[HomenShum]

A visitor following a narrated comparison could lose the active trace row because scrolling ran before the SDK mounted it. Parent updates also replaced existing row DOM, and rejected requests appeared as completed spans. This moves scrolling into the row, uses the SDK's existing stable render-function API, and maps error events to failed spans.

Updates the trace framework to exact @assistant-ui/react-o11y 0.0.47 and @assistant-ui/store 0.3.20, including the required tap/Radix closure. The provider uses AuiConfig/config. The old SDK retained a referenced MessagePort after SSR; the new real-wrapper regression requires 55 realistic renders and natural subprocess exit. Unrelated dependency records, model/provider behavior and CSS are unchanged.

Validation on the frozen five-file candidate:

• Clean install and dependency tree; all three typechecks and citations; build; 61 tests across 11 files; deterministic 100-turn comparison; full and production audits with zero findings.
• Foreground browser comparison: 100 turns per pane, stable row identity, current-turn visibility, rejected-request status and successful rerun. Separate sustained harness: 127 cycles, 20 remounts, final DOM empty and matched mount/cleanup counts.
• Original-wrapper scrolling control, old-SDK natural-exit failure and rejected-status regression failure retained as causal evidence. An earlier background-tab observation was confounded and is excluded from the scrolling cause claim.
• Independent Astra code, dependency-graph and simplification review approved the exact frozen tree and repeated all eight install/check/build/test/demo/audit commands. Browser evidence was inspected separately, not represented as an independent browser rerun.

Costs and limits: JavaScript grows by 5,036 gzip bytes; CSS is identical. Ordered local update-to-DOM observations do not show a speedup. The existing 390px layout's 20px trace viewport remains a separate visual issue. No production, model-quality or unlimited-memory claim; no paid provider evaluation. Model-refresh PR #16 remains separate.

Evidence qualification (2026-10-14): Windows npm ls --all --json exited 0 but reported the same six extraneous packages in the baseline and every candidate/reviewer capture; the saved dependency tree is not warning-free. The whole-app foreground 100-turn run used candidate2; final candidate3 separately proves rejected-request recovery and the private sustained lifecycle cases. The 55 SSR renders are one scenario within the 61 outer tests, not additional tests. No new runtime rerun or cleanup is implied by this reconciliation.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
local-collab-mvp	Ready	Preview	Oct 14, 2026 8:08pm UTC
nodevoice	Ready	Preview	Oct 14, 2026 8:08pm UTC