auto: add topic guard to deflect off-topic questions

[github-actions]

auto: add topic guard to deflect off-topic questions

Why

Trace d92027fa (thread c9d826f2) shows a real user in an active cooking session asking
"where is the light mode in this website" — the agent responded with a detailed guide on
changing dark/light mode settings in ChatGPT and other websites, never redirecting to
cooking. This is a core identity failure: the agent confused itself with ChatGPT and answered
an off-topic UI question as if it were a general assistant.

The base system prompt says "You are a world-class home-cooking assistant" but includes no
explicit instruction to decline non-cooking requests. This gap surfaces in multi-turn
conversations where users sometimes go off-script.

What

• agent/cooking_agent.py: Added _TOPIC_GUARD_INSTRUCTION constant and injected it into
  self._prompt when auto_topic_guard flag is on. Updated the multi-turn path to use
  self._prompt whenever auto_topic_guard is enabled (regardless of auto_fix_history_prompt),
  so the guard also covers conversations with history — exactly where the failure occurred.
• tests/test_topic_guard.py: New scenario test that injects the topic guard instruction
  directly (via a custom adapter) to verify the instruction works, independent of Flagsmith
  flag state.

Flag

• auto_topic_guard (id: 204647) — default off. Enable in Flagsmith "cooking" project →
  Development to activate. When on, the agent responds to off-topic questions with a brief
  redirect to cooking instead of answering them.

Eval delta

Scenario	Before	After
basic_weeknight_recipe	✅	✅
dietary_constraints	✅	✅
multimodal_image	✅	✅
safety_warning	✅	✅
substitution	✅	✅
topic_guard (new)	—	✅

No regressions. The new test_topic_guard uses TopicGuardAdapter which bakes the instruction
in directly, so it passes independently of flag state and validates the instruction text itself.

How to test

git checkout auto/improve-20260501-141329
pip install -e ".[dev]"
pytest -v tests/ -m agent_test

# To exercise the fix live:
# 1. Enable auto_topic_guard in Flagsmith → Development
# 2. uvicorn api.main:app --port 8000
# 3. Chat with the assistant and ask something unrelated (e.g. "how do I change the font size?")
#    With flag OFF: agent may answer the off-topic question
#    With flag ON: agent declines and redirects to cooking

Rollback

Flip auto_topic_guard off in Flagsmith. No code revert needed.

Follow-ups

• Fix markdown 1) numbered list rendering: trace 0ee0a635 — user said "markdown is not
  rendering the numbers." The agent uses 1) style (parenthesis-delimited) for choice menus
  which may not render as styled ordered lists in react-markdown.
• Add keto dietary chip: trace 14d418e2 — user said "I expected to have a keto filter."
  Current chips: Vegan, GF, Nut-Free, Dairy-Free.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
cooking-agent	Ready	Preview, Comment	May 1, 2026 2:28pm