DashClaw GEMINI.md

DashClaw Development Context

This file provides operational context for AI coding agents working inside the DashClaw repository.

Agents must treat this repository as a production open source infrastructure project. All modifications should prioritize stability, developer clarity, and maintainability.

DashClaw is infrastructure software. Avoid introducing complexity unless it clearly improves reliability or developer experience.

---

Project Overview

DashClaw is a governance runtime for AI agent decisions.

It governs AI agents before they execute real world actions by introducing a policy evaluation and approval layer.

Core decision flow:

Agent intent -> policy evaluation -> approval or block -> execution -> decision evidence recorded

DashClaw acts as the decision governance layer between AI agents and external systems.

The platform allows developers and organizations to:

• intercept risky agent actions
• enforce policy checks
• require human approval
• record verifiable decision evidence
• monitor agent behavior
• detect decision drift

DashClaw enables permissioned autonomy for AI agents.

---

Core Product Primitives

These primitives define the DashClaw architecture.

Guard

Evaluates policies before an agent executes an action.

Example usage:

const decision = await claw.guard({ actionType: "deploy", riskScore: 85 })

Guard responses determine whether actions are:

• allowed
• blocked
• escalated for approval

---

Action Records

Capture what the agent attempted to do.

Includes:

• action type
• parameters
• reasoning
• execution outcome

---

Assumptions

Tracks what the agent believed to be true when making a decision.

Used to detect decision drift and incorrect reasoning.

---

Approvals

Allows high risk actions to pause until a human operator approves or rejects them.

---

Evidence

Every governed decision produces verifiable evidence.

Evidence enables:

• debugging agent behavior
• compliance reporting
• post incident analysis