Simplify llm history manager

[ShaohongZ]

Overview

Simplify the LLM history manager for conversational use cases compatible with small local models. The original verbose, sensor-oriented formatting is replaced with concise, conversation-style formatting that reduces token usage and improves response quality.

Type of change

• Bug fix
• [v] New feature
• Documentation improvement
• Bounty issue submission
• Other:

Changes

• Simplified ACTION_MAP: Replaced verbose preambles ("**** felt: {}", "**** said: {}", "**** performed this motion: {}") with plain "{}" format strings. Added greeting_conversation action type.
• Conversational input format: Changed input formatting from sensor-style ("TestBot sensed the following: audio. User said hello | vision. Saw a person |") to conversational style ("User: User said hello Saw a person"). When no inputs match the current tick, uses "User: (no input)".
• Concise action messages: Changed action messages from "Given that information, **** took these actions: **** said: Hello" to "AgentName: Hello".
• JSON response extraction: Action values wrapped in {"response": "text"} are now unwrapped to plain text before storing in history.
• Summary prefix: Summaries are now prefixed with [Conversation summary - do not repeat] to prevent the model from echoing conversation summaries back to users.
• Concise summarization prompts: Replaced verbose summarization system prompt and command with shorter, conversation-focused versions tuned for small models (under 100 words, no markdown/tables).
• Added tests: 5 new tests covering ACTION_MAP contents, simplified format, JSON extraction, agent name format, and history_length=0 behavior.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ All tests successful. No failed tests found.

📢 Thoughts on this report? Let us know!