[drk-ram]

Summary

• Reject duplicate JSON object keys and NaN/Infinity constants in agentrust_capture_core.state.load_state (returning None) and pass allow_nan=False in agentrust_capture_core.hashing.sha_mapping.
• Verify seal.get("alg") == "SHA-256" when "alg" is present in agentrust_capture_core.seal.check_seal (returning INTEGRITY_BROKEN on algorithm mismatch).

Closes #366

Validation

• python -m pytest packages/agentrust-capture-core/tests/test_core.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 e8ff9802582d6738ac8e713c58ff0f5581ca928d in an isolated checkout.

The capture-core changes reject non-finite values before hashing, refuse duplicate/non-finite persisted JSON, and check a present seal algorithm while retaining the established legacy behavior when that field is absent. Existing valid seal/state controls and the new malformed-input cases pass. No blocking finding in the changed package.

Validation: all 56 package tests passed, scoped Ruff and diff whitespace checks passed, and all PR commits carry sign-offs. This approval does not establish completion of hosted CI or the maintainer policy gate.

Reviewed with OpenAI Codex.

[carloshvp]

Hosted CI follow-up at head e8ff9802582d6738ac8e713c58ff0f5581ca928d: after I approved the held capture-core run, its Python 3.9 job failed in actions/setup-python before dependency installation or tests: "The version '3.9' with architecture 'x64' was not found for Ubuntu 26.04." Job: https://github.com/agentrust-io/integrations/actions/runs/40037909203/job/120848920852.

The current actions/python-versions manifest publishes Python 3.9.25 x64 builds for Ubuntu 22.04 and 24.04, so pinning this compatibility job to ubuntu-24.04 is a concrete way to retain the declared 3.9 coverage: https://github.com/actions/python-versions/blob/9834b4e4911ca2950ff6c3b7547f09eacc006848/versions-manifest.json. This runner setup failure is separate from the package source review and its 56 locally passing tests.

[rajnisht7]

the overall changes looks good, but still some places need hardening:

• load_state still accepts exponent overflow, 1e309, -1e309, 1e999 etc. are parsed as inf/-inf instead of being rejected. parse_constant only catches literal NaN/Infinity. Need a validating parse_float with math.isfinite.
• check_seal can now raise, because of allow_nan=False in sha_mapping, a snapshot containing nan/inf makes check_seal raise ValueError. This breaks the documented “Never raises” contract. It should return INTEGRITY_BROKEN instead. Especially bad when combined with (1): load_state accepts 1e309 -> later check_seal raises.
• save_state still writes non-finite JSON, It uses the default allow_nan=True, so it can write files that load_state immediately rejects. Prefer allow_nan=False here too.

Current test covers duplicate key + literal NaN + wrong alg, but misses:

• 1e309 / -1e309
• nested overflow
• check_seal with in-memory nan/inf (must not raise, must return BROKEN)
• save_state with non-finite values

also rebase branch on main

[dinakarjs]

Source follow-up at e8ff980, supporting the distinct entry-point checks in Rajnish's latest review. The diff adds parse_constant rejection and allow_nan=False for hashing, but those changes need to preserve the public load/save/check contracts together.

Please use separate regression cases for exponent overflow (including nested values), direct in-memory non-finite input to check_seal, and attempted non-finite save_state output. Preserve the expected outcomes separately: corrupt loads return None; seal verification returns INTEGRITY_BROKEN without raising; a refused save should leave an existing valid destination intact. Retain a finite save/load/seal positive control.

The combined added test can stop at its first failure, so independently collected cases would make each boundary observable. This is source inspection and follow-up to the existing review, not a fresh execution or an additional independently reproduced defect. Reviewed with OpenAI Codex.

[drk-ram]

@rajnisht7 @dinakarjs @carloshvp Rebased onto origin/main and updated at 4e9cec9:

1. Exponent overflow in load_state (state.py): Added _finite_float (parse_float=_finite_float) using math.isfinite so float exponent overflow (1e309, -1e309, 1e999, including nested values) raises ValueError inside the parser and causes load_state to return None.
2. Preserve check_seal "Never raises" contract (seal.py): Caught (ValueError, TypeError) around state_digest(snapshot) in check_seal so an in-memory snapshot containing nan or inf returns INTEGRITY_BROKEN instead of raising ValueError.
3. Refuse non-finite JSON in save_state (state.py): Passed allow_nan=False to json.dumps in save_state so attempting to persist non-finite values raises ValueError before atomic_write touches the destination path.
4. Independent regression cases (tests/test_core.py): Split the regressions into separate parametrized test functions covering load_state exponent overflow (including nested), load_state non-finite constants, load_state duplicate keys, check_seal on in-memory nan/inf returning INTEGRITY_BROKEN without raising, save_state with non-finite floats raising ValueError while preserving an existing valid sealed baseline on disk, and a finite save_baseline/load_state/check_seal positive control (75 passed, 1 skipped).

[dinakarjs]

Re-reviewed exact head 4e9cec9 on Linux/Python 3.12.14.

The load/save/seal contracts now agree: corrupt and exponent-overflow loads return None, direct non-finite snapshots return INTEGRITY_BROKEN, and refused non-finite saves preserve an existing destination. Separate regressions and a finite round-trip control are present. Fresh hash-locked capture-core environment: 76 tests pass; syntax/error lint selection and whitespace pass. The earlier contract/coverage follow-up is addressed. This review covers the capture-core software change; unrelated example integrations and hosted merge gates remain separate.

Reviewed with OpenAI Codex.

[carloshvp]

Reviewed current head 4e9cec9d658692626973550f5825fd208a7c89a1 in an isolated checkout.

The persisted-state parser rejects duplicate members, literal non-finite constants and exponent overflow; saving non-finite in-memory data preserves the prior valid file. Seal checking returns a broken verdict rather than leaking a numeric-serialization error, and a present unsupported seal algorithm is refused. All 76 package tests and 60 root tests pass.

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.