[ViperJuice]

See #369. This implements the plan merged in #370, rev 2, with the two edits the ruling required.

What changes

All 42 nullable arguments are now advertised as anyOf: [X, {"type": "null"}]. Before, they used type: [X, "null"], plus null among an enum's values. Now null never appears as an enum value, a const, a default, or a member of a type array. The change is made in one place, src/pmcp/tools/schema.py::_canonical_nullable, and the snapshot diff is exactly those 42 nodes.

Hosts now receive (captured from the real binaries for the plan):

• OpenCode's Google provider: sends {type: X, nullable: true}. Before, it sent the string "null" in 4 enums and dropped null from all 42 arguments.
• Codex and Claude Code: pass the new form through unchanged.

The gate

The gate accepts and refuses exactly what it did before.

gate_error_for unfolds a nullable union's refusal into X's own errors before best_match ranks them. So a refused call reports the same message, keyword and path as 3.0, including calls with several failures. The one exception is that enum messages no longer list None.

validate_at_gate is the server's only gate call. The migration guide's checker uses gate_error_for.

Ruling edits (on top of the plan's patch)

• Fallback holds by construction. best_match now runs inside the fallback's try. If it fails, the gate uses jsonschema's own best_match, and if that also fails, the first raw error. A new test, test_a_refusal_survives_a_ranking_that_raises, forces best_match to raise. It checks that the call is still refused and that the scoped audit records exactly one audit.rejection with terminal_status: "invalid_arguments".
• CHANGELOG wording. In the amended api: advertised tool schemas drift from the pydantic models and unknown keys are silently ignored (C-10, M-02) #236 Changed entry, the "28 optional arguments … were previously rejected" clause now follows "accepts an explicit null … as the handlers always did".

Evidence (head 6439b21, base b9bf9d3, dev0)

• The patch is the proven one. The patch embedded in the merged plan is byte-identical to the proven spike's. Every file it touches matched the spike before the ruling edits, and the regenerated snapshot is byte-identical.
• Gates: ruff check, ruff format --check (182 files), mypy src/pmcp (53 files) and scripts/check_security_claims.py (131 cited node ids) all pass.
• Full suite, py3.10: 9101 passed, 5 skipped, 80 deselected.
• Targeted modules on 3.10, 3.11 and 3.12: 1129 passed on each. The modules are portability, schemas, migration_doc, scoped_advisor_audit, task_numeric_bounds and task_units.
• Differential fuzz: 69,000 calls with 1–4 sites each, compared against main's live schemas. 0 differences in what is reported, 0 accept/refuse mismatches.
• Mutants: 19/19 killed. These include:
  • the rev-1 rank-then-fold order (M15);
  • no fold fallback (M16);
  • best_match moved outside the try (M18, the rev-2 shape);
  • no last-resort first error (M19).

https://claude.ai/code/session_01Qq9UsYfWoENCKhQsuqRkNK

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.