clres

Browse and resume your Claude Code conversations from a tiny terminal picker.

Each conversation gets an emoji (keyword-based, stable), its first prompt as title, the project it belongs to, and its age. Hit Enter to resume it with claude --resume from its original working directory.

 clres · 19/36 conversations   ↑↓ move · Enter resume · t title · a all · q quit
 🖥️ ✨ Terminal browser for Claude Code conversations   my-tools          2m
 🐛 ✨ Fix flaky integration test on CI                 backend           3d
 📚 ✨ Sphinx docs landing page refresh                 docs-site         5d
 ...

Install (shell alias)

git clone https://github.com/BelCorentin/clres ~/git/clres
~/git/clres/install.sh   # adds `alias clres=...` to ~/.zshrc

Install (Claude Code plugin)

Adds a /clres slash command that lists conversations inside a session:

claude plugin marketplace add ~/git/clres
claude plugin install clres@clres

Usage

invocation	behavior
clres	interactive curses picker
clres --all	include tiny + headless convos (bare /model, sdk bots)
clres --index	haiku-title every untitled real conversation
clres --summarize	haiku-summarize every real conversation
clres --list	plain table (also used when piped)
clres --json	machine-readable dump
/clres	list inside a Claude Code session (plugin)

Keys

• ↑/↓ or j/k — move · g/G — top/bottom · PgUp/PgDn
• / — search (title + project + summary), Enter keeps filter, Esc cancels
• Enter — resume · s — summary popup · t — re-title · q — quit
• m — flag/unflag 🔖 "might come back" on the highlighted session
• c — toggle the focused view (🔖-flagged + recent ⇄ recent + old)
• a — show hidden (tiny/headless) too

Each row shows the session's git branch (⎇ branch, @worktree if a linked worktree) and a 🔖 when it's flagged to come back to.

What gets shown by default

The picker opens on the focused view: sessions you flagged 🔖 to come back to (via m), plus anything touched in the last 24 h — so active work is never hidden. c widens it to older sessions; a (or --all) additionally reveals tiny/headless ones, dimmed. The 🔖 flag is a marker file at ~/.claude/comeback/<session-id>, so it persists across runs and machines.

Titles prefer the live session-state goal (~/.claude/goals/<sid>.json, written by the session-state.sh hook) over the raw first prompt when no t-generated title is cached — the same clean goal shown in the statusline, ntfy, and ccview.

Generated titles & summaries

--index (or t on a row) sends the first user prompt + last assistant message of a conversation to claude --model haiku -p and caches the returned title in ~/.cache/clres/titles.json (marked ✨ in the list). --summarize (or s on a row) does the same for a 2-3 sentence summary, shown in the status bar / a popup and included in search. Only real conversations with ≥ CLRES_MIN_ENTRIES (15) transcript entries are auto-indexed. The titler's own headless sessions are corralled into a throwaway /tmp/clres-titler project and deleted, so they never pollute the list or claude --resume.

Tunables (env): CLRES_MIN_TITLE (20 chars — hide shorter titles), CLRES_MIN_ENTRIES (15), CLRES_MODEL (haiku).

Projects

Each conversation is tagged with a research-project slug (mindsentences, distraction, lppreadlisten, fusion, sevenT, syntax, config, obsidian, personal, else misc) inferred from its cwd → branch → title via the ordered PROJECT_RULES table (mirrors ccview). Classified rows show the project emoji + a coloured project label; p cycles a project filter (header shows the active one). Search (/) still matches the raw cwd basename too.

How it works

Reads ~/.claude/projects/*/*.jsonl transcripts (honors CLAUDE_CONFIG_DIR), takes the first real user prompt as the title (slash commands are unwrapped, hook/system noise skipped), sorts by file mtime. No dependencies beyond Python 3.10+ stdlib.