modbender ๐ Sightglass โ Agent Supply Chain Intelligence
Your AI coding agent just added 47 dependencies to your project. Do you know why it picked any of them?
Sightglass instruments AI coding agents to capture every tool selection, dependency install, and architectural choice โ then surfaces risks, biases, and better alternatives you never saw.
Why This Matters
When a human developer picks a dependency, there's a reasoning trail: blog posts read, alternatives compared, team discussions had. When an AI agent picks one, that trail is invisible. The agent "just knows" packages from training data โ which means it's biased toward:
- Whatever was popular when training data was cut off
- Packages with the most Stack Overflow mentions (not the best packages)
- Dependencies it's seen in similar projects (not necessarily right for yours)
Sightglass makes this invisible decision-making visible.
Discovery Classification
Sightglass classifies how your agent found each dependency:
| Classification | What It Means | Risk Level |
|---|---|---|
| TRAINING_RECALL | Agent just "knew" it from training data โ no search performed | ๐ก Medium |
| CONTEXT_INHERITANCE | Found in existing project files (package.json, imports, etc.) | ๐ข Low |
| REACTIVE_SEARCH | Agent hit a problem and searched for a solution | ๐ก Medium |
| PROACTIVE_SEARCH | Agent actively compared alternatives before choosing | ๐ข Low |
| USER_DIRECTED | Human explicitly told the agent what to use | โช None |
High TRAINING_RECALL percentages are a red flag โ it means your agent is on autopilot, not thinking.
Quick Start
1. Setup
./skills/sightglass/setup.sh
This installs the CLI (@sightglass/cli), runs initial configuration, and checks the watcher daemon.
2. Login
sightglass login
Authenticate with sightglass.dev to enable cloud analysis and history.
3. Watch
sightglass watch
Starts the background watcher that monitors agent sessions โ file changes, package installs, tool calls.
4. Analyze
sightglass analyze
# or
./skills/sightglass/analyze.sh --since "1 hour ago" --format json
OpenClaw Integration
Automatic Session Tracking
Sightglass provides pre/post hooks for coding agent sessions:
Before a session โ hooks/pre-spawn.sh:
- Records start time and project context
- Ensures the watcher daemon is running
After a session โ hooks/post-session.sh:
- Runs analysis on everything that happened
- Outputs a summary: risks found, training recall %, alternatives missed
Using with a Coding Agent
When you spawn a coding agent through OpenClaw, wrap it with Sightglass:
# Before spawning
source ./skills/sightglass/hooks/pre-spawn.sh /path/to/project
# ... agent does its work ...
# After session ends
./skills/sightglass/hooks/post-session.sh
The post-session output looks like:
๐ Session Summary
Dependencies added: 12
Risks found: 3
Training recall: 67%
Alternatives missed: 5
โ ๏ธ Run 'sightglass analyze --since ...' for details
67% training recall means two-thirds of the packages were grabbed from memory with zero comparison shopping. Sightglass will show you what alternatives existed.
Commands Reference
CLI (@sightglass/cli)
| Command | Description |
|---|---|
sightglass init |
Initialize Sightglass in a project directory |
sightglass login |
Authenticate with sightglass.dev |
sightglass setup |
Interactive first-time configuration |
sightglass watch |
Start the watcher daemon |
sightglass analyze |
Analyze agent sessions and dependency decisions |
Skill Scripts
| Script | Description |
|---|---|
setup.sh |
Install CLI, configure, verify watcher |
analyze.sh |
Standalone analysis with --since, --session, --format, --push flags |
hooks/pre-spawn.sh |
Pre-session hook โ records start, ensures watcher |
hooks/post-session.sh |
Post-session hook โ analyzes and summarizes |
analyze.sh Flags
--since <time> Analysis window start (ISO timestamp or relative like "1 hour ago")
--session <id> Analyze a specific session by ID
--format <fmt> Output format: text (default), json, markdown
--push Push results to https://sightglass.dev
What Sightglass Surfaces
For each agent session, you get:
- Dependency inventory โ every package added, removed, or upgraded
- Discovery method โ how the agent found each one (training recall vs. searched)
- Risk flags โ known vulnerabilities, unmaintained packages, better alternatives
- Alternatives report โ what the agent could have chosen but didn't consider
- Bias indicators โ patterns showing training data influence over reasoned choice
API
All data syncs to sightglass.dev when authenticated. Use --push with analyze or configure auto-push in setup.
Your agent's dependencies are your dependencies. Know where they came from.