulns; 30% ship it anyway — Checkmarx survey
70% of developers believe AI-generated code has significantly more vulnerabilities, yet 30% knowingly ship vulnerable code into production. The Checkmarx survey of 2,350 respondents also finds 93% of orgs suffered security breaches from vulnerable apps.

Anthropic Launches Claude Code Channels for Messaging from Telegram or Discord
Anthropic has released Claude Code Channels, allowing developers to message their AI coding sessions from Telegram or Discord while keeping code local.