Water Rescue Drone Design

Water Rescue Drone Design — Skill Registry

Overview

The water-rescue-drone-design skill is a production-grade AI agent harness that transforms LLMs into domain experts for Search-and-Rescue Drone Engineering for Water. It orchestrates a multi-step pipeline that combines real-time data aggregation, recognized engineering methods, and academic research into evidence-backed, risk-disclosed outputs.

Skill Registration

Registration Pattern

Skills are registered in the harness via the SkillRegistry class in water_rescue_drone/skills/__init__.py:

from water_rescue_drone.skills import SkillRegistry
from water_rescue_drone.skills.gather_requirements import GatherRequirementsSkill

registry = SkillRegistry()
registry.register("gather_requirements", GatherRequirementsSkill)

Skill Resolution

The agent resolves skills by:

1. Loading the skill's markdown template from skills/.md
2. Instantiating the corresponding Python class from water_rescue_drone/skills/
3. Passing the AgentConfig and LLMProvider to the skill constructor
4. Executing the skill via the BaseSkill.run(ctx) method

Skill Execution Flow

1. Load Template → 2. Build Prompt → 3. LLM Call → 4. Parse JSON → 5. Apply to Context

Each skill:

• Loads its persona and workflow from the markdown template
• Builds a structured prompt from the current AgentContext
• Calls the LLM provider with JSON mode enabled
• Parses and validates the structured response against a schema
• Merges the result back into the context
• Records a StepRecord for auditability

Registered Skills

1. gather_requirements

Purpose: Clarify object/scope/timeframe/inputs/language before fetching.

Input Schema:

{
  "type": "object",
  "properties": {
    "user_input": {"type": "string"},
    "context": {"$ref": "#/definitions/AgentContext"}
  },
  "required": ["user_input"]
}

Output Schema:

{
  "type": "object",
  "properties": {
    "object": {"type": "string"},
    "scope": {"type": "string"},
    "timeframe": {"type": "string"},
    "available_inputs": {"type": "array", "items": {"type": "string"}},
    "target_audience": {"type": "string"},
    "language": {"type": "string", "enum": ["en", "vi"]},
    "analysis_type": {"type": "string"}
  },
  "required": ["object", "language"]
}

2. evidence_collector

Purpose: Fetch authoritative real-time + reference data.

Input Schema:

{
  "type": "object",
  "properties": {
    "requirements": {"$ref": "#/definitions/Requirements"},
    "context": {"$ref": "#/definitions/AgentContext"}
  }
}

Output Schema:

{
  "type": "object",
  "properties": {
    "evidence_items": {
      "type": "array",
      "items": {
        "type": "object",
        "properties": {
          "source": {"type": "string"},
          "date": {"type": "string"},
          "tier": {"type": "string", "enum": ["1", "2", "3", "4"]},
          "content": {"type": "string"},
          "url": {"type": "string"}
        }
      }
    }
  }
}

3. core_analysis

Purpose: Design payload/sensors/endurance/path/reliability + scenarios.

Input Schema:

{
  "type": "object",
  "properties": {
    "requirements": {"$ref": "#/definitions/Requirements"},
    "evidence": {"$ref": "#/definitions/Evidence"}
  }
}

Output Schema:

{
  "type": "object",
  "properties": {
    "drone_type": {"type": "string"},
    "payload": {"type": "object"},
    "drop_mechanism": {"type": "string"},
    "sensors": {"type": "array"},
    "endurance": {"type": "object"},
    "wind_resistance": {"type": "number"},
    "water_ingress": {"type": "string"},
    "search_path": {"type": "string"},
    "reliability": {"type": "number"},
    "certification": {"type": "array"},
    "scenarios": {"type": "array"},
    "metrics": {"type": "object"}
  }
}

4. knowledge_updater

Purpose: Surface academic citations with tiers; flag gaps.

Input Schema:

{
  "type": "object",
  "properties": {
    "query": {"type": "string"},
    "context": {"$ref": "#/definitions/AgentContext"}
  }
}

Output Schema:

{
  "type": "object",
  "properties": {
    "citations": {
      "type": "array",
      "items": {
        "type": "object",
        "properties": {
          "authors": {"type": "array", "items": {"type": "string"}},
          "title": {"type": "string"},
          "year": {"type": "number"},
          "doi": {"type": "string"},
          "tier": {"type": "string"},
          "relevance": {"type": "number"}
        }
      }
    },
    "gaps": {"type": "array", "items": {"type": "string"}},
    "coverage_rating": {"type": "number"}
  }
}

5. advisor

Purpose: Synthesize risk-disclosed conclusion + evidence chain.

Input Schema:

{
  "type": "object",
  "properties": {
    "analysis": {"$ref": "#/definitions/CoreAnalysis"},
    "knowledge": {"$ref": "#/definitions/KnowledgeCitations"},
    "context": {"$ref": "#/definitions/AgentContext"}
  }
}

Output Schema:

{
  "type": "object",
  "properties": {
    "verdict": {
      "type": "string",
      "enum": ["RECOMMENDED", "CONDITIONALLY_RECOMMENDED", "NOT_RECOMMENDED", "INSUFFICIENT_DATA"]
    },
    "scenarios": {"type": "array"},
    "risks": {"type": "array"},
    "evidence_chain": {"type": "array"},
    "remediation": {"type": "string"},
    "disclosure": {"type": "string"}
  }
}

Quality Gates

The skill enforces quality gates at two levels:

Universal Gates (U1-U6)

