ibara
Install ibara

TEST RESULTS

Test results.

Acceptance runs on the test machines, newest first, including the ones that failed.

PassedPartial or openFailed

Updated

Runs#

DateTestResultDetailMachines
Focused rerun on 0.1.0-185 of the benchmark tasks, 3 agent setups, once on each machine.Passed21 of 30 runs passed, and no failure was ibara’s: 8 were a free model out of quota, 1 its malformed calls. Claude Code and Codex passed 20 of 20, including the file manager task 4 of 4. See Benchmarks.tater0, tater1
Workflow benchmarks on the releaseThe same 12 tasks on the published 0.1.0-17, with 7 agent setups and the packaged ibara skill.Partial132 of 157 runs passed (84 %). 13 failures were ibara’s: 10 were the file manager’s rename popup refusing all input, fixed in 0.1.0-18, and 3 the editor’s hidden settings; since 0.1.0-20 the editor uses your own settings. See Benchmarks.tater0, tater1
Live Video (Preview)One computer’s screen played as live video in another’s console.PartialOn 0.1.0-15 the video reached the console as H.264 at 480×270, and closing it stopped everything. The fleet card ran about 17 s behind; on 0.1.0-16, with the fix, the lag measured 1.5 s.tater0 → tater1
Workflow benchmarks with real agentsClaude Code, Codex, omp and opencode ran 6 desktop and 6 browser tasks, each from a clean start.Partial49 of 73 runs passed; 17 of the 24 failures were ibara defects, fixed since. See Benchmarks.tater0, tater1
100 save-a-file tasks in a rowThe same task 100 times, one after another.Passed99 of 100; median 12.6–12.9 s with no drift; ibara’s memory flat at 17.4 MiB. The one failure was a monitor reconnecting mid-typing, which ibara reported honestly. A fix is in, not yet rechecked on the test machines.tater0
Two agents on each of two computersTwo agents at once on each computer.Passed60 of 60 tasks that began passed in 7 minutes. There’s no queue, so a waiting agent waited up to 107 s.tater0, tater1
50 Take Control and Hand Back cyclesA person takes over and hands back, fifty times.PassedNo errors, no stuck keys, nothing left running.tater1 → tater0
20 restarts of ibara mid-taskibara is restarted while an agent is working.Passedibara 0.1.0-9. Every outcome honest, every app and its text survived, agents back in 6.5–9.2 s. For those seconds agents were told a person had the computer; fixed, and rechecked on 0.1.0-10 and 0.1.0-11.tater0
10 network drops of 10–90 sThe network is cut during a task.PassedNo lockout. Each call either waited and ran once, or said it may have run. The task survived 90 s offline.tater0
Files of 1–5 GBSending and receiving very large files.Failedibara 0.1.0-9. Not supported: the limit is 250 MB. At 100 MiB, sends ran at 2.4–3.8 MiB/s and receives at about 5 MiB/s. Sends of 200 and 256 MiB arrived whole but were reported failed; fixed, and rechecked on 0.1.0-10 (verified in 59 and 81 s). The 250 MB limit was rechecked on 0.1.0-11.tater0
Clicks land where aimedClicks checked against targets at known positions.Passed300 of 300 first-try hits, 0 px from center: page buttons in Chrome at scale 1 and 1.5 and zoom 100 and 125 %, and a GTK 4 app at scale 1 and 1.5. An idle mouse was plugged in during this test.tater0
Take Control and Hand Back, timedTimed from the choice in the console.PassedViewer on screen in 8.2 s median; your first key arrives 50 ms after it shows; the computer takes an agent again 0.7 s after Hand Back.tater1 → tater0
Your mouse while an agent worksYou move your own mouse while an agent is clicking.PassedYour pointer moves 17 ms median after you move the mouse, 15 of 15; the agent’s next click went through each time.tater0
Memory and CPU, full RAMEach state again on tater0, with all of its memory.PartialIdle 17.8 MiB; agent working 117.0 MiB; console open on the fleet 58.2 MiB, over its 40 MiB target.tater0
Idle soakHow ibara behaves left idle for a long time.OpenCut to about 5 minutes; an hour-long trend hasn’t been measured.tater0
First-time setup and take control, by handFrom a console with nothing paired yet: add a second computer, watch it, send files and take control.PassedRun by a person at the keyboard, not scripted.tater0, tater1
Four agents connect themselvesCodex, Claude Code, opencode and omp each get the one connect prompt, add ibara to their own settings, then open the text editor on another computer, type a line and save it.PassedAll 4 connected and all 5 tasks completed, each file read on disk; 3 model providers. Two defects found and fixed afterwards: choosing an item from the editor’s File menu failed (3 of 5 runs), and an editor opened without waiting for its window was left open (3 of 5).tater1
Memory on ibara 0.1.0-7Each memory state again, on a machine limited to 2 GB.FailedIdle, console not opened: 18.6 MiB (limit 40). Over the limit: agent working, console closed, 121.6 MiB (120); idle after every part of the console was opened, 48.8 MiB (40); agent working with the console open, 187.1 MiB (120). Accepted for this release.tater1
Browser tasks on signed-in browsersAn agent fills in and submits a web form: typing, choosing an option, ticking a box, and clicking a button that only responds to real clicks.Passed10 tool calls, 14.2–14.6 s. Every page event arrived as real input.tater0, tater1
Clicks land where aimedClicks on page buttons checked at different screen scales and zoom levels.Passed0 px off target in every run.tater0, tater1
Browser control on a fresh profileA brand-new browser profile becomes ready for agent control with no clicks.PassedReady in 2.0–3.1 s.tater0, tater1
Save a file, with Cua inputOpen a text editor, type a line, save it, hand the file back, with every keystroke through Cua.Passed7 tool calls, 11.4–11.5 s; 12.2 s on a 2 GB machine.tater0, tater1
Save a file, repeatedThe same task five times in a row.Partial1 of 5 runs failed typing into a Save As dialog that had just opened. The next three passed.tater0
Every agent connection routeEach console reaches each computer through its installed connection.PassedAll five routes.all
Memory: idleibara’s memory on a machine limited to 2 GB, before the console is opened.Passed16.5 MiB (limit 40).tater1
Memory: agent workingMemory while an agent works, console closed.Passed84.2 MiB (limit 120).tater1
Memory: idle after console useMemory once the console has been opened and closed.Failed41.8–45.4 MiB (limit 40).tater1
Memory: working, console openMemory while an agent works and the console is open.FailedOver the limit in 1 of 3 runs: 109.8–128.9 MiB (limit 120).tater1
Permissions and approvalsOne-time approvals, revoking and restoring access, and re-pairing, for agents and people.Passed7 agent cases per computer, 7 cases for people.tater0, tater1
Take control, stream and hand backA person takes over from an agent, watches the stream, and hands back, on both computers.PassedAlongside a save-a-file task: 7.9–9.9 s.tater0, tater1
Two desktops at onceA person controls one computer while an agent works on the other; then two agents at once; then one computer is stopped and revoked without affecting the other.PassedWorked on both computers. Its timings are replaced by the 28 September measurement: viewer in 8.2 s, Hand Back 0.7 s.tater0, tater1
RestartThe computer is restarted and its disk password typed at the keyboard. It must come back usable, with no control carried over from before the restart.PassedOne restart so far, on an earlier release that came back paused and was resumed by hand.tater0
Guided setup of a second computerA second computer set up entirely from the console, then paired with an on-screen code; a wrong code is refused.PassedAfter about 10 defects were fixed, including a permissions hole.tater1
File transfer safetyRefused links, resumed partial writes and uncertain transfers.PassedChecked against a test fixture.fixture
Handoff enforcementTake over from an agent, hand back, let another person replace you, and refuse viewers that reconnect afterwards.PassedWith enforcement on.tater0, tater1
Stream closed on revokeA viewer whose access is revoked is disconnected.PassedConnection closed in 0.59 s.tater0
Preview memory capA wall of 100 computer previews stays within its memory cap.PassedPeak 63.9 MiB (cap 64).your computer
Task notificationsA notification when an agent finishes, and none when a task only partly succeeds.PartialCompletion notices proved; the “stay silent” case is not yet reliably proved.tater1

