v1.11.0

Fixed

• tools_description_present_in_all now catches absent descriptions.
  The old predicate compared against the empty string only, so a tool with
  no description at all (or a whitespace-only one) passed the HIGH rule
  that exists to catch exactly that — the most common real-world defect
  earned the points. The rule now uses the same predicate as its
  prompt/resource siblings (absent, empty, and whitespace all count as
  missing). Scores move for servers whose tools omit descriptions; both
  regression tests were proven to fail against the old implementation.

Changed

• MCP Python SDK bumped to mcp==2.1.1 (from 2.0.0). Two client-side
  SDK fixes ride along and can move scores on affected live servers by
  turning listing failures into real audits: pre-2026
  sessions now drop later-revision cache-hint fields instead of failing
  list_tools(), and boolean sub-schemas in tool properties now parse.
  Servers whose listings previously died inside the SDK go from
  insufficient-data skips to actually being judged.

• The "Tools Quality" category is now "Primitives" in the README and
  docs — the MCP specification's own collective noun for tools, prompts,
  resources, and resource templates, which had outgrown the old name (tools
  are 18 of the category's 50 rules). Engine reports are unaffected: the
  category taxonomy lives in mcpscore.dev's report surface, and rule_ids,
  group_names, and scores are unchanged.

Added

• Smoke mode: --smoke invokes your own server's tools after the audit.
  The audit itself still never calls tools/call; smoke mode is the separate,
  opt-in surface for developers testing a server they operate (typically in
  CI). It verifies what listing alone cannot: that a declared outputSchema
  is honored by real structuredContent (per MCP 2025-11-25 Tools §Output
  Schema, a MUST — an explicit empty {} schema still requires the field to
  be present), that schema-invalid arguments are rejected with an isError
  result or a protocol error, and that unknown tool names get a JSON-RPC
  error rather than a result. For both rejection checks, any error code
  counts as a rejection except a hang, a closed connection, or -32603
  internal error — those are crashes, not rejections.
  The probed unknown name is derived deterministically to be
  provably absent from the server's own catalog; when the tool listing is
  missing or incomplete, that check skips rather than risk invoking a real
  tool. By default only tools annotated
  readOnlyHint: true are called; --call-all is the explicit consent for
  the rest — so unannotated tools visibly cost smoke coverage. Argument
  synthesis is deterministic (defaults / const / examples / enum-first / type
  zero-values — never an LLM) and tolerates broken schemas: malformed
  non-string required entries are ignored rather than aborting the smoke
  run. Verdicts are pass/fail/skip, and a tool whose
  upstream is down reports as skip: someone else's outage should not fail
  the build. Results land in a new smoke section of the --json report and
  never touch score/max_score; any smoke failure exits with the new code
  4 (a failed --fail-under gate still takes precedence with 3). The
  web service deliberately has no smoke mode — CLI (and the GitHub Action,
  in a coming release) only.

• Modern-only stdio servers are now fully auditable. When a local server
  rejects the legacy initialize handshake but answers MCP 2026-07-28
  stateless requests, mcpscore retains its resolved launch command, probes it
  over fresh stdio processes, reconstructs the catalog and server metadata,
  and emits a normal full audit with era: modern instead of exiting 2. A
  server that neither handshakes nor demonstrates modern support still exits
  2, preserving the CLI contract. This works for direct .py/.js targets
  and arbitrary-language --stdio commands.

• Three catalog connection-independence rules. Modern servers are checked
  that the complete tools, resources, and prompts identity sets do not vary
  across two independent connections carrying the same authorization context.
  Comparisons ignore ordering, traverse pagination completely, bound reported
  differences, and skip rather than fail when either observation is incomplete.
  Each requirement is HIGH and scoped to MCP 2026-07-28, where catalog sets
  MUST NOT vary per connection.