RLM: expose prompts and history as symbolic session objects

[Hmbown]

Why

RLM is strongest when large context stays outside the model's token stream and is manipulated through durable symbolic handles. Today the RLM REPL can hold an input handle and return bounded projections via helpers such as peek, search, chunk, sub_query_batch, and handle_read. The next step is to make CodeWhale's own prompt/session state addressable the same way.

Shape

Add first-class symbolic objects for the active session:

• current system prompt and mode overlays
• latest user prompt and composer history
• transcript slices by turn/cell/tool-call id
• saved session metadata and titles
• tool traces and large tool outputs through existing handle storage

These should be readable by RLM through bounded, structured APIs instead of by copying whole transcripts into prompts. The model should be able to ask for projections, filters, and exact references, while the REPL keeps large objects resident.

Acceptance criteria

• RLM sessions can inspect prompt/history/session objects through stable symbolic identifiers.
• The parent transcript receives compact handles/summaries, not raw prompt or transcript dumps.
• Sensitive surfaces respect existing trust/redaction boundaries.
• Tests cover at least one active-session prompt object, one transcript slice, and one large-tool-output handle.
• Docs explain the pattern without encouraging agents to paste their own prompt or whole history into context.

Notes

This should build on crates/tui/src/rlm/*, crates/tui/src/repl/runtime.rs, and crates/tui/src/tools/handle.rs; avoid creating a parallel context storage path unless the existing handle store cannot support the access pattern.

[Hmbown]

Design bar for this work:

This should not land as a bag of JSON handles or a hidden helper API. The user-facing shape should make symbolic context feel inspectable and powerful: prompt/history/session objects as named, navigable artifacts; compact object cards; a clear graph or stack of references; preview panes for bounded projections; and exact provenance for every slice.

The RLM should feel like it can reason over its own prompt/history as first-class mathematical objects while still keeping raw text out of the token stream. The implementation can be pragmatic, but the experience should be unusually sharp: beautiful, legible, and obviously more powerful than copy-pasting transcript text.

[Hmbown]

Fixed by #2047 (a6bd5ac): RLM now exposes active session prompt/history/transcript objects through stable session:// identifiers, rlm_open can load those refs, transcript cards stay compact for large tool output, and README/docs explain the pattern without encouraging raw prompt/history dumps. Verification included session object tests for prompt refs, transcript refs, large tool output compaction, rlm_open loading, and cargo check for codewhale-tui. Thanks @Hmbown for the design notes and acceptance criteria.