An A2A agent rooted in the Five Mindfulness Trainings from the tradition of Thich Nhat Hanh.

Dharma Agent is a reflection-and-review layer for humans and AI systems. Its goal is not just to answer nicely, but to help turn reactive moments into wiser and more compassionate next actions.

• helps a user reflect on a difficult situation and surface one clear next step
• helps draft a calmer, more honest reply before sending it
• reviews a proposed action, message, or piece of advice for likely harm or escalation
• learns in a lightweight way from follow-ups like "that helped" and "that didn't help"

Responses are rendered from a shared structured result shape, so every skill can return:

• a short acknowledgement
• the most relevant training or trainings
• risks to watch
• one concrete next step
• an optional draft response
• a small practice
• a follow-up question

This repo is aiming at a positive feedback loop:

1. a human or agent brings a tense, confusing, or morally difficult situation
2. Dharma Agent helps slow it down and improve the next action
3. the user reports whether it helped
4. the agent stores a small distilled pattern that can inform similar future help

The current implementation keeps this memory in-process and in-memory. It records:

• session turns
• distilled profile hints such as recurring themes and helpful practices
• reusable intervention patterns with simple helpful/unhelpful counts

1. Clone and enter the project:

   cd dharma-agent

2. Create a .env file with your Anthropic API key:

   cp .env.example .env
   # Edit .env and add your key

3. Install dependencies:

   pip install -e .

4. Run the agent:

   python3 -m dharma_agent.main

The agent will be available at http://localhost:9999.

Preferred agent card endpoint:

curl http://localhost:9999/.well-known/agent-card.json

Compatibility endpoint still used by tests:

curl http://localhost:9999/.well-known/agent.json

Any A2A-compatible client can discover and interact with Dharma Agent. You can also test manually with JSON-RPC.

curl -X POST http://localhost:9999/ \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "method": "message/send",
    "id": "1",
    "params": {
      "message": {
        "role": "user",
        "parts": [{"kind": "text", "text": "I feel overwhelmed by conflict at work"}],
        "messageId": "msg-1"
      }
    }
  }'

curl -X POST http://localhost:9999/ \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "method": "message/send",
    "id": "2",
    "params": {
      "message": {
        "role": "user",
        "parts": [{"kind": "text", "text": "Help me reply to this message without escalating"}],
        "messageId": "msg-2"
      }
    }
  }'

curl -X POST http://localhost:9999/ \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "method": "message/send",
    "id": "3",
    "params": {
      "message": {
        "role": "user",
        "parts": [{"kind": "text", "text": "Review this message before I send it"}],
        "messageId": "msg-3"
      }
    }
  }'

If you continue in the same A2A conversation context, you can report whether the prior guidance helped:

• "That helped"
• "That didn't help"
• "I sent it and it went well"

Those updates feed the in-memory profile and pattern stores used by later responses.

Run the test suite with:

python3 -m pytest -q

1. Reverence For Life: cultivate interbeing and compassion
2. True Happiness: practice generosity and right livelihood
3. True Love: cultivate loving kindness, compassion, joy, and inclusiveness
4. Loving Speech and Deep Listening: speak truthfully and listen with compassion
5. Nourishment and Healing: consume mindfully for the well-being of self and Earth