feat: extend tool_arg_contract to catch type mismatches (Closes #1528, #1529)

[Lexus2016]

Summary

Extends check_tool_args_contract() in agent/tool_arg_contract.py to validate basic types (string, integer, number, boolean, array, object) against the schema's type declaration, mirroring the existing validate_tool_args() logic in tools/tool_search.py so native tools get the same type guard already available to discovered tools.

Changes

• agent/tool_arg_contract.py (+75 insertions, -6 deletions):

  • Added ArgContractViolation.type_mismatch() classmethod
  • Added _SCHEMA_PY_TYPES dict and _check_type() helper (mirrors tools.tool_search)
  • Added type-checking loop in check_tool_args_contract() — runs after enum check, validates each present arg against its declared type
  • Updated module docstring to reflect type checking is now included
  • Default-OFF gating unchanged (tool_arg_contract_enabled() not modified)

• tests/agent/test_tool_arg_contract.py (+104 insertions, 0 deletions):

  • 13 new test cases covering: violation shape, all 6 correct types, 7 parametrized type-mismatch cases (including bool-for-integer edge case), number accepts int, array accepts tuple, union types, unknown-type fail-open, no-type-declaration skip, None skip, combined missing_required + type_mismatch, combined invalid_enum + type_mismatch

#1528 Audit Results

All 8 native tools (terminal, search_files, tool_call, write_file, patch, browser_console, execute_code, process) already have required arrays and type declarations on all properties. No schema fixes needed — all tools are already contract-check compatible.

Design Decisions

• bool is subclass of int — explicitly rejected for integer params (isinstance(value, int) and not isinstance(value, bool))
• Fail-open on unknown types — unrecognized type strings return True (don't block dispatch)
• Union types — value must match at least one variant in a ["string", "integer"] list
• Dependency-free — _SCHEMA_PY_TYPES and _check_type() kept local (not imported from tools.tool_search) to maintain the module's dependency-free design like its siblings agent.verify_policy / agent.policy_interceptors

Line count

185 total changed lines (75+6 + 104+0) — within 200-line self-merge cap.

Checks

• lint ✓ (ruff check)
• format ✓ (ruff format)
• tests ✓ (49 passed — 36 existing + 13 new)

Closes #1528
Closes #1529

Co-Authored-By: Hermes Evolution evolution@hermes.ai