Summary

Factor repeated OpenAI transport-stream test fixtures and reconcile dispatch-from-config with its shared harness, without changing behavior coverage.

Problem

src/agents/openai-transport-stream.test.ts is 13,865 lines. Roughly 15-20% is repeated model/chunk fixture construction, and a reasoning replay cluster duplicates the same test body across model variants. src/auto-reply/reply/dispatch-from-config.test.ts also duplicates fixture and setup behavior already owned by its sibling shared harness.

The duplication makes regression-heavy tests harder to review and maintain. Many cases carry issue-linked coverage, so assertion or scenario loss is not acceptable.

Proposed work

• Add typed completions/responses model factories and a completions chunk factory.
• Replace hand-built fixture literals while preserving meaningful overrides.
• Fold identical reasoning replay cases into named tables.
• Fold standalone completions parameter cases into the adjacent tables.
• Reuse the shared dispatch harness where semantics match; retain richer local setup with a boundary comment.
• Leave src/gateway/server-methods/agent.test.ts untouched.

Acceptance evidence

• Affected files pass on Blacksmith Testbox before and after.
• Exact before/after test counts match.
• Every existing assertion survives, modulo mechanical table/factory composition.
• Diff reports the test LOC reduction and no production-code changes.