wend

wend

Find, recover, and resume your past Claude Code sessions — from the terminal or from inside Claude Code itself. Fast, local, single binary, zero network.

You run dozens of Claude Code sessions across many directories and can't find the one you need. wend indexes your whole local history (~/.claude/projects) and lets you:

• find a past session by keyword — wend search "that firecrawl pricing chat"
• recover the pre-compaction history the live UI hides — wend show <id> --recovered ⭐
• resume an old session — wend resume <id> → cd … && claude --resume …
• read / label / map — wend show, wend name <id> "<alias>", wend tree

Everything is local and read-only over your transcripts; the only state it writes is its own index (~/.local/share/wend/index.db, 0600).

Install

One-liner (prebuilt binary, macOS + Linux):

curl -fsSL https://raw.githubusercontent.com/us/wend/main/install.sh | sh

Pin a version with WEND_VERSION=v0.1.0 or change the target dir with WEND_INSTALL_DIR=~/.local/bin. On macOS, curl|sh installs avoid the Gatekeeper quarantine bit (a binary downloaded via the browser would be quarantined unless notarized).

Prebuilt binary (manual): download the archive for your platform from GitHub Releases (macOS arm64/x64, Linux x64 gnu + musl static, Windows x64), unpack, and put wend on your PATH. These are the lean keyword-only build.

From source:

cargo install --path crates/wend-cli   # puts `wend` on your PATH
wend index                             # build the index (~15s for ~200 sessions, then incremental)
wend doctor                            # check status

# optional: semantic (meaning-based) search — heavier build (ONNX Runtime)
cargo install --path crates/wend-cli --features semantic
wend index --embed                     # downloads the e5 model once, embeds your prompts

Use

wend search "rust sqlite fts"           # keyword (BM25, stemmed, session-grouped)
wend search "fixing a crash" --semantic # meaning-based (hybrid keyword+vector); needs --features semantic build
wend search "auth bug" --json           # machine-readable (for scripts/skills)
wend show <id>                          # read a transcript (numbered messages, total shown)
wend show <id> --count                  # just how many messages
wend show <id> --range 10:20            # messages 10–20 (also --head N / --tail N)
wend show <id> --recovered              # surface pre-compaction history the UI hid
wend resume <id>                        # prints the cd + claude --resume command
wend name <id> "payment-spike"          # alias a session so you can find it later
wend tree [project]                     # worktree/session topology

Short session-id prefixes work everywhere (wend show f8bd399d); ambiguous prefixes list the candidates.

Use it from inside Claude Code (plugin)

claude plugin marketplace add ~/coding/wend
claude plugin install wend@wend

This adds:

• a find skill that auto-fires on natural language ("where was that chat about X?", "recover the compacted history", "nerede konuşmuştuk") and calls wend,
• a /wend:search command,
• a SessionStart hook that keeps the index fresh in the background.

(The plugin calls the wend binary, so install it on your PATH first.)

Status

Working today: index, search (keyword + optional semantic hybrid), show (+recovered, --range/--count, numbered messages), resume, name, tree, doctor — verified on a real 213-session / 178k-message corpus. Semantic search is opt-in (--features semantic): fastembed (ONNX Runtime) with the multilingual multilingual-e5-small model, embedded at the chunk level over your own prompts, cosine + RRF fusion with keyword (first index --embed downloads the model once, then incremental; thread use is capped so it won't pin the machine — override with WEND_EMBED_THREADS). Not yet implemented: subagent indexing (--include-subagents), export. See PLAN.md / ARCHITECTURE.md / IMPLEMENTATION.md.

MIT.