4.1 KiB
Sprint 163 Plan: Strict Grammar Foundation and Schema Normalization
Context
Current MCP grammar coverage is complete (347 tools), but strictness is not maximal.
The grammar generator still falls back to broad productions (any-object,
any-array, any-value) for many nested argument shapes, which permits syntactic
validity without full schema-level correctness.
This sprint builds the strictness foundation:
- define a measurable strictness score
- normalize per-tool schemas into a canonical, recursion-safe form
- refactor grammar generation so strict handling is possible for every schema node
No MCP tool names change. Tool contracts stay backward-compatible at the API level.
Goals
- Add strictness metrics for all tool schemas and generated grammars
- Introduce canonical schema normalization for recursive codegen
- Eliminate fallback productions at top-level tool arguments
- Keep generated artifacts deterministic across runs
Steps
Step 1859: Strictness audit utility and baseline report (10 tests)
Add tools/mcp/audit_grammar_strictness.py.
Input:
tools/mcp/whetstone_tool_schemas.jsontools/mcp/grammars/per_tool_schemas.jsontools/mcp/grammars/*.gbnf
Output:
tools/mcp/grammars/strictness_report.json- per-tool metrics: fallback count, nested coverage, enum coverage, unsupported keywords
Tests (10): deterministic output, counts align with 347 tools, missing-file handling, invalid schema handling, fallback detection, enum detection, nested object detection, array-item strictness detection, summary rollup correctness, JSON schema for report.
Step 1860: Canonical schema normalizer (12 tests)
Add tools/mcp/schema_normalizer.py.
Normalize each tool schema into a canonical AST-like structure:
- explicit node kinds: object, array, string, integer, number, boolean, null
- preserved constraints: required, additionalProperties, enum, const, pattern, min/max, minItems/maxItems, oneOf/anyOf/allOf
$refexpansion within document scope
Tests (12): nested object normalization, array-of-object normalization, enum/const
preservation, required/optional split, additionalProperties preservation, oneOf,
anyOf, allOf flattening semantics, $ref resolution, cycle guard, stable key order,
unsupported keyword capture, malformed schema fallback with diagnostics.
Step 1861: Refactor grammar generator to use normalized schema IR (10 tests)
Update tools/mcp/generate_tool_grammars.py to consume normalized schema IR from
Step 1860 rather than ad-hoc field inspection.
No strict keyword expansion yet; this step is structural refactor only.
Tests (10): generated grammar parity for simple tools, deterministic ordering, dispatch includes all tools, per-tool files still generated for all tools, no regression in existing enum behavior, integer/number separation preserved, boolean/null handling preserved, optional field handling preserved, required field ordering stable, script exit code semantics unchanged.
Step 1862: Top-level strictness gate (8 tests)
Disallow any-object/any-array at top-level argument properties unless explicitly
annotated with x-whetstone-allow-broad: true in schema metadata.
Tests (8): gate fails on broad top-level property, gate passes when annotated, gate passes when strict object spec exists, failure message includes tool+field, report includes waiver count, waivers are deterministic, unknown annotation ignored, CI-friendly nonzero exit on violation.
Step 1863: Sprint 163 Integration Summary (8 tests)
Add editor/src/Sprint163IntegrationSummary.h.
Record: steps_completed=5 (1859-1863), strictness_baseline_established=true,
top_level_broad_fallback_blocked=true, success=true.
Tests (8): constructable summary struct, steps_completed==5, baseline artifact exists, normalizer available, generator refactor active, strictness gate active, no tool-count regression (347), success==true.
Architecture Gate
- No MCP runtime behavior changes; this sprint is tooling + generation pipeline only
- Generated grammar artifacts remain in
tools/mcp/grammars/ - Deterministic generation is mandatory (stable ordering)
- All strictness checks must be machine-readable JSON