sightglass
by davidgeorgehope
Coding Agents & IDEsBrief
Your AI coding agent just added 47 dependencies to your project.
Description
What This Skill Does
Sightglass monitors AI coding agent sessions to track how dependencies get selected, classifying each one by how the agent discovered it — training recall, context inheritance, search, or user direction. It surfaces dependency risks, missed alternatives, and bias patterns after each session. Data syncs to sightglass.dev when authenticated.
Unlike manual post-hoc dependency review, Sightglass captures the discovery method at session time, making it possible to distinguish reasoned choices from autopilot training recall.
When to Use It
- Auditing which packages an AI agent added without explicit instruction
- Checking training recall percentage after a large agent-driven refactor
- Comparing what alternatives the agent ignored when picking a library
- Flagging unmaintained or vulnerable packages introduced by an agent session
- Reviewing dependency history across multiple agent sessions on a project
Example Workflow
User asks: analyze what my agent just did in the last hour
- 1
Run pre-spawn.sh before starting the coding agent session to record start time and ensure watcher is running
- 2
Agent completes its work, installing dependencies and making architectural choices
- 3
Run post-session.sh to trigger analysis of the completed session
- 4
Review the summary output showing dependencies added, risk count, training recall percentage, and missed alternatives
- 5
Run sightglass analyze --since '1 hour ago' --format json for full detail
Session report listing 12 dependencies added, 3 risks, 67% training recall, and 5 alternatives the agent never considered
Requirements
- sightglass.dev account (for cloud analysis, history, and --push sync)
Installation
clawhub install davidgeorgehope/sightglassRun the CLI command, or paste the prompt into Claude Code, Cursor, Codex or whichever coding agent you use.