Skip to contentdocs.Growther.si
Download C5

Common issues

Start here when something is wrong. If your problem is not listed, run the checkup:

bash
growther doctor

It checks your whole setup and usually names the cause. See Running doctor.

Installing and starting #

"growther: command not found" #

Your terminal cannot find C5 yet.

Close the terminal window and open a new one, then try again. Shells only pick up PATH changes in new sessions.

If that does not work, reinstall — see Installation.

C5 will not start #

Check whether it is already running:

bash
growther status

If it says running but you cannot reach it, stop and start it:

bash
growther stop
growther

"BOOT ABORTED: CRITICAL LICENSE FAILURE" #

C5 stops rather than run unlicensed, so it exits instead of staying up in a half-working state. Read the line under the heading — it says which case you are in.

"bound to a different deployment key" — your licence is fine; the key on this computer is not the one it expects. This happens after restoring from a backup, cloning a machine, or moving to a new disk. Give this computer a new key, keeping the same licence:

bash
growther rekey

If that reports the key is gone or is not the bound one, use recovery, which confirms in your browser:

bash
growther rekey --recover

"GROWTHER_C5_LICENSE_SEED is missing" — no licence file. If this computer was never paired, run growther activate. If it was paired and the file has gone, growther rekey --recover restores it without creating a second deployment.

Do not run growther activate to fix a key problem on a computer that is already paired. Activate creates a new deployment and leaves the old one stranded; rekey keeps the one you have, along with its history.

Not sure which you are looking at? growther doctor reports the licence and key state in plain terms.

"Port already in use" #

Something else on your computer is using port 4299. Use a different one:

bash
PORT=5000 growther

The page will not load in my browser #

Confirm C5 is running with growther status, then go to http://localhost:4299 directly. If you changed the port, use that number instead.

If growther status says C5 is running but the browser still cannot reach it, try http://127.0.0.1:4299 and http://[::1]:4299. localhost is two addresses, and C5 listens on both — but if one of them was unavailable when C5 started, the boot log says so:

text
[bootstrap] no ::1 listener (EAFNOSUPPORT) — C5 is reachable on 127.0.0.1 only. If a browser resolves localhost to ::1 it will need http://127.0.0.1:4299 instead.

Use the address that line names. See Why C5 listens on two addresses.

Restart is greyed out or C5 is offline #

When C5 is not running, Restart is disabled because there is no running server to reboot.

To start C5 again:

  • Click the green Start C5 button in the sidebar menu, on the Settings → System page, or on the offline recovery screen.
  • Double-click the Growther.si C5 shortcut on your desktop or Applications folder.
  • Run growther start (or growther) in your terminal.

Once C5 starts, your browser tab reconnects on its own within a few seconds and the controls become active again.

Self-help: Missing "Start C5" button on offline screens? The green Start C5 button on the offline screen and sidebar menu is governed by an administrative RBAC toggle. If your organization has disabled Sidebar: Restart, Quit, Start C5 in Settings → Access, this button remains hidden for standard users. Ask your administrator to enable the permission for your account.

"Forbidden: Guardian Vetting is admin-only" #

The save touched a setting only an administrator can change, named as Settings names it — see Settings only an administrator can change. Nothing in that save was stored. Settings greys these out for anyone else, so this usually comes from a page opened before an update: reload it, and ask an administrator to make the change at the machine running C5. The related refusals, for Remote Control, remote voice, a policy lock, a paired device or an empty value, are under When a change is refused.

Models #

"No model configured" #

You have not connected a provider yet. Go to Settings → Integrations and add one. See Model providers.

"Invalid API key" #

Nearly always a copy-paste problem. Copy the key again, watching for a trailing space. If it still fails, make a new key with the provider.

"Rate limit reached" #

You are asking your provider for more than your plan allows. C5 slows down and retries by itself, so this usually resolves.

If it keeps happening, raise your limit with the provider or route some work to a different model.

OpenRouter requests fail after updating #

From v327, OpenRouter's price caps and routing settings take effect; before, C5 sent the caps under names OpenRouter does not read. A Max Prompt Price or Max Completion Price below the price of the model a request is routed to now makes it fail, and 0 allows free providers only. Check them on OpenRouter's Config tab in Settings, and keep or clear a stored 0 there. See OpenRouter routing settings.

