Skip to main content

teammate_skill.md

OpenClaw Hackathon — Skill / Agent Integration Layer

Your Scope: Build the OpenClaw Skill (Agent-Facing Layer)​

I’m building the filtering + sanitization server. You’ll build the OpenClaw skill that consumes sanitized events and turns them into useful agent behavior.

We are strictly separating concerns:

Filtering Server (me) → Sanitized JSON → OpenClaw Skill (you)

The agent must NEVER directly consume raw webhooks.

1. Goal

Create an OpenClaw skill that:

  • Receives sanitized email events from the filtering server
  • Stores or queries recent events
  • Allows the agent to answer questions like:
    • “What are my newest emails?”
    • “Summarize today’s emails”
    • “Any risky emails?”
  • Displays risk flags clearly
  • Never processes raw unsanitized content

2. Data Contract (What You Receive)

The filtering server will POST this JSON to your skill endpoint:

SanitizedEmailEvent:

{ "event_id": "...", "provider": "generic", "received_at": "...", "from_addr": "...", "to_addrs": [...], "subject_sanitized": "...", "body_sanitized": "...", "attachments_sanitized": [...], "risk": { "flags": ["html_detected", "injection_detected"], "injection_detected": true, "truncated": false }, "meta": { "original_sizes": {...}, "sanitizer_version": "v1" } }

Important:

  • You must treat this as already sanitized.
  • Do not attempt to "re-sanitize".
  • Use risk flags in decision logic.

3. Skill Responsibilities

3.1 Ingestion Endpoint​

Expose endpoint:

POST /events/email

Store events in:

  • memory (hackathon)
  • or lightweight DB (sqlite/postgres)

Must support:

  • idempotency by event_id
  • retrieval by time
  • filtering by risk flags

3.2 Query Functions​

Expose callable functions/tools to agent:

  1. list_recent_emails(limit=10)
  2. list_unread_or_recent(limit=10)
  3. summarize_recent_emails()
  4. list_risky_emails()
  5. search_emails(query_string)

The skill should:

  • Only operate on sanitized_body
  • Never store raw HTML
  • Respect truncation metadata

3.3 Agent Behavior​

The skill must:

  • Include risk flags in responses when relevant
  • Warn if injection_detected is true
  • Highlight truncated content
  • Never claim content is “safe” — only “sanitized + flagged”

4. Storage Schema (Suggested)​

Table: sanitized_emails

  • event_id (primary key)
  • received_at
  • from_addr
  • subject
  • body
  • risk_flags (json)
  • injection_detected (bool)
  • truncated (bool)

Indexes:

  • received_at
  • injection_detected

5. Demo Flow​

  1. Filtering server receives email
  2. Filtering server sanitizes
  3. Filtering server POSTs to your skill
  4. Agent can now answer:
    • “What are my latest emails?”
    • “Any suspicious ones?”
    • “Summarize today’s inbox”

6. Stretch Goals​

  • Add simple vector search over sanitized_body
  • Add risk-based ranking
  • Add “safe preview” vs “full sanitized text”
  • Add filtering by sender domain

7. Deliverables​

  • Skill service running
  • Endpoint for sanitized event ingestion
  • Queryable interface for agent
  • Clean logging
  • Basic persistence
  • Demo-ready

You’re building the intelligence layer. I’m building the safety layer.

Together: Safety → Skill → Agent