Projects created before GNS3 3.1 have no created_by field. The project
list filtered them out for every non-superadmin user, so all old
projects disappeared from the list after an upgrade.
Fall back to the pre-3.1 ACE-only visibility rule when created_by is
not set. Projects owned by another user stay hidden.
pytest 9.2 reads the old value of a monkeypatched attribute from the
instance __dict__. Starlette State keeps attributes in an internal
dict, so the old value looks unset and the teardown deletes the
attribute for the rest of the test session. The first test that
touches app.state.exiting passes but removes it, and every later
monkeypatch.setattr on it fails with AttributeError.
Use plain assignments instead.
By default uvicorn waits for open connections forever. A long-lived
streaming response keeps the server waiting after Ctrl+C. One common
case is the MCP Streamable HTTP transport: the client opens a
standalone GET request and the server answers with an SSE stream
(not the old SSE transport). A second Ctrl+C sets force_exit. Uvicorn
then skips the lifespan shutdown, so project and controller cleanup
does not run.
Add timeout_graceful_shutdown. After the timeout uvicorn cancels the
remaining tasks and runs the lifespan shutdown. Also make the four
notification WebSocket handlers stop when app.state.exiting is set.
The HTTP stream handlers already do this.
A field log showed one request stayed silent for 604s (the OpenAI SDK
default timeout), then the SDK retried on a fresh connection and got
200 OK in under a second. The abort signal did not cancel the in-flight
HTTP call, so the turn could look hung for ten minutes.
- set timeout (default 120s, request_timeout in the LLM config wins),
max_retries=1 and stream_usage=True for the providers that accept them
(openai, deepseek, anthropic, xai). ollama and google-genai reject
these kwargs, so they keep the old behavior.
- log the elapsed time of each LLM call. A slow first token or an SDK
retry now shows up in the log line.
- add request_timeout to the LLM config schemas and docs (JSON column,
no migration).
- tests: cover the defaults, the override, invalid values, provider
filtering and the title model.
The graph passed remaining_steps=20 (50 for injection) in the run inputs.
RemainingSteps is a LangGraph managed value derived from recursion_limit,
so the input was ignored and the real budget was 25 (100 for injection).
No one could see those numbers, and a turn that hit the cap ended on a
bare tool result with no explanation.
- Remove the dead remaining_steps input.
- Add resolve_recursion_limit(): max_steps from the LLM model config wins;
otherwise 25 teaching_assistant, 50 lab_automation_assistant, 100
troubleshooting_injection.
- When fewer than 4 steps remain after a tool run, route to the new
budget_wrapup node. The next llm_call binds no tools and gets a wrap-up
instruction, so the turn ends with a short summary of what was done and
what remains.
- should_continue skips the title call after a wrap-up (no step left).
- budget_exhausted is reset at the start of every user turn.
- Add max_steps to the LLM config schemas. The config is a JSON column,
so no migration is needed.
- Tests: 16 new tests cover the resolver, the routing, the tool-free
wrap-up call and the removed input. Docs updated.
The bridge made 49 project-scoped MCP tools available, but the mode
tables still mirrored the old copilot capability set. Expand them:
- lab_automation_assistant: node_console_exec (raw console for nodes
without a device_type tag), node_list, snapshot_list/create/delete/
restore
- teaching_assistant: node_list/node_get, link_list, link_marker
(read-only diagnostics plus traffic-insight markers)
- troubleshooting_injection: link_capture_start/stop/download and
marker_definition (inject -> capture -> verify loop)
Also fix the tools_by_name registry: it was built from the lab and
injection tables only, which worked while teaching was a strict subset
of lab. Teaching-only tools would have raised KeyError in tool_node
when the LLM called them; the registry now unions all three mode
tables. A new test pins each mode's exact toolset, checks that no
administration tool leaks in, and asserts every referenced name is
resolvable.
The copilot and the MCP service maintained two parallel wrapper layers
(LangChain BaseTool classes vs @mcp.tool() functions) over the same
handlers in gns3_client/api_handlers.py. Descriptions and parameters
drifted apart (compute_id and link_type were missing on the copilot
side). This makes the MCP registration the single source of truth and
removes the duplicated wrappers.
tools_v2/mcp_bridge.py adapts the MCP server's tool registration into
LangChain tools: names, descriptions and argument schemas come from the
MCP tools. The pool is filtered to project-scoped tools
(MCP_ALLOWED_TOOLS) because the copilot runs inside one open project —
project management, compute/appliance/image administration and server
info tools stay out. The bridge carries the copilot's request JWT into
the MCP contextvars (username/token_version decoded from the token so
ticket-minting tools work) and runs the async MCP functions from the
copilot's sync tool node on either thread flavor.
Mode toolsets (teaching / lab_automation / troubleshooting_injection)
are now composed from adapted MCP tools plus a few copilot-native ones
that add value the MCP versions lack: link_create (port-name
resolution), wait_seconds, packet_analysis(+skills), and the topology
reader. Eight orphaned LangChain wrappers are deleted.
The device command tools had the dependency inverted (the MCP handlers
instantiated the copilot LangChain classes and round-tripped through
JSON strings). The Nornir/Netmiko classes moved verbatim (git mv) into
the shared layer as plain runners — DeviceConfigCommandRunner,
DeviceShowCommandRunner, VpcsCommandRunner — with _run(json_string)
becoming run(params, gns3_ctx). The optional Jinja2 template rendering
moved from the MCP handlers into the runners (and into
utils/device_configs.render_device_template), so both entry paths get
it; the MCP handlers are now pure delegation. This also drops the
json.dumps/json.loads detour the MCP path used to take.
test_tool_params consistency gate extended: a handler that forwards the
whole params dict to a callee is a wildcard (same rationale as the
params.items() case) — the per-key check cannot see through delegation.
websocat's text mode (-t, and also the default) appends a newline after
every received WebSocket message and turns CR/LF inside a message into
spaces. The compute ships console output as one BINARY frame per
arbitrary read(1024) burst, so text mode diced the stream at every frame
boundary — prompts split into pieces, short "narrow-column" lines, lost
carriage-return redraws (field report: Password: broken in three).
Measured against a local WS server sending known frames:
-t: A\nB\nC\na b\na b\n ("a\rb" -> "a b")
-b: ABCa\rba\r\nb (byte-exact, both directions)
Binary command frames stay console data on the compute side (only a line
that is itself a JSON {cols,rows} object would be taken as a resize
frame). The regression test fails on -t and passes on -b.
Replaces node_console (which handed a URL + ticket to the client for a
bash websocat chain) with a tool that runs the whole session server-side:
spawn `timeout N websocat -t --no-close <console-ws-url>`, feed the
command lines through stdin (paced, CRLF), read stdout until the timeout
kills websocat, return the collected output as plain text.
- plain text replies: "output: ..." or bare error text — no JSON
envelope, no status field (the server cannot judge command success)
- raw output: UTF-8 decode with replacement only, no telnet/ANSI
stripping; control keys are real characters (description teaches
\u0003, no \x03 translation)
- output funnel: grep=<python regex> filter; <=50 lines return as-is;
beyond that a 10-line sample plus hints — regrep=true re-filters the
node's cached capture (no re-run), full=true returns the last 400
lines / 64 KiB, and the full capture is written to the node's files
(console-capture.txt) with a ticketed curl download URL
- get_node_console_info_handler removed with its hashlib/TTL imports;
the node-bound console ticket redemption path stays (comment updated,
no current producer)
The skills repository carries per-protocol fault-injection knowledge
(symptoms, injection commands, verification steps) that only the
copilot could read through InjectionSkillsTool. Add the MCP counterpart
with the same multi-call protocol:
list (context REQUIRED) -> category index -> single fault
The list step is topology-filtered: context matches a skill's category,
key and declared protocols (case-insensitive), so aliases such as ebgp,
ospfv3 and stp select the right category instead of relying only on key
substrings. Fault bodies stay gated behind detail/issue requests.
Fault category keys are derived from repository filenames
(injection_<file>_issues). Correct the stale injection_ospf examples in
the registry docstrings, the loader comment and the copilot docs, which
pointed at keys that never existed.
Docs: 84 tools; new Fault Injection Skills section.
SkillsManager refused to start when the local skills directory was not a
git repository (cloning into a non-empty path fails), and reload_skills()
returned early when injection/ was missing, so device and feature skills
never loaded. reload_packet_analysis_protocols() had the opposite bug:
an empty result wiped the registry while still reporting success.
- initialize(): accept a non-empty directory without .git as a local-only
source (skip automatic updates, keep loading files); degrade to local
files when GitPython is unavailable.
- _update_if_needed(): re-create/clone when the local directory vanished
or moved, instead of reusing a stale handle forever.
- reload_skills(): injection, device and feature load independently; a
missing category keeps its previously loaded entries.
- reload_packet_analysis_protocols(): an empty/missing directory keeps
the existing registry and reports failure.
- reload_skills_repository(): drop the console guard cache so
config/console_guard.txt applies on reload.
MCP device_skill_get now exposes the category filter, documents the
single-file (topic-less) layout, and returns a sections list for
index/summary of skills that have no topics/issues. An uninitialized
skills repository fails with a clear error instead of an empty list.
The skills repository carries per-device operational knowledge the
generic tools cannot know (FRR's console lands in vtysh, its ping has
no -c, scripted config needs 'do' prefixes; VPCS syntax; IOS paging
pitfalls), but only the embedded copilot could read it — MCP's sole
consumer was console_guard.txt.
device_skill_get reuses the copilot's framework-free core
(get_skill/list_available_skills) rather than the LangChain tool shell,
so both surfaces share the same access protocol by construction:
list -> device topic index -> one targeted topic body, bodies only on
an explicit topic request. The caller passes the skill key exactly as
the list returned it; there is no name mapping or node-based resolution
in this layer.
The tool description also tells clients that support subagents to
delegate the discovery steps on long tasks, keeping the full skill
bodies out of the main conversation's context.
The ".gitignore" rules "lib" and "lib64" (from the Python template, meant
for a virtualenv at the repository root) match a directory with that name
at any depth. They kept two folders of the bundled web UI out of git:
- assets/vnc-console/novnc/vendor/pako/lib/ (12 files, zlib for noVNC)
- assets/xpra-html5/js/lib/ (20 files)
Anything built from a git checkout (GitHub archives, Docker images) ships
without them. The server answers a missing asset with the web UI page, so
noVNC's module graph fails to load (a module served as text/html) and the
VNC console stays on "Initializing VNC Console..."; the xpra console loses
its scripts the same way.
Anchor both rules to the repository root and add the 32 files, taken from
gns3-web-ui v3.1.0a7 (the bundled version); every blob is identical to the
one in that tag. A test walks the VNC console's module imports and the xpra
pages' scripts and fails on any file that is not in the package.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Without the AI dependencies the LLM model configuration router is replaced
by a stub that answers 501. It was a catch-all ("/{path:path}") included
under the /access prefix, so it also answered every /access route included
after it: creating an API key (POST /v3/access/api-keys) failed with "AI
Copilot is not available", which left API keys unusable on servers without
the AI extras.
The LLM stub now covers only the LLM model configuration paths, under
/access/users/{user_id}/ and /access/groups/{group_id}/. The stub routers
are built by build_ai_stub_routers() so tests can check them with or
without the extras, and add_stub_routes() takes the paths to stub.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review feedback: 3.1 is still in alpha and the GNS3 VM ships a matching
uBridge, so a hard minimum is simpler than a version-based degrade path,
and it also covers Docker endpoints and the builtin Ethernet switch,
which never went through the anchor probes. Raise minimum_required_version
to 1.3.1 and drop supports_kernel_datapath and its uses; the relay
fallback for capability problems (no CAP_NET_ADMIN) is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The TAP netdev kept QEMU's default vnet_hdr, so virtio-net, e1000e, igb
and vmxnet3 guests offload TCP/UDP checksums (and TSO) to the TAP: their
frames carry only the pseudo-header sum in the checksum field. A kernel
bridge link finishes them, but the uBridge relay (cloud, NAT, hub, ...)
reads the TAP with libpcap and forwards the frames as they are, so DNS
and TCP through a cloud fail while ICMP and DHCP work.
With vnet_hdr=off QEMU stops offering the offloads and the guest (or the
emulated NIC) sends finished frames, as with the 2.x socket netdev.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A server killed before close() leaves the switch's kernel bridge behind
with its ports still enslaved: a peer's persistent anchor TAP and cascade
veth ends outlive uBridge. uBridge's brctl delete refuses a bridge that
still has ports (EBUSY), so the best-effort recovery delete in
_ensure_bridge() failed silently and brctl create hit EEXIST: the switch,
and the project holding it, could no longer be opened (HTTP 409 "Could
not create bridge gns3xxxxxx: File exists").
