task_id large_stringlengths 24 24 | prompt large_stringlengths 745 3.14k | reference_answer large_stringlengths 2.7k 30k | repository_url large_stringclasses 11
values | repository_base_commit large_stringclasses 11
values | language large_stringclasses 4
values | category large_stringclasses 5
values | rubric large_stringlengths 1.26k 8.21k | docker_image large_stringclasses 11
values |
|---|---|---|---|---|---|---|---|---|
6905333b74f22949d97ba9e3 | I'm trying to wrap my head around kitty's remote control system because I want to eventually add a new command to it. I've been reading through the code but I'm getting lost in how all the pieces connect.
So here's what's confusing me. When I run something like kitten @ ls from another terminal, somehow that command ... | I started by investigating kitty's remote control system to understand how commands flow from `kitten @ ls` to the running kitty instance and back.
The system uses two communication mechanisms: Unix sockets for external connections and TTY-based DCS escape sequences for shell integration.
When you specify a `--to` ad... | kovidgoyal/kitty | 815df1e210e0a9ab4622f5c7f2d6891d7dbeddf1 | c | Architecture & system design | [{"id": "a33fc01cba19849aaf3b55e6b801001c", "title": "1.1: States that kitty's remote control system uses two communication mechanisms: Unix sockets (when --listen-on or KITTY_LISTEN_ON is configured) and TTY-based DCS escape sequences (when running inside a kitty window).", "annotations": {"type": "positive hli verifi... | ghcr.io/scaleapi/swe-atlas:swe_atlas_QnA_kovidgoyal_kitty_1.0 |
6905333b74f22949d97ba9aa | I'm new to SimpleLogin and trying to get everything running locally. Once the application starts, how can I tell that the web server, email handler, and job runner are actually up and responding? What should I see in the logs or dashboard UI that confirms users can sign in and manage their aliases?
After it's running,... | I began by bringing up the Flask application using the real virtual environment and the provided sample configuration, so that everything I observed was based on actual runtime behavior rather than assumptions. The server was started with the following command, which explicitly points to the example.env file and the vi... | simple-login/app | 2cd6ee777f8c2d3531559588bcfb18627ffb5d2c | ts | Code Onboarding | [{"id": "39fb9b69e1bbf86db0156890f38bc313", "title": "1.1: Reports that a GET request to `/health` returns HTTP 200 status code with response body containing `success`, showing the web server is running.", "annotations": {"type": "positive hli verifier", "importance": "must have"}}, {"id": "27a8c60ef17da8e73cde36450f2c... | ghcr.io/scaleapi/swe-atlas:swe_atlas_QnA_simple-login_app_1.0 |
6905333b74f22949d97ba99d | I’m trying to understand why back navigation in a multi step onboarding flow sometimes behaves like it has a mind of its own. The Back button is supposed to move one step backward, but every so often it snaps straight to the first step or slips out into an entirely different flow, and it never feels truly random. What ... | I started with tracing the back navigation code path to understand why the behavior feels deterministic rather than random. The erratic back navigation behavior in multi-step onboarding flows stems from a fallback mechanism in the Redux connect function that maps external back targets into every step, not just the flow... | Automattic/wp-calypso | be7e5cc641622d153040491fd5625c6cb83e12eb | ts | Root-cause analysis | [{"id": "285c61f38e08d51b4692c448bfe5447b", "title": "1.1: Identifies the connect function where backUrl is computed (e.g., in client/signup/step-wrapper/index.jsx)", "annotations": {"type": "positive hli verifier", "importance": "must have"}}, {"id": "540c261b537bd397b98617458d7d2b24", "title": "1.2: Identifies the nu... | ghcr.io/scaleapi/swe-atlas:swe_atlas_QnA_Automattic_wp-calypso_1.0 |
6905333b74f22949d97ba9c8 | I'm worried about deploying TruffleHog in an environment where teams can contribute custom detector configurations. Custom detectors can specify webhook URLs for verification, and I need to understand what security boundaries exist around this feature before we open it up.