How agents were tested#

Until Sep 27, most full-task runs were driven by a scripted client that calls ibara’s MCP tools the same way any agent does. On Sep 28, real agents ran the full benchmark: Claude Code, Codex, omp and opencode, with models from Anthropic, OpenAI, DeepSeek, xAI and OpenCode Zen, 73 runs in all. Google Gemini hasn’t been tested yet. Benchmarks. Agents connect themselves from one prompt, copied from the console (Connect an Agent, then Copy Prompt) or printed by ibara prompt. On Sep 27, Codex, Claude Code, opencode and omp each connected themselves from the prompt of that build (0.1.0-8) and saved a file in the text editor on tater1. The prompt in 0.1.0-17 adds two steps, linking the ibara skill and adding a note to the agent’s instructions file; the four agents haven’t all run it yet. Agents and harnesses.

Not yet proven#

  • Memory is over target while the console is open: 58.2 MiB idle on the fleet (target 40), and up to 67 MiB over while an agent works with it open (the last measurement of that state, on 0.1.0-7); accepted for this release (resource use).
  • A power cut hasn’t been tested, and only tater0 has been restarted in a test. When a monitor reconnected during a task on tater0, the typing stopped part way; the fix, which keeps typing going through a display change, hasn’t been checked on the test machines yet.
  • Pages inside iframes and shadow DOM aren’t read yet.
  • An hour-long idle soak hasn’t been run.
  • Start without the disk password is proven in a virtual machine, not yet on real hardware.
  • Google Gemini hasn’t been tested.
  • The full benchmark of all 12 tasks hasn’t been run since 0.1.0-17. 0.1.0-18 had a focused rerun of 5 tasks; 0.1.0-19 and 0.1.0-20 haven’t been benchmarked.