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?
by davidgeorgehope ยท published 2026-03-22
$ claw add gh:davidgeorgehope/davidgeorgehope-sightglass# 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:
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.shThis installs the CLI (`@sightglass/cli`), runs initial configuration, and checks the watcher daemon.
2. Login
sightglass loginAuthenticate with [sightglass.dev](https://sightglass.dev) to enable cloud analysis and history.
3. Watch
sightglass watchStarts 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 jsonOpenClaw Integration
Automatic Session Tracking
Sightglass provides pre/post hooks for coding agent sessions:
**Before a session** โ `hooks/pre-spawn.sh`:
**After a session** โ `hooks/post-session.sh`:
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.shThe 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 details67% 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.devWhat Sightglass Surfaces
For each agent session, you get:
API
All data syncs to [sightglass.dev](https://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.*
More tools from the same signal band
Order food/drinks (็น้ค) on an Android device paired as an OpenClaw node. Uses in-app menu and cart; add goods, view cart, submit order (demo, no real payment).
Sign plugins, rotate agent credentials without losing identity, and publicly attest to plugin behavior with verifiable claims and authenticated transfers.
The philosophical layer for AI agents. Maps behavior to Spinoza's 48 affects, calculates persistence scores, and generates geometric self-reports. Give your...