AttackGen Assistant: apply refinements to the downloadable files, not just the chat

[mrwadams]

Problem

The AttackGen Assistant (pages/4_💬_AttackGen_Assistant.py) refines the scenario and — since #44 / #57 — the Detection & Response narrative, but its edits live only in the chat history. The downloadable files offered on the scenario pages (the scenario .md, the Detection & Response .md) always reflect the originally generated content, so a user who refines an output in the Assistant can't download the refined version — they'd have to copy text out of the chat by hand.

This is the "flow-back" nice-to-have explicitly deferred in #44:

Consider letting edits flow back so the updated narrative can be re-downloaded (nice-to-have; the current Assistant doesn't persist edits back to the source page either).

Why (current behaviour)

Downloads are built on the scenario pages from persisted session state (core/scenario_page.py):

• Scenario markdown — st.download_button(..., data=cleaned, ...) (~scenario_page.py:362)
• Detection & Response .md — data=defense_state["download_md"] (~scenario_page.py:454), where download_md = assemble_defense_document(deterministic_md, narrative_md, title=...) (scenario_page.py:286).

The Assistant only reads the cross-page handoff keys (last_scenario_text, last_defense_narrative) and appends model replies to assistant_messages_<target>. Nothing is ever written back to last_scenario_text, the narrative_md / download_md in f"{page_id}_scenario_defense", or the per-page text_key the download button reads.

Desired behaviour

After refining in the Assistant, the user can download the refined scenario and/or Detection & Response narrative. Covers all three Assistant modes: Scenario, Detection & Response, and Scenario + Detection & Response (the combined mode — one "apply" should update both consistently).

Sketch of the work

Two viable shapes:

1. Self-contained (recommended, most contained): add download buttons on the Assistant page itself for the current, refined artifact(s). No changes needed on the scenario pages.
2. Write-back: propagate the refined text into the session-state keys the scenario pages read, so returning to the scenario page shows updated downloads. Heavier — the Assistant holds the cross-page last_* keys but not the page-scoped text_key / f"{page_id}_scenario_defense" dict, so those would need threading through the handoff.

Either way, an explicit "Apply changes" step is needed: the chat is conversational prose, not a clean document replacement, so blindly writing the last reply back won't work. A dedicated model call that returns the full revised artifact (or a fenced block the UI captures) is the reliable approach.

Constraints / considerations

• Never LLM-edit the deterministic STIX reference. The Detection & Response download is deterministic_md + narrative_md. Only narrative_md may change; on apply, rebuild download_md = assemble_defense_document(deterministic_md, edited_narrative_md, title=...) so the deterministic join stays byte-for-byte intact.
• Signal edited vs. original (e.g. button label / filename suffix), and consider keeping the original available or confirming before overwrite.
• Combined "Both" mode should apply consistent edits to both artifacts in one action.

Context

Follow-up to #44 / #57 (Assistant editing the Detection & Response narrative) and the purple-team companion in v0.13 (#42). Builds naturally on the in-progress "Scenario + Detection & Response" combined Assistant mode.