[App][Remote] Stale active-turn state and repetitive short-word output after reopening a task

Summary

An existing task reopened through Codex Remote displayed older progress as if it were still active, together with a stop control and repetitive short-word output. Newer activity was present in the desktop's persisted history.

Observations

• Visible progress messages matched an earlier turn rather than the latest persisted activity.
• A thread-reading interface also returned the older turn as inProgress, despite subsequent turns in the local history.
• No explicit terminal event for that older turn was found in the inspected history; newer turns followed it.
• Short fragments such as Tool., Go., Do., End. and Stop. repeatedly appeared in the mobile chat view. They were not found in the inspected readable assistant-message entries.

Expected behavior

• Preserve chat history without automatically inheriting or resuming a previous run. New work should receive a separate run context; resumption should require explicit user action.
• After an execution process restarts, previously persisted RUNNING states should become interrupted unless explicitly resumed. After a client reconnects, verify live execution ownership before presenting a run as active; otherwise show an uncertain or interrupted state.
• Display historical messages, the current assistant response and tool-execution status separately. Keep internal analysis and protocol/state fragments out of ordinary chat text.
• Apply bounded repetition and no-progress protection, deduplicate replayed stream events, and avoid unbounded exception/retry loops without disrupting valid tool calls.

Investigation request

Please inspect active-turn reconstruction when a turn lacks a terminal event but newer turns exist, stream replay/rendering, tool-call parsing, stop conditions and exception/retry handling.

Scope and privacy

This describes a past observation, not a reliably reproducible test or a claim that the issue remains active. The exact cause is unconfirmed. The short fragments' origin is unknown, and there is no evidence that opening Remote restarted older work.

This report is intentionally abstract. It includes no identifying metadata, original screenshots, session exports, prompts, reasoning content or raw logs. Environment and account details are withheld. No attachments are provided.

No files or tests are identified. Start by tracing active-turn reconstruction, stream replay/rendering, tool-call parsing, stop conditions, and exception/retry handling. Done means the behavior is understood well enough to separate historical and live state, prevent stale or repetitive output, and define a reproducible validation path.