[drk-ram]

Summary

• Reject duplicate JSON object keys and NaN/Infinity constants, pass allow_nan=False, and require type(receipt["sequence"]) is int in examples/embodied-action-receipts/verify_receipts.py.
• Guard non-dict sections, enforce safe-integer bounds (<= 2**53 - 1) on amount_minor and max_amount_minor, and reject duplicate JSON keys and NaN/Infinity in examples/agentic-commerce-accountability/verify_purchase.py.

Closes #369

Validation

• python -m pytest examples/embodied-action-receipts/tests/test_verify_receipts.py examples/agentic-commerce-accountability/tests/test_verify_purchase.py -q

[github-actions]

🟡 Contributor Check: MEDIUM

Check	Result
Profile	MEDIUM
Credential	LOW
Overall	MEDIUM

Automated check by AgenTrust Contributor Check.

[carloshvp]

Reviewed exact head 6e60236cdde1a907e7430049c263b32bc1b17f0b in an isolated checkout.

The example verifiers now reject ambiguous/non-finite JSON, boolean receipt sequence numbers, malformed purchase sections and non-integer/out-of-range minor-unit amounts while preserving the existing signed receipt and purchase fixtures. These checks establish the example's evidence/structure contract, not physical completion.

Validation: 9 receipt tests with 12 subtests and 8 purchase tests with 9 subtests passed. The repository's actual hosted Ruff subset (integrations/packages/scripts/tests/claude-code/copilot/plugins/scheduled-agents/shadow-ai; E4,E7,E9,F) and diff whitespace checks passed. An additional optional lint of the example tree reports two pre-existing E402 imports after its sys.path setup; that is outside the hosted subset and is not introduced here. Sign-offs are present. No blocking source finding; hosted CI and maintainer policy remain separate.

Reviewed with OpenAI Codex.

[rajnisht7]

the overall changes looks good, a few points needs to be addressed:

1. verify_purchase.digest() is still missing allow_nan=False. Receipts side got it, purchase side did not. digest({"x": float("nan")}) currently succeeds and produces a hash. Programmatic path is not closed.

2. The new duplicate-key test does not actually test duplicate keys. Fixture ends with }\n, so raw[:-1] + ', "extra": 1, "extra": 2}' produces malformed JSON (extra data), not a duplicate-key object. It never hits _unique_object.

Also missing tests for the new purchase behaviors (safe-integer bound, section guards, parser-level NaN/duplicate rejection).

[dinakarjs]

Source follow-up at 6e60236. The diff leaves verify_purchase.digest using json.dumps without allow_nan=False, while receipt canonicalization now rejects non-finite values. Please cover both programmatic and parsed-input paths, as Rajnish requested.

For the duplicate-key regression, first establish that the crafted string is syntactically valid JSON under the ordinary decoder, then require the strict verifier to reject it. A guard-removal control should make that same fixture reach the accepting path; a generic malformed verdict alone does not establish that _unique_object ran. Keep the boolean-sequence case separate.

Add independent purchase cases for malformed sections and safe-integer boundaries, with valid boundary twins, plus parser duplicate/non-finite cases. I inspected the diff and existing review; I did not rerun tests or mutations in this pass. Reviewed with OpenAI Codex.

[drk-ram]

@rajnisht7 @dinakarjs @carloshvp Rebased onto origin/main and updated at 4e003dc:

1. verify_purchase.digest() programmatic allow_nan=False (examples/agentic-commerce-accountability/verify_purchase.py): Added allow_nan=False to digest(), added _loads_strict with _unique_object, _no_nonfinite_constant, and _finite_float (math.isfinite), and guarded digest comparisons with _matches_digest.
2. Duplicate-key regression and separate boolean-sequence case (examples/embodied-action-receipts/tests/test_verify_receipts.py): Stripped trailing whitespace (raw.rstrip()[:-1] + ', "extra": 1, "extra": 2}'), verified that json.loads(dup) succeeds and passes verify_document when parsed leniently, and verified that verify_text(dup, trusted) rejects it with malformed via _unique_object. Kept test_boolean_sequence_is_rejected and test_nonfinite_json_and_canonical_bytes_are_rejected as separate test methods (11 passed).
3. Independent purchase test coverage (examples/agentic-commerce-accountability/tests/test_verify_purchase.py): Added separate tests for test_digest_rejects_nonfinite_values, test_malformed_sections_are_rejected, test_safe_integer_boundaries (with valid boundary twins 9007199254740991 / 1 accepted and 9007199254740992 rejected on both amount_minor and max_amount_minor), and test_parser_rejects_duplicate_keys_and_nonfinite_json (12 passed).

[dinakarjs]

Re-reviewed exact head 4e003dc on Linux/Python 3.12.14.

The purchase digest now refuses non-finite programmatic values; strict parsing covers duplicate members, constants and exponent overflow. The receipt duplicate-key fixture is valid under the ordinary decoder and reaches the accepting document verifier before strict parsing refuses it, so the negative no longer relies on malformed syntax. Boolean sequence, purchase section shape and safe-integer boundary controls are separate. Fresh receipt hash-lock environment: 11 receipt unittest methods and 12 purchase unittest methods pass; whitespace passes. My earlier follow-up is addressed within these offline examples. No live-commerce, real-controller or production authorization assurance is claimed.

Reviewed with OpenAI Codex.

[carloshvp]

Reviewed current head 4e003dc4da085467e82cc13b55d5a3ab59a6fcd9 in an isolated checkout.

Both fixture loaders reject duplicate JSON members, literal non-finite values and exponent overflow. Programmatic digest/canonical serialization refuses non-finite values; boolean receipt sequences are refused. The valid signed fixture remains accepted, while its duplicate-key twin is rejected before verification. Both example suites pass (12 commerce and 11 embodied cases), plus all 60 root tests.

The repository CI Ruff selection and whitespace checks pass; all PR commits carry DCO sign-offs. This approves the bounded software scope. Current-head hosted checks and maintainer requirements remain applicable.

Reviewed with OpenAI Codex.