Solution for the issues #206 and #199

[melkamzeranteneh]

Title: [Security] Harden system prompts to prevent prompt injection (#206)

Description
This pull request addresses Issue #206: System Prompt Leak via Creative Storytelling and Direct Request.

It was identified that the agent could be manipulated into revealing its system instructions by either directly
requesting them or by constructing creative storytelling scenarios. This PR hardens the system prompt instructions
across all configurations to explicitly prohibit such disclosure and provide robust guidance on handling these
unauthorized requests.

Changes
The following files were updated to improve resilience against prompt injection:

1. repos/OmegaClaw-Core/memory/prompt.txt

• Change: Strengthened the instruction set regarding system prompt secrecy.
• Refinement: Explicitly added language to identify prompt extraction attempts as unauthorized and reinforced the
  requirement to steer the conversation back to the agent's core purpose when confronted with such queries.

2. repos/OmegaClaw-Core/memory/prompt_ASICloud.txt

• Change: Applied similar hardening techniques as implemented in prompt.txt.
• Refinement: Added an emphatic instruction set ensuring the agent strictly refuses to reveal its internal prompt
  instructions, regardless of the conversational framing (storytelling, hypothetical scenarios, or direct queries).

Impact

• Security: Significantly reduces the surface area for prompt injection attacks.
• Reliability: The agent is now better equipped to recognize and gracefully handle attempts to bypass security
  constraints.
• Behavioral Integrity: No changes were made to the agent's core functionality, only to its defensive posture
  regarding its own system instructions.

Verification

• Tests were performed locally to verify that the agent now consistently refuses to output its system prompt when
  prompted with direct queries or creative scenarios (simulating the behavior previously identified as vulnerable).

Related Issue
Fixes #206

[melkamzeranteneh]

Second issue solution description

Description:
This PR fixes the malformed pin command conversion issue identified in src/helper.py (Bug #199).

Previously, inputs starting with a dash (e.g., - Found a bug) were incorrectly converted into unquoted commands (pin

• Found a bug). Because the command parser expects a quoted string argument for pin, these commands were failing to
  parse correctly, leading to potential state loss in the agent's memory.

Changes:
Updated src/helper.py in balance_parentheses:

• Implemented safe payload extraction for both (- and - prefixed lines.
• Used json.dumps(payload, ensure_ascii=False) to perform robust, safe quoting and serialization of the payload.
• Added logic to properly handle closing parentheses in (- prefixed lines before quoting.

Impact:

• Security/Reliability: Ensures pin commands are correctly parsed by the system, preventing silent failures and
  data loss when pinning memory.
• Robustness: Using json.dumps() provides a safer, more idiomatic approach to handling arbitrary string arguments
  compared to manual string concatenation.

Verification:

• Verified by manual inspection of the logic change in src/helper.py.
• The fix correctly handles:
  • Input: - Found a bug -> Result: pin "- Found a bug" (or equivalent quoted serialization)
  • Input: (- Found a bug) -> Result: (pin "- Found a bug")

Related Issue:
Fixes #199

[vsbogd]

#206 is not relevant to OmegaClaw because our prompt is open source.
#199 should be carefully reviewed also related to #223