fix(history): stop shipping cumulative request history in session detail; page-scope list payload lookups

changelog.d/activity-monitor-slowness.md

@@ -0,0 +1,12 @@
+---
+category: Fixes
+---
+
+**Fix admin dashboard slowness from response payload bloat**: the session-detail
+API no longer re-sends the full cumulative conversation history inside every
+turn (response payload is now O(total messages) instead of O(turns^2); the
+dedup the activity-monitor frontend previously did client-side moved
+server-side via `request_delta_start`), and the session-list queries fetch
+preview/model payloads only for the page's sessions instead of probing every
+stored request payload in the table. Markdown/JSONL exports also stop
+repeating prior turns' history inside each turn.

src/luthien_proxy/history/models.py

@@ -65,6 +65,17 @@ class ConversationTurn(BaseModel):
     original_response_messages: list[ConversationMessage] | None = None
     # Request params (everything except messages/system, which are already parsed)
     request_params: dict[str, Any] | None = None
+    # Index into request_messages where THIS turn's new messages begin.
+    # Agent clients (e.g. Claude Code) re-send the full conversation history on
+    # every request, so raw per-turn request payloads are cumulative. To keep
+    # the response payload O(total messages) instead of O(turns^2), the service
+    # strips the re-sent prefix from request_messages for unmodified turns
+    # (request_delta_start stays 0 and request_messages holds only the new
+    # messages). For policy-modified turns the full arrays are kept so
+    # original-vs-final diffs still line up, and request_delta_start marks
+    # where the new messages start. Clients should render
+    # request_messages[request_delta_start:].
+    request_delta_start: int = 0
 
 
 class SessionSummary(BaseModel):