agents-support-ops

A typed companion library for the OpenAI Agents SDK that turns traces, tool outcomes, approvals, and costs into support-engineering artifacts.

It composes the SDK's native human-in-the-loop, RunState, retries, tracing processors, and sessions. It does not fork or replace those primitives.

What it adds

• Redacted JSONL audit trails.
• Tool retry budgets and circuit breakers.
• Deliberate failure injection for regression tests.
• Versioned replay bundles with canonical digests.
• Evidence-linked escalation packages in JSON and Markdown.
• A synchronous trace processor compatible with the Agents SDK processor interface.

Installation

Install the model-neutral operational primitives:

pip install agents-support-ops

Install the optional OpenAI Agents SDK integration:

pip install "agents-support-ops[openai]"

from agents_support_ops import AuditLog, Redactor, SupportTraceProcessor

audit = AuditLog("audit.jsonl", redactor=Redactor())
processor = SupportTraceProcessor(audit)
# agents.add_trace_processor(processor)

See docs/sdk-composition.md for the integration boundary. Maintainers should follow docs/releasing.md for the OIDC-based TestPyPI and PyPI release process.

Compatibility

Capability	Supported versions
Core package	Python 3.11, 3.12, and 3.13; no agent SDK required
OpenAI integration	openai-agents>=0.6,<1
AI Escalation Lab live adapter	openai-agents>=0.18.2,<1

Replay schema 1.1 records typed run, tool, model, prompt, token, cost, and integrity provenance. Schema 1.0 bundles remain readable for backward compatibility. Canonical replay digests exclude only the bundle integrity record and the modeled runtime timestamps (created_at, evidence observation times, approval decision times, and model generation time). Metadata-like keys in ticket, tool, model, and recorded output payloads remain integrity-covered. ReplayStore accepts only portable replay IDs and verifies that resolved paths stay inside its configured root.

Escalation packages require field-specific evidence citations in narrative_evidence for summary, impact, reproduction_steps, suspected_component, and recommended_next_action. Each field must cite at least one evidence ID present in the package. JSON exports retain this mapping and Markdown exports display the citations beside each narrative field.