Wake

A native desktop app that gathers every coding-agent session on a machine into one place. It opens the private directories the agents already write to — read-only, and it never writes back — and turns them into one library that can be browsed by agent or project, searched, and resumed in a terminal at the original project directory. The supported-agent table has 22 rows, and between them they cover JSONL trees, dated rollout directories beside a SQLite state file, three table schemas across two database files from one product, a key-value store, zstd-compressed logs whose file name carries a format generation, and one agent whose transcripts are encrypted and can therefore only be listed as metadata. Search is SQLite FTS5 over a trigram index, so a Chinese phrase and a code substring such as useEffect( take the same path, and a hit opens the transcript at the matched message. The same index is handed back to the agents through a bundled read-only MCP server with five tools, and a command line that prints byte-identical output. Around that sit an Insights statistics page, remote hosts mirrored over SSH, a read-only memory page, and a cleanup page that moves session files to the system trash behind a marker so they stay deleted. Rust on GPUI, MIT, macOS first with experimental Linux and Windows builds.

Who built itAn independent developer whose own main tools are Claude Code and Codex, and who works mainly in Chinese. He wrote 181 of the repository’s 200 commits; the rest come from a GitHub account named claude with 10, YTwsy with three, and five others with one each. Of the 129 co-author trailers, 125 name a Claude model — Fable 5.1 on 63, Fable 5 on 36, Opus 5 on 16 and Opus 5.5 on nine.

1. 01

   Forty-two versions in forty-one days, and a release line that routed around a proxy

   The repository was created on 2026-08-18 and 0.1.0 shipped the same day with seven agents. In the forty-one days to 0.8.5 on 2026-09-27 the changelog gained 42 version sections, and the 200 commits split 76 in August and 124 in September. The busiest day is 2026-08-25 with five versions; 2026-09-03 carries four, and 2026-09-01 and 2026-09-25 three each. The API list in the recon report stops at twenty entries, from v0.4.2 to v0.8.5, because that is one page; the changelog goes back to the first release. Beside it sit 1,387 stars, 88 forks, 2 watchers, 6 open issues and an MIT licence. Eight accounts have committed: the owner wrote 181, an account named claude wrote 10, YTwsy three, and five others one each. The release workflow is worth reading on its own. Pushing a v* tag creates the GitHub Release with notes cut out of the changelog — the script takes everything between ## [X.Y.Z] and the next version heading — and then three jobs build for macOS, two Linux runners and Windows in parallel. They run on native machines because, in the author’s own comment at the top of the file, GPUI’s shader compilation cannot be cross-compiled. Assets already on a release are skipped and never overwritten, so whatever was uploaded first, including from the maintainer’s own laptop, is the published artifact. The same comment records a workaround: tag pushes and manual dispatch both came back 403 from the proxy the development environment pushes through, so pushing a branch named release/vX.Y.Z triggers the workflow instead and the tag is stripped out of the branch name.

2. 02

   One trait, twenty-six adapter files, and the formats the agents invented

   The data layer is wake-core, and everything that differs between agents is confined behind one AgentAdapter trait. The README states the payoff in a comment — 「add an adapter, get the whole UI for free」 — and adds that remote.rs wraps any adapter over a synced cache, so remote hosts reuse the same code rather than a parallel implementation. crates/wake-core/src/adapters/ holds 26 files: 22 agent modules plus mod.rs at 61 KB and parse_utils.rs at 41 KB. The formats they absorb have almost nothing in common. Claude Code writes one JSONL per session under a per-project directory. Codex writes rollout-*.jsonl into a dated tree and keeps a parallel state_5.sqlite. Cursor keeps two stores that never meet: CLI transcripts under ~/.cursor/projects, and the IDE’s Chat and Composer history inside state.vscdb, a key-value database where one key gives the ordering and per-message keys give the bodies; a chat that also has a full transcript stays with the transcript so it lands under the right project. OpenCode ships three schemas across two database files, both scanned. DeepSeek Harness writes session[.vN].jsonl[.zstd], where the file name carries a format generation and the logs are decompressed on the way in — exactly the shape that broke in issue #45, where an adapter still looked for the older name and indexed zero sessions out of 112 on disk. Devin is a plaintext two-table database whose visible transcript is a forest walked from main_chain_id back along parent_node_id, with a documented fallback to every node in write order. Cursor’s directory slugs are the smallest and nastiest case: a single - stands for a path separator, a hyphen and a space, so My Project was decoded as My/Project until issue #38.

