Skip to contentdocs.Growther.si
Download C5

Updating C5

C5 can update itself. One command gets you the newest version.

Update now #

bash
growther update

This downloads the latest release, checks that it is genuine, and installs it.

Before updating, C5 checks whether it is currently running:

  • If C5 is running, it installs the update and restarts C5 for you automatically.

  • If C5 is not running, it installs the update and asks:

    text
    Start growther? Y/n

    Press Return (or type y) to start C5 right away. Type n or press Ctrl-C to exit without starting.

Flags for updates #

You can tell growther update what to do in advance:

bash
growther update --only     # only update; do not restart or ask to start
growther update --start    # update and automatically start or restart without asking
growther update --check    # only check if an update is available
  • --only (or -o): Downloads and installs the update, but leaves C5 alone. It never asks to start C5 and never restarts a running server. Use this in automated maintenance or when you want to update now and start later.
  • --start (or -s): Ensures C5 is running after the update without asking. If C5 was stopped, it starts it; if C5 was already running, it restarts it.
  • --check: Shows the latest version and tells you if an update is available, without downloading or changing anything.

If there is a THIRD-PARTY-NOTICES.txt file beside the program — as there is in a copy unpacked from a release archive — the update brings it up to date with the new release too. C5 does not create one where there was none. See After an update or a rollback.

If an update was interrupted, run growther update again. If the new program was already in place, it answers ✓ Already up to date. and puts the files beside the program back in step with it. It restores the notices file from the program you run, and rewrites the program's checksum file (<program>.sha256, from a website download) only when the signed build record on this machine vouches for the program. Otherwise it leaves the checksum file as it is and warns:

text
  ! <program>.sha256 does not match the running binary, and no signed build manifest on this machine names the binary's hash, so it was left as it is. Run `growther verify`: the binary may have been altered on disk.

A checksum that does not match is also exactly what a damaged or altered program looks like, so C5 never makes the file vouch for a program it cannot vouch for itself. Run growther verify (see Verify what you run). A copy that a package manager installed is left to the package manager.

Just check, do not install #

Want to see if there is something new without changing anything?

bash
growther update --check

It tells you your version, the latest version, and whether you should update.

Update and restart #

When C5 is already running, growther update restarts it for you once the new version is in place:

  • If C5 was running in a terminal or in the background, it gracefully stops the old process and starts the new one.
  • If you run C5 as a background service on macOS or Linux (using growther service install), it restarts that service.
  • On Windows, the update is staged and applies the next time C5 restarts.

To prevent automatic restarts or prompts (for example during a maintenance window), pass --only or set GROWTHER_UPDATE_RESTART=0 in your environment. See growther update for each case.

Update the way you installed #

If you used Homebrew:

bash
brew upgrade growther-c5

If you used the install script, run the same one-line command from Installation again. It will replace your copy with the newest one.

Go back to the older version #

If a new version gives you trouble, you can step back to the one you had before:

bash
growther rollback

Your data is not touched. Only the program file changes, along with the THIRD-PARTY-NOTICES.txt beside it, if there is one, which goes back to the one for the older version. C5 first checks that the older program it kept is still exactly the file it kept and that it still runs; if either check fails, nothing is replaced, and the message says so.

If your secrets are held in a keystore rather than in c5.yaml (see Secrets and keys), an older version may not be able to read them. When your licence is held there, rollback refuses and says why (refusing to roll back: …): run growther secrets revert first, on the version you are running now, then roll back. When only API keys are held there, rollback goes ahead but warns that the providers using them will fail to authenticate on the older version; run growther secrets revert to avoid that. Use growther rollback --force only if you know the older version can read the keystore references (ref: values) in c5.yaml.

The record of the kept program also names its version. So if an older version of C5 later replaced the kept copy without updating that record, rollback can tell a stale record from a damaged copy: it starts the copy, and if the copy reports a different version from the one recorded, it goes ahead and says the record was out of date. A copy that reports the recorded version but does not match is damaged, and rollback refuses. If the install had no signed build record before the update, rollback leaves none afterwards, so growther verify does not judge the older program against the newer release's record.

Downloaded models stay where they are. If the older version has no local classifier, its model simply sits unused in the classifier folder; see Running out of disk space if you want the space back.

