Sightglass

What This Skill Does

Sightglass is an advanced observability tool designed specifically for AI coding agents. It provides transparency into the 'black box' of AI decision-making during software development. When your agent installs a dependency or chooses an architectural pattern, it often relies on outdated training data or biased popularity metrics. Sightglass intercepts these actions, logs the 'reasoning' (or lack thereof), and classifies the source of the decision. By monitoring tool calls and package installations, it generates a transparency report that helps developers understand if their AI is acting on intent or simply hallucinating through training recall.

Installation

To integrate Sightglass into your OpenClaw workflow, use the following commands:

1. Install via the OpenClaw Hub: clawhub install openclaw/skills/skills/davidgeorgehope/sightglass.
2. Execute the setup script to configure the CLI and local watcher: ./skills/sightglass/setup.sh.
3. Authenticate your CLI instance: sightglass login to sync analysis with your cloud account.
4. Initialize the background watcher: sightglass watch.

Use Cases

• Dependency Auditing: Verify if your AI agent is choosing libraries based on actual requirements or just popular training data markers.
• Risk Mitigation: Identify high-risk TRAINING_RECALL patterns that indicate the agent is guessing rather than searching for the latest documentation.
• Architectural Validation: Audit why the agent chose a specific tech stack over alternatives by reviewing the PROACTIVE_SEARCH logs.
• Team Onboarding: Use analysis summaries to explain to team members why certain dependencies were introduced during automated sessions.

Example Prompts

1. "Analyze my last coding session with Sightglass to see if the agent performed any proactive searching before adding the new database driver."
2. "Sightglass, summarize the risk level of the dependencies added in the current session and highlight any that relied solely on training recall."
3. "Show me the last hour of agent activity and identify which dependencies were suggested via context inheritance versus proactive search."

Tips & Limitations

• Always use hooks: Ensure you trigger pre-spawn.sh before starting your agent session to guarantee that all telemetry is captured effectively.
• Review High Recall: If your agent exhibits a high TRAINING_RECALL percentage, it is likely on autopilot. Consider forcing the agent to use a web-search tool explicitly for package selection.
• Security: While Sightglass captures data, it does not prevent installations. You still retain final authority over what enters your package.json.