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.
TypeScript 3d-modelingagent-frameworkagent-harnessai-agentscadclaude-codecodexcoding-agentsdeveloper-toolsesp32game-developmentmcp-servermulti-agentopen-hardwareopencodeopenharnesspcbroboticsterminaltmux
docs: separate app and agent performance accounting
deec1e749da6dde
diff --git a/README.md b/README.mdindex 3b65701c2..3d13e5714 100644--- a/README.md+++ b/README.md@@ -82,16 +82,19 @@ actual machines. Each row includes 600 echoes per workload across three runs:The route matters. These round trips exclude UI rendering; adding independentp95 values would not produce an end-to-end p95.-Idle resource use with 1,000 seeded lines per retained terminal:+Desktop app resource use with 1,000 seeded lines per retained terminal:-| Retained terminals | Foreground / hidden CPU | Foreground / hidden footprint |+| Retained terminals | App foreground / hidden CPU | App foreground / hidden footprint ||---|---:|---:|| 1 | 0.08% / 0.08% | 213 / 217 MiB || 16 | 0.09% / 0.10% | 304 / 306 MiB || 48 | 0.10% / 0.08% | 559 / 559 MiB |Thirty-second process snapshots; 100% CPU means one core. Memory is median-physical footprint, excluding daemons and agents. Hidden fixtures still recorded+physical footprint of the **desktop app alone**, including its terminal renderer+and scrollback. The 213 MiB one-terminal result contains no Claude Code, Codex,+shell, tmux, or Harness daemon process memory. Those need separate process+measurements; their usage was not measured in this fixture. Hidden fixtures still recorded12–14 interrupt wakeups/s; these measurements do not establish zero backgroundwork in a connected app.diff --git a/docs/performance/2026-09-23-core-experiences.md b/docs/performance/2026-09-23-core-experiences.mdindex e142bd202..1895ba30f 100644--- a/docs/performance/2026-09-23-core-experiences.md+++ b/docs/performance/2026-09-23-core-experiences.md@@ -177,12 +177,62 @@ longer before sampling than the other fixtures.| core-48-background-idle | 0.08% | 558.5 / 564.0 MiB | 13.5 || core-48-foreground-idle | 0.10% | 559.2 / 596.1 MiB | 16.0 |-The idle samples include the fixture and benchmark bookkeeping, not daemon or-agent memory. A background window still has measurable wakeups. The synthetic+The **212.8 MiB** one-terminal foreground result is entirely the desktop fixture+process. It includes the app, terminal parsing/rendering, scrollback, and benchmark+bookkeeping. It contains no Claude Code, Codex, shell, tmux, or Harness daemon+process memory. This process counter cannot separate the app shell from its+in-process terminal renderer; that requires allocation profiling or controlled+empty-app versus terminal fixtures. The independent 1/16/48-terminal snapshots+are not a causal per-terminal allocation measurement.++Future connected resource measurements must report these separately:++| Resource owner | What belongs in the measurement |+|---|---|+| Desktop app | UI, terminal parsing/rendering, retained scrollback, in-process caches |+| Harness daemon | Connection handling, encryption, terminal transport, registries and watchers |+| Terminal infrastructure | tmux/PTY host and shells, counted once per process |+| Agent CLI and its children | Claude Code, Codex, etc., with tool subprocesses attributed to their session |++Sample all groups over the same interval and show both group usage and an+explicitly defined total. Keep remote-machine processes on their own machine's+scorecard. Record process exits, missing permissions, agent activity, and shared+processes rather than silently counting them as zero or attributing them twice.+Also measure the desktop's incremental retained memory as terminal count and+scrollback grow, and memory after repeated open/close cycles and a long soak.++A background window still has measurable wakeups. The syntheticfixture has no real connections, so these results do not support a claim of zerobackground work in the connected application. Raw samples include residentbytes, CPU deltas, disk I/O, and package-idle wakeups as well.+## Performance priorities and missing measurements++Prioritize frequent developer actions and their slow tails. Every latency result+needs a verified usable outcome, a workload, a sample count, failure/timeout+counts, and p50/p95/p99 with maxima. A fast frame with incorrect focus or a lost+keystroke is a failure. Do not pool different transports or workload sizes.++| Priority | Experience | Measurement boundary |+|---|---|---|+| 1 | Local and remote typing | Input to matching PTY echo; then separately input to visible echo. Compare verified direct P2P, Cloudflare TURN, and Harness WebSocket relay on the same target, idle and during output. |+| 1 | Open, switch, create | Cmd+O selection, tab/pane switch, and Cmd+N submission to a usable terminal and first accepted input. Keep opening the creation UI separate from creating a session or starting an agent. |+| 1 | Busy-terminal responsiveness | Typing, scrolling, search, resize, and zoom with sustained and bursty output; frame stalls, achieved throughput, queue growth, and correctness. |+| 2 | App startup | Cold/warm process launch to first usable window, restored active session, and accepted input. First paint alone is insufficient. |+| 2 | Recovery | Sleep/wake, connection loss, and route migration to resumed input/output; lost/duplicated bytes, stale state, and unsuccessful recoveries. |+| 2 | Resource efficiency | The separate process groups above: CPU, physical footprint, wakeups, disk/network traffic, long-session growth, and idle versus output-heavy behavior. |++Install time is outside this scorecard. Startup remains in scope. Agent/model+time to first token must be kept separate from Harness's session and transport+overhead; external model latency must not mask an app regression.++To keep improvements, establish repeated baselines on a controlled runner before+setting per-workflow regression budgets. Gate common paths on tail latency and+correctness, retain raw observations and outliers, and require the same workload+before claiming a speedup. Treat memory growth and hidden-window wakeups as+regressions even when a shortcut median remains fast. Budgets are a next step,+not a claim that these workstation snapshots already enforce an SLO.+## Native navigation workNavigation changes rebuild the History menu even when it is closed in the