Repository navigation
Sequence 0.1.8.4, the full public release.
Download
- Windows:
Sequence.Setup.0.1.8.4.exe(installer) orSequence-0.1.8.4-win.zip(portable). - macOS, Apple silicon:
Sequence-0.1.8.4-arm64.dmgorSequence-0.1.8.4-arm64-mac.zip. Intel:Sequence-0.1.8.4-x64.dmgorSequence-0.1.8.4-x64-mac.zip. - Linux:
Sequence-0.1.8.4.AppImageorSequence-0.1.8.4-amd64.deb.
Every file is listed with its SHA-256 in SHASUMS256.txt. Check your copy against it.
The installers are not code-signed or notarized. macOS 14 and earlier: right-click the app and choose Open. macOS 15 and later: open the app once, then System Settings > Privacy & Security > Open Anyway. Windows: SmartScreen shows More info > Run anyway.
Release status, stated plainly
- The Windows installer and zip were built and tested on a Windows PC:
pnpm test:releasepasses 15 of 15 there. On the GitHub Windows runner, the same packaged-app check failed two items: window-shown at 2.8 s against a 2 s budget, andoutside-home showed no refusal. The Windows files on this page are the PC-built ones. - The macOS dmg and zip were built on a macOS runner. They are unsigned and were not launched on a Mac.
- The Linux AppImage and deb were built and launch-smoked on the GitHub Ubuntu runner.
Sequence 0.1.8.4
Highlights
- A new sidebar. The rail has navigation rows (Workflows, Components, Profiles, Skills) that open their own views, each with a hover card that previews an item and, for a workflow, launches it. Every chat lives under its workspace folder.
- Spaces are per chat. A board records the chats it belongs to, so a chat opens with its own board.
- Workflows you launch are watched live on the Agents surface: the run you started is the one the Workflows tab and the overview show.
- The chat overview and the file button float over the pane with no big boxes around them. Focused or Detailed tool calls is one control, in Settings > General.
- Coming back to the Architecture tab no longer shows the cards jumping. The last arrangement stays on screen while the new one is computed: 12 different arrangements before the final one, now 0; 446 ms to the final one, now 99 ms (numbers under Faster).
- The desktop app starts its engine before it builds the window, and no longer waits for the first project read before it shows the app.
New
- Rail navigation sections with hover previews; Skills sit under the Skills row instead of a folding group.
- Workflow preview cards (hover to read, Launch to start) and a live view of the launched run on the Agents surface. The Agents tabs follow the pane width: on a narrow pane they show icons only.
- Per-chat Spaces: a board belongs to the chats that made it.
- Floating chat overview and file buttons; a light edge instead of a heavy shadow on the floating panel.
- Focused/Detailed tool-call detail moved to Settings > General (it is no longer in the status bar).
- Agent usage in Settings > General: the total tokens your agents reported, and a contribution-style year of them (one square per day, darker where more was spent). It counts only turns where the agent reported usage, so the total is a floor, and the card says so.
- The Bastion backdrop is drawn at the first input or idle moment, not at boot.
Fixed
- Reopening the Architecture tab: the board started from its seed pile and settled twice. It now keeps the last arrangement of the same set of cards (
canvas/layoutCarry.ts) and does not draw an identical result twice. - The board's tab strip is a real tab list: it holds only tabs, so a screen reader reads "tab 1 of N" and not the rename field or the add button as tabs.
- The open sidebar drawer and its scrim now stack above the board's dock on a narrow window; before, the dock's buttons showed through.
- The status bar is announced as a status region, not as a page footer.
- The narrow Settings navigation fades at the edge where more is scrolled out of sight, on browsers that support scroll-driven animation.
- Stop really closes the provider connection on Node 20 and 22, not only on Node 24; a provider that kept generating after Stop no longer keeps the call open (and billed).
- The scanner no longer draws three empty message-topic nodes for the engine's own network diagnostics when it reads Sequence's source.
- Compose and Kubernetes scan evidence uses
/in paths on every operating system; on Windows a backslash appeared in evidence and warning paths. - Scans list directories in one order on every file system; on POSIX a Windows drive path in a read is refused and backslashes read as separators;
run_commandkills its whole process group on abort or timeout. - The release check keeps its timer alive, so a stalled response can no longer let the process exit unsettled.
- A long tool call is no longer reported as an agent that stopped responding. A command that runs for minutes without sending frames (a 240-second sleep, a long build or test run) keeps the calm "Running for 3m 57s" note with Stop, and never turns into "isn't responding" with Retry. Only a turn with no tool running is marked stalled after 40 seconds.
- A slash command can be typed inside a prompt, not only at its start:
/after a space opens the palette, and a skill or the agent's own command is inserted where you typed it, with the rest of the sentence kept. A path such aspackages/web2still never opens it.
Faster
From docs/bench/PERF-0.1.8.4.md. Before is the 0.1.8.4 integration snapshot, after is this release; same machine, one tree after the other, medians (one CPU-only container with the browser throttled, so absolute numbers on your hardware will differ).
- Reopen the Architecture tab: arrangements shown before the final one 12 to 0; click to the final arrangement 446 ms to 99 ms (1x CPU), 1,145 ms to 604 ms (4x throttle).
- First board, cold start, board drawn: 1,441 ms before, 1,445 ms after (the same), with 33 boot requests either way.
- Streaming an answer, chat switches and the JavaScript heap: unchanged within the run-to-run spread. The first-page script grew by 554 bytes over the snapshot.
The first-page script (2,655,231 bytes) is 48,567 bytes above the previous recorded ceiling; 48,013 of those came with the program's own features before this round and were accepted by the owner. The reason is recorded in tools/measure/perf-ratchet.json.
Known limits
- Update checksums are not signed: there is no signing key yet, so updates are verified against an unsigned
SHASUMS256.txtonly. The installers are not code-signed or notarized. - Built on Linux x64 (Node 22) and on GitHub runners. The Windows installer and zip were built on a Windows PC and pass
pnpm test:release15 of 15 there; on the GitHub Windows runner the packaged app's window-shown check took 3.1 s against a 2 s budget. The macOS dmg and zip were built on a macOS runner, are unsigned, and were not launched on a Mac. The Linux AppImage and deb launch-smoke on the runner. The Playwright end-to-end tier and a real agent turn were not run for these notes. - The demo clips and stills use a stand-in model whose answers were written by hand; the app, the scan and the board are real. They are not model output.
tools/measure/switch-bench.mjstimes out about half the time at 1x CPU (on this release and the one before it); the figures above are from the runs that finished.- Auto-drawing the board when a project opens is still left out on purpose: it would spend an agent turn on every open.
Sources
In this repository: docs/bench/PERF-0.1.8.4.md (before and after, with the limits), tools/measure/perf-ratchet.json (the boot ceiling and why it moved), tools/measure/board-reopen.mjs (the reopen measurement) and docs/bench/REPRODUCE.md (how to reproduce).