The proposal card lets the owner correct the text before confirming.
POST /api/chats/changes/{id}/confirm now accepts an optional JSON body
{"content": "..."}. The text has the same size limit and text check as a
proposal, the file and the audit hash hold what was written, and a
refused text leaves the proposal waiting.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The Memory settings page needs to know which index and topic files the
host loads into the prompt. GET /api/chats/context-files walks the
registered context sources, expands the agent and channel folders that
exist, adds the agent profiles and reports each file with its size, the
line count of the index, and the caps. It only reads; the page edits and
deletes through the Files API.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The Memory page reads the context-file sources that master added.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
model.unload calls Free over gRPC and keeps the backend process alive.
llama-cpp does not implement Free, so the call answers Unimplemented,
nothing is released, and the next backend.install for the same replica
reuses the live process. The second LoadModel on that process dies a
second after the slots are initialized, and the controller sees EOF or
Unavailable until the following install starts a fresh process.
When Free answers Unimplemented, stop the process after the unload so
the next install starts a clean one. Backends that implement Free keep
their process, so reloads stay cheap for them.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Open the proposal store beside the chats (auth database, file or memory)
and document the mechanism for add-on authors: the wire record, the
limits, the gate, the audit, the REST routes and how a cluster shares
it. Update the registry and actor-scan specs for the new tools.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
List, read, confirm, dismiss and undo the changes a model proposed, and
read their audit. The routes sit under /api/chats with the other chat
routes and need the chats feature. A person reaches only their own
proposals; another person's id is a 404.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
When a scale-up failed, the reconciler tried again on every busy or
pressure tick, which logged a warning and sent a new backend install each
time. The "attempt" field in that warning was the loop index, so it always
read 1 and hid the repeats. A model that no request has loaded yet logged
"no load info" at warn level on every tick.
Track consecutive scale-up failures per model in memory. After a failure,
wait 30s, doubling on each further failure up to 10 minutes, before the
next try. A successful scale-up resets the count. Log the failure count
and the wait. Mark the missing load record with ErrNoLoadInfo and log it
at debug level, because only a first request can fix it.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Say that the agent can attach an existing file, what limits apply, and
that the host sends every file through the send_file op so an add-on needs
no code for it. Ask add-ons to return a short plain message when the
platform refuses a file.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The turn block of a connector chat said that files made in a turn are
attached, but not how to give a file that already exists. Name the
attach_file tool in the block. The block stays under 700 characters.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The host sent only the files a turn made or changed. A person who asked
in a messaging app for a file that already existed got nothing.
Add the attach_file tool. It is offered only in a chat that a connector
started whose binding sends files, so a chat in Studio has no such tool.
It takes a path as the file tools show it, checks the file against the
chat's access (a scoped agent reaches only its folders), the type
allowlist, the size limit and the count, and marks it. When it refuses, the
result says why in plain words so the model can tell the person.
sendFiles now sends the files the turn made, then the attached ones, each
path once and at most five in all. It reads them with the chat's access
resolved at send time. Delivery is the existing send_file op, so every
connector add-on that implements it gets the feature with no code of its
own. A failed send says "I could not send X" and adds the reason the add-on
reports, cut to one short line.
The marks are kept in the memory of the process, like the other
per-turn state of the connector manager.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
read_file and edit_file answered "X.docx is not a text file (43963
bytes)" and gave the model nothing to go on.
Make the error typed and let the chat's tool list add the hint. If the
chat has a skill for the file type (docx, xlsx, pptx, pdf), the error
names it. Otherwise it suggests a script run with the commands tool, with
a library for office types, when the chat can run commands. A chat with
neither is told that no tool opens the file. The skills are resolved only
when a read fails.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
A model asks to change the person's memory or an agent profile with
propose_change. The host stores a proposal and writes nothing until the
owner confirms; the write is one workspace change set, so revert_change
and the Files history can undo it. A proposal expires after 24 hours,
is refused when the target changed meanwhile, and is capped at 10
pending per person and 5 per hour per agent.
