A demo of darner, with no model and no sandbox: nothing you type leaves this browser. dejima.page

dn-example

dn-zzzzz. Started by web, 0m ago. running

Everything marked as the agent's was written by the agent. Treat it as untrusted.

Nothing yet.

What you send is typed into the session and not kept, so it won't appear here.

Started with

Config sets
recipe-service
Blurbs
recipe-service, search
Harness
cline
Backend
self-hosted
Model
qwen3.5-4b
Effort
high
Context window
120000
Domains
pypi.org, recipes.example.com
Internal domains
recipes.example.com
Dependencies
pip:requests
Files
recipe-service-api-token, add_recipe.py

dn-example

dn-zzzzz. Started by web, 0m ago. running

Everything marked as the agent's was written by the agent. Treat it as untrusted.

Output the agent's

This is a demo of darner, and there's no model on the other end. No sandbox was started, and nothing you typed left your browser.

On a real darner, your task would now be running as an isolated dejima (https://dejima.page) session on the dejima host, with the settings you chose. Its page would show its output once the agent wrote ~/shared/RESULT.md and signalled it was done, with everything else it left in ~/shared as a zip. Until then, you could send it a message, upload a file into its ~/shared/uploads, and allow or deny the domains it asks for.

This output came the way a real session's does: from the stub launcher this demo was built with, through darner's API, cleaned and shown as text written by an agent, which is why it's labelled as the agent's. Self-test sends text written like a hostile agent's down the same path, to show what this page does with it.

Download the zip 610 bytes

What you send is typed into the session and not kept, so it won't appear here.

Started with

Config sets
recipe-service
Blurbs
recipe-service, search
Harness
cline
Backend
self-hosted
Model
qwen3.5-4b
Effort
high
Context window
120000
Domains
pypi.org, recipes.example.com
Internal domains
recipes.example.com
Dependencies
pip:requests
Files
recipe-service-api-token, add_recipe.py

dn-zzzzz

Ended. Its output is kept here until it's cleared. Its ~/shared is kept on the dejima host too, since it holds files the zip doesn't.

Everything marked as the agent's was written by the agent. Treat it as untrusted.

Ended

Its output, zips and ~/shared on the dejima host are deleted when it's cleared.

dn-zzzzz

Ended. Its output is kept here until it's cleared. Its ~/shared on the dejima host was removed, since everything in it is in the zip.

Everything marked as the agent's was written by the agent. Treat it as untrusted.

Output the agent's

This is a demo of darner, and there's no model on the other end. No sandbox was started, and nothing you typed left your browser.

On a real darner, your task would now be running as an isolated dejima (https://dejima.page) session on the dejima host, with the settings you chose. Its page would show its output once the agent wrote ~/shared/RESULT.md and signalled it was done, with everything else it left in ~/shared as a zip. Until then, you could send it a message, upload a file into its ~/shared/uploads, and allow or deny the domains it asks for.

This output came the way a real session's does: from the stub launcher this demo was built with, through darner's API, cleaned and shown as text written by an agent, which is why it's labelled as the agent's. Self-test sends text written like a hostile agent's down the same path, to show what this page does with it.

Download the zip 610 bytes
Its output, zips and ~/shared on the dejima host are deleted when it's cleared.

History

  1. Ended

    0m ago

dn-zzzzz

Ended. Its output is kept here until it's cleared.

Everything marked as the agent's was written by the agent. Treat it as untrusted.

Seems to be waiting: its screen hasn't changed for 5m the end of its screen, the agent's

Frontend self-test. This isn't an agent: it's text written like a hostile agent's, to check that this frontend shows it safely. Each check says what you see if it's safe (PASS) and what you see if it isn't (FAIL), then gives its test line. The checks describe markup in words, so they read the same however the test lines come out.

Check 1, raw HTML. PASS: test line 1 shows b tags, in angle brackets, around its words. FAIL: no angle brackets show, and the words are bold.

Test line 1: <b>these words must not be bold</b>

Check 2, scripts. PASS: test line 2 shows script and img tags, in angle brackets, and no dialog or picture appears. FAIL: the line is empty or missing its tags, a dialog pops up, or a picture or broken-image icon appears.

