Add contextual continuation hints and ephemeral refresh prompts

Summary

Codex should expose a continuation-hint system after assistant messages, errors, and idle periods, with ephemeral hints/prompts and a shared annotation pass that decides what should be suggested next.

Problem

There is currently no coherent system for helping the user continue a thread with minimal friction when the next move is obvious.

Useful cases include:

• after an assistant message, suggesting a contextually meaningful next-key action
• after Escape, surfacing what can be resumed or continued
• after an error, hinting the most relevant recovery/continue action
• after long session duration, hinting a summarization/refresh prompt

These all want the same substrate: a unified annotation/intention pass that can identify likely continuation endpoints and express them as concise UI hints.

Requested behavior

• Add contextual continuation hints after assistant messages, errors, and escape/cancel moments
• Support keyed hints such as 1/2/3/4/5/6/7/8/9 when those keys have contextual meaning in normal mode
• Detect teleological endpoints like would you like to... / enumerated next-step affordances and turn them into hints without requiring the user to restate them manually
• Add idle/session-age refresh hints, including summarization prompts after long-running sessions
• Render those refresh prompts as ephemeral messages that disappear automatically on the next user message, assistant message, or tool event

Architecture direction

This should not become a pile of separate per-feature LLM calls.

Instead:

• have one global generic annotation pass that tags a message/turn with actionable continuation metadata
• use that shared pass for loom endpoints / continuation hints / prompt necessity detection
• reserve separate extraction tasks only for heavier cases that would overload a single pass, especially if smaller models are used

Success criteria

• After relevant assistant turns, the UI can suggest a meaningful continuation key/action
• Error and escape flows can surface recovery/continue hints automatically
• Long sessions can hint a summarization/refresh move without permanently polluting the transcript
• Ephemeral hint messages self-remove on the next conversation/tool activity
• The system is driven by one coordinated annotation layer rather than many disconnected ad hoc calls

Related context

A narrower companion request covers a direct input gesture for continuation via double-Enter.