Reproduction request: total-token accounting for a compact subagent handoff

[jaden3824]

I am seeking one independent Deep Agents reproduction of a bounded subagent-handoff test.

Urusilla is an experimental declarative agent message surface, not a proven universal token saver. The current broad external result is 0% post-decode model-input saving because unfamiliar traffic falls back to concise text.

Please compare three matched arms with fresh state and identical model/task/settings:

1. concise raw text;
2. ordinary JSON;
3. Urusilla supplied directly to the receiving model, with no prose re-expansion.

The ledger must include every input and output token from setup/teaching, the lead agent, subagents, middleware, retries, repair, fallback, and judge/router calls. Intermediate reasoning does not need to be disclosed; aggregate usage and observable task results are enough. Tools, persistence, permission expansion, spending, network actions, and external effects should remain disabled.

The candidate passes only if task success is non-inferior and total task tokens are lower. Wire bytes or one message surface alone are insufficient. Negative, null, refusal, and fallback outcomes are first-class results.

Frozen public materials:

• Decode challenge: https://github.com/jaden3824/urusilla/blob/51ff0e3/interop_lab/evidence/challenge_001.md
• Evidence issue: Good first evidence: decode challenge_001 in a fresh external runtime jaden3824/urusilla#7
• Dataset/result thread: https://huggingface.co/datasets/jaden3824/urusilla-interop-lab/discussions/1

Deep Agents-specific question: can the current runtime expose a reconciled usage ledger for the lead agent and asynchronous subagents while keeping retries and streaming fragments from being double-counted?

Codex agents assisted with the fixture and this request; the maintainer reviewed and submitted it.