The host holds no platform code. The origin of the chat is recorded as
the source, so a Studio chat and the connector chats of any platform
use the same path. A gate decides if a chat may propose; the default
refuses a chat that runs a saved agent until grants fill it. Every step
is audited with a content hash, not the content.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Describe how a component declares a source (scope, path, cap), the index
format, what a turn carries and the limits, and how a chat of a saved agent
reads its own notes without gaining access to the files. Add the agent
profile file to the page on making LocalAI yours.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Each turn the chat prompt gets the agent profile and the indexes of the
registered context sources, after the persona and before the skill index.
The scope of a source is derived from the chat record: the saved agent, and
the binding when a connector started the chat. The text is the same at each
turn, so the transcript does not grow, and it feeds the refresh key so the
harness restarts only when a file changes.
memory_read reads a topic next to an index as the user, outside the chat
fence. A scoped agent reads its own notes and the user notes and gains no
access to the files. The service exposes ContextSources and
SetContextWriter so other parts of the host can register a source or plug in
the save path.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
A source declares a scope (user, agent or channel), a home-relative index
path and a byte cap. The loader reads the indexes that apply to a chat and
returns one bounded block under a fixed header, plus a key that changes with
it. Details of a topic are read next to the index; the path is built from
the registry and the chat, so the model can only name a topic.
The package only reads. A Writer interface lets a later change plug in the
save path, and CheckIndex and CheckProfile return the current entries when a
write would pass a cap, so the model merges them.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Memory and profile changes write to fixed places in the owner's home;
the target a model names is checked against a small pattern so it
cannot leave Memory/. Other features register their kinds. The gate is
the hook the agent grants fill; by default it refuses a chat that runs
a saved agent.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The host must write a confirmed change as the owner, in one titled
change set whose revision it can undo later. Add ReadHome and
WriteHome, which resolve the path inside the home and refuse a path
that leaves it.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Keep what a model asked to change, until the person answers, in a
store with memory, JSON file and GORM implementations and one shared
conformance suite. Create checks the pending cap and the hourly rate in
the same step as the insert, Settle is a compare-and-set so two
frontends cannot both answer, and the audit holds hashes, never
content.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
A model can ask to change the person's own state (a memory note, a
profile, a skill). Add the versioned wire record, its kinds, scopes,
statuses, source and limits, so every surface that runs a model, and
any add-on, shares one type without platform code.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
kong derives a flag name from the Go field name, so the P2PEnableLimits
fields were exposed as --p-2-p-enable-limits. The docs and the help text
describe --p2p-enable-limits. Set the name on the four commands that take
it (run, explorer, and the llama.cpp and MLX P2P workers) and parse the
documented spelling in the CLI boundary test.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Describe the 1.2-second case as success within its configured timeout,
not as exceeding the actual 60-second connector default.
Remove trailing whitespace from the P2P command example.
Assisted-by: nib:gpt-6.1-sol
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Pass the resource-limit choice through application and node configuration.
Keep only the exact string true enabled, including worker and explorer
entry points. Leave other legacy environment reads unchanged.
Assisted-by: nib:gpt-6.1-sol
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude Code:claude-sonnet-5-5
Name the MCP tool explicitly so the Ginkgo-only lint rule does not
mistake its Run method for testing.T.Run.
Assisted-by: nib:gpt-6.1-sol
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The function templates stored the few-shot JSON sample inside a Go raw
string that still had an extra pair of braces. text/template printed
that literal, so the prompt showed {{"example_name": "example_value"}}
and models copied invalid JSON.
Fixes#12890
Signed-off-by: leilei3167 <imleilei123@gmail.com>
(cherry picked from commit 72f1dc25fc)
Root cause: #5889 changed the comparison to `!= "true"` while the result
still feeds ResourceLimit.Enable directly, so the libp2p resource manager
was enabled when the variable was unset and disabled when it was set to
"true" -- the opposite of the docs, and of the debug example that the same
PR added.
