Wake — past coding-agent sessions

Wake indexes every coding-agent session on this machine — Claude Code, Codex, Cursor, Gemini CLI, OpenCode and a dozen more — and wake-cli reads that index from a shell.

Reach for it whenever the answer lives in an earlier conversation rather than in the code: what was discussed, what was decided and why, where previous work stopped, which approaches were already tried and rejected, whether an error has been seen before. Git history and the working tree do not hold any of that.

Everything below is read-only. Nothing here can modify or delete a session.

Finding the binary

command -v wake-cli || ls /Applications/Wake.app/Contents/MacOS/wake-cli

Use whichever resolves. On Linux it is on PATH after a deb install, or at ~/.local/bin/wake-cli from the tarball; on Windows it sits next to Wake.exe.

Two things can be missing, and they need different answers:

• No binary. Say so — the user needs Wake. Do not try to install it yourself.
• No index (no Wake index at … — launch Wake once to build it). Wake is installed but has never run. Run wake-cli index once; it builds the index in a few seconds without opening the app, then retry your query.

If the freshness line at the end of a listing is well behind and Wake is not running, wake-cli refresh updates the index in a few seconds; it does nothing while Wake has its window open or is still scanning. Retry the query afterwards.

Commands

Always scope to the current repository with --project "$PWD" unless the user clearly means "everywhere". Without it, every project on the machine is in scope.

What happened in this project recently

wake-cli sessions --project "$PWD" --limit 10

Add --agent codex (repeatable), --since 7d, or --starred to narrow it.

Where was this discussed

wake-cli search "rate limiter" --project "$PWD"

Terms are ANDed. CJK text and code substrings like useEffect( work. Drop --project when the user asks about something they might have worked on elsewhere.

Read a session

wake-cli show 'claude-code:1b2c3d4e-…'

Add --tools for tool inputs and outputs, --thinking for recorded reasoning. Both are verbose — start without them. Work done by subagents (Claude Code sidechains, Cursor subagents) is not in the main transcript; its ids are listed at the end of the output — read one with --subagent ID, or --subagent '*' to list them all. Sub-agents that run as sessions of their own (Codex spawn_agent, Grok) are listed at the end too, by key; sessions only lists parents, so this is where to find them.

Which projects have history

wake-cli projects --limit 20

What an agent already remembers about this project

wake-cli memories --project "$PWD"

The memory files agents keep for themselves — Claude Code's auto-memory, Codex's memories, ZCode's project memory — plus the instruction files kept for them (CLAUDE.md, AGENTS.md, GEMINI.md, .cursor/rules …), read-only, grouped by project, user memory (notes that apply to every project) last. Each entry ends with a wake://memory/… reference; pass it to show to read the file. Reach for it before re-deriving a project's conventions or decisions from scratch.

Keys, references and paging

sessions and search print a key for every session — <agent>:<id>, or <agent>:<host>:<id> for a session mirrored from a remote machine. Pass a key to show.

search also prints ref: wake://session/<key>#<seq> under each snippet. Pass the whole reference (quoted) to show and the transcript starts at that message — that is the fast path from "this term was mentioned" to "here is the surrounding conversation".

Long transcripts are paginated. When the output ends with a from_seq hint, run again with --from <that number> to continue.

Reading the results

• Exit code 0 means it ran — including "no matches" and "no such project". Do not treat an empty result as a failure; report that nothing was found.
• Exit 1 means it ran and failed — an unknown or ambiguous key, or a transcript that would not parse. Exit 2 means the command line was wrong, or there is no readable index yet. Both print an explanation to stderr.
• Listings usually end with the index freshness (Index covers activity up to …) — the newest activity Wake has indexed, not the current time. A stale-looking value usually means Wake is not running, but some agents (Copilot, OpenCode, Antigravity, Hermes, OpenClaw) are stored in SQLite and only refresh when Wake launches, on its Refresh button or on wake-cli refresh, so say what the line shows rather than concluding the user did nothing. show does not depend on that line — it reads the agent's own file rather than the index. The exception is a session from a remote host (<agent>:<host>:<id>): that file is a local mirror, current as of Wake's last successful sync. Say so rather than reporting it as the other machine's latest state.
• Values out of range are clamped, not rejected: --limit 999 returns the maximum.

Sensible defaults

• Start with sessions --project "$PWD" when the user says "last time" or "yesterday", and with search when they name a topic, an error or a decision.
• Read one or two transcripts, not ten — show output is large.
• Quote keys and wake:// references: they contain characters the shell will eat.