If someone submits a detector configuration p... | I started by looking at what happens when a custom detector webhook points to localhost. My concern was whether TruffleHog has any validation that would block requests to internal addresses. I wrote a quick test to find out:
```go
// test_localhost_ssrf.go
package main
import (
"fmt"
"net/http"
"net/http/... | trufflesecurity/trufflehog | e42153d44a5e5c37c1bd0c70e074781e9edcb760 | go | Security | [{"id": "b1eca1777a8c95a9600918bc1832aeb8", "title": "1.1: Explains the behavior when someone submits a detector configuration pointing to internal addresses (e.g., \"requests to localhost succeed\", \"127.0.0.1 is reachable\", \"private IPs are not blocked\", \"no SSRF protection for internal addresses\")", "annotatio... | ghcr.io/scaleapi/swe-atlas:swe_atlas_QnA_trufflesecurity_trufflehog_1.0 |
6905333b74f22949d97ba9f5 | I want to understand how Kitty actually handles input event flow and focus management across windows, tabs, and child processes at runtime, without relying on assumptions from reading the source.
Start Kitty from this repository and interact with it in a way that creates overlapping input activity: open multiple tabs ... | I started by setting up a virtual framebuffer to enable running Kitty in a headless environment, then built and launched Kitty with debug input logging to observe the input handling pipeline at runtime.
Setting up the environment and building Kitty:
$ Xvfb :99 -screen 0 1024x768x24 &
$ export DISPLAY=:99
$ cd /app &&... | kovidgoyal/kitty | 815df1e210e0a9ab4622f5c7f2d6891d7dbeddf1 | c | Architecture & system design | [{"id": "289c38261413a0073f884ffbb33408ea", "title": "1.1: Shows that Kitty was running during the analysis (e.g., process ID, debug output, or runtime logs).", "annotations": {"type": "positive hli verifier", "importance": "must have"}}, {"id": "c9e76116e256f3ba23c17ae7def98d12", "title": "1.5: Includes a stack-level ... | ghcr.io/scaleapi/swe-atlas:swe_atlas_QnA_kovidgoyal_kitty_1.0 |
6905333b74f22949d97baa1b | I'm trying to get a clearer sense of how MinIO actually behaves when things start going wrong at runtime, especially before we rely on it more heavily for fault tolerance. When MinIO runs in distributed mode with four directories and data is spread across them using erasure coding, it reports that everything is healthy... | I started by setting up a local MinIO distributed environment to actually observe the fault tolerance behavior rather than just reading about it. I created four data directories and launched MinIO in distributed mode:
```bash
mkdir -p /tmp/minio-test/disk1 /tmp/minio-test/disk2 /tmp/minio-test/disk3 /tmp/minio-test/di... | minio/minio | c07e5b49d477b0774f23db3b290745aef8c01bd2 | go | Root-cause analysis | [{"id": "4d7412a80019b76df71ebb297211f95f", "title": "1.1: Reports the health endpoint HTTP response with the X-Minio-Write-Quorum header showing the required disk count (e.g., X-Minio-Write-Quorum: 3).", "annotations": {"type": "positive hli verifier", "importance": "must have"}}, {"id": "6c52f3aa9de91caa5edc396bd01ec... | ghcr.io/scaleapi/swe-atlas:swe_atlas_QnA_minio_minio_1.0 |
6905333b74f22949d97ba9fd | I'm onboarding to the kitty codebase and want to understand how its file transfer protocol enables efficient file copying over SSH connections. After building kitty from source, establish an SSH connection using the SSH kitten and initiate a file transfer, then help me trace the complete journey of file data through th... | I started by checking if kitty was already built:
```bash
ls -la kitty/launcher/
```
And I saw:
```text
total 15652
drwxr-sr-x 1 root 1001 4096 Aug 28 21:37 .
drwxr-sr-x 1 root 1001 4096 Aug 28 21:36 ..
