You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
New: OpenAI-compatible nodes are first-class (LM Studio, llama.cpp llama-server, vLLM, LiteLLM). A node takes "openai": ["http://HOST:PORT/v1"] beside or instead of "ollama", validated with the same shape rules and
errors. Each URL is probed with GET <base>/models under the existing single deadline, and the snapshot records
per endpoint whether it answered, the latency and the model ids (node.openai[]). Before, {"openai": [...]} was
a usage error ("nothing to probe"), and the older "kind": "openai" form reported a dead server as ollama.unreachable, an empty server as worst: OK, and could not authenticate.
New: "api_key_env": "NAME" names the environment variable holding a node's bearer token. The value is sent
to that node's openai URLs only, dropped if a redirect leaves the host, and never written to the snapshot, the
state folder or any output; a pasted key in place of a name is refused without being echoed.
New rules:openai.unreachable (FAIL), openai.no_models (WARN) and openai.slow (INFO, openai_slow_ms,
default 2000), graded by run history like the Ollama rules, with stack_explain entries. model.not_loaded
counts a model an OpenAI endpoint lists. Loaded-model expiry (model.expiring) and --ttft stay Ollama-only.
Unchanged: the "kind": "openai" / "endpoints" form still works and means what it meant, ollama.unreachable
included. The MCP tools and their descriptions are unchanged (642 tokens, budget 650). dsh-stack-guardian is unchanged.
Releases are published by trusted publishing (OIDC). Publishing a GitHub Release runs .github/workflows/release.yml: it fails unless the tag is v + the pyproject.toml version, runs the tests,
builds and twine checks the sdist and wheel, and uploads them to PyPI with a short-lived credential and signed
attestations. No stored PyPI token. Owner setup and release steps: docs/RELEASING.md.
The release workflow asks PyPI first. Its verify job (which also runs on a pull request that touches pyproject.toml, as a dry run) now fails if the version in pyproject.toml is already on PyPI, and otherwise
prints stack-guardian X.Y.Z is new on PyPI (latest there: ...). A PyPI that cannot be reached is a warning.
Support kit: bug and feature issue forms (blank issues off, a private security-reporting link), SUPPORT.md and CODE_OF_CONDUCT.md, with a test that keeps them in place. Docs and repository metadata only.