Going straight back to v319 #

v319 is the last version without the local classifier. If you go back to it directly from v320 or v322 — with growther rollback after updating from v319, say — v319 may treat the classifier's internal process as an ordinary extension and keep starting it. You see a Local classifier entry under Settings › Extensions. Usually it fails to start and is retried every minute. If the newer C5 program is still installed elsewhere on the machine, v319 can start it outside its own control and offer its tools to your agents.

What to do: after going back, open Settings › Extensions in v319 and switch Local classifier off. Its badge then reads DISABLED, and v319 stops trying to start it. If you switch it back on in v319, the problem returns. Running v320 or v322 again also switches it back on, so do this each time you go back to v319 from one of them.

If you reach v319 some other way than growther rollback — a copy from the release archive, say — you can instead start a version after v322 once before you switch to v319. It switches that entry off, and v319 respects that. This does not help with growther rollback: updating to a newer version replaces the copy that rollback goes back to.

You can skip all this if you are leaving a version after v322: those versions always leave that entry switched off, because they start the classifier another way.

The older version and your databases #

Your data stays as the newer version left it, and a newer version may have upgraded the databases. If you went back to a version after v322, it notices this when it starts: it carries on as usual, and prints a line like this for each such database, to its output and its error log:

text
[migration] growther: the database is at schema v<N>, newer than this build knows (v<M>). It was last opened by a newer C5; upgrade, or restore a backup taken before the upgrade.

growther doctor shows the same on its Schema row. Update C5 again when you can, or restore a backup taken before the newer version first ran (see Backups and recovery). v322 and earlier say nothing about it, so after going back to one of them nothing warns you.

Remote Control Terms after a rollback #

If a newer version brought newer Terms of Service and you accepted them, rolling back to a version after v322 keeps that acceptance: Remote Control does not ask again. v322 and earlier accept only the exact Terms version they carry, so after rolling back to one of them the Settings › Remote Control tab asks for its Terms again at the next save, and the newer version then asks again after you update. See If the Terms changed after you accepted them.

What happens on the first start after an update #

Most updates change nothing you need to act on.

If your shell exports OPENCLAW_GATEWAY_URL for OpenClaw's own command-line tool, versions after v322 read it too: C5 started from that shell dials that gateway, and the OpenClaw card's Host and Port lock. See Local model servers and gateways.

Updating to the release with Terms v1.13 #

The Terms of Service are now version 1.13, effective October 28, 2026. No section is added or renumbered. The only change from version 1.12 is in Section 14, which now says Settings > Classifier in the four places it said Settings > Learning: the classifier's controls moved to a tab of their own (below). Nothing you may do, and nothing C5 does, changes with it.

If Remote Control is on, it keeps running and nobody is signed out. The Settings › Remote Control tab says that version 1.13 has not been accepted on this machine, and the next save on that tab asks someone at the machine to accept it. See If the Terms changed after you accepted them.

The classifier has its own tab, Settings › Classifier, right after Learning, which keeps its other five sections. The classifier's settings now wait for the tab's Save button, as other settings do: the Built-in model switch, each kind of decision's mode, Models allowed from, Use as fallback and the chosen model's own switch. The questions that go with them still come when you move the control. Choosing, testing and removing a model, deleting the built-in model, fitting a calibration, and Act on M's answers still happen as soon as you confirm them. See Saving your changes.

Updating to the release with Terms v1.10 #

The Terms of Service are now version 1.10, effective October 22, 2026. No section is added or renumbered. What changes depends on the version you update from.

From v327, v328 or v329, which carry version 1.9, two sentences change:

  • Section 14: when a classifier model from your providers cannot be reached, the built-in model answers in its place only while its use as a fallback has not been turned off, in Settings › Learning or by your organisation's policy.
  • The Privacy Policy, under Local by default: with a name and certificate you supply, nobody but you and your machine is in the path; a tunnel you supply puts its provider in the path, under your agreement with it.

From v326 or earlier, which carry version 1.8, you also get the changes version 1.9 made:

  • Section 14 also covers a classifier model from your providers, if an administrator chooses one (below). The Privacy Policy says the same: those questions go to that provider, not through Growther.si.
  • Section 3 says that C5 downloads Cloudflare's cloudflared for any Cloudflare Tunnel Remote Control runs, including one of your own.
  • Section 15 says that a tunnel you supply yourself puts its provider in the path.

