Codex Desktop should expose exact current and past chat transcripts to the assistant when the user asks

What feature would you like to see?

Codex Desktop should provide a first-class, user-authorised way for the assistant/model to read the exact transcript of:

• the current active chat, from the first character to the latest turn
• older Codex Desktop chats when the user asks for them

The important requirement is exact transcript access, not only summaries, compacted context, UI previews, or best-effort memory. The assistant should be able to inspect the conversation character by character when the user explicitly asks it to audit, count, quote, review, or reconstruct what happened.

Why this matters

Today, after compaction or long-session degradation, the assistant may only have a summary of earlier turns. That means it cannot reliably:

• audit exactly what was said earlier in the same chat
• count or inspect raw transcript characters
• verify whether a summary lost a detail
• produce accurate end-of-chat handoffs from the real transcript
• review old sessions when a user asks it to find context across prior work
• distinguish between UI-visible chat history and model-visible context

This is especially important for long-running project work where Codex Desktop is used as a working notebook, planning surface, and implementation coordinator.

Suggested capability

A safe implementation could expose transcript-reading tools with explicit user intent and appropriate limits, for example:

• read_current_thread_exact(start/end or page cursor)
• read_thread_exact(thread_id, start/end or page cursor)
• list_threads/search_threads metadata helpers
• clear markers for omitted tool output, binary attachments, or redacted secrets
• optional exact byte/character counts returned by the tool

The key point is that the model should be able to access the stored transcript as data when the user asks, rather than relying on whatever survived context compaction.

Safety and privacy expectations

• The user should remain in control of when old chats are read.
• The tool should make it clear which thread is being read.
• Secrets or sensitive tool outputs may need existing redaction policy support.
• This should not silently inject all old chat history into every prompt; it should be deliberate retrieval.

Duplicate check performed

I searched existing issues first. Related but not duplicate:

• #5781: chat export support using existing Codex logs. Useful for user-facing export, but this request is specifically for assistant/tool-readable exact transcripts during a session.
• #22220: conversation compaction telemetry/context health. Related because compaction causes loss of model-visible detail, but this request is for exact transcript retrieval, not telemetry.
• #18993: VS Code extension cannot open past conversation history. Related to UI history loading, but this request is for model-readable transcript access when the user asks.

Additional information

This would also make chat-end/handoff workflows more reliable: instead of producing handoffs from compacted summaries, the assistant could read the exact active transcript and any user-selected old chats, then produce a faithful summary with clear evidence boundaries.