AI developer tooling / Jul 2026 - Present
Codex Session Visualizer
A VS Code extension that makes AI coding-agent behavior inspectable by reconstructing Codex sessions into file activity, execution traces, and multi-agent workflow graphs.
VS Code Extension API / TypeScript / AI Agent Observability / Multi-agent Systems
Highlights
- Added file and folder activity states for Created, Read, Update, Deleted, and Found, making the agent's interaction with a repository visible directly inside VS Code.
- Built readable session traces covering prompts, responses, tool calls, file operations, and execution flow instead of requiring developers to inspect raw session logs.
- Reconstructed parent and sub-agent relationships to visualize how work is delegated across multi-agent Codex sessions.
- Combined repository activity, chronological execution, and agent hierarchy into a single debugging and inspection workflow.
Problem
AI coding agents can make dozens of decisions during a single task: exploring directories, reading files, invoking tools, modifying code, running commands, and delegating work to other agents.
The final code diff shows what changed, but it does not explain how the agent arrived there.
Raw session logs contain much of that information, but they are dense and difficult to follow. When multiple agents are involved, understanding which agent performed which work becomes even harder.
Codex Session Visualizer turns that hidden execution history into something a developer can inspect directly inside the coding environment.
What I built
I built a VS Code extension that transforms raw Codex session activity into three complementary views of agent behavior.
Repository activity
Files and folders are labeled according to how the agent interacted with them:
Found / Read / Edited / Created / Deleted
This provides a quick visual map of where the agent explored and where it actually made changes without requiring the developer to reconstruct that information from logs.
Session execution
The extension parses raw Codex sessions into a readable chronological trace containing the agent's prompts, responses, tool usage, file operations, and execution flow.
Instead of treating the session as an opaque request followed by a code diff, the developer can inspect the sequence of actions that produced the result.
Multi-agent workflows
For sessions where Codex delegates work, the extension connects the main session with its sub-agent sessions in a graph.
This makes it possible to see where work was delegated, which agents participated, and how the overall session was structured.
Technical decisions
The extension deliberately presents the same coding session from three different levels of abstraction.
File activity answers:
Where did the agent work?
Session traces answer:
What did the agent do, and in what order?
Agent graphs answer:
Which agent did the work, and how was it delegated?
Rather than building a single massive log viewer, I separated those concerns so developers can move from a quick repository-level signal to increasingly detailed execution information when they need it.
The underlying Codex session data remains the source of evidence; the extension's job is to transform that evidence into representations that are easier to inspect while working inside VS Code.