dracon @_draconnn

Install in Claude Code: /plugin marketplace add draconnn/session-profiler In Codex: codex plugin marketplace add draconnn/session-profiler Read-only and local: it serves the report on 127.0.0.1 and uploads nothing. MIT. github.com/draconnn/sessi…

What you get: • every agent on one timeline: subagents, and side sessions too • the slowest steps ranked, so the one command eating the session stands out • background builds and tests timed until they actually finish • click any bar to see the exact command

My coding agent took 49 minutes to merge a branch. Where did the time go? I built session-profiler to find out: ask Claude Code or Codex to "profile this session" and it opens a timeline of every agent, tool call and background job, from your local logs. 🧵

Why the Diff pane isn't enough: • in a worktree it also shows changes Claude didn't make • it's a separate view, so I lose which explanation goes with which change • it lists files, not the order of changes Related issue: github.com/anthropics/cla…

@bcherny request for Claude Code desktop: a "Normal + code changes" transcript view. Tool calls stay collapsed, but Edit/Write diffs show inline (with the scrollable diff viewer), also inside folded tool groups. Today each change is 2 clicks away: expand the group, then the edit.

@thsottiaux Also it would be nice to evaluate how model behaves with or without a skill/different version of agents. md/etc

@thsottiaux Have an easy way to use different harnesses to be able to compare them. In my case I removed my whole harness and gpt 6 astra worked much better - but to do this I had to backup my codex folder.

@thsottiaux Also it would be nice to evaluate how model behaves with or without a skill/different version of agents.md/etc

@thsottiaux Have an easy way to use different harnesses to be able to compare them. In my case I removed my whole harness and gpt 6 astra worked much better - but to do this I had to backup my codex folder.

@finkd When will it be released in Europe?

@ClaudeDevs As I understand I could have infinite session for this project? How context is handled in this case?

Code: github.com/draconnn/millw…. Fair warning, it's a reference implementation, not a turnkey tool — each repo needs its own guard script. Questions welcome.

Things that bit me: - /bin/sh never re-reads a running script. Edited the daemon, old version kept running for a day. - killing a phase mid-run held the worker lock for an hour - the worker ignored its time budget until I started injecting the wall-clock deadline into its prompt

I have two AI agents working on my side projects around the clock. Not a demo — a daemon on my Mac that's been running unattended for weeks. The setup, including the parts that broke: