Session History Table

Lists Claude Code sessions from the local session store with a one-line description for each, so the user can find a session ID to resume or recover.

How it works

Claude Code writes one JSONL transcript per session under ~/.claude/projects/<encoded-cwd>/<session-uuid>.jsonl. Each transcript contains an aiTitle record (Claude's own auto-generated session title) and the user's prompts. The bundled list_sessions.py reads these, skips subagent and workflow transcripts, and prints a table sorted by last-active time.

Run it

python3 ~/.claude/skills/sessions/list_sessions.py $ARGUMENTS

Falls back to python if python3 isn't on PATH (e.g. some Windows setups) — the script itself has no version-specific dependencies.

Arguments ($ARGUMENTS)

• [days] — look-back window in days (default 7).
• --project SUBSTR — only sessions whose cwd contains SUBSTR (e.g. --project myproject).
• --md — emit a GitHub-flavored markdown table (default: aligned monospace text). Use --md when presenting the result inline to the user.
• --all — ignore the day window; list every session on disk.

Examples

• /sessions → last 7 days
• /sessions 14 --md → last 14 days as a markdown table
• /sessions --project myproject --all → every session ever, for a given project folder

Presenting results

Render the output as a markdown table for the user (pass --md, or reformat the text output). Columns: last-active, turns, session ID, description. Point out that a session is resumed with claude --resume <SESSION ID> — this works from any directory, so the user does not need to cd anywhere first. If several sessions share an identical timestamp, note that — it usually means a crash or shutdown flushed them together.

Notes / gotchas

• Namespaces are already unioned — do not "fix" this. Transcripts are split across one directory per launch folder (home, ~/Projects, one per project, …). list_sessions.py globs projects/*/*.jsonl, so the * covers every namespace in one pass. Same for session-search.py, which iterates the same root.
• Resume is not namespace-scoped; the picker is. claude --resume <ID> with an explicit ID resolves across namespaces and reattaches to the original transcript wherever it lives. Only --continue and the bare --resume picker are limited to the current directory — those are the cases where launch folder matters.
• Description fallback chain when a session has no aiTitle yet: first typed prompt → first slash-command invocation (<command-name> + <command-args>, skipping config commands like /effort and /model) → the last-prompt record.
• Turn count = non-meta user prompts, including slash-command turns. 0 turns genuinely means an empty session. (Earlier versions undercounted slash-command-only sessions as 0 and labeled them "(no prompt)".)
• Any secret-scanning hook that blocks shell head/cat/heredoc pipes over these JSONL files should be respected — always read them inside a Python script (as this tool does), never via a bash pipe.