v0.2.0

Highlights

• Preserve deterministic workflow results when the optional LLM dependency is unavailable.
• Report DEGRADED consistently across workflow execution, orchestration, evaluation, and metrics.
• Bound in-memory tool invocation, evaluation case/run, and knowledge document/chunk stores.
• Add a bounded evaluation-case list API (limit parameter) and configurable capacity environment variables.
• Add regression tests and synchronized English/Chinese documentation.

Verification

• ./tools/check-format.sh — 121 files checked, no violations
• Java 17: ./mvnw --batch-mode --no-transfer-progress clean verify
• 112 tests passing
• CodeQL, dependency review, build, and analysis checks passed on PR #17

This release is based on the squash merge of PR #17 into main.

Delivery Harness v0.1.0

This is an educational MVP and reference implementation, not a production
delivery or compensation service. It uses synthetic business data, in-memory
state, advisory model output, and no authentication, authorization, tenant
isolation, or rate limiting. Do not expose it to the public internet or use it
with real personal or order data without adding those controls.

This release provides two synchronous workflows for abnormal-order analysis
and compensation suggestions. Compensation amounts remain owned by the
deterministic rule engine; model output explains the decision and can trigger
human review, but no payment is executed. The release includes the runnable
harness-api fat JAR and its SHA-256 checksum.

Highlights:

• 109 automated tests with end-to-end workflow coverage against a stub model.
• Rule-base-driven compensation conditions and actions.
• Deterministic fallback when the model endpoint is unavailable.
• Trace IDs, evaluation status accounting, bounded feedback, and advisory
  guardrails.
• English and Simplified Chinese documentation plus four architecture records.