Skip to main content
Current release: 1.8.2 · What’s New → Agent Journal is a Java library for recording agent runs as typed events and analyzing them after execution. It stores immutable execution facts separately from derived analysis, and it can project Claude Code, Gemini CLI, Grok, Codex, and Antigravity results into a shared journal model. Release 1.8.0 adds tool-trajectory capture for Grok, Codex, and Antigravity, each parsed from that CLI’s own durable JSONL stream into ordered tool and cost events. The three new modules carry no vendor SDK dependency and run on Java 17. The change is additive on the existing event and trace contracts.

Execution journal

LLM calls, tool calls, state changes, metrics, git operations, and custom events in append-only JSONL.

Derived analysis

Cost attribution and step outcomes in a separate analysis.jsonl stream.

Evaluation seam

EvalSubject adapters expose recorded behavior to evaluators such as Agent Judge.

​
Modules

journal-core uses Jackson 2 and the SLF4J API. The capture modules also depend on their respective Java SDKs, whose published bytecode requires Java 21.

​
Install

The six released artifacts use the io.github.markpollack group ID and are available from Maven Central.

​
Documentation

Getting started

Record and persist a first run with journal-core.

Core concepts

Understand experiments, runs, event streams, analysis streams, feedback, and storage.

Capture SDK sessions

Project Claude Code and Gemini CLI results into Agent Journal.

Analyze runs

Query generated journals with DuckDB.

API reference

Stable 1.8.2 entry points, events, storage operations, and capture options.
Journal files and trace files can contain prompts, tool inputs and results, file contents, and command output. Treat them as potentially sensitive operational data. JsonFileStorage is local file storage, not a multi-tenant security boundary.
  • Agent Workflow provides the separate workflow-journal integration module.
  • Agent Judge consumes evaluation subjects derived from recorded behavior.
  • Markov fingerprinting is one downstream analysis technique for tool-call sequences.

Source code

Agent Journal on GitHub (BSL 1.1)