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.
Rename five copilot tools to the MCP <object>_<verb> naming so both
agent surfaces expose the same names:
- create_gns3_node -> node_create
- create_gns3_link -> link_create
- create_gns3_text_drawing -> drawing_create_text
- create_gns3_rectangle_drawing -> drawing_create_rectangle
- create_gns3_ellipse_drawing -> drawing_create_ellipse
Also close two capability gaps found during the comparison:
- node_create: accept optional per-node compute_id (passed through to
the shared handler; before, every node landed on local)
- link_create: accept optional per-link link_type (ethernet default,
serial supported; before, serial links were impossible)
All tools keep calling the same shared handlers in
gns3_client/api_handlers.py, so the rename has no effect on execution
semantics. Old names appeared nowhere else (prompts, tests, mode tool
lists are clean); docs updated accordingly.
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 send_command read loop polls read_channel (netmiko decorates it
with a session-lock wrapper that only __init__ can satisfy), so the
old read_until_pattern mock died with AttributeError on instances
built via __new__.
The pcname kwarg was added to VPCSTelnet but never passed by the only
production instantiation path, so a node renamed away from PC<n> never
matched the prompt pattern and every command stalled the full read
timeout. Pass pcname via the netmiko connection extras (the host key is
the node name).
send_command now owns its read loop instead of delegating to netmiko's
read_until_pattern: that helper raises ReadTimeout discarding accumulated
output, so a command slower than the timeout returned an empty string
that looked like success. The partial output is returned instead.
Aligns with uBridge's tc.md correction: TCA_NETEM_PRNG_SEED (uapi slot 14,
added in Linux 6.6 — not the nonexistent TCA_NETEM_SEED at "5.1") is what
the seed filter rides on; pre-6.6 kernels parse qdisc options
non-strictly and silently drop the attribute, so seed determinism is
unavailable there while every other netem knob applies. Stated in both
filter tables and the seed section of packet-filters.md, the seed rows of
the MCP and copilot tool descriptions, and the kernel-impairment spec.
The device tools (device_config_send / device_show_run / vpcs_config_set)
built their Nornir inventory with only the console port and inherited the
group hostname, which is always the controller host. Console ports of
nodes on remote computes are bound on the compute's own host, so those
nodes were unreachable from the MCP tools - only local-compute nodes
worked.
Pass the per-node console_host through the project inventory and inject
it as a host-level Nornir hostname (which overrides the group's
controller-wide default). Missing or wildcard (0.0.0.0/::) values keep
the group inheritance, so single-machine setups behave exactly as
before. VPCS builds its own host entries, so it carries the hostname
over explicitly.
The text drawing tool was MCP-only; the copilot needs the same
capability and its area-annotation helper carried the same Web UI
compat bugs. Split into three layers with a single implementation:
- gns3server/utils/drawings.py: pure SVG builders (text, rectangle,
ellipse) with full input validation (ValueError) — the Web UI
format knowledge lives in core utils next to picture.py
- api_handlers.py: shared REST handlers (create_text_drawing_handler
moved here, create_rectangle/ellipse_drawing_handler added) — the
layer both consumers already share
- thin shells: MCP tools (drawing_create_text unchanged, plus new
drawing_create_rectangle/_ellipse) and three LangChain tools
(tools_v2/gns3_create_drawing.py) registered in both copilot mode
lists
Shape builder details:
- rectangle: solid fill or "none", stroke block emitted only when a
stroke is given, rx corner radius, dasharray validated against the
Web UI's Qt remap table (QtDasharrayFixer rewrites those patterns)
- ellipse: bounding-box API; cx/cy derived as offsets inside the
drawing box (the Web UI treats them as box-relative, not canvas
coordinates); odd sizes round down and the actual size is returned
gns3_drawing_utils now delegates to the shared builders, fixing the
px-blind sizing, the ignored text-anchor/x/y centering and the
missing XML escaping in the area-annotation path.
Compute:
- Track the anchor that owns uBridge's single AF_PACKET capture slot: a
second kernel port's capture is refused with a clear error (its flag
rolled back), a failed start no longer wedges the port, and only the
owning port's stop issues the process-wide `capture stop_kernel`
(docker/qemu/iou/dynamips/switch, via the mixin's slot helpers).
- Unwind failed starts instead of leaking uBridge and the anchor TAPs:
QEMU's stop body runs under try/finally, QEMU and IOU start the datapath
inside the launch try, Dynamips undoes a half-attached port (the retry
used to be short-circuited into a silent black hole), and Docker deletes
the veth pair when the namespace move fails.
- Restore the relay filter guard on IOU (kernel-only filters raise
NodeError, an uncompilable BPF line degrades to a warning), validate the
eBPF/quota/window/bpf filter shapes before touching the bridge (short or
non-numeric values surfaced as 500s after partial state), and refuse a
second NIO on an occupied Docker/QEMU port (the guard IOU and Dynamips
already had).
- Probe fixes: `_probe_bridge_tap` survives a failing uBridge start, and
the probe caches key on ctime too, so a chmod/setcap remedy re-probes.
