[mavericksea-ai]

Summary

Tool-argument rules currently compare formatted text, so paths: ["sales 2025.csv"] matches paths: ["sales", "2025.csv"]. A wrong file request passes the rule and can make CI succeed. In an offline benchmark, this reports 100% versus 100% and zero lift; structural comparison correctly reports 0% versus 100%, −1.00 lift, and exit code 1.

Related issues

Closes #303.

Changes

• Compare each declared argument using JSON value encoding, preserving array boundaries, object members and scalar types.
• Preserve matching across ordinary YAML integer and JSON float representations, and allow extra top-level actual arguments as before.
• Cover both tool_called and tool_called_in_turn, plus structured collisions and compatibility cases; update the changelog.

Test plan

• make test passes (race-enabled Go tests and 22 plugin tests)
• make verify passes (fmt + vet + lint), using Go 1.25.13 with the repository's pinned tools
• make fmt run
• Offline real-CLI reproduction on main and published v0.12.0; eight regression subcases fail before and pass after

Notes for reviewers

The current Go 1.27 compiler produces export data unsupported by the pinned lint tool; make verify passes with the supported Go 1.25.13 toolchain. No module files or lint rules were changed. E2E-tag tests are not required for this change because no runner or e2e files changed; the separate reproduction runs the CLI end to end.

[CLAassistant]

All committers have signed the CLA.

[zpzjzj]

LGTM