ONEOFFICEAUTOMATION

The Java 21 production platform lives in platform/ (see its README). The Python code in reference/ is the behavioural specification it must match.

Python reference implementation

Contracts and the Python reference implementation of the ONEOFFICEAUTOMATION platform: the OOA Flow format, Package SDK, Runner, AI Gateway and agent runtime. Built from the Product Specification v1.1 and Technical Architecture (18 Oct 2026). Architecture section 1 makes this Python slice the behavioural reference; the Java 21 production Runner must pass the same conformance suite.

Layout (architecture section 15)

contracts/flow-schema/   OOA Flow v1 JSON Schema (the shared contract)
platform/                Java 21 production platform (Gradle): SDK, Runner, packages, adapters, CLI
contracts/conformance/   language-neutral Runner test cases
reference/ooa/sdk/       Package SDK: decorators, manifests, signing, types, errors
reference/ooa/runner/    Runner: expressions, validation, interpreter, events, core package
reference/ooa/packages/  first-party packages (string, json, datatable, file, excel, rest, genai, agents)
reference/ooa/ai/        AI Gateway, PII guard, budgets, governance log, provider adapters
reference/ooa/agents/    agent runtime, tools from bots, approvals, MCP server
reference/ooa/governance.py  audit chain, RBAC with scopes, versioned repository
examples/refund/         spec journey 9.6: refund agent with bots as tools
examples/excel/          invoice summary TaskBot
docs/                    llms.txt and requirements traceability

Quick start

cd reference && pip install -e ".[test]" && python -m pytest -q

Run the refund agent journey (spec 9.6) with the deterministic scripted model:

cd examples/refund
export OOA_MODELS=models.json
ooa validate bots/*.json refund_agent.flow.json
ooa publish bots/*.json refund_agent.flow.json -m "refund v1"
ooa --user ops.anna --roles "Automation Runner,AI Admin" agent run refund_agent \
    -i "customer_email=Order ORD-10042 arrived cracked. Please refund."
ooa approvals list
ooa --user fin.lead approvals approve <approval-id> --comment "Checked"
ooa --user ops.anna agent resume <agent-run-id>
ooa audit verify

To use a real model, point OOA_MODELS at models.anthropic.json (needs ANTHROPIC_API_KEY) or models.litellm.json (needs OOA_LITELLM_URL). State lives in .ooa/ (repository, audit chain, AI governance log, approvals, agent checkpoints).

Enterprise controls in this slice

Signed packages with publisher allow-lists; secrets as SecureString that never reach logs, events, text variables or models; schema validation before any run; deterministic plan hashes; typed errors with suggested fixes; agent tool allow-lists, argument validation, step and cost ceilings; human approval with separation of duties; PII masking on every model call; role-based model access and team budgets; a hash-chained audit trail; permission scopes enforced for people, agents and MCP clients alike.

Not in this slice

This Python slice is the reference implementation. The product itself is the Java platform under platform/ (One AgentStudio, One CommandCenter, OneRunner, the recorder and UI automation, SQL Server and PostgreSQL, single sign-on, external vaults, the task inbox). Still open there: voice, business KPI dashboards and the agent marketplace. See docs/TRACEABILITY.md (the Java table) for status by requirement ID.