Extract newNodeConfig from newNodeOpts so the environment parsing can be
covered by a test.
Fixes#13031
Assisted-by: Claude:claude-fable-5-1
Signed-off-by: Vivek1898 <viveksingh27795@gmail.com>
(cherry picked from commit 404f86194c)
Root cause: #9124 moved the llama.cpp P2P worker under `p2p-worker` and
made `worker` the generic distributed worker, but the hint printed after
`run --p2p`, the Swarm page's command builder and the docs kept the old
`local-ai worker p2p-llama-cpp-rpc`, which the parser now rejects with
"unexpected argument p2p-llama-cpp-rpc". The MLX shard command had the
same stale prefix.
The hint becomes a constant that a test parses against the real CLI
tree, so the printed command cannot drift from the command set again.
Fixes#13032
Assisted-by: Claude:claude-fable-5-1
Signed-off-by: Vivek1898 <viveksingh27795@gmail.com>
(cherry picked from commit 3c11ab96a0)
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
A tool call through an MCP connector was cut after the connector
package default of 60 seconds, while the model's own MCP servers allow
360 seconds. Research and document tools behind a connector take
minutes, so a saved agent lost them mid-turn with "connection closed:
calling tools/call: client is closing" and the model fell back to a
weaker tool.
The connector service now gets the MCP tool timeout the rest of
LocalAI uses (DistributedConfig.MCPToolTimeoutOrDefault, 360 s by
default). The package default and the per-call output cap are
unchanged.
Tests in core/application start the connector service on JSON stores
against a slow MCP server: the timeout defaults to the MCP tool
timeout, a tool slower than the old cap but within the timeout
returns, and a tool past the timeout is still stopped.
Docs: chats-connectors.md, "How chats use them".
Assisted-by: Claude Code:claude-opus-5-5
Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de>
(cherry picked from commit 99f663a777)
getAMDGPUMemory only asked rocm-smi. ROCm 7+ images no longer ship it
(only amd-smi, whose Python module is missing in the image), so on an
AMD Strix Halo (gfx1151, ROCm 10.1 image) GET /api/resources returned
"gpus": null and aggregate.gpu_count 0, and the Models page said
"No GPU found. Fit is measured against 122.7 GB of memory" while
inference ran on the GPU.
The amdgpu driver exposes the same figures in sysfs, readable inside the
container:
/sys/class/drm/card0/device/vendor 0x1002
/sys/class/drm/card0/device/mem_info_vram_total 536870912
/sys/class/drm/card0/device/mem_info_vram_used 240300032
/sys/class/drm/card0/device/mem_info_gtt_total 128849018880
/sys/class/drm/card0/device/mem_info_gtt_used 33055911936
When rocm-smi is absent or reports nothing, walk the DRM tree with the
existing scanSysfsGPUs (cardN only, connectors skipped, vendor 0x1002)
and report every card that has mem_info_vram_total. The GPU report is
built by one helper shared with the rocm-smi parser, so the
shared-memory APU rule from #12058 (GTT >= 8x VRAM folds GTT into the
totals) applies identically. GetResourceInfo, GetGPUAggregateInfo and
the fit check all read GetGPUMemoryUsage, so they now see the GPU;
DetectGPUVendor already consulted sysfs.
Assisted-by: Claude Code:claude-opus-5-5
Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de>
(cherry picked from commit bf222cadba)
A chat started from a messaging app has its answer read in that app, and
the host attaches the files a turn makes to it. The model did not know
this: it said it could not send files and quoted a sandbox path.
Add a short block to the turn prompt of connector chats. It names the
platform, says that files saved in the chat folder are attached after the
answer, gives the limits the host enforces, and says that files are not
sent when the binding turns them off. If the binding cannot be read the
file part is left out. Other chats are unchanged.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Lists the specs added by the connector error, upgrade state and wire
payload changes in the reports of the gateway, artifacts, lifecycle and
apps suites, so verify.sh finds the inventory equal to the reports. The
counts were checked against fresh Ginkgo reports of both packages.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
AddonRegistryProvider called requests.setScope(actor) and replaced
authority.current in the render body. React may start a render and throw
it away (concurrent rendering, StrictMode), so a discarded render for a
new session epoch aborted live reads and reset the list of the actor
that is still on screen.
