--- title: "Twenty-one Releases in Thirteen Days" description: "What the Foxl v0.7 series shipped: arrangeable cloud terminals with per-CLI cost readouts, saves 10x faster, search 50x smaller, and four limits turned back off." author: "Foxl Team" date_published: "2026-09-15" date_modified: "2026-09-15" canonical_url: "https://foxl.ai/blog/v0-7-series-what-shipped" markdown_url: "https://foxl.ai/blog/v0-7-series-what-shipped/index.md" image: "https://foxl.ai/blog/v0-7-series-cover.png" social_image: "https://foxl.ai/blog/social/v0-7-series-what-shipped.png" content_type: "Release note" topics: ["Cloud terminals","Cost visibility","Performance","Model providers","Release engineering"] products: ["Foxl Code","Foxl Agent","Foxl Notes"] --- # Twenty-one Releases in Thirteen Days > Foxl v0.7.0 shipped on September 3 and v0.7.20 on September 15. Two other posts cover what went wrong in that window; this is what changed for somebody using the product. Cloud terminals became arrangeable, stopped killing their agent when you close a pane, and started showing what they spend - read out of the accounting each coding CLI keeps for itself rather than estimated, so Kiro reports the credits you have left and Hermes on Bedrock reports no price at all. Saving got ten times faster and search returned fifty times less data, both measured on a real 1.3 GB database. GPT-6 Astra arrived on Bedrock the day AWS made it available, and the reasoning effort you pick on that route now takes effect instead of being accepted and ignored. And four request limits shipped in v0.7.1 were switched off a day later, because the thresholds sat inside normal use. - Author: Foxl Team - Published: 2026-09-15 - Last reviewed: 2026-09-15 - Reading time: 8 minutes - Canonical HTML: [https://foxl.ai/blog/v0-7-series-what-shipped]() ![A tear-off calendar with thirteen days marked beside a stack of twenty-one sheets: twenty-one releases in thirteen days.]() Foxl shipped twenty-one releases in thirteen days, from v0.7.0 on September 3 to v0.7.20 on September 15. Two posts already cover what went wrong in that window: [Every Gate Was Green]() is about the first three releases, one of which had to be withdrawn, and [A Citation Is Not a Measurement]() is about a vendor guarantee we had quoted accurately and read too broadly. This one covers what changed for somebody using the product, with the numbers behind the parts that are claims rather than opinions. ## The cloud terminal became somewhere you can work Most of the series went into Foxl Code. A cloud terminal was already there in v0.6; what it lacked was everything around it. **Panes can be arranged, not just split.** The Terminals menu has five named layouts: Columns, Rows, Main on top, Main on the left, and Grid. None of them opens or closes a terminal, because rearranging what is in front of you should not be a decision about which agents keep running. **Closing a terminal closes the window and stops nothing.** Until v0.7.7 the X on a pane you had opened yourself also cancelled the agent inside it. The old reasoning was defensible: that pane *is* its session, so closing it leaves a cloud machine running with nothing attached. The problem is that X sits next to the zoom button, asks for no confirmation, and is the reflex gesture for "get this off my screen". A gesture made that casually should not cancel running work. **Foxl Code has its own reasoning depth.** It carries the same Low through Max effort levels as Foxl Agent, as a separate setting. Before, it read the Agent's setting and had no control of its own, so turning up how hard your coding agent thought also turned up your chat. **Files opens a real editor.** A PDF, a Word file, a spreadsheet and a deck preview the way they do in Foxl Agent, an `.html` file in your workspace opens as a page rather than as source, and shared documents can be edited in place. Late in the series the transport underneath all of this was rewritten, because keystrokes had never been reaching the machine they were addressed to. That is the subject of the citation post; it is worth reading in this order, because the arrangement and the editor were built on top of a terminal that could not yet be typed into. ## Every terminal shows what it is spending A coding CLI running in a terminal puts no usage on the wire we can read. So the figure in the top-right corner of a Foxl Code pane is the accounting the CLI keeps for itself, read back and displayed, rather than an estimate of ours. Context, tokens sent and received, and the assistant's own cost figure, in the same formatting Foxl Agent's titlebar uses, because it is the same measurement in a different window. Five assistants report, and they do not report the same thing, because they are not billed the same way: - On pay-as-you-go the corner shows a price. On a subscription it shows subscription usage instead, because a dollar figure for work covered by a monthly fee is a number nobody is paying. - A Kiro terminal shows how many credits you have *left*, not how many you have spent: the balance, the share of the plan gone, the plan's name, and when the cycle resets. The unit label comes from Kiro too, so a plan metered in something other than credits still reads correctly. - opencode and Hermes show tokens, with opencode's own cost figure beside them. A Hermes session on Bedrock shows *no* price, because Hermes records that it did not price that session, and we would rather show nothing than print `$0.00` over work that really cost money. ## Speed, as numbers v0.7.3 was a performance release, and the figures below are measured on a real 1.3 GB database rather than a fixture. **Saving is roughly ten times faster.