Test line 2: <script>alert("selftest")</script> <img src="https://pixel.selftest.invalid/x.png" onerror="alert('selftest')">

Check 3, HTML links. PASS: test line 3 shows an a href tag, in angle brackets, with evil.selftest.invalid in it. FAIL: it shows only the words "click here", as a link.

Test line 3: <a href="https://evil.selftest.invalid">click here</a>

Check 4, markdown links. PASS: test line 4 shows evil.selftest.invalid. FAIL: it shows only the words "click here", as a link.

Test line 4: [click here](https://evil.selftest.invalid)

Check 5, markdown images. PASS: test line 5 shows pixel.selftest.invalid, and no picture or broken-image icon. FAIL: a picture or icon appears, or the address doesn't.

Test line 5: ![a picture](https://pixel.selftest.invalid/x.png)

Check 6, mentions. PASS: nobody gets a notification, and nothing on test line 6 becomes a highlighted name or pill. FAIL: anyone is notified, or anything on it is highlighted.

Test line 6: @room @here @channel @everyone <!here> <!channel> <@U0SELFTEST> @selftest:selftest.invalid

Check 7, Slack links. PASS: test line 7 shows evil.selftest.invalid. FAIL: it shows only the words "click here", as a link.

Test line 7: <https://evil.selftest.invalid|click here>

Check 8, invisible characters. PASS: test line 8 ends with abc, in that order, with nothing between the letters. FAIL: the letters are in another order, such as cba, or something shows between them.

Test line 8: abc

Check 9, long addresses. PASS: test line 9 shows the whole of its link's address, ending in evil.selftest.invalid/x, wrapped if it must be, without scrolling sideways. FAIL: the address is cut off before evil.selftest.invalid, or only docs.selftest.invalid shows.

Test line 9: [https://docs.selftest.invalid/](https://docs.selftest.invalid.mirrorcdndownload0123456789abcdefghijklmnopqrstuvwxyz0123456789.evil.selftest.invalid/x)

Check 10, code fences. PASS: check 11 appears in this same message, after test line 10. FAIL: the message ends at test line 10, or check 11 appears outside it.

