Serialization Audit

Serialization Audit Skill

Overview

The serialization audit workflow generates branch-specific JSON snapshots of API behavior and compares them to detect behavioral differences. It exercises the full pipeline: API submission → queue → job processing → Elasticsearch storage → API response.

Workflow

1. Run Audit Script on Both Branches

The audit runner lives at .agents/skills/serialization-audit/scripts/audit-api-surface.ps1. It:

• Submit events with different JSON casing conventions (snake_case, PascalCase, camelCase, mixed)
• Capture the raw request body, Elasticsearch stored document, and API response
• Save each to audit-output/{audit-run-id}/{branch-name}/{scenario}/ with files like:
  • request.json — what was submitted
  • elastic.json — what was stored
  • response.json — what the API returned

# Start Exceptionless locally first:
aspire run

# On main branch:
git checkout main
pwsh .agents/skills/serialization-audit/scripts/audit-api-surface.ps1 \
  -BranchName main \
  -AuditRunId live-serialization

# On feature branch:
git checkout feature/system-text-json-v2
pwsh .agents/skills/serialization-audit/scripts/audit-api-surface.ps1 \
  -BranchName feature-system-text-json-v2 \
  -AuditRunId live-serialization

Requirements: API and Elasticsearch must be running locally through Aspire.

2. Diff the Output

diff -r audit-output/main/ audit-output/feature-system-text-json-v2/ | head -100

Or for structured comparison:

# Compare specific test outputs
diff audit-output/main/events-post-snake-case/elastic.json \
     audit-output/feature-system-text-json-v2/events-post-snake-case/elastic.json

3. Categorize Differences

Common difference categories:

4. Write Targeted Tests

For each difference found, write a unit test that reproduces it in isolation:

// In tests/Exceptionless.Tests/Serializer/CasingCompatibilityTests.cs
[Theory]
[InlineData("reference_id")]   // snake_case - should always work
[InlineData("ReferenceId")]    // PascalCase - must also work
[InlineData("referenceId")]    // camelCase - must also work
public void Deserialize_ReferenceId_MatchesAllCasings(string key)
{
    string json = $$"""{"type": "error", "{{key}}": "test-ref-123"}""";
    var ev = _serializer.Deserialize<PersistentEvent>(json);
    Assert.Equal("test-ref-123", ev.ReferenceId);
}

5. Implement Fixes

Common fix patterns:

• Multi-word property casing: Add fallback in IJsonOnDeserialized.OnDeserialized() to check ExtensionData for alternate casings
• Date-only string parsing: Check for time separator ('T') before calling TryGetDateTimeOffset in ObjectToInferredTypesConverter
• Naming policy mismatches: Use [JsonPropertyName] attributes or TypeInfo modifiers

6. Re-run Audit

After fixes, run the audit into a new output directory (or the same branch directory — it overwrites):

pwsh .agents/skills/serialization-audit/scripts/audit-api-surface.ps1 -AuditRunId post-fixes

Compare again to verify differences are resolved.

Key Files

Design Principles

1. Backwards compatibility first: Any payload that worked with Newtonsoft must still work with STJ
2. Snake_case output, any-case input: Serialize as snake_case, but accept PascalCase, camelCase, and snake_case on deserialization
3. Preserve user data types: Don't expand date-only strings to DateTimeOffset — users may store non-date strings that happen to look like dates
4. Test the full pipeline: Unit tests catch the bug, integration tests prove the fix works end-to-end