List the leftover bridge's ports from /sys/class/net/<bridge>/brif and
release each with brctl delif before the delete. uBridge's brctl show
reports no ports, and the switch already reads sysfs on the same host.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
uBridge 1.2.3 already has the tap module, so the persistent-TAP probes
(_probe_persistent_taps on the node, probe_tap_support and siblings for
/capabilities) pass on it. QEMU, Dynamips and VPCS nodes then anchor on
TAPs and the controller builds kernel links that need 1.3.x commands:
kernel markers and the eBPF tc modes (1.3.0) and per-interface kernel
captures (`capture stop_kernel <ifc>`, 1.3.1). On 1.2.3 a MikroTik CHR
node sent frames but never got DHCP past DISCOVER/OFFER, and stopping
a capture failed with "Bad number of parameters".
Add Hypervisor.supports_kernel_datapath (version >= 1.3.1) and refuse
older builds in the node-side probe and the four anchoring capability
probes (tap, iol tap, bridge tap, vxlan). The existing fallback then
applies: nodes keep the legacy relay datapath with the usual warning,
which now names the installed version, and links stay on the relay.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Creating a node from a template whose name has characters invalid in a
hostname, such as "MikroTik CCR1072-1G-8S+", failed on the compute with
"'MikroTikCCR1072-1G-8S+-1' is an invalid name to create a Qemu node":
the template name was only stripped of spaces.
Replace the other invalid characters by hyphens, following the rules the
compute validates node names with (RFC 1123 for Qemu and Docker, IOS for
Dynamips and IOU), as project import already does for existing nodes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When only a port is renamed in a template's custom adapters, the adapter is
saved with "adapter_type": null. The emulators read it with
custom_adapter.get("adapter_type", self._adapter_type), which returns None
for a key set to null: QEMU was then started with "-device None,..." and
died ("'None' is not a valid device model name"), GNS3 only reporting
"Could not start Telnet QEMU console". VirtualBox and VMware read the
adapter type the same way.
Drop the null settings in _get_custom_adapter_settings so the node's own
adapter type (and MAC address) applies.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When the compute is itself a Hyper-V guest without an invariant TSC (e.g.
the GNS3 VM on hosts whose CPU lacks TSC scaling), Linux falls back to the
hyperv_clocksource_tsc_page clocksource and the guest timekeeping KVM
derives from the TSC is broken with recent kernels:
- kvmclock runs thousands of times too fast: Linux guests (MikroTik
RouterOS, ...) panic with "watchdog: BUG: soft lockup - CPU#0 stuck for
29604s" and reboot in a loop right after boot;
- TSC deadline timers fire far too late: guests with -cpu host hang.
The only workaround so far was -machine accel=tcg, i.e. no KVM at all.
Add a kvmclock setting to the [Qemu] section (auto, enabled, disabled).
In auto mode, KVM guests started on a host whose clocksource is a Hyper-V
one get kvmclock and tsc-deadline turned off: on the -cpu given in the
additional options (unless it sets these features itself), otherwise on
QEMU's default CPU model.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Add per-drive IDE/SATA selection with backward-compatible IDE defaults
- Persist interfaces in nodes, templates, and appliance imports
- Add database migration support for versioned and unversioned databases
- Prevent partial updates when rejecting live interface changes
- Validate interfaces and support SATA media changes through the monitor
Marker pcaps are written by uBridge on the compute that hosts the capture
node. The tag replay only looked in the project folder on the controller.
So a source from a remote compute showed count 0 and no frames, with no
error. Its frame detail returned 404 and said nothing ever matched (#3010).
Each source now gets its pcap through the compute of its capture node.
Local files are read as before. Remote files are fetched over HTTP on
every acquire into a small cache (32 idle files, cleared at shutdown).
A file is replaced only when its bytes changed. This keeps the old mtime,
so the sharkd session for that file stays alive.
When a compute cannot be reached, its source now shows an error message
in the timeline instead of just showing nothing. The detail endpoint now
returns an error instead of 404.
