template_list was annotated as dict[str, Any] but _run_handler_sync always
returns a list envelope ([{"type":"text",...}]). Other tools use
list[dict[str, Any]] consistently. This caused a Pydantic dict_type
validation error on the client.
When MCP client authenticates with API key, _resolve_token now
returns a newly generated JWT. This JWT is stored in the ContextVar
and used by all subsequent tool handler API calls - zero extra bcrypt.
New format: gns3_<api_key_id>_<random_secret>
- Auth extracts api_key_id from token → single DB query by UUID → one bcrypt
- No more scanning all keys (was O(n) with bcrypt per key)
- bcrypt.checkpw offloaded to thread pool to prevent event loop blocking
- Legacy gns3_<random> format removed (compatibility break)
bcrypt.checkpw is CPU-bound (~1.3s per call) and was running
synchronously inside the async event loop, blocking ALL concurrent
requests. With 5 API keys and 10 concurrent requests, this caused
~13s delay before any handler could start.
Reading the entire DB file into OS page cache eliminates the 8-14s
cold-start penalty on the first ORM query (users/templates/api_keys).
Previous SELECT 1 only warmed the connection pool, not the file cache.
This is a non-fatal issue that occurs during concurrent project close
when multiple nodes try to write their md5sum cache while the project
directory is being cleaned up.
- Expose optional 'fields' parameter to AI via tool description
- List all available node fields in the parameter description
- Mention inherited template_id in batch mode
- Batch mode now inherits top-level template_id as default per-node
- Both single and batch modes filter response to minimal fields by default
(node_id, name, node_type, status, console)
- Optional 'fields' parameter overrides the default field set
Dynamips node creation (e.g., Cisco 7200 with multiple adapters)
can exceed the previous 10-second timeout, causing MCP tools to
fail with Read timed out errors.
- Add class-level caches for loader and default RAM/NVRAM values
- When multiple IOU nodes use the same image, only the first one runs
the ld-linux --verify and iou-image -h subprocesses
- Subsequent nodes reuse cached results, saving ~2 subprocess calls per node
- Cache is populated only on successful subprocess execution to avoid
overwriting explicitly set ram/nvram values with stale cached defaults
The _iou_id_lock must cover _create_node() because get_next_application_id()
checks in-memory nodes (self._nodes), which are only registered after
_create_node() completes. Without this serialization, concurrent IOU
node creation produces duplicate application IDs.
- Narrow IOU lock scope to only cover application_id allocation,
allowing concurrent IOU node creation via Pool(concurrency=5)
- Parallelize link creation during project.open() using Pool(concurrency=5)
instead of sequential processing
- Add batch UDP port allocation endpoint on compute to allocate N ports
in a single HTTP call
- Pre-allocate UDP ports per compute before link creation during project
loading, reducing HTTP round-trips
- UDPLink.create() falls back to individual port allocation if no
pre-allocated port is available
YueGuobin
Jeremy Grossmann