"C5 cannot read your OpenRouter routing limits …" #

A chat or task request to OpenRouter was refused because C5 could not read your price caps and provider limits, and had not read them once since it started. C5 never sends such a request without them. Once it has read them, a later fault sends the limits it last read instead. Open OpenRouter's Config tab in Settings to check that they load, and try again. See If C5 cannot read these settings.

A coordinator shows "Skipped — repeated failed generations" #

Its recent answers came back empty or broken, so C5 tries it last for a while — a minute at first, doubling each time it fails again, up to 15 minutes — and asks the others first. It is tried again on its own when the time runs out; saving the provider again does not change that, and restarting C5 clears it. If it keeps happening, check the model it runs. See When a model is unavailable.

A coordinator shows "Tried last — repeated gateway errors" #

Recent requests to it got no answer at all: its gateway returned server errors, refused the connection, or was down. C5 tries it after every other coordinator for a while, on the same clock as Skipped. Its status check can still answer, which is why the card may read Online beside the chip. Check that the gateway or server behind it is running and healthy. The chip goes on its own once the block ends; restarting C5 clears it too.

Results got worse #

Check what changed. Analytics shows quality over time — find the day it dropped and think about what you changed then. A model switch is the most common cause. A change in how C5 grades a role can move its scores too, as the analyst's rubric did in v327; see Reading the charts.

A "Scoring check" notification #

Scoring check: steps fail "" almost every time means nearly every recent step of that role failed one check of its rubric. That usually says the check does not fit what the role is asked to do, not that the work is bad. Nothing changes on its own; see When one check fails almost every step.

Downloaded models #

C5 downloads up to five models from Hugging Face, each only for the feature that needs it: three for memory search (about 2.26 GB), one for offline voice (about 60 MB), and one for the local classifier (about 812 MB). Each is pinned to one exact file and checked against its size and sha256 before C5 uses it. The full list, with each model's licence, is in Open-source licences and downloaded models.

Memory search says "query limited" or "reduced capability" #

A startup line such as [mcp] QMD query limited without download requirements — running (bundled runtime), keyword search and document tools work; … means C5 is running normally. Keyword search works; semantic search is waiting for one or more of its models. The rest of the line says which model is missing, still being checked, or not the right file.

  • Missing: C5 fetches it in the background after startup. If your network cannot reach Hugging Face, place it by hand — see Placing the models by hand.
  • Being checked: wait. On the first start after an update, C5 reads each model file once to check it, which can take a minute or two on a slow disk.
  • Not the right file: see the next entry.

growther qmd-run pull fetches whatever is missing and says exactly what it did.

A model file was renamed to .rejected #

C5 found a file at a model's name that is not the model this release records — a different upload, a cut-off copy, or a page a proxy saved in its place. It renamed the file to <file>.rejected so it is never loaded, and said so in the log. Nothing was deleted.

Where downloads are allowed, C5 fetches the right file in its place by itself. If automatic downloads are switched off with a GROWTHER_*_NO_AUTO_DOWNLOAD variable, run growther qmd-run pull (memory search) or growther voice pull (voice). If the egress posture is Offline, place the right file by hand — the memory search and voice pages give each file's address, size and sha256. Once the feature works, delete the .rejected file to get the space back.

If the log says C5 could not set a file aside, stop C5, move the file away yourself, and start C5 again. On Windows this happens when another program has the file open.

The local classifier works differently: a wrong file at its model's name is left where it is and never loaded, and the next download — automatic, or growther classifier pull — replaces it with the right one. Settings › Classifier says so: The file at the model's path is not the model (…), so C5 does not use it. On an offline install nothing replaces it, so delete it (an administrator at the machine can press Delete the file on the card while the classifier is off) and copy the right file in; see Offline and air-gapped installs. A symbolic link at the model's name is never downloaded over: point it at the model, or remove it.

