Remote delegation routing reproduction

Run against each checkout after vp i. Place drive.py, codex-replay.mjs, and claude-replay.mjs together in an empty scratch directory. Set REPRO_PARENT_MODEL and REPRO_TARGET_MODEL to distinct synthetic identifiers (their values are omitted from public evidence), then run:

python3 drive.py before <base-checkout>
python3 drive.py after <head-checkout>

The script starts two real unmodified T3 servers using fresh isolated homes, enables the real Codex and Claude drivers with replay CLI executables, pairs the environments using the real CLI/OAuth path, creates one synthetic project through each app, and starts a Codex parent through the CLI. The replayed parent calls orchestrator_capabilities and delegate_task over the real provider-session MCP endpoint. The target names the linked environment, project, Claude driver, and target model, with no providerInstanceId. The destination receives a real thread launch and provider turn. Only the CLI provider processes are replayed; no T3 services, HTTP handlers, database or adapter implementations are mocked. Existing recorded protocol frames supply required response envelopes.

The terminal outputs selected fields directly from delegate_task and the destination's real t3 thread list --json response. Model values are omitted; modelMatchesRequest compares the returned model with the request in memory. Raw protocol and server logs remain private and are not included. Recording uses the same 1500x1000 terminal capture, white theme, script and settings on both sides. Still captures were taken 500 ms apart and had zero differing pixels.

Base: 5c6f2f1229514881a18af286b8a63e15ab4bef70
Head: 826ea80c5293011c97cddd154e0d6d065c105ef8

This proves server routing and adapter selection, not live model inference or provider-specific authentication. Synthetic identifiers are advertised as custom models to both replay drivers; rejection of a model belonging to another instance is outside this proof.

AI-assisted; verified with the tests and tools listed above.