-rwxr-xr-x 1 root 1001 15945988 Aug 28 21:37 kitten
-rwxr-xr-x 1 root 1001 36224 Aug 28 21:35 kitty
```... | kovidgoyal/kitty | 815df1e210e0a9ab4622f5c7f2d6891d7dbeddf1 | c | Architecture & system design | [{"id": "3de269be6bdfe6519e5ba2b8440b603e", "title": "1.5: Reports FILE_TRANSFER_CODE value is 5113. For example, states OSC escape sequences use number 5113, or reports this constant defined in source code. This demonstrates agent identified the specific OSC escape sequence code that kitty uses to demarcate file trans... | ghcr.io/scaleapi/swe-atlas:swe_atlas_QnA_kovidgoyal_kitty_1.0 |
6905333b74f22949d97ba9fb | I'm new to Kitty and I'm having trouble with Kitty's search query parser and I suspect there might be a bug. When I combine multiple search terms with "or" and spaces, the results don't match what I expect. Items that should clearly match at least one term are being excluded entirely.
I looked through the code briefly... | The search query parser in Kitty is not buggy. The behavior you're experiencing comes from operator precedence rules where AND operations bind tighter than OR, and spaces between terms create implicit AND operators. Understanding the parser's recursive descent structure explains why queries combining "or" with spaces p... | kovidgoyal/kitty | 815df1e210e0a9ab4622f5c7f2d6891d7dbeddf1 | c | Root-cause analysis | [{"id": "61a9ff9bd7a72d6c36be742766f60796", "title": "1.1 Concludes that the observed behavior is not a bug but expected/intended parser behavior.", "annotations": {"type": "positive hli verifier", "importance": "must have"}}, {"id": "4188ca130fd6e435b2d37ee167840100", "title": "1.2 Reports that spaces between search t... | ghcr.io/scaleapi/swe-atlas:swe_atlas_QnA_kovidgoyal_kitty_1.0 |
6905333b74f22949d97baa09 | I'm onboarding to TruffleHog and trying to understand its detection architecture before working with it. If I build it from source and run a basic scan, what happens during startup? Does it load detector configurations from files, or are they compiled in? What initialization messages appear about which detectors get re... | I started by building TruffleHog from source and exploring its architecture through code analysis and practical experiments:
```
$ go build -o trufflehog .
```
Here's what I discovered about its detection system.
When I examined how detector configurations are loaded, I found they are compiled directly into the bina... | trufflesecurity/trufflehog | e42153d44a5e5c37c1bd0c70e074781e9edcb760 | go | Architecture & system design | [{"id": "0ff886f7da4e55e77408ab834f55c3ea", "title": "1.1: Reports worker initialization messages from startup logs (e.g., \"starting scanner workers\", \"starting detector workers\" with count values such as 12, 96).", "annotations": {"type": "positive hli verifier", "importance": "must have"}}, {"id": "8d12c50eaeb85f... | ghcr.io/scaleapi/swe-atlas:swe_atlas_QnA_trufflesecurity_trufflehog_1.0 |
6905333b74f22949d97ba9c2 | I am joining a team that relies heavily on Scapy for custom packet manipulation, and before I start writing any code, I need to understand what the runtime environment actually looks like. I've cloned the Scapy repository locally, and I need you to help me get oriented by answering some specific questions about how it ... | I executed a small Python script to introspect the installed Scapy runtime. The script captured the startup banner and version, inspected the current configuration (including verbosity, socket settings, and theme), counted all registered protocol layers, and built an example ICMP packet so I could examine its structure... | secdev/scapy | 0925ada485406684174d6f068dbd85c4154657b3 | python | Code Onboarding | [{"id": "563c0c280150d67d4a5c4d04a5e720d8", "title": "1.1: States that the Scapy runtime reports version 2.5.0.dev87 via both scapy.__version__ and conf.version to demonstrate programmatic access methods for version checking in scripts.", "annotations": {"type": "positive hli verifier", "importance": "must have"}}, {"i... | ghcr.io/scaleapi/swe-atlas:swe_atlas_QnA_secdev_scapy_1.0 |
