backtrack

A read-only desktop app for re-reading your old Claude Code and Codex sessions. It shows only the top-level conversation — what you said and what the agent said back — with tool calls, results, thinking, and system noise stripped out. Sessions are grouped by project, with fast full-text search across everything.

It never restarts or modifies sessions. It only reads the logs on disk.

Data sources

• Claude Code — ~/.claude/projects/<encoded-cwd>/*.jsonl
• Codex — ~/.codex/sessions/YYYY/MM/DD/*.jsonl

The Rust backend (src-tauri/src/sessions.rs) parses both formats on launch and holds an in-memory index. Project grouping uses the session's recorded cwd.

Run

bun install
bun run tauri dev

Build a distributable .app

bun run tauri build
# → src-tauri/target/release/bundle/macos/backtrack.app

Usage

It opens straight to your most recent session, so "what was I just doing" is answered immediately. Built around recency + precision: find the exact thing an agent said a couple hours / a session or two ago without rehydrating a dead session.

• Search (⌘K): every message across all sessions. Returns one result per matching message — clicking lands you on that exact line, highlighted.
• Filters (under the search box): who spoke (you / agent), source, project, time window. Combine with a query to narrow the haystack.
• Find in session (⌘F): in-transcript find with n/N (or ⏎/⇧⏎) to step matches and a live count.
• Badges: CC = Claude Code, CX = Codex.

Keyboard

Key	Action
⌘K	Search everything
⌘F	Find within the open session
Tab	Switch between list and transcript
↑ ↓ / j k	Move in list · jump message-to-message
→ / ⏎	Open selected (expand project / open session)
←	Collapse project
[ ]	Previous / next session (global recency order)
⏎ / ⇧⏎	Next / previous find match
Esc	Close find · clear search · back out
?	Toggle the shortcut legend

Moving the list selection auto-previews the session (Mail.app style), so you can scan fast.

Notes / next

• The index is built once at startup. Use the reindex command (already wired in the backend) to refresh without restarting — not yet exposed in the UI.
• Sessions with no real top-level prose (command-only / empty logs) are skipped, so the indexed count is lower than the raw file count.
• Search ranking is recency-first; match-count ranking is a planned follow-on.