Developer Experience

AI Coding Trace

A local trace viewer that converts Codex and Claude Code JSONL sessions into redacted context and token visualizations.

traceredactioncontext windowdeveloper tools

Problem

Long AI coding sessions are hard to debug because the useful state is hidden in logs: context window pressure, user turns, cache behavior, and token growth.

Workflow

01Parse Codex and Claude Code session JSONL logs into a common timeline.
02Redact user message text and keep only safe structural signals.
03Visualize context window, cumulative tokens, user-message markers, and per-message deltas.
04Use the trace to discuss where an agentic coding workflow becomes hard to operate or review.

Evidence

Codex trace

Redacted demo view for a Codex-style session.

Claude trace

Redacted demo view for a Claude Code-style session.

Interactive HTML trace

Static demo HTML generated from anonymized fixture data.

Open artifact

Boundary

This is a local trace/replay prototype, not a full observability platform.
The public demo uses anonymized fixture data, not private working-session logs.
It proves product and mechanism understanding, not ownership of a production Agent Harness platform.

Role Mapping

AI coding / developer tools: makes invisible session state inspectable.
Agent Harness adjacent roles: provides trace, context pressure, and failure-analysis language.
LLM Eval: supports behavior review across long task chains.