agentdebugger in development

How it works

Agent Debugger reads the log your coding agent wrote and shows what it did, step by step.

What it reads

The session log files that three agents write on your computer:

  • GitHub Copilot. The Copilot CLI's ~/.copilot/session-state/<session>/events.jsonl, and the transcripts of Copilot Chat in VS Code.
  • Claude Code. ~/.claude/projects/<project>/<session>.jsonl, with the logs of its subagents in the session's folder.
  • Codex. ~/.codex/sessions/<year>/<month>/<day>/rollout-*.jsonl.

The home page shows where to find them on your system.

What it flags

Seven kinds of wasted effort. Each finding names the steps involved, what they cost in tokens (and whether that number is measured or estimated), and a fix to try.

Re-reads
A file read again with nothing changed, by the agent's file tool or through cat or sed -n. Example: config.py read 4 times, same content each time.
A whole data file for one value
A YAML, JSON, CSV or similar file read in full, about 2,000 tokens or more. Example: settings.yaml read whole, ~6,000 tokens.
A tool's source read instead of its help
The agent opens a script to learn how to run it, then runs it. Example: scripts/export.py read to learn how to run it, then run.
Retries and error loops
The same call failing again and again, or the same error coming back after edits. Example: build.sh failed 3 times in a row, then worked.
Large outputs
One result of about 5,000 tokens or more, which every later model call is sent again until the context is compacted. Example: a test run returned ~12,000 tokens, carried through 40 later model calls.
Many searches in a row
Eight searches or more without opening any result. Example: 9 searches in a row without opening any result.
A skill loaded again
The same skill loaded twice or more with no compaction in between. Example: skill deploy loaded 3 times without a compaction in between.

What each log records

Logs differ. The session page marks every estimated token count "est."

  • Claude Code and Codex record the context size on every model call, so the chart's line is measured. The breakdown by kind of content and what each call added are estimated (characters ÷ 4).
  • Copilot CLI records the real context size only at compactions and at the end of the session: the chart and per-call sizes are estimated (characters ÷ 4); the totals it records at shutdown are measured.
  • VS Code transcripts have no tool output. What a tool result added to the context is unknown, so the page counts messages only.

Each agent counts a little differently (Claude Code's size includes its system prompt and tool definitions), so don't compare sizes across agents.

Pins and highlights

Pin a log to find it again in the list on the home page, and highlight a step, a piece of text or a finding to add a note to it. Your pins, notes and the list of logs you opened stay in this browser; Privacy has the detail.

The sample

Try a sample opens a real Claude Code session (version 2.0.76) in which an agent works on a Rust no_std C library. It is from the public repository winksaville/rlibc-x, under its Apache-2.0 licence. Only the home path and user name were replaced with user; the rest is as Claude Code wrote it.