The changes to Sections 3 and 15, and the Privacy Policy's tunnel sentence, describe what C5 already did. See Privacy and Cloudflare.

If Remote Control is on, it keeps running and nobody is signed out. The Settings › Remote Control tab says that version 1.10 has not been accepted on this machine, even where version 1.9 was, and the next save on that tab asks someone at the machine to accept it. See If the Terms changed after you accepted them.

The Classifier card is laid out around the chosen model. With a model from your providers chosen, its panel now comes first, tagged Primary while the built-in model stands behind it, and has a switch of its own that turns it off without removing it. Stop using it is now Remove. Under the built-in model's switch, Use as fallback (on by default) decides whether the built-in model answers in the chosen model's place when that model cannot. Nothing changes until someone uses them: a model chosen before the update stays on, and the built-in model stands in for it as before. See Turn the chosen model off for a while and Use as fallback. The policy key classifier_builtin_fallback sets Use as fallback across a fleet; write your templates again to get it (see Managed configuration).

The classifier is measured more fairly. Nothing you set changes, but its numbers move:

  • The promotion gate counts the last 200 labelled answers, reaching back up to 180 days when the last 30 do not hold 200, and the gate line says over how many days it measured and how long 200 take at the current pace. Teach the classifier asks about task categories up to 180 days old.
  • Speed counts only for Task categories, the one kind of decision something waits for, and it now counts the questions that timed out as slower than the budget, out of every question asked, so a gate that passed on its answers alone can fail Speed. A question dropped from the background line every kind shares is not counted, and a model that replaced another starts its count where it took over.
  • Only a new label keeps a kind verified. A question answered again on a label it already had no longer re-verifies a kind that acts, so a kind nobody labels any more now enters its 60 days of grace, and administrators are told so. Answer its questions in Teach the classifier to keep it acting. A raise is judged on the newest measurement, the nightly re-check's included.
  • Step quality's agreement line names the reward model where it, not the rubric, gave the usual score.
  • Step quality no longer sees the rubric's score it is compared with, so its gate starts again from the steps recorded after this update.
  • Agreement with the usual model is shown beside what always giving its commonest answer would score.
  • Teach the classifier starts one category question in four blank, and asks about a few of the tool calls the guardian allowed (Tool calls to check).
  • The built-in model stays loaded for an hour after its last question on a machine with 32 GB of memory or more, instead of 10 minutes.
  • Ops › Classifier counts what became of every question from this update on, across restarts, and shows the usual model's own calls.
  • Each kind of decision says where its labels come from, and how many of its questions wait for you, with a button to the HITL tab. Ops' Labels panel puts Askable now beside Unlabelled answers, which used to be called Answered states waiting.
  • A gate with no answers to compare reads nothing to measure yet, not not enough evidence yet.
  • With Use as fallback off, the questions the built-in model is not asked are counted, on its Health line and in Ops, as not asked: Use as fallback is off.
  • Ops, Analytics and Labs can show a chosen model's own record. While a model from your providers is chosen, Whose answers beside the kind of decision switches from the built-in model's charts to that model's, read-only.

See When it starts to act and Teaching it.

Security settings are for administrators. What you set stays as it was, but only an administrator at the machine running C5 can change it from now on: Tool & Web Escalation, System Control, Egress Data-Loss Prevention, the read and write Permitted Directory Paths, Master Tool Authorizations, the State Machine HITL Decay Threshold, Guardian Vetting and Back up secret files for the GitHub backup. Before, anyone who may change settings could change some of them, from a paired device too; now Settings shows them greyed out to a teammate and a paired device, saying why, and a refusal names the setting as Settings does. The same goes for answering a tool request and for the HITL window's Rules and Training buttons. See Settings only an administrator can change.

Remote Control's restrictions are loosened only by an administrator. A teammate can still turn Read-only for remote devices or Read-only restriction for non-admin accounts on, or Enable Remote Control for non-admin accounts off, but only an administrator can undo them. A teammate can now save the tab on an install no administrator has saved yet, which was refused before, and Voice on remote devices is turned on only by an administrator now, though anyone who may change settings can still turn it off. Turning Enable Remote Control for non-admin accounts off now takes effect as soon as it is saved: before, a non-administrator's paired device kept working until C5 restarted. See Who can change these switches.