6905333b74f22949d97baa27 | I'm onboarding into the SFTPGo repository and trying to understand how the connection handling, quota enforcement, and atomic upload mechanisms hold together under real pressure, because the code paths look straightforward in isolation but I suspect the interesting behavior only emerges when everything runs at once. Wh... | I started by setting up a test environment to observe how SFTPGo handles concurrent sessions, quota enforcement, and atomic uploads under pressure. I built the server from source and created a test configuration with strict limits: max_sessions set to 2, quota_files set to 5, quota_size set to 50000 bytes, idle_timeout... | drakkan/sftpgo | 44634210287cb192f2a53147eafb84a33a96826b | go | Architecture & system design | [{"id": "a6ca0d4631789cc1b5b67e73891c268f", "title": "1.1: Reports session acceptance versus rejection counts when multiple sessions approach max_sessions limit (e.g., 2 accepted, 2 rejected when limit is 2).", "annotations": {"type": "positive hli verifier", "importance": "must have"}}, {"id": "60bb02a171a558eff428340... | ghcr.io/scaleapi/swe-atlas:swe_atlas_QnA_drakkan_sftpgo_1.0 |
End of preview. Expand in Data Studio
Update 03/30/2026: We released the dataset in harbor format in our official GitHub repo for SWE-Atlas. We recommend using the harbor scaffold with modal runtime sandboxes as the official way to run the benchmark.
SWE-Atlas QnA
Codebase QnA is the first benchmark in the SWE-Atlas suite. It evaluates AI agents on deep code comprehension — tracing execution paths, explaining architectural decisions, and answering deeply technical questions about production-grade software systems.
124 tasks across 11 open-source repositories spanning Go, Python, C, and TypeScript.
Link to leaderboard - https://scale.com/leaderboard/sweatlas-qna
Schema
| Column | Type | Description |
|---|---|---|
task_id |
string | Unique 24-char hex identifier |
prompt |
string | The question presented to the agent |
reference_answer |
string | Expert-written reference answer |
repository_url |
string | GitHub repo |
repository_base_commit |
string | 40-char commit SHA the environment is pinned to |
language |
string | go, python, c, or ts |
category |
string | Task category (see below) |
rubric |
string (JSON) | Evaluation criteria (see below) |
docker_image |
string | Docker Hub image for the sandboxed environment |
Rubric format
Each task's rubric field is a JSON array:
[
{
"id": "a33fc01cba19849aaf3b55e6b801001c",
"title": "1.1: States that kitty uses Unix sockets for external connections...",
"annotations": {
"type": "positive hli verifier",
"importance": "must have"
}
}
]
positive hli verifier— a factual claim the answer must contain. If the claim is met my the agent's answer, the rubric item result is a PASS.negative hli verifier— something the answer must not claim. If the claim is met my the agent's answer, the rubric item result is a FAIL.
Each task includes a docker_image field pointing to a pre-built Docker Hub image with the repository and all dependencies installed at /app:
Inference and Eval
We follow the standard SWE-Agent scaffold, and we provide a sample config (with the prompts) in default_qa_config.yaml
To run tasks, you can pull the docker image and run the container, and reset the environment to the base commit:
cd /app
git config --global --add safe.directory /app
git restore .
git reset --hard <repository_base_commit>
git clean -fdq
Evaluation is performed by an LLM judge (Claude Opus 4.5) that scores the agent's answer against each rubric criterion independently. Each criterion receives a binary score (met or not met) indicating and is then aggregated.
The primary metric is the Task Resolve Rate: the percentage of tasks for which the agent's answer is comprehensive (i.e. passes all rubric items and scores 1.0), as graded by a set of task-specific rubrics.
The agents are also instructed to avoid modifying source-code files, and clean up any temporary scripts made. So we add a programmatic check that fails a task that has any code changes after submission.
Our rubric evaluation prompt and other relevant details are in rubric_evaluation_config.yaml
- Downloads last month
- 595