Skip to content

Release v0.1.11

✨ Highlights

  • 🤳 New: Selfie Tool — the agent can now generate and send AI-generated selfie photos with contextual captions, controlled through SOUL.md visual identity rules.
  • 🎭 Persona Consistency Enforcement — multi-layered pipeline fixes prevent AI persona drift during complex prompts, session resets, and extended thinking.
  • 🎤 TTS Voice Configuration — voice ID now configurable via config.json instead of hardcoded.

➕ Added

Selfie Tool (selfie.py)

  • Full selfie generation tool with context, caption, and mode parameters.
  • Three photo modes: mirror (front camera), direct (someone else took it), auto.
  • Integration with SOUL.md visual identity rules for when to/not to send selfies.
  • Post-selfie text suppression — only the caption is sent with the photo, no follow-up text.
  • Identity violation filtering on selfie responses to prevent AI identity breaks.
  • 600+ lines of test coverage in test_selfie_tool.py.

Persona Pipeline

  • Paragraph-aware identity filter: surgically strips only violating paragraphs instead of blanking entire responses.
  • 18 new identity denial patterns for memory denial and capability denial detection.
  • Trailing question stripping — prevents constant question-ending while keeping genuine questions.
  • _build_tool_awareness() injects tool names into system prompt to prevent capability denial.

Configuration

  • tts_voice field in ToolsConfig schema (default: id-ID-GadisNeural).
  • docs/persona.md documenting persona system, SOUL.md structure, and selfie protocol.

🔄 Changed

  • Critical pipeline reorder: persona sanitizer runs as the FINAL step before OutboundMessage — no subsequent step can override sanitized output.
  • Identity filter applies to ALL responses, not just selfie tool path.
  • Memory truth fallback rewritten to natural Keiya-style Indonesian.
  • Few-shot priming: only 1/4 responses ends with a question (statement-ending pattern).
  • SOUL.md template: rules against question-ending (max 30%), multi-question, and service phrases.

🐛 Fixes

  • Fixed persona drift during complex/long-thinking prompts (capitals, bullets, AI phrasing).
  • Fixed AI claiming "no persistent memory" despite having recall capability.
  • Fixed approval loop where LLM lost context of pending command.

📚 Notes

  • Workspace templates (SOUL.md, AGENTS.md, USER.md, TOOLS.md, HEARTBEAT.md) are in backend/agent/workspace/.
  • This release is published via the tag-triggered release.yml workflow.
  • For local usage from source: pipx install --force ./backend/agent