3. 03

   What the file boundary keeps, what it hides, and what it leaves visible on purpose

   Codex writes its background threads into the same sessions directory as real conversations, and the first version of Wake indexed them: issue #30 reported guardian auto-review threads appearing as Untitled entries, some of them blank and some showing a reviewer’s approval JSON. Both shapes announce themselves on the first line, in a session_meta record carrying a thread_source and a structured source object, and the decision was taken at the file boundary rather than after parsing, so the watcher and the incremental path reach the same verdict as a full scan. v0.6.6 excluded the whole family. Then issue #42 asked for spawn_agent sub-agents back, noting that codex.rs already carried a comment saying their transcripts were held back for now and could later be reattached through parent_links. v0.8.0 answered it: spawned children are indexed again, nested under the session that spawned them and titled with their task name, the parent history Codex copies into each child is folded into one stand-in so it is not indexed twice, and the guardian, /review, compaction and memory-consolidation threads stay out. The README states the tie-break for anything it cannot classify: such a file stays visible rather than risk hiding a real conversation. The same instinct shows up with Craft Agents, which runs other engines and leaves a copy of each conversation in Claude Code’s own history; Wake lists the Craft session and hides the engine copy only while the original exists, and the copy reappears if the Craft session is deleted.

4. 04

   Search over every agent, and two reports that turned a session id into a command-line flag

   Search is one SQLite FTS5 table with a trigram index, which is why a Chinese phrase and a query for useEffect( are the same operation; below three code points it falls back to a LIKE scan, and 0.8.0 fixed a crash where a one- or two-character term hit a message containing a character such as Ω or İ before the match. A hit jumps to the matched message rather than the top of the transcript, on a seq contract that the index and the reader both honour, and 0.6.5 added title matching and a small preference for recently active sessions. That same release had to stop an agent’s own wake_* lookups from appearing as hits inside the session that made them. Resume is where the sharpest reports landed. Issue #50 showed that a session id was concatenated into the agent’s command line with no shape validation, and that the quoting helper’s fast path treated - as a safe character, so a session file named --dangerously-skip-permissions.jsonl became a real option on that command line. Issue #49 went the other way: on Windows the Command Prompt path wrapped the id in double quotes without escaping anything inside them and filtered only %, while the ids of several adapters come from a database rather than a file name, so the assumption the code was written on did not hold. 0.8.4 fixed both by validating ids before opening a terminal, opening a desktop app, or copying an SSH command.

5. 05

   GPUI: no symbols, no zoom, no Dock hiding, and a window layer CI only compiles

   The app is GPUI 0.2 with gpui-component 0.5, and several small constraints follow from that and are written down. GPUI has no SF Symbols, so every chrome icon is a bundled Lucide SVG and every agent identity is a bundled brand PNG rendered in its original colours, never tinted, not even in a selected row. There is no page zoom, because there is no web view: issue #51 asks for readable text and lists what does not exist — no font-size setting, no working Cmd +, and a Window menu zoom that only maximizes the window — and the only workaround offered is macOS Accessibility Zoom, which magnifies the whole screen. Issue #43 found a second closed door: the author had tried LSUIElement in the app bundle to keep Wake out of the Dock, and it does nothing, because GPUI sets the activation policy itself at launch. Two more limits are recorded as rules rather than bugs. A pointer callback must not synchronously dispatch a keyboard event, because opening the search palette from an AppKit mouseUp re-enters GPUI’s event dispatch; the root has to save the current focus before the palette opens, or ⌘K stops working once it closes. And the theme turns the component library’s focus ring off, because that ring is an absolutely positioned child two pixels outside the element’s box and gets clipped by dialog layers and scrolling lists — focus is carried by tinting the border instead, which takes no space and cannot be cut in half. Continuous integration builds the app on macOS, Linux and Windows, but the window layer is gated only by compiling there; the memory page, the cleanup page and the workbench itself ship with no windowed test.

6. 06

   A design document written as numbers, beside a list of what the author could not verify

   DESIGN.md is 26,541 characters of specification rather than guidance: colours may only change in theme.rs, icons must be registered in assets.rs where a missing file extension is a silent blank, every interactive element gets an .id() before anything is bound to it, and every window root must mount the overlay layers after its content in an order the document calls a contract. Some of it reads like measurement notes. The sidebar’s centreline is the macOS traffic light’s red centre at x = 26.75, and everything else descends from it, down to a 7.75-pixel leading inset; the title inset of 9 was reverse-engineered from the width of the letter W, with a warning that changing the font size or weight invalidates it immediately. One gap was unified at 6 pixels on 2026-09-22 and is deliberately off the four-pixel grid, because brand PNGs carry no padding while Lucide icons carry about 1.5 pixels of it. A third note records that the rasterisation step on a 2x display is 0.5 pixels and that the digits past the decimal point should be left alone: the header inset had been 12.125 and landed at −0.125, and rounding it to 12.25 moved it to +0.375, a different physical pixel. The same document bans rem-based spacing outright after a migration finished on 2026-08-24, because rem is pinned to 14 pixels and the utility classes it replaced rendered at sizes such as 10.5 or 12.25. Underneath that, PRODUCT.md keeps the other kind of record: which adapters were checked against a real session on the author’s machine and which were inferred from source code, that OpenClaw, CodeBuddy and Reasonix had no local sessions to check against at all, that Qoder’s token counters are zero and its Credits are a different unit that will not be converted, and that from 2026-08-19 every screenshot and demo uses a synthetic home directory built by a 34 KB script, because real project names are private.