Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

docs: where a new user's first 30 seconds go (#1125)

Dee8f29b28c31cc
diff --git a/docs/research/2026-10-08-activation-sim.md b/docs/research/2026-10-08-activation-sim.md
index e2d056337..1c87e92e7 100644
--- a/docs/research/2026-10-08-activation-sim.md
+++ b/docs/research/2026-10-08-activation-sim.md
@@ -469,6 +469,22 @@ times on the starter task (`opencode run -m <model>`, a fresh folder and home ea
The slow runs were the default model under load at that moment. Pinning another model would risk
a first session failing when OpenCode changes its free list, for no steady gain.
+**Where a new user's first 30 s go** (final main, fresh VM, from the app's and CLI's logs, 2026-10-09):
+
+| After opening | Step |
+|---|---|
+| +4 s | app launched |
+| +7 s | tmux in place (managed runtime) |
+| +7 → +20 s | Harness's Node runtime and the CLI downloaded and installed: the longest step |
+| +15 s / +18 s | OpenCode downloaded (10.4 s, in parallel) / its first run done |
+| +22 → +25 s | `harness start`: the daemon is ready |
+| +27 → +30 s | the first workspace planned (1.9 s) and its three panes opened (4.5 s) |
+
+OpenCode is off the critical path, so finishing it in its pane as Codex and Claude Code do would
+gain nothing. The one large step is the Node runtime and the CLI. Shipping them in the app would
+save about 13 s, against the documented choice that `install.sh` alone installs them
+(desktop/CLAUDE.md). That is the owner's decision. Everything else left is 1–4 s.
+
**Rig note:** the VM has no developer tools, so it raises Apple's dialog for anything that runs
`git`. Returning users of Claude Code or Codex have them, so a dialog seen there is the rig's, not
theirs.