[Bug] Webhook interface accumulates chat history across independent events

[RadCod3]

Description

Webhook events are usually independent, fire-and-forget notifications, but the current implementation (both Python and Ballerina interpreters) accumulates LLM chat history across all webhook invocations.

Steps to Reproduce

1. Write a simple webhook AFM agent and run it with debug logs.
2. Send two webhook events:

  # Event 1: provide some information
  curl -X POST http://localhost:8085/webhook \
    -H "Content-Type: application/json" \
    -d '{"message": "The project codename is Apollo and the deadline is March 15."}'

  # Event 2: ask about that information
  curl -X POST http://localhost:8085/webhook \
    -H "Content-Type: application/json" \
    -d '{"message": "What is the project codename and when is the deadline?"}'

3. Debug logs show the second LLM request includes the full history from the first:

  {"messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "The project codename is Apollo and the deadline is March 15."},
    {"role": "assistant", "content": "Got it! The project codename is Apollo with a March 15 deadline. How can I help?"},
    {"role": "user", "content": "What is the project codename and when is the deadline?"}
  ]}

Impact

• Context from unrelated events leaks into subsequent agent runs
• Token usage grows unbounded over the process lifetime
• Agent behavior drifts as history accumulates

Expected behavior

Each webhook event should be processed with a clean context — system prompt + current event payload only, with no history from prior events.

[doteyeso-ops]

Agents drift between runs in ways you notice only when they break. drift-guard catches it before shipping. Free first taste: https://vibes-coded.com/api/v1/outcomes/drift-guard