Master Pause has its own permission. Pausing or resuming all orchestration needs Edit on Orchestration Master Pause Button in Settings › Access, which administrators always have. That permission alone is now enough, and the right to change other settings no longer includes it. See Pausing everything.

Always Allow is for the agent that asked. Choosing Always Allow for a whole tool no longer sets that tool's switch under Master Tool Authorizations to On, which let every agent use it without asking, over any other agent's Always Deny, and made a 1 hour or 10 uses answer permanent. An Always Deny now turns an On switch Off and leaves Prompt alone. A switch an earlier version turned On this way stays On: an administrator should look at Master Tool Authorizations in Settings › Extensions after updating, and set to Off any tool that should ask. See When an agent asks.

Check the Egress Data-Loss Prevention mode. Up to v329, a save of any switch on Settings › Tools stored Off for a mode nobody had chosen, which turned the check off. The card now shows the mode C5 enforces, and a save stores only what changed. If an administrator saved that tab before, open the card: see Data protection.

Tool approvals keep your rules. Approving an agent's request for a tool it lacks no longer turns your Always Deny into a grant, cuts a permanent Always Allow down to a budget, or resets a grant that still has uses left; spent and expired grants are cleared every night. Choosing Always for a tool no longer resets the other tools' switches under Master Tool Authorizations when C5 cannot read them, and a read_url that attaches a page's image asks for approval in an unattended run. See When an agent lacks a tool.

OpenRouter keeps your limits. Nothing you set changes:

  • A chat or task request never goes out without your price caps and provider limits. If C5 cannot read them, it sends the ones it last read, or, with none read since it started, refuses the request and says why. Earlier versions sent such a request with no limits.
  • A save from OpenRouter's Config tab keeps plugin entries with options, which earlier versions deleted. Check Raw JSON if you had set some.
  • Fleet › Assignment never erases OpenRouter's settings when it cannot read them first.
  • The classifier's model list offers the models OpenRouter lists. Its choice no longer has to be one of your Model Selection models, and never changes them. The list is read only where such a model could be chosen, and offers only the models your organisation's policy allows.
  • No agent is put into JSON mode by its name. With the Config tab saved, a planner's requests were asked for JSON, which broke its plans; now they are asked for whatever the work needs.

See OpenRouter routing settings.

Review buttons follow who may decide. A task's own Approve and Deny are offered only to its owner or an administrator, with HITL Edit; the HITL window's Rules and Training buttons only to an administrator at the machine; and Variants to whoever has the Variants grant under Labs in Settings › Access, HITL Edit or not. The Retrospective Loop's switch and settings need HITL Edit, which the permission to change settings no longer stands in for. An Edit grant on any row of Settings › Access now opens what it covers too. See Human review (HITL).

Spending limits work in the installed app. Up to v329 the installed app filed each day's use under a date the Hard Budget Cap and Ops › Token Budget could not read in order: the cap left out most of the spending, the classifier's billed questions included, and could count older days in every period, and today's use read nothing. In a time zone ahead of UTC every period also began a day early, so a daily cap counted yesterday's spending too. On the first start after updating, C5 refiles the use those versions recorded, and the dates its nightly jobs last ran, under the right day, so the cap's total for the current period can change. See How the cap counts.

Nightly upkeep C5 missed is caught up. If C5 was off, asleep or restarting at 23:55, the expiry of unreviewed rule proposals and findings and the Scoring check run once when it next starts, or within the hour, and never twice in one day; before, a night C5 missed was skipped. In the installed app, the Daily Output count of completed tasks, which read 0 there, counts again.

A restart keeps the work under way. A step a model was working on when C5 restarted or quit is left for the next start, never failed or counted as one of its tries, and a guardian check the restart cut short no longer leaves a false entry in Guardian blocks. See Starting, restarting and quitting C5.

From v326 or earlier: the classifier can use a model from your providers. The section in Settings › Learning is now called Classifier, and the local classifier's switch is now Built-in model. An administrator at the machine can choose a model that one of your providers serves — llama.cpp, Ollama, LM Studio, vLLM or a cloud API — to answer the classifier's questions before the built-in model. Nothing changes until someone does: none is chosen, and Models allowed from starts at This machine. See A model from your providers. To keep these models off across a fleet, or to allow them while the built-in model stays off, see Classifier models from your providers.

