perf(schema): drop the draft-07 $schema header from every tool

[VickyXAI]

What

Removes the "$schema": "http://json-schema.org/draft-07/schema#" header from every tool's inputSchema.

Clients validate the schema either way and the Anthropic API ignores it, but it ships in the tool block the model carries in its context on every turn.

The number

Measured against the built server — real MCP handshake (initialize → notifications/initialized → tools/list), tokenizing the model-visible projection {name, description, input_schema} with the mcp__blockrun__ prefix the host adds, o200k_base. Before and after measured the same way, not subtracted:

Profile	Tools	Before	After	Saved
full (default)	20	13,200	12,900	−300
trading	9	5,689	5,554	−135
media	7	5,541	5,436	−105
research	6	3,114	3,024	−90
chat	3	1,969	1,924	−45

Exactly 15 tokens per tool.

This is one item from a wider tool-schema audit. The larger prize — route catalogues and per-model pricing tables living in tool descriptions, which are ~53% of the block — is a separate batch, because it changes what the model reads and needs its own review.

Why it isn't a one-line deletion

It isn't our string. The SDK's zod→JSON-Schema conversion stamps it on, and exposes no option to suppress it — verified on @modelcontextprotocol/sdk 1.29.0 for both the zod v4 branch (z4mini.toJSONSchema) and the zod v3 branch (vendored zod-to-json-schema).

So stripJsonSchemaDialect() wraps the tools/list handler as it is installed, using only the public setRequestHandler. Two deliberate properties:

• Never load-bearing. The wrapper is identity-matched against ListToolsRequestSchema. If a future SDK stops routing through it, the wrapper stops matching and the header simply comes back. Nothing breaks.
• Degrades to a no-op. It skips when there is no low-level server to wrap — the tool-annotations and apps suites pass a minimal fake McpServer, and an earlier version that assumed server.server exists broke 8 of them.

Tests

test/schema-dialect.test.ts (2 tests). Because the wrapper is identity-matched, an SDK bump could silently restore the header with nobody noticing — an optimization that reverts unobserved is worse than not having done it. These tests are what notices.

No behaviour change

• 425/425 tests pass (423 before, +2 new).
• annotations on all 20 tools and _meta.ui on the 2 app tools are untouched.
• Live stdio check: schema validation still rejects a missing required argument and a mistyped argument (-32602 both), and a free tool call still succeeds. No paid tool was called.

Note on the figures

An earlier revision of this PR quoted 13,765 → 13,465. Those numbers were 4.4% high: the measuring script used Python's json.dumps, which defaults to ensure_ascii=True and re-encodes every —, → and · as \uXXXX — six ASCII characters where the model sees one glyph. Independently reproduced with a JSON.stringify-based harness, which matches the wire. The saving is unaffected at −300, since the removed string is pure ASCII; only the totals it is measured against moved.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WUL3ExR4Nz7uebxKaKwjDi