** The local database was running in SQLite's default journalling mode, which flushes to disk twice for every row. A burst of 500 agent events went from 337 ms to 31 ms, and 200 chat messages - which also update the search index - from 188 ms to 24 ms. Across 6,456 samples over three minutes the slowest 1% of requests went from 17 ms to 2.8 ms. **Searching your conversations returns fifty times less data.** The dropdown shows one clipped line per result and was being sent every matching message in full: 1.5 MB for a single search, including one 1.2 MB message. It now sends the snippet it displays, which is 31 KB, and the search itself is five times faster. From a phone that is 1.5 MB no longer downloaded to draw the same dropdown. The overview page's rollup became roughly fifteen times cheaper, and its cache started working. None of these three is a new feature, which is why they are easy to leave undone: nothing was broken, it was just slower than it had any reason to be. ## Two new models, and an effort control that was being ignored GPT-6 Astra runs on Amazon Bedrock as of v0.7.9, on Foxl credits for Pro and Ultra and with your own AWS account under Settings, with a 1,050,000-token context. It had been reachable only through a ChatGPT subscription or an OpenAI API key; AWS made it generally available on September 8 and it was offered the same day. The same release made Astra use the reasoning effort you pick. The parameter Foxl was sending on that route was accepted by the service and ignored, so a picker reading Max was running the model at its own default. An ignored parameter is harder to catch than a rejected one, because a rejection produces an error message and this produced a picker showing the setting you chose. Claude Fable 5.1 became available under the Claude Code sign-in in v0.7.5, and GPT-6 Astra under the ChatGPT sign-in in v0.7.3. ## The browser extension stopped asking you to carry a secret Foxl used to print a 64-character pairing code in Settings and ask you to copy it into the extension. That existed because a browser extension runs in Chrome's sandbox and cannot read a file, so a person was the only channel a secret could travel through. Since v0.7.5 the desktop installs a small connector that *Chrome* starts, and Chrome permits only the Foxl extension to use it. The code is gone from Settings. This covers the desktop side only. Measured on September 10, the published Chrome Web Store build is still the older one, which predates the connector and uses the previous path, so the pairing code is gone from the app before it is gone from every install. The extension update that completes it ships from its own repository, on its own review cycle. ## What we switched off v0.7.1 added four request limits to Foxl Code, and v0.7.3 turned all four off because the numbers were too small. A chat turn was capped at 20 a minute, the autonomous loop's next run at 10, adopting a GitHub installation at 5, and typing into a cloud terminal at 600 a minute. Every one of those is reachable on a busy afternoon, which means the limit you would actually have met was your own use rather than any runaway. A rate limit whose threshold sits inside normal use works like a delayed outage rather than a safety feature, so we took the limits back out instead of explaining them. One change went the other way. In v0.7.19 Workspaces and Channels became part of the paid plan: Channels, the Workspace settings section, and shared files no longer appear for a free or signed-out account. Those are the surfaces that spend a workspace's credits. Nothing already created is deleted, and a plan that includes the Workspace brings all of it back where it was. ## One typeface, one product menu Foxl has its own typeface since v0.7.3. The app, foxl.ai and docs.foxl.ai render in Geist for body text, Instrument Serif for display headings and JetBrains Mono for code, self-hosted so they are the same on macOS, Windows and Linux instead of falling back to whatever each system ships. Korean uses Pretendard, loaded only when Korean is on screen. Agent, Notes and Code share one product menu and one sidebar shape as of v0.7.10. Model pickers went to compact single-line rows in v0.7.15 and, on the desktop, into a small native popup that can extend past the window edge; on a phone the picker became searchable in v0.7.13. v0.7.12 added a visible conversation menu for pinning and renaming, Export and Import for custom tools, Cmd+D to delete the open note in Foxl Notes, and Cmd-click on a file path inside a tool's *result* rather than only in the model's prose. ## What the shape of it says Shipping twenty-one releases in thirteen days makes a wrong decision cheap to undo: the four limits were live for one day. It also lets a defect sit underneath a fortnight of work built on top of it, which is what happened with the terminal transport. Both of those are visible in the changelog rather than summarised out of it. The per-release detail, including every fix this post skipped, is at [foxl.ai/changelog](). ## References and further reading 1. [Foxl changelog]() - Release 2. [Foxl Code]() - Reference 3. [Download Foxl]() - Reference