The local classifier says this machine cannot run it #

Settings › Classifier shows Unsupported for its Built-in model and a reason, growther doctor shows the same reason in its Local classifier row, and C5 downloads nothing. This is not a fault: C5 works exactly as it did before the classifier existed, and an administrator can still choose a model from your providers to answer the classifier's questions instead. The reasons you may see:

Reason What to do
No local classifier build ships for your platform, or C5's release for it is experimental (Linux on ARM) Nothing; the classifier is not offered there
This machine has less than 8 GB of memory (C5 counts the memory the system can use, rounded to a whole gigabyte: 7.5 GB or more passes. Most 8 GB machines qualify; one that keeps more for built-in graphics, such as one showing 8 GB memory (7.4 GB usable), does not) Nothing, unless you can add memory
The classifier's runtime needs a newer macOS, or a newer glibc on Linux, than this machine has Update the operating system, then restart C5
The classifier's runtime needs a libstdc++ with a newer GLIBCXX_ version (Linux) Update the system's C++ runtime (usually by updating the distribution), then restart C5
The classifier's runtime needs the Microsoft Visual C++ 2015-2022 Redistributable (Windows) Install it from the link in the message, then restart C5

Run growther classifier status to see the same verdict in a terminal. More on each reason is in Machines it cannot run on.

The local classifier is on but has no model #

The card shows No model or explains why the model is not being downloaded on this machine. The usual causes:

  • The egress posture is Offline. Nothing is downloaded. Place the model by hand: the card and growther doctor give its download address, exact path and sha256, and Offline and air-gapped installs walks through it.
  • C5 runs with GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD=1. Run growther classifier pull.
  • There is not enough free disk space. C5 needs room for the remaining download plus 64 MB to spare. Free some space, then restart C5 or turn the classifier off and on.
  • The model's path is a symbolic link whose file cannot be reached. The card says The model's path is a link whose file cannot be reached (a share that is not mounted?). C5 uses the model once that file can be reached, and never downloads over the link. Mount the share or restore the file, or remove the link to have C5 fetch the model.

After growther classifier pull or placing the file by hand, you do not need to restart C5: open Settings › Classifier, and C5 checks the new file and starts the classifier. (C5 looks for a new model file whenever that page asks for the classifier's status.) Otherwise it picks the file up at its next start.

Model downloads or updates fail behind a corporate proxy #

C5 sends update checks, update downloads, Mothership calls, the Remote Control tunnel program and every model download through the proxy and certificate authorities your organisation configured under Settings › Enterprise › Network.

Updates. When growther update fails, the message says why. In your terminal, a proxy that refuses reads like this:

text
the proxy http://proxy.example:3128 refused to open a tunnel to api.growther.si:443: it answered CONNECT with 403 Forbidden; its policy does not allow this host — ask for it to be allowed
The proxy answered What to do
407 Proxy Authentication Required Put the credentials in the proxy URL (http://user:password@host:port). C5 supports Basic proxy authentication only
403 Forbidden Ask for the host named in the message to be allowed
502, 503 or 504 The proxy could not reach the host itself; try again later or ask your network team

C5's own background check writes the same reason to its error log, and an update started from the app shows it — there, the line says the proxy without its address. To see which proxy C5 uses, run growther doctor (its Proxy reachability row) or, as an administrator, open Settings › Enterprise › Network. See The proxy address stays out of logs and alerts.

Model downloads. A model download that fails on the network says why in the same way: what the proxy answered, or the certificate C5 does not trust. Memory search and voice give it after fetch failed:; the local classifier gives it straight after Could not download the classifier model:. Memory search's line in C5's own output names the proxy, and so do growther qmd-run pull, growther voice pull and growther classifier pull in a terminal:

text
[mcp] QMD model download did not complete (embedding: fetch failed: the proxy http://proxy.example:3128 refused to open a tunnel to huggingface.co:443: it answered CONNECT with 403 Forbidden; …). …

The card in Settings › Classifier leaves the proxy's address out, and says what starts the download again once the proxy is fixed:

text
The download failed: Could not download the classifier model: the proxy refused to open a tunnel to huggingface.co:443: it answered CONNECT with 403 Forbidden; its policy does not allow this host — ask for it to be allowed. Once that is fixed, turn the classifier off and on, restart C5 or run `growther classifier pull` to download it; or fetch … on a connected machine and place it at … (sha256 …).

If the reason says C5 refused to connect … the configured proxy … cannot be used, the proxy setting itself is broken: the Network card under Settings › Enterprise reads Proxy unusable — connections refused. Fix or remove the proxy URL; see Corporate proxy.

Check that the proxy allows huggingface.co and its download hosts — see Hugging Face redirects to a download network. An administrator can press Test connectivity under Settings › Enterprise › Network to ask each host C5 contacts, those included, through the proxy and see which answered.

A certificate error behind a proxy that inspects TLS means C5 has not been given the proxy's root certificate. See Behind a TLS-inspecting proxy.