The custom 14789 was picked to keep our demux socket off the ports other
VXLAN software listens on, but a private default surprises operators and
blocks interop with gear that pins the standard port. Default to the
IANA-assigned 4789 instead; the gns3_server.conf override is unchanged
for computes where swarm overlay/flannel-style VXLAN already owns 4789.
Mixed-version deployments must pin the port on every compute until they
all run the new default, or the two tunnel ends will disagree.
The scenario needed root for its network namespace and its fabric subnet
turned out host-specific: policy routing (a proxy's table consulted
before main) sends any non-RFC1918 destination out the default route, so
the peer compute was unreachable behind the veth — unfixable by picking
another subnet, and the failure paths leaked namespaces and processes.
Too environment-coupled to keep in the suite.
Unit coverage is unchanged (selection matrix, tunnel spec derivation,
VNI retry/degrade/self-heal, device lifecycle, capability probe); the
feature doc's Verification section now carries the manual two-compute
checklist instead of the e2e reference.
Cross-compute links can ride kernel VXLAN tunnels instead of the uBridge
UDP relay: each endpoint creates one VXLAN device (uBridge's rtnetlink
vxlan module) enslaved into the same per-link Linux bridge the same-host
datapath uses, so filters, markers, capture and suspend keep running on
the anchor unchanged. Opt-in (Server.enable_vxlan_cross_compute, default
off) because the tunnel caps the usable MTU at the underlay minus the
outer headers; the relay stays the default and the fallback.
The controller derives what the pair must agree on: one VNI per link
(upper VNI space, per-link random salt) and deterministic device names
(gvl{link}{node}) that make crashed-run leftovers claimable
(show -> delete -> create). Same-host compute pairs, switch endpoints
and serial/unix-socket containers stay on the relay.
A VNI the kernel refuses self-heals on every path it can surface:
interactive create retries a fresh salt then degrades to the relay;
project open re-dispatches refused batch entries the same way; node
start recognises the refused device, stops the half-started node,
re-wires both ends on a fresh VNI (or degrades the link) and retries
the start. The device MTU is derived from the lower device (dev=) and
the per-link bridge inherits it - no configuration. The capability
probe (ubridge_vxlan in /capabilities) runs one scratch create/delete
on a scratch port that never equals the configured dstport.
The e2e scenario runs two computes on one machine (a network namespace
joined by a veth pair as the fabric) and needs root.
Updating any Docker node property destroys and recreates the container
(the route calls DockerVM.update() as soon as the property diff is
non-empty), so it discards everything written inside it. On a running
node that happened silently, as a side effect of an unrelated edit: the
user only saw the node restart.
Refuse the update while the container is up and say what stopping it is
for, like QEMU, IOU, Dynamips, VPCS, VirtualBox and VMware already do
for their own restart-requiring settings. The check runs before anything
is applied — a refused update leaves the node untouched — and only when
the PUT really changes a property, since Web clients send the whole node
back on every edit.
is_running() is not reusable for the question: it reports a paused
container as not running and stops crashed containers as a side effect,
so DockerVM gains is_up() for "running or paused".
A compute 409 (for instance "must be stopped to change the console type
to vnc") left every value that had already been applied locally on the
controller node, because the update loop sets them before the PUT.
Nothing was dumped at that point, but project.dump() serializes the
whole topology: the next successful change anywhere in the project wrote
the refused values to disk, and the next project open recreated the node
with them — this time accepted by a stopped node. A refused change
silently landed later.
Restore the applied properties, the node properties, custom adapters,
missing images and the derived ports before re-raising, the way the
degraded-image create() path already cleans up after its own failure.
When a project is opened, Docker nodes are created in parallel. The free
display was found by checking whether /tmp/.X11-unix/X<n> exists, but that
socket only appears once Xvnc has started, so several nodes got the same
display. Only the first Xvnc started; the others exited with "Server is
already active for display N" ("The vnc process has stopped with return
code 1"), and their containers drew on the screen of another node.
- reserve the displays given to the nodes of this server, and also skip a
display whose lock file belongs to a live process
- restart a crashed Xvnc on the display the container was created with,
after removing its stale socket, instead of picking a new display the
container cannot reach (black screen)
- on close, only remove the X socket of our own Xvnc, never the socket of
the node that owns the display
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>