The test
The experiment asked whether architectural work could survive a hostile handover boundary: different vendor, different harness, no session transcript, and no bespoke briefing. The durable handover consisted of a ten-line agent pointer, versioned native documents in Git, and the same public CLI available to any user.
- Claude Code grew a live product architecture through the stateless design interview.
- The session ended with 119 open questions. No conversational state was transferred.
- Codex oriented from the repository pointer, ran
yarramate design, and resumed at the model’s recomputed top question. - Atomic apply batches either compiled as a complete workspace or wrote nothing.
- The foreign agent reduced the open set from 119 to 56 and left 673 reviewable lines of model enrichment.
What the handover exposed
Two defects escaped the happy path
Codex encountered a retired-requirements gap and a flow-item authoring failure that had survived more than 300 tests and several days of single-harness use. Both were reported, fixed, and released in the same session as YarraMate v0.8.1.
The verifier graded its own authors
Evidence reconciliation had already caught the original session overclaiming two subjects as current. Later, the model caught an approval-path regression before release.
“The model was not merely documentation.”
The important result is not that either agent was flawless. It is that mistakes became explicit, source-located, and reviewable instead of becoming silent handover loss.
The public receipt
Model at resume: 230 concepts, 344 relationships, 6 documents
Handover: one ten-line pointer file plus the published CLI
Read the canonical case study in the product repository for the complete chronology, qualifications, and links.
What this does not prove
This is one adversarial case study, not a controlled estimate of productivity. It demonstrates resumability, deterministic write gates, and defect discovery under a real handover. It does not prove that a checked model always beats a strong design document.
YarraMate’s separate spec-build benchmark found no measurable convergence advantage over a good design document under a strong external specification. That negative result remains part of the product record.
Try the same boundary
Install YarraMate, initialize a repository, and let the repository—not the chat session—carry the next question.
npm install --global yarramateyarramate init .