v4.2.1

Fixed

• LMv1 authentication: the signature is now base64 of the hex HMAC-SHA256 digest, as
  LogicMonitor documents. Base64 of the raw 32-byte digest is rejected with 401, so every
  portal configured with LM_ACCESS_ID/LM_ACCESS_KEY failed to authenticate; bearer-token
  portals were unaffected. The tests now check a fixed reference vector computed with LM's
  recipe.
• LMv1 request bodies: request() and ingest_post() serialize a JSON body once
  and send those exact bytes, so the signature covers what goes on the wire. The
  signature was computed over json.dumps with default separators while httpx sent
  compact JSON, so every POST, PUT and PATCH with a body failed with 401 on LMv1
  portals.
• Stdio no longer fails at startup with UnboundLocalError when LM_MULTI_PORTAL
  and TF_WORKSPACE_DIR are both set: logging is imported at module scope.
• /readyz no longer calls the LogicMonitor API. With
  LM_HEALTH_CHECK_CONNECTIVITY on it ran the connectivity check on every probe,
  although the result could never change its answer, so an unauthenticated caller
  could drive portal traffic (#56).
• LM_MAX_RETRIES is honored on both transports. Stdio and HTTP built the
  single-portal client with the default of 3 whatever the setting.
• get_api_tokens and get_api_token were uncallable: their handlers required an
  admin_id that neither schema declared. admin_id is now optional and declared
  in both; omitted, tokens list portal-wide and a token resolves by id
  (filter=id:<id>), and an unknown id returns NOT_FOUND. get_website_data's
  schema marks checkpoint_id required, as the handler always did. A test holds
  every tool's schema to its handler signature.
• Tool errors are reported as errors. Error results carry a type (ErrorContent),
  call_tool returns them as CallToolResult(isError=True), and the write audit
  log and session history take success from the result. A 403 on a write, or a
  refusal with writes disabled, was audited as success=True. The text of every
  error is unchanged.
• The mcp floor is 1.19.0, the first release whose low-level server passes a
  CallToolResult through; with 1.18 and earlier a tool error reached the client
  as a pydantic validation message instead of its text.
• Configuration validation errors no longer echo the rejected value, which
  previously put a too-short LM_HTTP_AUTH_TOKEN or LM_BEARER_TOKEN into the
  startup traceback, container logs, and the health endpoint's response body.
• lm_mcp.__version__ synced to the release version (the published 4.2.0 wheel
  self-reports "4.1.0" at the HTTP root endpoint); the foundation test now
  asserts __version__ against the installed package metadata so a bump that
  misses __init__.py fails the suite instead of shipping.
• HTTP transport tools/list now includes the Terraform tools when a Terraform
  runner is configured. They were dispatchable over HTTP but invisible to HTTP
  clients, unlike stdio which listed them correctly.
• Documentation: removed the LM_ANALYSIS_TTL_MINUTES row from the README
  (not a config field), corrected the LM_CORS_ORIGINS default, and dropped a
  .gitignore negation for a file that does not exist.
• server.json's version fields are held to pyproject.toml by a test, and the
  README and documentation/client-setup.md are held to the registry's tool count
  (306); the client setup guide said 280.

Added

• LM_HTTP_AUTH_TOKEN: opt-in inbound bearer authentication for the HTTP
  transport. When set, /mcp and /api/v1/* require
  Authorization: Bearer <token> (constant-time comparison) and return 401
  with WWW-Authenticate: Bearer otherwise; /, /health, /healthz, and
  /readyz stay open for container healthchecks and orchestrator probes.
  Unset preserves previous behavior and logs a startup warning. Minimum token
  length 16; a blank value means unset, and surrounding whitespace is stripped
  so a token read from a secret file still matches. The scheme is matched
  case-insensitively per RFC 7235, and probe paths match with or without a
  trailing slash. Stdio is unaffected.
• CONTRIBUTING.md: fork-and-pull-request workflow, development setup, PR
  expectations, and the tool-surface regeneration commands.
• deploy/k8s/deployment.yaml: vendor-neutral Kubernetes Deployment and
  Service for the published container image, with credentials read from a
  user-created secret and probes wired to the health endpoints.
• .pre-commit-config.yaml: local hooks mirroring the blocking CI gates (ruff
  check and format on src/ and tests/, file hygiene, private-key guard).
  Install with uv run pre-commit install; pre-commit is now a dev dependency.
• .github/pull_request_template.md: prompts contributors for a summary,
  changes, test plan, and rollback note.
• CI step that regenerates the tool contract fixture and documentation/tools.md
  and fails with the exact regeneration commands when either is stale.

Changed

• deploy/docker-compose.yml now publishes the plaintext port on loopback
  (127.0.0.1) instead of every interface. That port is published whether or
  not the tls profile is active, so a TLS deployment was still exposing an
  unencrypted listener that would carry LM_HTTP_AUTH_TOKEN in cleartext.
  Caddy is unaffected, since it reaches the service over the compose network.
  Set LM_HTTP_BIND=0.0.0.0 to serve plaintext to other hosts deliberately.
• deploy/docker-compose.yml no longer defaults LM_CORS_ORIGINS to *.
  Deployments that relied on the implicit wildcard should set an explicit
  origin list rather than restoring *: CORS runs with
  allow_credentials=true, so a wildcard lets any site a browser visits call
  /mcp with the visitor's credentials.
• With LM_HEALTH_CHECK_CONNECTIVITY on, /health runs the connectivity check
  only for callers that present LM_HTTP_AUTH_TOKEN, or when no token is
  configured; an unauthenticated probe gets the local checks only.
• CI runs each matrix leg on the Python it names (every leg ran 3.11), installs
  from uv.lock with --locked, pins every action to a commit SHA, and runs with
  a read-only token. A new job installs the huggingface extra with CPU torch and
  runs its tests. Dependabot covers uv, GitHub Actions and Docker weekly.
• Releases are gated. A tag must be on main and match pyproject.toml, both
  server.json versions, lm_mcp.__version__ and a CHANGELOG section
  (scripts/check_release_tag.py), and lint, the format check, the suite on
  Python 3.11 to 3.13 and the generated files must pass before PyPI publishes; the
  container and the GitHub release wait for PyPI.
• mypy is a ratchet: CI fails when the error count rises above
  mypy-baseline.txt, in place of a step that could not fail.
• The container installs from uv.lock with pip --require-hashes, pins its base
  images by digest, drops uv and curl from the runtime image (the healthcheck
  runs on Python), and runs as UID 10001. deploy/k8s/deployment.yaml sets
  runAsUser and runAsGroup to 10001, pins image 4.2.1 in place of latest,
  and mounts an emptyDir at /tmp under its read-only root; compose uses the
  Python healthcheck and a digest-pinned Caddy.