Repository navigation
fix(server): agents drive simulators through the device tools instead of ssh #25661
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Merged
gabrielelpidio
merged 30 commits into
t3/device-manager-evals-api
from
t3/device-agent-guidance
Oct 28, 2026
+2,280
−214
Merged
Changes from all commits
Commits
Show all changes
30 commits
Select commit
Hold shift + click to select a range
23706b7
fix(server): device tools find a device id on any host
gabrielelpidio 48a4c0d
fix(server): device_screenshot says why it failed
gabrielelpidio 981f92a
fix(server): device_open returns an agent-device command bound to the…
gabrielelpidio 3558e7e
feat(server): every provider session with device access gets the devi…
gabrielelpidio 03fc169
fix(server): device_open's description asks for the hostId
gabrielelpidio 4c6a350
fix(server): the bound agent-device launcher reads flags only before --
gabrielelpidio 1224d82
fix(cursor): resumed and compacted conversations get the T3 instructions
gabrielelpidio 7ae08c5
fix(server): device launchers last only as long as their session
gabrielelpidio 3f52d3b
fix(server): Claude's native subagents get the device instructions
gabrielelpidio ffba1bd
fix(server): a device closed mid-open leaves no agent-device launcher
gabrielelpidio db1a006
fix(server): device_open still returns its target flags and session
gabrielelpidio 5f93761
fix(server): the device block leaves legitimate ssh work open but kee…
gabrielelpidio e58fb16
fix(server): device_open binds the session it opened and says when it…
gabrielelpidio 8e0e3e2
fix(server): an interrupted device close still removes its launcher
gabrielelpidio d28a055
fix(server): device_open says how to save full-resolution iOS screens…
gabrielelpidio b59828f
fix(evals): the mcp-tool-use eval's Claude prompt keeps the device block
gabrielelpidio 610bb2b
fix(server): the device failure rule covers the command device_open r…
gabrielelpidio b189472
refactor(server): DeviceService resolves a device id across hosts
gabrielelpidio b416ae2
fix(server): device guidance starts the returned command with open
gabrielelpidio 3e4168a
fix(server): device guidance opens the app before snapshots, screensh…
gabrielelpidio b91fa6d
fix(server): the device-access error says where the user turns access on
gabrielelpidio c4852f5
fix(server): the device-access error says access applies when the age…
gabrielelpidio b5b76e5
fix(server): the device failure rule never reuses or closes another s…
gabrielelpidio be619d5
fix(server): the bound agent-device command handles commands that ref…
gabrielelpidio fe52096
fix(server): ending a device session ends its agent-device session
gabrielelpidio fe52e45
fix(server): the bound agent-device command finds the command past le…
gabrielelpidio 3cb6969
fix(server): device connection errors cover network-namespace sandbox…
gabrielelpidio e1a0e5a
fix(server): a device_close that powers off releases every session on…
gabrielelpidio 621de28
fix(server): the bound command's connection check uses --daemon-base-…
gabrielelpidio 3b7a32e
test(server): raise the agent context budget for device guidance
gabrielelpidio File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Uh oh!
There was an error while loading. Please reload this page.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Derive the launcher path from
%~dp0in the.cmdwrapper.Line 60 puts the literal
launcherPathinto the batch script.launcherPathcomes fromstateDir.cmd.exetreats a single unmatched%in that path as a parameter or variable reference. The path is then misread and the launcher fails to start.agent-device.cmdandagent-device-launcher.mjsare in the same directory.%~dp0avoids the problem for this path. Thenodepath stays literal. It comes fromPATH, so it is less likely to contain%.Add a Windows regression test with a
%in the state directory name.🛠️ Proposed fix
Based on learnings: in
.cmdlaunchers, avoid interpolating raw paths that may contain%; derive self-referential paths with%~dpn0/%~dp0.Committable suggestion
Source: Learnings
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Agreed that a
%in the state directory or node path would break the.cmd. That pattern predates this PR: main's unbound shim already writes the literal launcher path intoagent-device.cmd. Here the session directories aret3-<hash>and add no%. I'd rather fix it for both launchers in the sharedwriteLauncher(%~dp0agent-device-launcher.mjs) together with a Windows test that actually runs the.cmd. No test does today, so I'll do that as a follow-up instead of an untested change here.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@gabrielelpidio Agreed. The diff confirms that the unbound shim already used the literal launcher path. A follow-up in shared
writeLauncher, with a Windows test that runs the.cmd, is reasonable. This remains deferred, not fixed.Would you like me to open a GitHub issue for the follow-up?
🧠 Learnings used
You are interacting with an AI system.