Tasks #

A task is stuck as "Blocked" #

The card says why.

  • It is waiting for your decision. The card shows what it is asking, with a note box, Reject and Approve. Write a note, then choose; see A decision will not submit. Answer it rather than dragging the card.
  • It stopped. The reason is in a box under the card's title. If it is cut short, hover or tap it to read the rest; the task window shows all of it under Why this is blocked. Many reasons also say what to do next. Otherwise, once you have dealt with the cause, drag the card to In Progress to pick the work up now, or to To Do to send it back through the queue.

If the card says nothing — one blocked before v340, for example, or one you dragged there yourself — open it and read its latest comments. Work that passed C5's review and is waiting for your approval goes to Review instead, with Human Review Required inside it.

If its comments end with a 🩺 note, C5 found the run stuck for good: 🩺 Run run_… cancelled: nothing in it could ever run again., followed by why — for example, review steps waiting on steps that were never created. C5 checks for such runs every minute, once they are an hour old, and cancels them. A task that was still in progress is then Blocked for a person to decide; a task waiting to start is picked up again with a fresh run, when it is one C5 runs on its own. The reason is also under Latest Error in the job's details in Monitor › Queueing, and Ops › Audit lists what C5 did, under Stuck-run recoveries in its Source column.

A decision will not submit #

Approving or rejecting a task needs a reason. A sentence saying why is enough. Until you write one:

  • On a blocked card, in the task window and on the Tasks tab of the HITL window, Reject and Approve stay greyed out, and pointing at Approve says Enter a review note before approving. Approve All (N) waits for a note on every task.
  • For a task in Review, the card's Review… menu keeps Deny and Approve greyed out the same way, and Human Review Required in the task window answers Reasoning required — please enter your reasoning before submitting.
  • In chat, put the reason after the decision: /c5 review 2 approve It covers every point we asked for.

Without permission to decide, the HITL window's note box and Human Review Required say so instead: Needs HITL edit permission to approve or reject tasks. In the task window, a task someone else owns may say Only this task's owner or an administrator can approve or reject it. See Who can use it.

A task keeps failing #

Open it and read the error. The usual causes:

  • A file it needed moved or was renamed
  • An expired key
  • A budget limit reached
  • A website or service it needed is down

Fix the cause, then click Retry.

"Task created, but its pipeline hand-off failed …" #

You may see this after Create Task: Task created, but its pipeline hand-off failed. The pipeline will start on its own within a few minutes. The task was saved; only the request that starts its work did not get through. You do not need to do anything. If you created the task in another column, its work already started when you created it. If you created it in To Do, C5 starts the work itself within a few minutes.

Leave the task in To Do until then. C5 starts it only while it is in To Do with AI Enabled on: if you move it, turn AI off, archive it or delete it, C5 leaves it to you.

If it is still in To Do after ten minutes, see Everything is queued and nothing starts.

A sub-task never starts #

A sub-task — a card with a yellow id badge — never starts on its own:

  • It is still in the Pending sub-tasks block at the bottom of To Do. Nothing there runs. Drag it up into To Do to approve it.
  • It is approved, but AI is off. Sub-tasks are created with AI Enabled off, and an approved one with AI off waits in To Do for people. Open it, turn AI Enabled on, and click Save; it is then picked up in the normal task order.

See Sub-tasks.

No sub-tasks appear #

Orchestration only creates sub-tasks while Create Sub-Tasks is on in Settings › General › Tasks, and it is off by default. See the Tasks panel.

"That task is not available …" #

You followed a link to a task — for example the Parent task link on a sub-task's id badge — that you cannot open: That task is not available. It may have been removed, or you may not have access to it. The task was removed, or it is one you are not allowed to see. Nothing else changes; ask its owner or an administrator if you need it.

Everything is queued and nothing starts #

Open Monitor → Queueing.

If nothing is moving, check that you have a working model provider, that you have not hit a budget limit, and that no task is waiting on your approval.

System Jobs shows only Expired. That is normal after an update: a background job still waiting 7 days after it was queued expires without running, and stays in the list. Expired jobs never run. See When work seems stuck.

Only one task step runs at a time #

Open Settings › General › Adaptive Parallel Agents. Concurrent Agents is how many steps may run at once on local models and gateways; Cloud Concurrent Agents adds steps on paid providers, but only while a paid provider is first in Task Priority (Settings › Integrations › Coordinator Priority). If a paid provider is connected but not first, the cloud slider's card says so and it adds nothing.

