# Runtime evidence for #15313 PR: [#15313](https://github.com/pingdotgg/t3code/pull/15313). Before: [044f07257c](https://github.com/pingdotgg/t3code/commit/044f07257c). After: [53ede9f0c4](https://github.com/pingdotgg/t3code/commit/53ede9f0c4). Proof-audit pass 1 confirmed an empty UI matrix: server MCP/service/tests only, no renderer changes. Required rows 0; captured matrix rows 0 (0/0). Server-only exception: red test plus real-runtime proof; viewport/theme fixtures and visual diffs are not required. ## Reproduction and result Exact prompt, submitted once: > Switch this thread to GPT-6.1-Sol using provider instance codex_alt and low reasoning with t3_thread_configure, then write a four-line poem about changing your mind. Do not edit files or use subagents. [Before runtime transcript](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a60b641573b8991f/before-runtime.json): selection saved as codex_alt / gpt-6.1-sol / reasoningEffort low, but the original run and its sole attempt ended interrupted at 19:45:10 UTC, without continuation or poem. [After runtime transcript](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1b95a09339198773/after-runtime.json): prompt at 20:12:20.135 UTC on 2026-10-22; attempt 1 superseded at 20:12:44.073; persisted agent-created continuation at the same timestamp; attempt 2 started at 20:12:46.276 using codex_alt. The same run completed at 20:13:01.529 with gpt-6.1-sol and reasoningEffort low. The transcript contains exactly one human-created message; the continuation is agent-created despite its user role. The final assistant message is the requested four-line poem. Both JSON files parse and were reviewed in full for secrets before this upload step; a credential-pattern scan was also clean. Assertions checked the completed run, selection/options, both attempt states, shared run ID, one human message, four-line final poem, and interrupted baseline. The before transcript was already public when supplied. ## Test evidence and attribution [Captured red-test summary](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2df4f99d3a29d631/red-test.txt), supplied by the primary: baseline 044f07257c with test commit [ff3739bb45](https://github.com/pingdotgg/t3code/commit/ff3739bb45), command `vp test run apps/server/src/orchestration-v2/SelectionRestart.integration.test.ts`. The saved-selection assertion passed; expected completed run status was actually interrupted. Seven other tests passed. The summary reports that temporarily applying the existing PR implementation made this test pass, then reverting it before the test-only commit. This is the captured green test fact; no separate head test log was supplied, and this evidence step did not rerun tests. The after real-runtime JSON independently verifies completed behavior on the supplied head. ## Captured visuals Before — interrupted, no poem: ![Before: model selection saved but run interrupted without a poem](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2c1098929c17d28a/browser-screenshot-localhost-mvjyb7xv-8f3c1a84.png) After — agent continuation and completed poem: ![After: agent continuation and completed four-line poem with GPT-6.1-Sol and Low selected](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bda77cc5c46be934/browser-screenshot-localhost-mvjzaxiq-ca832b93.png) [After recording (49.573 seconds)](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ec909957aa0b61df/mcp-launch-client-3a1523f540-1b0b-41f8-9024-81ebd4895c7e-fwd-3a438301f78c51385a6-f2443c51-14ec-4636-a997-5f21e27bbb3b-mp4.mp4). Approximate recording checkpoints, inspected from sampled frames (not an exact wall-clock synchronization): - 00:00: submitted prompt is visible; initial work starts. - 00:10: initial attempt working. - 00:20: agent announces model update and capability lookup is visible. - 00:30: superseded attempt, agent steer, and GPT-6.1-Sol / Low selection are visible; replacement is working. - 00:40: replacement recovers the prior task context. - 00:48: completed four-line poem and Done status are visible. The before recording supplied to the primary begins only at the interrupted state; it is not full-interaction evidence. No full-before-recording claim is made here. The meaningful visible difference is the after agent continuation and completed poem versus the interrupted before state without a poem. Thread title, time, and an additional sidebar thread differ incidentally. ## Environment, limitations, and friction Primary captured the integrated real T3 Browser panel / real Codex run on fs-03, Chromium on Linux, reported viewport 1280x800. DPR was not recorded and remains unknown; the encoded recording is 2560x1600, which does not by itself establish DPR. Supplied metadata said dark, but the inspected screenshot and recording visibly appear light; theme metadata is therefore inconsistent. Native desktop/mobile clients and other provider binaries were not manually run. No renderer, responsive, theme, or cross-client visual claim is made. The screenshot includes incidental sidebar thread titles already present in the supplied capture; JSON contains synthetic reproduction content and internal identifiers, with no credentials found. Friction: evidence metadata — reported dark conflicts with visibly light capture; DPR unknown. Evidence completeness — before recording is partial, and no standalone head green-test log was supplied. Tooling/access — no blocker. This step started no browsers, servers, or tests and edited no source; scratch files remain under /tmp/issue15300-evidence for parent cleanup.