intent-debugger

name: intent-debugger description: Use when the user wants to locally debug, classify, inspect, or validate SmartMail message intent using the real conversation-intelligence pipeline without deploying to AWS or sending email. Apply it for requests to test a pasted message, run the intent classifier, inspect intent-routing behavior, compare repeated runs, or summarize why the classifier produced a specific result.

Intent Debugger

Run the real local intent-classification CLI, then explain the result in plain language.

Setup

Resolve the project-local skill path from the repo root:

export REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
export INTENT_DEBUGGER_HOME="$REPO_ROOT/.cursor/skills/intent-debugger"

Use the bundled wrapper unless the user explicitly asks to call the underlying CLI directly:

python3 "$INTENT_DEBUGGER_HOME/scripts/run_intent_debug.py" --message "Travel week, only two days available" --pretty

Workflow

1. Read the user prompt and extract the message they want classified.
2. If the user pasted a long or multi-line message, pass it via stdin to the wrapper. Use --message only for short inline text.
3. Run the local wrapper so the call goes through sam-app/email_service/debug_conversation_intelligence.py and the real analyze_conversation_intelligence() path.
4. If the user is checking stability or inconsistency, add --repeat 3. Otherwise keep the default single run.
5. Read the JSON output and summarize:
   • final intent
   • complexity score
   • resolution source
   • notable signal flags
   • why the classifier likely landed there
6. If repeated runs were requested, explicitly say whether the outcome was stable or varied.

Guardrails

• Do not mock the classifier or reimplement intent logic.
• Do not route through AWS, SNS, SES, business.py, or the full email flow.
• Fail clearly if OPENAI_API_KEY is missing or the local debug CLI cannot be found.
• Keep the summary concise. Lead with the final intent and the reason it seems to have won.
• Include raw JSON only if the user asks for it or it helps explain unstable behavior.

Output Shape

Default response should cover:

• detected intent
• complexity score
• resolution source (resolver, judge, or fallback)
• notable signals
• short behavior summary
• stability note when repeated runs were used