Five required elements

Five required elements

1. Files to read first — list in order, with purpose

Read these files before writing anything:
- bots/shared/guardrails.py      (style reference — follow this pattern)
- bots/tech-support/rtvi_server.py  (the file you are extending)

2. Files to create — exact absolute paths, no ambiguity

Create exactly these files (no others):
- /Users/.../bots/shared/working_memory.py
- /Users/.../bots/tech-support/memory_server.py

3. File ownership — what the agent must NOT touch

Do NOT modify any existing file.
Do NOT touch: bots/shared/vector_memory.py  (owned by a parallel agent)

4. Output contract — exact class names, method signatures, key behaviors

Class ConversationSummaryProcessor(FrameProcessor):
  - intercepts LLMMessagesFrame between user_agg and llm
  - fires when len(messages) > compress_threshold (default: 10)
  - calls gpt-4o-mini, replaces old messages with summary system message
  - logs: [MEMORY] ConversationSummaryProcessor: compressing N messages → M

5. Style anchor — point to an existing file that exemplifies the expected pattern

Follow the exact style of bots/shared/guardrails.py:
- Same logger format: logger.warning(f"[TAG] ClassName ACTION | ...")
- Same docstring structure (class docstring explains FrameProcessor vs BaseObserver)
- Same copy-extend pattern for server files (copy guardrails_server.py, add imports)

Anti-patterns to avoid

Anti-pattern	Problem	Fix
"Implement the memory module"	Too vague — agent guesses	Specify every class and method
Relying on conversation context	Agent has none	Inline all necessary context
"Agent B reads Agent A's file" when parallel	A's file may not exist yet	Describe A's API in B's prompt by spec
Long inline prompt string	Shell quoting breaks on special chars	Always write to temp file with heredoc

---

Back to orchestration → /cmux-agents For isolation decisions → /cmux-agents-isolation