LLM Call run event shows 'No new context' incorrectly + missing load-history action

[rowan-stein]

User request

In the run events UI, LLM Call items often show "No new context for this call."

Problems reported:

1. It should not be possible to have LLM calls without new context. Every LLM call contains new tool outputs and/or user messages.
2. Even in case there is "No new context for this call.", we should display a button to load history as if there were new context items.

Current behavior (from code)

In packages/platform-ui/src/components/RunEventDetails.tsx, the LLM Call Context panel displays only context entries with __agynIsNew === true. When there are no __agynIsNew entries, it renders the empty-state text "No new context for this call.".

The only button currently available is "Load more", which is shown only when there are older (non-new) items present in event.data.context to reveal client-side. This does not cover cases where:

• the backend provides no context items, or
• newness tagging is missing/incorrect,
  so the user cannot load history in those cases.

Specification / acceptance criteria

1. For every LLM Call event, the UI must provide an explicit action to load/view context history (label can be "Load history" or "Load context"). This action must be visible even when the UI would otherwise show an empty state.
2. The empty state must not misleadingly claim there was "no new context" when the delta is unknown/omitted. We need an explicit status signal (e.g. contextDelta.status in {available, empty, unavailable, redacted, first_call}) rather than inferring from context.length === 0 after filtering by __agynIsNew.
3. When delta is truly empty, show copy like "No new context added for this call." but still allow loading history.
4. Add/update UI tests so that in the "no new context" scenario, the load-history action is still present.

Notes

Relevant code/tests mentioning the string:

• packages/platform-ui/src/components/RunEventDetails.tsx
• packages/platform-ui/__tests__/components/RunEventDetails.context.test.tsx
• packages/platform-ui/src/components/__tests__/RunEventDetails.llmOutputs.test.tsx

[rowan-stein]

Update from investigation: root cause for repeated "No new context" is the UI attempting to resolve all llmCall.inputContextItems[].contextItemId via GET /api/agents/context-items?ids=... which can exceed proxy URL limits (observed 414 Request-URI Too Large). Even switching to POST still leaves an unbounded design.

Proposed robust fix (Option A): add an event-scoped, paginated endpoint so the client never sends thousands of ids.

Proposed endpoint

GET /api/runs/{runId}/events/{eventId}/llm-context?cursor=<optional>&limit=<optional>

• Default limit: 50
• Semantics: return newest-first by default; next page returns older items.
• Response includes items (renderable context records) + nextCursor (for older page) + highlightIds (or per-item isNew).

Acceptance criteria

• UI initial load fetches only latest page (<=50 items) for LLM call context.
• "Load more" loads older pages until exhausted.
• No request includes unbounded id lists.
• When fetch fails, show explicit "Failed to load context" state + Retry (not "No new context").

Implementation should be done in the existing PR #1323.