AgentProbe — AI skill for Claude Code

AgentProbe

Session-history collectors for AI coding agents.

Development is spread across a growing set of agents — Claude Code, Antigravity, Gemini, Copilot, Cursor — and each stores sessions differently. Claude Code writes JSONL transcripts and fires lifecycle hooks. Cursor keeps conversations in SQLite inside workspace storage. Copilot and Gemini bury logs in extension directories. It is all local, none of it is searchable, and most of it is rotated away or lost when a machine is rebuilt.

AgentProbe attaches one probe per agent. Each probe understands a single tool's storage format, extracts completed sessions, and normalizes them into a shared schema: the full transcript, every command run, every file read or edited, plus timing, project and user metadata. The result is pushed to Firestore, where it becomes permanent and queryable.

The pipeline is **deterministic by design** — no language model anywhere in the collection path. Capture costs nothing per session, adds no latency, and never invents detail that was not in the original record.

AgentProbe is the collection half of a pair. **AgentContext** is the dashboard that reads what it writes, rendering your work day by day with search, filters, and selective public sharing.

---

Probes

Provider	Source	Method	Status
Claude Code	~/.claude/projects/*/*.jsonl	SessionEnd hook	✅ working
Antigravity	workspace storage	poll + diff	planned
Gemini CLI	~/.gemini/tmp/	poll + diff	planned
Copilot	VS Code extension storage	poll + diff	planned
Cursor	workspaceStorage/*.vscdb	SQLite read	planned

Only Claude Code exposes real lifecycle hooks. The rest are poll-and-diff, which is why the collector needs a watcher daemon rather than just a hook script.

Pat