session-index-viewer

English · 简体中文

Local web viewer for AI coding CLI sessions on this machine. Browse recent sessions from Claude Code, Codex, Devin, and Grok in one place, search across tools, inspect token/tool usage when metadata exists, and resume any session in a new Terminal window with a single click.

CLI resume lists (claude --resume, codex resume, grok --resume, …) show ids and timestamps — not what was said. This viewer shows the opening prompt and last reply of every session so you can spot the one you want and pick up where you left off.

macOS only. Uses launchd for autostart and AppleScript / open -na to launch Ghostty, iTerm, or Terminal.app. Auto-detects whichever is installed, in that order.

Run

./install.sh            # install as launchd agent (run at login, keep alive)
open http://localhost:7333

Or run in the foreground: python3 server.py.

Frontend (optional rebuild after UI changes):

cd frontend && npm install && npm run build

server.py serves frontend/dist/ when present, otherwise sessions-index.html.

Supported sources

Source	Default path	Resume	Usage notes
Claude Code	~/.claude/projects/*/*.jsonl	claude --resume <id>	Lifetime sums from assistant message.usage; context ≈ peak input+cache
Codex	~/.codex/sessions/**/rollout-*.jsonl	codex resume <id>	Last cumulative token_count; context ≈ max per-turn total
Devin	~/.local/share/devin/cli/sessions.db	devin -r <id>	Aggregated from message_nodes
Grok	$GROK_HOME/sessions (default ~/.grok/sessions)	grok --resume <id>	Context size only on disk (signals.json); headless one-shots (is_non_interactive) are skipped

Usage on the card is a short chip (ctx · out · tools · turns). Click it (or press u on the active card) for a modal with overview KPIs, token mix, activity counts, and a short note on how that source measures tokens.

Context is peak / window occupancy — not a full-session billing total. Grok often only has that size signal (labelled size only on the chip).

Keyboard

Key	Action
j / k or arrows	Move active card
Enter	Resume in Terminal
c	Copy resume command (not ⌘C — system copy still works)
p	Pin / unpin
u	Open usage modal (when usage exists)
⌘K / Ctrl+K	Command palette

Pieces

• server.py — thin entry point (python3 server.py / launchd).
• siv/ — stdlib-only backend on 127.0.0.1:7333:
  • GET / serves the viewer.
  • GET /api/sessions?limit=100 scans session files live, with an mtime/size cache so only changed files are re-parsed.
  • POST /api/resume opens a Terminal window with a validated cd <cwd> && <tool> resume <id> command. If the recorded cwd belongs to another machine, the home-dir prefix is remapped to this machine.
  • Host badges come from the username in each session’s cwd (/Users/<name>/... or /home/<name>/...).
  • Parsers: siv/sources/ (claude, codex, devin, grok).
• frontend/ — React card UI (search, source/host filters, pin, usage modal).
• sessions-index.html — legacy single-file viewer (fallback if dist is missing).
• install.sh — launchd plist + optional frontend build. Logs: ~/Library/Logs/session-index-viewer.log.
• index-sessions.sh — legacy shell indexer; superseded by server.py.

Multi-machine setup

If you sync session trees with Syncthing (or similar):

Path	Sync-friendly?
~/.claude/projects	Yes (per-file jsonl)
~/.codex/sessions	Yes (per-file jsonl)
~/.grok/sessions	Yes if you sync only session dirs and ignore session_search.sqlite / *.lock
Devin sessions.db	No — single SQLite DB does not merge across machines

No extra viewer config is required: each session is labelled by the username in its recorded cwd, and the host filter picks those labels up. Sessions that share a username across machines appear under the same host.

Avoid writing the same session id on two machines at once (conflict copies).