Stats
Actions
Available In
Tags
claude-replay
The undo button you forgot to ask for.
claude-replay reads the raw session JSONLs that Claude Code already wrote to disk and uses them as a deterministic event stream. It can rebuild the git history you forgot to make, recover a single file Claude deleted, resurrect an entire rm -rf'd project — even export every tool call as a structured event log for AI-agent research.
Byte-perfect. Zero dependencies. Works on any language, any project type.
Four reasons this exists
🚑 1. Rescue mode — you forgot to commit
git status is a wall of red. 165 uncommitted files across a week of sessions. You can't tell "the v2 fix you regret" from "the v1 you want to keep." Hand-writing 20 semantic commits from a snapshot means inventing history.
You don't have to. Claude already wrote it down. Every Edit, every Write, every Bash command was streamed to ~/.claude/projects/-{cwd}/*.jsonl (and subagents/*.jsonl for sub-agents). Replay them in timestamp order against the right baseline commit and you reproduce the exact byte sequence — including every intermediate state that was later overwritten.
$ /replay
↳ status: 1274 ops across 6 sessions
↳ plan: 20 commit boundaries from docs/*.md timestamps
↳ execute: built sandbox in 23s, byte-equal to working tree (0 diff)
↳ apply: real repo HEAD now: 1205576 (previous tagged as backup)
After: a clean git log --oneline. git revert one feature, git bisect to find a regression, push 20 PRs. Your forgotten week became 20 reviewable commits.
🪢 2. Lifeline — Claude (or you) deleted an important file
You ran rm in the wrong terminal. Claude refactored a file out of existence "as cleanup". A Bash one-liner went rogue. Whatever the reason, an important file is gone — and you wrote it through Claude.
$ claude-replay recover-file --file /home/me/repo/important.py --list-versions
=== /home/me/repo/important.py (10 ops) ===
2026-05-07T04:55:22 Write (content: 28375 bytes)
2026-05-07T04:57:11 Write (content: 26973 bytes)
2026-05-07T05:07:59 Edit (replace 48 chars)
...
$ claude-replay recover-file --file /home/me/repo/important.py --output-dir /tmp/restored
RECOVERED /tmp/restored/important.py (54874 bytes from 10 ops)
Recover the latest version, or any historical version with --at-ts 2026-05-06T14:00:00Z. Every version Claude ever wrote is reachable.
⚰️ 3. Back from the dead — your entire project is gone
Worst case: you rm -rf'd the project. Or moved a disk and forgot to copy. Or a script ate it. Whatever — the working tree is gone. But ~/.claude/projects/ is untouched.
If the project was built primarily through Claude, the JSONLs contain enough to reconstruct the entire file tree:
$ claude-replay recover-project \
--project-name -home-me-myapp \
--output-dir ~/recovered-myapp \
--git-init
Session dir: /home/me/.claude/projects/-home-me-myapp
Original project root (inferred): /home/me/myapp
Files with recoverable ops: 174
Recovered 174 files into ~/recovered-myapp
Initialized git repo at ~/recovered-myapp with one snapshot commit.
Then chain into commit-history reconstruction (use case #1):
cd ~/recovered-myapp
claude-replay execute --apply
The result is a fully-recovered project with semantic commit history. Information loss limited to: (a) files Claude never touched (typically initial scaffolding or vendored deps), (b) files whose first op was Edit rather than Write — Claude saw them but never wrote them from scratch.
🔬 4. Research mode — you study AI agent behavior
Claude Code's session JSONLs capture months of real human-agent collaboration on real codebases — including the messy parts: failed edits, sub-agent escalations, context compactions, mid-task pivots, multi-day workflows. claude-replay analyze is your structured tap.
$ claude-replay analyze --format stats | head -15
Total events: 2874
By origin: main=2280 subagent=594
