Claude Code Trace Visualizer

Claude Code Trace Visualizer

**See exactly what Claude Code did on a task** — every tool call, in order, with timing, tokens, dollars, network requests, and where it got stuck.

Point it at a session and it renders one **self-contained, offline HTML dashboard**: a timeline of tool calls, a read/explore → execute/write phase view, per-turn token & context growth, a tool-call breakdown, a file-access table, a **network-activity** panel, a **benchmark-validity audit** (flags an agent fetching a fix's provenance, instance-id leaks, work stranded in `git stash`, and writes to the graded tests — see FINDINGS finding 11), a file co-access graph, detected retry loops, a **repeated-work** panel (identical or near-identical calls the agent re-issues — a caching/optimization signal), and an errors list. No instrumentation, no services, no dependencies — it just reads the transcript Claude Code already writes.

---

Quick start (about 60 seconds)

No dependencies — **Python 3.9+ standard library only**.

git clone https://github.com/fairozkhan10/claude-code-trace-visualizer
cd claude-code-trace-visualizer

# 1. see a sample dashboard (synthetic data, safe to open)
open examples/example-report.html          # macOS  (Linux: xdg-open)

# 2. profile your most recent real Claude Code session
python3 -m cc_trace --latest --open

That's it. Step 2 finds your latest session under `~/.claude/projects/`, profiles it, and pops the report open in your browser.

On systems where `python` already means Python 3, you can drop the `3`. If you prefer a short command, `pip install -e .` gives you a `cc-trace` binary.

---

Everyday usage

# list the sessions available to profile (newest first)
python3 -m cc_trace --list

# profile a specific session (by id prefix) or a transcript path,
# and also dump the parsed data as JSON for your own analysis
python3 -m cc_trace 20814a75 -o reports/run.html --json --open

Transcripts live at `~/.claude/projects/