v2.45.2

Dashboard Provider Error Visibility

Provider and API failures now remain visible in dashboard transcripts instead of leaving unexplained, empty thinking blocks. Terminal failures also surface on fleet cards, while transient failures transition cleanly into retry state.

Added

• Inline provider errors. Failed assistant attempts display Error: <message> alongside any partial text or thinking already produced.
• Fleet error state. Terminal provider failures show an error chip and reason on the corresponding fleet card.
• Recovery parity. Error metadata and current terminal state survive hydration, refresh, and SSE resync.
• Retry-aware snapshots. Refreshing during automatic-retry backoff restores the failed attempt and retry warning without incorrectly marking the session terminal.

Reliability

• Preserves assistant/tool/assistant ordering and trailing partial output when a provider fails around tool calls.
• Suppresses finalized whitespace-only thinking placeholders while retaining active streaming placeholders.
• Clears provisional provider errors during automatic retry and overflow compaction without hiding eventual terminal failures.
• Prevents duplicate terminal status after exhausted retries.
• Correctly recognizes Claude 5 adaptive thinking and xhigh support while keeping dated Claude 3.x models on budget-based thinking.
• Keeps GPT-5.6 available at its existing xhigh level; first-class max and ultra support remains tracked separately.

Documentation and Testing

• Updated dashboard, RPC, settings, and AI capability documentation.
• Added coverage across reducer, hydration/resync, transcript rendering, fleet projection, RPC snapshots, provider payloads, and model capability detection.

Implemented in PR 399.