find_invalid_x_mcp_header: never repr a non-string annotation value

Merged
MA
5mo ago
+64-24

Follow-up to #2974, on find_invalid_x_mcp_header: make its reason strings total over any inputSchema value, and pin the schema-position walk's applicator keywords independently of the implementation.

Motivation and Context

find_invalid_x_mcp_header is meant to return a reason string or None for any decoded inputSchema, but two of its checks fused two different failures — "the value isn't a string" and "the string is invalid" — into one branch whose message interpolated the value with !r either way. repr() of an arbitrary schema value is not total: an int with more digits than sys.get_int_max_str_digits() (default 4300) raises ValueError. That propagates out of ClientSession.list_tools() instead of the offending tool being dropped, and it is reachable from the public API: the in-memory Client(server) path has no JSON decode step in front of the client-side filter, so the value reaches it as a live Python object and await client.list_tools() raises under default interpreter settings.

Each fused check is now two checks with two messages, so a value only reaches an interpolation once the preceding branch has proven it is a str (where repr is total). No guard and no helper — a non-string value structurally never reaches a repr — and the messages get more precise: x-mcp-header must be a string, not int rather than pretending an int is a malformed token.

Separately, the reject test for "annotation under a non-properties applicator" hand-picked 6 of the walk's 19 applicator keywords, so the other 13 — including the chain-breakers the spec names anyOf / not / then / else, plus additionalProperties — could be removed from the walk without any test failing. The cases are now a literal keyword → shape table covering all 19 (deliberately not derived from the implementation's keyword sets, so dropping a keyword fails its case rather than shrinking the parametrization), with one equality test against those sets so an added keyword must gain a literal case. The hand-picked six are subsumed; one accept-side symmetry gap is filled (enum was the only instance-data keyword named in the source comment with no lookalike test).

How Has This Been Tested?

  • Two new reject params (oversized-int-header, oversized-int-type) fail on main with ValueError: Exceeds the limit (4300 digits) for integer string conversion — one at each message site — and pass here (checked by stashing the src/ change and re-running them).
  • End to end at the public surface: a lowlevel server advertising those schemas, driven through Client(server).list_tools(). On main that call raises ValueError out of inbound.py; here it returns, logs one warning per dropped tool, and keeps the valid one. Reason strings for string-valued annotations are unchanged.
  • ./scripts/test: full suite, 100% coverage, strict-no-cover clean. pyright: 0 errors. No new pragma / type: ignore / noqa.

Breaking Changes

None. Accept/reject decisions are identical for every input; only the reason-string text for non-string values changes (e.g. (got {'not': 'valid'}) becomes (... the type keyword is dict, not a string)), and that string is only ever logged.

Types of changes

  • Bug fix (non-breaking change which fixes an issue)
  • New feature (non-breaking change which adds functionality)
  • Breaking change (fix or feature that would cause existing functionality to change)
  • Documentation update

Checklist

  • I have read the MCP Documentation
  • My code follows the repository's style guidelines
  • New and existing tests pass locally
  • I have added appropriate error handling
  • I have added or updated documentation as needed

Additional context

The walk itself is untouched. A deeply nested container value (the other non-total repr input) cannot reach this code from any transport — the in-memory server's own result serialization and the HTTP JSON parser both reject it first — so the oversized int is the one shape with a live path, and it is the one the regression tests pin.

AI Disclaimer