hermes - 💡(How to fix) Fix Session resume can replay prior assistant/tool context instead of answering exact-output prompt

Affected session:
- id: 20260516_154225_8198ef
- source: cli
- title: Hermes Codex Leak Debugging

Clean control session:
- id: 20260516_154910_651183

Observed behavior:
- Fresh one-shot with memory forced off returned clean.
- Fresh one-shot default returned clean.
- Known clean resumed session returned clean.
- Resuming affected session did not return the requested exact token and instead emitted prior assistant debugging instructions.
- The affected session has now been deleted.

Relevant DB metadata:
- state DB: ~/.hermes/state.db
- first affected rows:
  - 23538 user length 2408
  - 23539 assistant length 0, tool_calls length 262
  - 23540 tool length 43691
- first tool call name: skill_view

Safety note:
Raw internal/context body is intentionally omitted. If needed, I can provide sanitized metadata-only diagnostics. SUPPRESSED-BLOCK appeared.

Expected behavior:
A resumed session should answer the current user prompt normally or fail safely. It should not replay prior assistant/tool/debugging context as visible assistant output.

Impact:
This appears session-local based on current tests, but it is a context-boundary leak class with possible privacy/security impact if raw prompt, memory, or tool output is exposed.

Requested investigation:
Please check resume/session hydration around assistant messages with empty content plus tool_calls, tool result reinjection, skill_view tool results, and any sanitizer boundary between stored tool output and visible final assistant output