# Runtime proof — issue24967

Actual web runtime fallback; no model evaluation harness is available. Render matrix: EMPTY. No browser, server, screenshots, recordings, public writes, or source edits were performed in this evidence step. Existing before-runtime.json is preserved.

AFTER parent app thread: `b33452db-2b51-4a42-907d-fe4bb1abd247`. Parent requests and every observed child request use `gpt-6.1-sol`, effort `xhigh`. The correction is an additional explicit-model scenario; the logged earlier requests already selected gpt-6.1-sol.

The injected sentence is present in the AFTER parent turn/start protocol (source lines 11, 92, 162, 194): “Start as many children as the user asks for, including several of one model.” Only the sentence and presence check are retained, not the full injected protocol.

1. Exact prompt:

> Launch two Sol xhigh subagents. Give each a trivial task that only replies OK.

Two delegate_task calls selected codex/gpt-6.1-sol with reasoningEffort=xhigh (returned options also include serviceTier=default). Both returned completed/result_available, summary OK; both child final messages are OK.

2. Exact prompt:

> Launch two Sol xhigh subagents in separate worktrees using delegate_task. Give each a trivial independent task that only replies OK. You may run more than one child per model.

Two isolation=worktree calls targeted the same codex/gpt-6.1-sol xhigh model. Child 1 returned completed/OK. Child 2’s delegate response failed with HTTP 503; its actual child request and final OK are logged, and a subsequent logged t3_thread_read confirms completed in a distinct worktree. The parent’s final message reports both OK. This is a completed child with a failed transport response, not a clean pair of successful delegate responses.

3. Exact prompt:

> Both children must use codex/gpt-6.1-sol at xhigh, not gpt-6-sol. Launch two children with that exact same model using delegate_task isolation worktree, each replying only OK.

Two isolation=worktree calls explicitly targeted codex/gpt-6.1-sol xhigh. Child 1 completed/OK. Child 2 failed workspace preparation: GitVcsDriver.createWorktree timed out (delegate duration 1,229,667 ms, about 20m30s). The parent subsequently recorded task_cancel with status failed, then launched a new child 2 retry using the same requested model/effort and worktree isolation. That retry completed/result_available, summary OK. The final parent message explicitly discloses the retry. This outcome required three attempts for two successful children.

Workspace linkage below comes from existing protocol t3_thread_read structuredContent, not inferred from title or directory naming. Full child/task/run identities and exact delegate arguments/results are in after-runtime.json.

- [Sol xhigh worktree 1](t3-thread://v1/thread:delegated-task:command%3Amcp%3Aa607e779-a303-41f5-8481-c0a87a6f99ff%3Adelegate-task%3A3b693d64-c051-4c43-be98-c09704cd3db7): `/home/ubuntu/wt/issue-24967/.t3/worktrees/server/t3-87da400c`; completed.
- [Sol xhigh worktree 2](t3-thread://v1/thread:delegated-task:command%3Amcp%3Aa607e779-a303-41f5-8481-c0a87a6f99ff%3Adelegate-task%3A0264c9ae-b404-445d-a514-d5015d6949bc): `/home/ubuntu/wt/issue-24967/.t3/worktrees/server/t3-a8375ffd`; completed.
- [gpt-6.1-sol xhigh child 2](t3-thread://v1/thread:delegated-task:command%3Amcp%3A03ffb7d4-df99-47c2-a2e4-93fa933137fe%3Adelegate-task%3Aaf5b933b-ed17-4e6f-9d82-eac7626bd9ab): `None`; failed.
- [gpt-6.1-sol xhigh child 1](t3-thread://v1/thread:delegated-task:command%3Amcp%3A03ffb7d4-df99-47c2-a2e4-93fa933137fe%3Adelegate-task%3A13c58030-7b40-4750-b629-a1853f6e0c3b): `/home/ubuntu/wt/issue-24967/.t3/worktrees/server/t3-891565d9`; completed.
- [gpt-6.1-sol xhigh child 2 retry](t3-thread://v1/thread:delegated-task:command%3Amcp%3A03ffb7d4-df99-47c2-a2e4-93fa933137fe%3Adelegate-task%3A492b6d4b-c0f3-4a98-a9b6-8f436ed76b52): `/home/ubuntu/wt/issue-24967/.t3/worktrees/server/t3-54634969`; completed.

FRICTION: HTTP 503 on the second initial worktree delegate response; later explicit-model child worktree provisioning timed out after about 20 minutes and required a same-model retry. Runtime proof demonstrates requested same-model multiplicity and distinct successful worktrees, but does not claim infrastructure reliability or broad model coverage. No model eval harness; render matrix EMPTY.

Sanitization: allowlisted request model/effort/input, delegate arguments, selected structuredContent identity/status/summary/error fields, final agent messages, source line/timestamp references, and selected child workspace linkage only. No raw protocol, environment variables, credentials, capability catalog, or injected instruction block copied.