Render now only reads the ref: for a new actor it shows an empty list at
once. A layout effect commits the new scope and the ref. It runs at
commit, before the browser paints and before any passive effect, so no
read of the old actor can publish in between.
The hook harness of the unit tests gains useLayoutEffect and a
discarded-render mode, and a test shows that a discarded render leaves a
live read alone. The private-access form test expects the phrase in the
owner patch, as the owner now checks it.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
suites.json kept the required cases of a Ginkgo suite twice: once at the
top level and once per report. The two copies drifted in #13034 and broke
the CI gate on master.
For a suite with reports, verify.sh now builds the required cases from the
reports, and refuses a suite that still has a top-level list. Suites with
a single report keep their one list. The redundant lists are removed from
suites.json, and prerequisites_test.sh covers the refusal.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The test for "the same pinned transition under another operation id" was
written out three times in manager.go and upgrade_operation.go, and the
"completed or failed" phase test five times. They are now sameTransition
and isActive.
OwnerHandler rejected a request that carried a second routing payload with
a hand-written list of nil checks per branch, each list slightly
different. wireRequest.exclusive counts the payloads once. The earlier
branches return, so every branch rejects the same requests as before.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
upgradeTransition ended in panic for a phase or event it did not know.
The phase of a submitted operation is read from the journal on disk, so a
corrupted file would crash a background goroutine. The reducer now
returns errInvalidUpgradeTransition, and submit marks the operation as
failed and returns ErrOperationFailed. The current call paths only pass
known phases, so this is a guard, not a fix of a known crash.
upgradeViews also grew without bound, and every persist walked it. Views
of finished upgrades are now kept for the newest 256 only; the journal
still answers for older ones. The journal itself is rewritten on each
persist, so the walk stays proportional to what is written anyway.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
ConnectorCall, localConnectorCall and Carrier.connectorRequest each set
their own 45 second timeout. On a frontend the first and third nested
inside each other, and on the owner the first and second did. Each hop
now sets it once (connectorTimeout): the forwarding carrier on a frontend,
and localConnectorCall on the owner, which must bound the work whichever
door the call came through. ConnectorCall no longer adds a third one.
The kind lookup in localConnectorCall uses slices.IndexFunc instead of a
loop that kept scanning after a match.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
localConnectorCall and the kinds lookup turned every error from the
connector source into a bare ErrUnavailable. VerifyTokenWithMe treats
that as "could not tell" and saves the binding, so a definite refusal by
the add-on was indistinguishable from an outage, and the cause was lost.
An add-on refusal (invalid request, or denied for this person) now keeps
its type through Refused. Other errors still read as ErrUnavailable but
wrap the cause, so errors.Is and logs see it. The forwarding carrier maps
the owner's 400 back to ErrInvalidRequest instead of ErrUnavailable.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
With add-ons enabled /api/settings rejects chats_egress_allow_private, so
the typed phrase check there never ran on the only write path, the
host-settings-update action. The page asked for the phrase but dropped it
from the patch, and a direct admin call could turn the switch on without
it.
The patch now carries chats_egress_private_confirm and hostsettings.apply
rejects turning the switch on without it. The phrase is hashed into the
operation receipt like the rest of the patch and is never stored. The e2e
fixture enforces the same rule and the egress spec asserts the phrase in
the owner patch again.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Add a table of what runs in the combined, owner and frontend roles, and
explain that an owner must join distributed mode with the frontends.