Test line 10:
```
Check 11, the end. PASS: you can read this line. FAIL: this line is missing, so the message was cut short.
Its output, zips and ~/shared on the dejima host are deleted when it's cleared.

History

  1. Ended

    0m ago

  2. Help request the agent's

    0m ago

    Frontend self-test. This isn't an agent: it's text written like a hostile agent's, to check that this frontend shows it safely. Each check says what you see if it's safe (PASS) and what you see if it isn't (FAIL), then gives its test line. The checks describe markup in words, so they read the same however the test lines come out.

    Check 1, raw HTML. PASS: test line 1 shows b tags, in angle brackets, around its words. FAIL: no angle brackets show, and the words are bold.

    Test line 1: <b>these words must not be bold</b>

    Check 2, scripts. PASS: test line 2 shows script and img tags, in angle brackets, and no dialog or picture appears. FAIL: the line is empty or missing its tags, a dialog pops up, or a picture or broken-image icon appears.

    Test line 2: <script>alert("selftest")</script> <img src="https://pixel.selftest.invalid/x.png" onerror="alert('selftest')">

    Check 3, HTML links. PASS: test line 3 shows an a href tag, in angle brackets, with evil.selftest.invalid in it. FAIL: it shows only the words "click here", as a link.

    Test line 3: <a href="https://evil.selftest.invalid">click here</a>

    Check 4, markdown links. PASS: test line 4 shows evil.selftest.invalid. FAIL: it shows only the words "click here", as a link.

    Test line 4: click here (https://evil.selftest.invalid)

    Check 5, markdown images. PASS: test line 5 shows pixel.selftest.invalid, and no picture or broken-image icon. FAIL: a picture or icon appears, or the address doesn't.

    Test line 5: a picture (https://pixel.selftest.invalid/x.png)

    Check 6, mentions. PASS: nobody gets a notification, and nothing on test line 6 becomes a highlighted name or pill. FAIL: anyone is notified, or anything on it is highlighted.

    Test line 6: @room @here @channel @everyone <!here> <!channel> <@U0SELFTEST> @selftest:selftest.invalid

    Check 7, Slack links. PASS: test line 7 shows evil.selftest.invalid. FAIL: it shows only the words "click here", as a link.

    Test line 7: <https://evil.selftest.invalid|click here>

    Check 8, invisible characters. PASS: test line 8 ends with abc, in that order, with nothing between the letters. FAIL: the letters are in another order, such as cba, or something shows between them.

    Test line 8: abc

    Check 9, long addresses. PASS: test line 9 shows the whole of its link's address, ending in evil.selftest.invalid/x, wrapped if it must be, without scrolling sideways. FAIL: the address is cut off before evil.selftest.invalid, or only docs.selftest.invalid shows.

    Test line 9: https://docs.selftest.invalid/ (https://docs.selftest.invalid.mirrorcdndownload0123456789abcdefghijklmnopqrstuvwxyz0123456789.evil.selftest.invalid/x)

    Check 10, code fences. PASS: check 11 appears in this same message, after test line 10. FAIL: the message ends at test line 10, or check 11 appears outside it.

    Test line 10:

    Check 11, the end. PASS: you can read this line. FAIL: this line is missing, so the message was cut short.

  3. Output the agent's

    0m ago

    Frontend self-test. This isn't an agent: it's text written like a hostile agent's, to check that this frontend shows it safely. Each check says what you see if it's safe (PASS) and what you see if it isn't (FAIL), then gives its test line. The checks describe markup in words, so they read the same however the test lines come out.

    Check 1, raw HTML. PASS: test line 1 shows b tags, in angle brackets, around its words. FAIL: no angle brackets show, and the words are bold.

    Test line 1: <b>these words must not be bold</b>

    Check 2, scripts. PASS: test line 2 shows script and img tags, in angle brackets, and no dialog or picture appears. FAIL: the line is empty or missing its tags, a dialog pops up, or a picture or broken-image icon appears.

    Test line 2: <script>alert("selftest")</script> <img src="https://pixel.selftest.invalid/x.png" onerror="alert('selftest')">

    Check 3, HTML links. PASS: test line 3 shows an a href tag, in angle brackets, with evil.selftest.invalid in it. FAIL: it shows only the words "click here", as a link.

    Test line 3: <a href="https://evil.selftest.invalid">click here</a>

    Check 4, markdown links. PASS: test line 4 shows evil.selftest.invalid. FAIL: it shows only the words "click here", as a link.

    Test line 4: click here (https://evil.selftest.invalid)

    Check 5, markdown images. PASS: test line 5 shows pixel.selftest.invalid, and no picture or broken-image icon. FAIL: a picture or icon appears, or the address doesn't.

    Test line 5: a picture (https://pixel.selftest.invalid/x.png)

    Check 6, mentions. PASS: nobody gets a notification, and nothing on test line 6 becomes a highlighted name or pill. FAIL: anyone is notified, or anything on it is highlighted.

    Test line 6: @room @here @channel @everyone <!here> <!channel> <@U0SELFTEST> @selftest:selftest.invalid

    Check 7, Slack links. PASS: test line 7 shows evil.selftest.invalid. FAIL: it shows only the words "click here", as a link.

    Test line 7: <https://evil.selftest.invalid|click here>

    Check 8, invisible characters. PASS: test line 8 ends with abc, in that order, with nothing between the letters. FAIL: the letters are in another order, such as cba, or something shows between them.

    Test line 8: abc

    Check 9, long addresses. PASS: test line 9 shows the whole of its link's address, ending in evil.selftest.invalid/x, wrapped if it must be, without scrolling sideways. FAIL: the address is cut off before evil.selftest.invalid, or only docs.selftest.invalid shows.

    Test line 9: https://docs.selftest.invalid/ (https://docs.selftest.invalid.mirrorcdndownload0123456789abcdefghijklmnopqrstuvwxyz0123456789.evil.selftest.invalid/x)

    Check 10, code fences. PASS: check 11 appears in this same message, after test line 10. FAIL: the message ends at test line 10, or check 11 appears outside it.

    Test line 10:

    Check 11, the end. PASS: you can read this line. FAIL: this line is missing, so the message was cut short.

dn-zzzzz

This session has ended and been cleared, or never existed here. darner keeps nothing about a session once it's cleared.

Back to sessions