From v326 or earlier: Tool-call safety and Skill routing are recorded every time. On a machine where the built-in model runs on the CPU, or is marked slow, versions v320 to v326 recorded only one question in four for these two kinds as well as for Step quality. Now only Step quality is sampled there; see On a machine where it runs on the CPU.

From v326 or earlier: old background work expires, and stuck work is closed. On the first start, and every five minutes after, C5 expires the background model jobs — delegation autopsies, skill crystallisation, sparring matches, notification reviews — still waiting 7 days after they were queued. They never run, and stay listed as Expired under Monitor › Queueing › System Jobs; on an install whose queue had stalled, that can be most of them. Requests to improve a task's description never expire. Ops › Triage has a Background LLM Jobs card with the count (see Budgets & costs). A run that could never move again is cancelled with a 🩺 comment on its task (see A task is stuck as "Blocked").

From v326 or earlier: background self-improvement runs again. A check that asked "is real work waiting?" counted jobs that could never run, so on some installs alert triage and the background learning jobs had not run for weeks. They now run in the gaps between real work, so the tokens C5 spends in the background can rise.

From v326 or earlier: unreviewed proposals expire. Rule proposals nobody reviewed within 30 days, and findings about steps that have since been pruned, expire every night at 23:55, or at the next start or within the hour when C5 was not running then. They are kept, never deleted; see Human review (HITL).

From v326 or earlier: the analyst has its own rubric. Analyst steps are graded on a checklist of their own (quality rubric v4), so their scores rise — on the one install measured, from about 20 to about 40 on average — and the analyst's variants in Labs › Variants start again unmeasured. See Analytics.

From v326 or earlier: OpenRouter's price caps and routing take effect. A cap below the price of a routed model now makes the request fail, and a stored 0 waits until you keep or clear it. See OpenRouter routing settings.

From v326 or earlier: llama.cpp models are listed. v326 never listed a llama.cpp server's models, so they could not be chosen; now they can, for agents and for the classifier.

From v326 or earlier: an old debug log may hold the start of task goals. If your C5 folder has logs/categorization-debug.log, earlier versions wrote the first 80 characters of a task's goal there whenever a categorisation failed. C5 now stops writing goals to it and, on macOS and Linux, narrows it to your own account, but does not delete it; an administrator may. See Where your data lives.

From v326 or earlier: one run per scheduled occurrence. A scheduled task now starts once each time it is due, not two or three times.

Going back to an earlier version. v327 to v329 have neither the chosen model's own switch nor Use as fallback: there a chosen model you turned off is asked again, and the built-in model answers in its place whenever it cannot, while it is on. v326 and earlier do not know about a model from your providers: they ignore the choice, and the built-in model answers as it did there. An acceptance of version 1.10 for Remote Control also counts in any version after v322, so going back to one of those asks nothing (see Remote Control Terms after a rollback). v326 and earlier do not know that a job, a rule proposal or a finding expired: they never run an expired job, but Monitor › Queueing shows it with the clock of a waiting one, and expired proposals and findings come back into the Rules and Training queues of the HITL window. In the installed app, v329 and earlier file each day's use under the old dates again, so their Ops › Token Budget does not show the days this version refiled, and an Always Allow for a whole tool given there sets its switch to On for every agent, as it did.

Updating to the release with the local classifier (Terms v1.8) #

If you are updating from a version that had no local classifier, the first start does three things once.

C5 checks the models already on your disk. The memory-search and offline-voice models are now pinned to exact files and checked by size and sha256 fingerprint before C5 uses them. On the first start, C5 reads each model file you already have, once, and records the result beside it. On a slow disk this can take a minute or two. Memory search waits for the check; the rest of C5 starts as normal. A file that turns out not to be the right one is renamed to .rejected, never deleted, and where downloads are allowed C5 fetches the right one in its place. See Memory search and Voice.

The local classifier comes on. This release adds the local classifier, a model of about 812 MB that answers some of C5's routine questions on your own machine. It is on by default, so on this first start C5 switches it on, unless:

  • someone on this install had already switched it off — it stays off; or
  • your organisation's policy sets classifier_enabled — the policy decides.

