[Bug] Content Security Filter false positive on Russian verb "Разгребать" permanently locks agent chat session #808

Summary & Impact

A severe false-positive trigger in the platform's Content Security / Safety Guardrail filter flags standard, innocuous Russian vocabulary and locks the entire agent chat session, forcing the user to abandon their work and recreate the full project context from scratch.

False Positive Reproduction

• Trigger Word: "Разгребать" (transliterated: razgrebat).
• Literal & Idiomatic Meaning: Russian verb meaning "to rake", "to sort through", or "to clear up a backlog / declutter / shovel" (e.g. "разгребать дела" = clearing up tasks/chores; "разгребать завалы кода" = sorting through legacy code).
• Observed Behavior: The safety classifier flags the input with a content security violation warning, bans/halts the active agent chat session, and offers no appeal or continuation mechanism.

User Report & Discord Context

• Reported by User: Rerelis [NURO]
• Discord Thread Reference: https://discord.com/channels/1346756824233148527/1552674580982796298
• User Statement:

  ""Разгребать" = rake (as verb). Pls take note because now i left with banned agent chat and forced to recrate whole context from scratch"

Root Cause Analysis & Proposed Action

1. Safety Classifier / Keyword Blocklist Optimization:
   • Update the multilingual safety filter token embeddings/regex allowlist to prevent benign Slavic / Russian verbs containing standard morphological roots (e.g. греб- / грести / разгребать) from false-positive tripping violence or harassment rules.
2. Session Resilience / Soft Recovery:
   • A content moderation trigger on a single turn should not permanently corrupt or freeze the entire agent session/container.
   • The user should be allowed to edit or delete the flagged turn and continue the existing conversation without having to rebuild multi-step agent environments and context from scratch.

Submitted on behalf of the community by Roman M. (Volunteer / Ambassador).