Controller:
- Parse GNS3_UNIX_SOCKET_NIO like the compute does (line-anchored, truthy
value) and mirror the compute's class selection in kernel_anchor_type.
- Read partial capability reports defensively; make BridgeNIO.bridge
required-but-nullable so an omitted key is a 422 rather than a KeyError.
- Strip the packet_loss correlation slot even when it is zero: the relay
filter takes exactly one argument and rejected `packet_loss 10 0`.
Ethernet switch: skip ports whose absorbed anchor does not exist yet (the
join is deferred by design) and roll the ports_mapping back when the VLAN
reconcile fails, so a retry still diffs against the old mapping.
Agent: reject device config entries without a device_name instead of
silently dropping them (a malformed batch used to execute partially while
reporting success).
Also split the netem builder to clear C901 and drop the two per-file ruff
ignores the PR had added, keeping the remove-only policy.
The rebase onto upstream 3.1 (mypy + ruff I001/F401 waves) left the tree
short of the newly enforced baselines in three ways:
* contextlib / UbridgeError imports dropped as false duplicates during
conflict resolution left NameErrors in iou_vm and qemu_vm (F821);
the Dynamips router grabbed the compute-level NIOTAP instead of the
hypervisor one, breaking TAP anchoring (test failure).
* unused imports (KERNEL_UNSUPPORTED_FILTERS, UbridgeError, List) and
unsorted blocks left behind where the stack's later commits moved the
last usage elsewhere — ruff I001/F401/RUF022 now enforce these.
* mypy, now baseline-free: cache dict annotations in tc_probe, the
locators candidate list, the resolved-params None branch in the MCP
runner, and validate_bpf_syntax keeps the upstream Dict[str, Any]
return (the strict Optional[str] was wrong for the "valid" bool entry).
The e2e raw-frame probe resolves its compiler via shutil.which (S607)
and zips with strict= (B905); the remaining subprocess call is S603 by
nature and gets a per-file ignore like other test files.
ruff check/format clean, mypy at upstream parity (no new errors),
full unit suite 2167 passed.
A batch carrying the same device twice lost all but the last entry's
commands: the per-device command map kept only the last entry while
result processing iterated every original entry, so each row reported
the last entry's output under its own command list.
c3c78f99a consolidated the merge into the MCP template renderer, but
that only runs when a template is passed — the MCP direct-commands
path and gns3-copilot's direct _run path were left uncovered.
Normalize once in each tool's _run (single point covering all entry
paths, restoring the c42b3a59b coverage) and keep the renderer merge
as boundary normalization. Blocked-command info now also accumulates
across same-device entries instead of last-winning in the per-device
map.
The tool descriptions have always advertised kernel_datapath as a
requestable field, but the handler's whitelist did not contain it: asking
for it was rejected as an invalid field, so the MCP clients could not see
which links run on kernel interfaces at all.
Also record what seeding a netem experiment on IOU costs: IOL emits its own
CDP/keepalive frames and they draw from the netem RNG stream, so a
seeded run only reproduces after `no cdp run` (verified bit-identical pcaps).
Three fixes from live-calibrating the kernel impairment filters on an
isolated instance (Docker-to-Docker veth link, per-sequence evidence):
- update_filters skipped an unchanged filters dict as a no-op, so a
one-shot window_drop outage that had already passed could never be
re-armed: re-applying the same values produced no datapath command at
all. A PUT still carrying window_drop now reconciles, restarting the
schedule from the apply as documented. Measured: identical re-apply
moves the outage window from seq 50-70 to 70-91 (arm at +8s of the
new apply).
- gemodel parameter semantics were documented backwards in the
validation rules, available_filters, the MCP/copilot descriptions and
the feature docs. Measured against the kernel (per-direction counters,
offered = sent + dropped): p is the good-to-bad transition chance, r
the bad-to-good transition, 1-h the loss chance in the bad state (the
good state loses nothing) - mean one-way loss p/(p+r)*(1-h), e.g.
[10,30,100] -> 24.8%/25.7% measured vs 25% predicted, [5,45,100] ->
10.8% both ways vs 10%.
- kernel_datapath was silently stripped from every REST/MCP link payload
by the response model (LinkBase carried no such field), so clients
could not tell a kernel link from a relay one.
Also verified on the same setup: the 3-argument window_drop form works
exactly as specified ([3000,5000,100] -> a single seq 31-55 outage,
[20000,6000,100] -> seq 115-144), and both generations of the command
builder emit the parameters unpadded.
Adds an API test for the field and controller tests for the no-op skip
and the window_drop re-arm exemption.
link_update, link_available_filters and the copilot packet-filter tool
still advertised the pre-kernel-datapath five types. List the full
thirteen with positional parameter formats, ranges and kernel-only
marking, point at link_available_filters as the per-link authority
(topology + compute capability gated), and document window_drop's
apply-relative schedule restart. Also add kernel_datapath to the link
field lists.
Description-only: handlers pass filters through unchanged.