When C5 switches it on, administrators get one notification saying so, and what will happen on this machine:

This machine What happens
Can run it, and can reach Hugging Face The notification says C5 downloads the model once (about 812 MB)
Already has the checked model The notification says nothing is downloaded and it is ready to use
Has the egress posture set to Offline The notification says nothing is downloaded, and to place the model by hand
Runs with GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD=1 The notification says nothing is downloaded automatically, and to run growther classifier pull
Has too little free disk space The notification says how much room it needs and where. Free some space, then restart C5 or turn the classifier off and on
Has a symbolic link at the model's path The notification says C5 never downloads over a link. If the file it points to cannot be reached (a share that is not mounted), C5 uses the model once it can be. If the link points to a file that is not the model, it stays unused until the link points at the model or is removed. A link C5 has not checked yet is checked before it is used
Cannot run it (platform, memory, operating system or system libraries) No notification, and nothing is downloaded

The download happens in the background after startup and goes through your proxy settings. At first the classifier only records its answers beside the usual ones, for every kind of decision it can answer — task categories, skill routing, step quality, tool-call safety and alert triage each start in Shadow — and it decides nothing until it has been measured on your install. On a machine where it runs on the CPU, step quality, tool-call safety and skill routing record only one question in four, so it stays out of your agents' way. To turn it off, or delete its model, go to Settings › Classifier — see Turn the classifier off. The FAQ answers what it is and what it sends (nothing to Growther.si).

The Terms of Service became version 1.8. Section 14, Models C5 downloads and runs on your machine, is new, so Remote Control moves to Section 15. If Remote Control is on, it keeps running and nobody is signed out. The Settings › Remote Control tab says that the new version has not been accepted on this machine, and the next save on that tab asks you to accept it. See Turning it on and pairing a phone.

Updating from a version that already had the local classifier #

In this release every kind of decision the classifier answers starts in Shadow: asked in the background, its answer recorded beside the usual one and never used. Before, only task categories did, and the other four were Off.

A mode someone on this install chose is kept, Off included, and nothing is raised into a mode that acts. Earlier versions saved every kind's mode as soon as anyone changed one, so the kinds nobody chose could be saved at the old defaults. On the first start, once, C5 looks in its audit log: a kind still at its old default that nobody chose is reset to today's default. The reset is recorded in the audit log as a change made by system. If the audit log no longer reaches back far enough to tell (to 27 September 2026), C5 changes nothing, and while the classifier is on it logs [classifier] site modes kept as stored: the audit trail no longer shows which were chosen, so none is taken for a default. You can set any kind's mode yourself in Settings › Classifier. See What it does at first.

In versions after v322, the promotion gate also checks Telling cases apart: whether the classifier's answer actually changes when the right answer does. If Task categories was set to Assisted or Authoritative before this update and fails this check, the first nightly re-check moves it back to Shadow, and administrators get a notification saying why. Its answers keep being recorded. If an administrator raised it even though its gate had not passed, this check does not move it back. See When it starts to act.

Updates through a corporate proxy #

The update check and the update download both go through the proxy and certificate authorities your organisation set under Settings › Enterprise › Network, like the rest of C5. If certificate pinning is on, it still applies when the connection goes through the proxy.

When a check or download fails, growther update says why — a certificate the pin did not accept, a missing certificate authority, or a proxy that refused, with the answer it gave:

text
the proxy http://proxy.example:3128 refused to open a tunnel to raw.githubusercontent.com:443: it answered CONNECT with 407 Proxy Authentication Required; it wants credentials — put them in the proxy URL (http://user:password@host:port), C5 speaks Basic proxy authentication only

C5's own background check writes the same reason to its error log, and an update started from the app shows it if it fails — in both, the line says the proxy without its address, because anyone signed in can read them. See The proxy address stays out of logs and alerts.

See Network and egress for the hosts updates use, and Common issues for what each proxy answer means.

Every update is checked #

Every release is published with a fingerprint and a signature. Before C5 installs an update, it checks both. If either one does not match, the update stops and nothing is installed.

This means you can only ever end up running a version that Growther.si actually built and signed. See Verify what you run.

Release channels #

Most people should stay on the stable channel, which is the default. If you want early versions, you can switch:

bash
GROWTHER_CHANNEL=beta growther update
Channel Who it is for
stable Everyone. Tested and ready for daily use.
beta People who want new features early and can hit bugs.
dev Testing only. Expect rough edges.
Edit this page on GitHub ↗Last updated Nov 20, 2026