The LLM Queue badge on Productivity shows what is in use: one pair behind a server icon for local models and gateways, and, once a paid provider is first, a second pair behind a cloud icon. If the badge shows 1/1 with a slider set higher, the slider is in automatic mode and C5 has backed off — the card shows the live value — or a paid provider is not first.

A schedule did not run #

Schedules need C5 running. If your computer was asleep or off, the run was missed — C5 catches up when it wakes.

If C5 was running and it still did not fire, open the schedule's history for the error.

A "Use Recipe" schedule never did anything #

In v341 and earlier, a Use Recipe schedule did nothing when it was due: it created no task and reported no error. Later versions fix this. Each time it is due, it creates a task on the board, as Create Task does, and that task follows the recipe chosen in the schedule. Your existing repeating Use Recipe schedules need no change: each starts working at its next run after you update. A one-off that already ran will not run again; see What each action does.

Performance #

Everything is slow #

Check Monitor → Health. If your fleet is large, your computer may be overloaded — try running fewer agents. See Your agent fleet.

Running out of disk space #

Open Monitor → Database only if you suspect the databases. They are usually not what filled the disk. The bulk is almost always downloaded models and caches. They sit in five folders under your C5 folder (~/.growther by default), outside the databases entirely:

Folder What it holds Typically If you delete it
qmd/ The memory-search engine, its three models (in qmd/cache/qmd/models/), its search index of your folders, and C5's own checked copies of the models (in qmd/model-view/checked/) about 2.3 GB, and up to about 2.3 GB more for the checked copies on a disk that cannot clone files. growther cache show does not count the copies To free model space, delete only qmd/cache/qmd/models/: C5 downloads the models again at its next start, unless downloads are off. Deleting all of qmd/ also deletes the search index, and memory search finds nothing in your folders until they are indexed again: an administrator presses Sync Now on the QMD Data Folders card under Settings › Locations. See What is safe to delete
classifier/ The local classifier's model about 812 MB Downloaded again while the classifier is on. Delete it from Settings › Classifier instead — see below
speech/ The offline speech model about 60 MB Downloaded again at the next start while Built-in (on-device) voice is on
cache/ Working files varies Recreated as needed
lib/ Unpacked runtime libraries varies Unpacked again at the next start

Delete these folders only while C5 is stopped. "Unless downloads are off" means the egress posture is Offline, or C5 was started with one of the GROWTHER_*_NO_AUTO_DOWNLOAD switches; on such an install a deleted model stays gone until you place it again.

To remove the local classifier's model, use Settings › Classifier. An administrator, signed in at the machine itself (not from a paired phone), can press Delete the model (… MB) on the card while the classifier is off, or Turn off and delete the model in the question that turning it off asks. (For an unfinished download the button reads Delete the partly downloaded model, and for a model placed as a link to a file elsewhere, Remove the link to the model.) Turning it off first is what keeps C5 from downloading it again. If your organisation's policy keeps the classifier on, the delete is refused, because C5 would only download the model again. See Delete the model for who can delete it and what is removed.

growther cache show reports what each folder is using, and growther cache relocate --to <path> moves all five to another disk without touching your work. If the databases really are the problem, shortening how long history is kept is the biggest lever — see Backups and recovery.

Costs #

I spent more than expected #

Open Ops and look at spending by task and by model.

The usual culprits are a frequent schedule, a large fleet, or an expensive model doing simple work. Set a budget so it cannot happen again.

Serious problems #

"Integrity check failed" #

Danger The program on disk is not the one Growther.si signed. Stop using it, reinstall from the official installer, and if it fails again ask for help before running it. See Verifying releases.

An update broke something #

Go back to the previous version:

bash
growther rollback

If you are going straight back to v319 from v320 or v322, also switch off Local classifier under Settings › Extensions in v319. After any rollback, the older version may warn that a database is newer than it knows. That is expected. See Go back to the older version.

Then tell us what happened — see Getting help.

I think I lost data #

Do not keep working in that install — that can overwrite what is recoverable.

C5 keeps automatic backups. See Backups and recovery for how to restore, and test the restore into a separate folder first.

Edit this page on GitHub ↗Last updated Nov 20, 2026