← Back to Daily Builds

Daily build note · August 3, 2026

Agent Trace Receipt Bridge

A local packaging layer that imports OpenAI Agents SDK trace JSON into the existing Agent Receipts Desk contract, produces internal and redacted client-safe evidence, and feeds the existing...

AI Tools Build note published Public demo coming soon

What shipped

A self-contained Python 3.11+ offline CLI with three commands:

  • sample creates stable healthy and risky trace fixtures.
  • import-trace maps a local trace to four review artifacts and a retained receipt index.
  • validate checks the receipt workspace, redaction boundary, and downstream field contract.

The healthy fixture demonstrates tool, handoff, passed guardrail, approval linkage, verification, timing, usage, and unknown-span evidence. The risky fixture deterministically produces all seven required trace risk codes. verify.sh proves the full workflow and four negative cases without installed packages or network access.

Architecture

The adapter is a linear standard-library pipeline: defensive shape detection, provider-neutral TraceEvent normalization, explicit event reduction, allowlist rendering, relative-path index rebuilding, and strict validation. JSON writes use same-directory temporary files followed by replacement. Receipt IDs derive from stable trace timestamps and IDs, making repeated imports deterministic.

The smallest stable pieces carried forward from Agent Receipts Desk are its schema version 0.1.0, required receipt field shape, receipt/index directory layout, deterministic Markdown concept, recursive pattern-redaction approach, and risk flag structure. They were adapted into this independent package; the parent build is not a runtime dependency and was not modified.

The downstream AgentOps Evidence Pack loader contract was copied into a focused assertion that accesses receipt_id, title, created_at, agent, review, repo.changed_files, tools_used, commands, verification, and open_questions from inputs/agent-receipts/runs/*/receipt.json without translation.

Trimmed scope

This smoke test intentionally excludes live trace ingestion, provider SDK installation, model-generated summaries, replay/evaluation, dashboards, databases, authentication, hosting, broad historical SDK compatibility, and performance analytics. It maps one local JSON trace at a time.

Limitations

  • Input compatibility covers the fixture shape and documented common aliases, not every Agents SDK release.
  • Sensitive-tool classification is a conservative name-based heuristic plus explicit approval_required evidence.
  • Redaction catches documented secret/path forms but cannot discover every kind of sensitive prose or identifier.
  • Internal artifacts contain concise synthesized evidence, not cryptographic provenance or signed attestations.
  • Re-importing the same deterministic trace replaces that run rather than creating a duplicate.
  • Validation establishes structural compatibility and tested leakage boundaries, not compliance certification.

Suggested next steps

  1. Test sanitized exports from the specific Agents SDK version used by the first buyer and add fixture-backed aliases only where observed.
  2. Add a user-owned redaction configuration file with testable custom patterns and path prefixes.
  3. Add optional source-file hashes or signed manifests for stronger evidence provenance.
  4. Pilot three golden traces through AgentOps Evidence Pack and collect reviewer feedback before adding ingestion or UI scope.