• U1: ≥3 sources cited, ≥1 academic/authoritative
• U2: Safety/risk/limitation disclosure present BEFORE recommendation
• U3: Evidence hierarchy stated per source (Tier 1–4)
• U4: Language matches user preference
• U5: Output uses declared output template
• U6: Every claim traceable to ≥1 cited source OR flagged as judgment

Domain Gates (G1-G4)

• G1: Payload capacity validated against drop mechanism
• G2: Endurance calculation accounts for weather/wind
• G3: Sensor coverage adequate for target area
• G4: Reliability estimate includes redundancy analysis

Gate Enforcement

from water_rescue_drone.gates import QualityGateRunner

runner = QualityGateRunner(config)
results = runner.evaluate_all(context)
failed = runner.summary(results)["failed"]

if config.strict_gates and failed > 0:
    raise GateFailure(f"{failed} gates failed")

Tool Definitions

The skill uses the following tools with their respective schemas:

WebSearch Tool

Purpose: Search for real-time domain data.

Input Schema:

{
  "type": "object",
  "properties": {
    "query": {"type": "string"},
    "max_results": {"type": "number", "default": 10}
  },
  "required": ["query"]
}

Output Schema:

{
  "type": "array",
  "items": {
    "type": "object",
    "properties": {
      "title": {"type": "string"},
      "url": {"type": "string"},
      "snippet": {"type": "string"},
      "date": {"type": "string"}
    }
  }
}

WebFetch Tool

Purpose: Fetch and parse web content.

Input Schema:

{
  "type": "object",
  "properties": {
    "url": {"type": "string"},
    "format": {"type": "string", "enum": ["markdown", "text"], "default": "markdown"}
  },
  "required": ["url"]
}

Read Tool

Purpose: Read local files (knowledge base).

Input Schema:

{
  "type": "object",
  "properties": {
    "file_path": {"type": "string"},
    "offset": {"type": "number"},
    "limit": {"type": "number"}
  },
  "required": ["file_path"]
}

Hooks System

Pre-Skill Hook

@hook("pre_skill")
def before_skill(skill_name: str, context: AgentContext):
    # Log skill start, validate prerequisites
    pass

Post-Skill Hook

@hook("post_skill")
def after_skill(skill_name: str, result: SkillResult, context: AgentContext):
    # Record metrics, update analytics
    pass

Error Hook

@hook("on_error")
def handle_error(skill_name: str, error: Exception, context: AgentContext):
    # Implement custom error handling, fallbacks
    pass

Graceful Degradation

The skill implements 5 degradation levels (0–4):

| Level | Description | Banner | |-------|-------------|--------| | 0 | Full operation | None | | 1 | Minor fallbacks | "LIMITATION: Some secondary features unavailable" | | 2 | Primary failures | "LIMITATION: Operating with reduced capability" | | 3 | Knowledge-only mode | "LIMITATION: Knowledge base only, live data unavailable" | | 4 | Critical degradation | "LIMITATION: Critical failures - review required" |

Language Support

The skill automatically detects and supports:

• English (en) — Default
• Vietnamese (vi) — Full localization

Language detection uses keyword matching and can be overridden via WRD_LANGUAGE environment variable.

CLI Usage

# Run with default deterministic provider (offline)
python -m water_rescue_drone "design a coastal SAR drone for 1 km range"

# Run with OpenAI provider
WRD_LLM_PROVIDER=openai WRD_LLM_MODEL=gpt-4 python -m water_rescue_drone "analyze UAV for lake rescue"

# Run with Anthropic provider
WRD_LLM_PROVIDER=anthropic WRD_LLM_MODEL=claude-3-opus python -m water_rescue_drone "design river rescue drone"

# Strict mode (raise on gate failures)
WRD_STRICT_GATES=true python -m water_rescue_drone "evaluate search drone"

Development

Project Structure

water-rescue-drone-design/
├── water_rescue_drone/        # Runnable Python package
│   ├── agent.py               # Main orchestrator
│   ├── config.py              # Configuration management
│   ├── context.py             # Typed context
│   ├── errors.py              # Error types
│   ├── gates.py               # Quality gates
│   ├── knowledge.py            # Knowledge integration
│   ├── cli.py                 # CLI entry point
│   ├── llm/                   # Provider abstraction
│   └── skills/                # Sub-skills
├── skills/                    # Markdown templates
│   ├── main.md                # Harness specification
│   └── sub-*.md               # Sub-skill specifications
├── tools/                     # Utilities
│   ├── knowledge_updater.py   # Knowledge crawl
│   ├── run_test_scenarios.py  # E2E validator
│   └── validate_project.py    # Contract validator
├── config/                    # Configuration files
├── references/                # Domain knowledge
├── assets/                    # Static resources
├── scripts/                   # Automation scripts
├── hooks/                     # Lifecycle hooks
├── SECOND-KNOWLEDGE-BRAIN.md  # Knowledge base
└── SKILL.md                   # This file

Testing

# Run unit tests
pytest -q

# Run E2E scenarios
python tools/run_test_scenarios.py

# Validate project contract
python tools/validate_project.py

License

MIT License — See LICENSE file for details.

Version History

• 1.2.0 (2026-07-27) — Enhanced architecture: hooks system, modular directories, structured logging, tool schemas, configuration management, domain references
• 1.1.0 (2026-07-13) — Initial production release with core harness and quality gates

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dungnotnull
• Source: dungnotnull/water-rescue-drone-design-agent-skill
• License: MIT

Install and usage instructions live in the source repository linked above.