Describe pull and push connector delivery, the one-connection-per-Slack-app
rule and the reinstall needed for new Slack scopes. State what the Sandbox
needs to report an isolated backend, and that gVisor needs no /dev/kvm.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
sdk.ui gains ModelPicker (multiple, allowCustom, clearable, field
variant, alias tags). HomeModelPicker and ModelSelector stay as names for
the same picker so existing add-ons keep working. The add-ons guide
lists the new component and its props.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Settings, config fields, the failover and router editors, the agent task
form, the scheduling rule sheet and the agent job form now use
ModelPicker. SearchableModelSelect, ModelMultiSelect and the scheduling
ModelPicker are removed; ModelSelector stays as the name add-ons use and
draws the shared picker. Aliases come from the capabilities listing, so
the scheduling sheet no longer builds its own hints.
A broken alias is listed, marked and not selectable. The picker's style
sheet now loads with the picker, so the e2e specs that read a settings
group right after opening it wait for its first field.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Move HomeModelPicker to components/ModelPicker and give it what the
other pickers had: multiple selection with chips, free text, a clear
button, a search box on demand, a form-field look, and loading and
empty states. Every entry shows 'alias of X', disabled, pinned or
broken-alias tags from the capabilities listing, and entries that
cannot be used are not selectable. Auto-select skips them.
Add-ons keep HomeModelPicker as a name for the same component.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
An alias entry now takes its capabilities, backend, MCP, voice cloning,
upscale scale, 3D operations and loaded state from the model it points
to, and keeps its own name, disabled and pinned state. An alias that
cannot be resolved (unknown target or a chain) is logged at warn and
listed with alias_of and alias_error instead of vanishing.
The chat model checker refuses a broken alias when it is saved.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The controls sit in a fieldset, which drew the browser border and
padding around the rows. The rows also showed raw setting keys such as
chats_egress_ceiling next to their labels.
Reset the fieldset, hide the raw keys in the add-on controls, and make
the row label fall back to the key in words, so a snake_case key is never
a label. Add e2e and unit tests that fail when a raw key is drawn.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
A smooth scroll to the bottom from far up reports a distance above the
pin threshold on its way down. The scroll handler read that as the reader
scrolling up and unpinned the thread, so the new message and the answer
that streamed in after it stayed out of view.
Only a move up unpins the thread now. A reader who scrolls up while an
answer streams is still left where they are.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The Network row of the Tools menu could never show the Off, Registries and
All switcher: the host stopped setting `applies` because isolation alone does
not prove network-policy enforcement, and nothing else set it.
Let the provider report enforcement in its typed execution status
(`networkPolicyEnforced`). The host sets `applies` only when that report is
true and the status is available and isolated. The Sandbox reports it for the
gVisor and KVM backends, whose guests reach the network only through the proxy
that asks the host. Providers that do not report it keep the "not enforced"
text.
When the host cannot read the status, give the unknown state a reason
(`unknown_reason`) in the API and the Tools menu: owner unreachable, not
reported, stale, invalid, not selected, no answer or reported unknown. Log the
class of each failure, rate limited. After a failure that a restart of the
add-on owner can cause, let a long turn bind and read again, at most once every
five seconds, instead of keeping the first failure for the whole turn.
The Sandbox now reports "unavailable" with a reason when it started without
settings, and explains a backend it does not know.
When the Sandbox document skills are hidden because the Sandbox is not known to
be isolated, say so in the skill list, in the turn note and in the
`load_skill` error, instead of "This chat has no skills now".
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
With mcp_connectors_allow_private on, a connector used a plain transport
with the environment proxy and no address check, so its URL could reach
cloud metadata addresses. Build the private path on the safehttp guard
with private addresses allowed: it resolves once, checks every answer,
and keeps the metadata addresses blocked. An optional DenyAddrs hook
refuses further addresses, for example the server's own.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Add the Azure wire server (168.63.129.16), Oracle (192.0.0.192) and
Alibaba (100.100.100.200) addresses to the metadata list. The Azure one
is public, so web_fetch reached it. Check the IPv4 address inside a
64:ff9b::/96 address too, so private addresses and metadata addresses
cannot be reached through NAT64 when private addresses are allowed.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
A call with no timeoutSeconds ran for the 120 second default even when
an administrator set a lower maximum. Use the smaller of the two.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Ettore Di Giacinto