Release 0.29.0

Co-authored-by: openhands <openhands@all-hands.dev>

.github/workflows/deploy-docs.yml

@@ -38,8 +38,6 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: '3.12'
-      - name: Generate Python Docs
-        run: rm -rf docs/modules/python && pip install pydoc-markdown && pydoc-markdown
       - name: Install dependencies
         run: cd docs && npm ci
       - name: Build website

.github/workflows/eval-runner.yml

@@ -1,139 +0,0 @@
-name: Run SWE-Bench Evaluation
-
-on:
-  pull_request:
-    types: [labeled]
-  workflow_dispatch:
-    inputs:
-      reason:
-        description: "Reason for manual trigger"
-        required: true
-        default: ""
-
-env:
-  N_PROCESSES: 32 # Global configuration for number of parallel processes for evaluation
-
-jobs:
-  run-evaluation:
-    if: github.event.label.name == 'eval-this' || github.event_name != 'pull_request'
-    runs-on: ubuntu-latest
-    permissions:
-      contents: "read"
-      id-token: "write"
-      pull-requests: "write"
-      issues: "write"
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-
-      - name: Install tmux
-        run: sudo apt-get update && sudo apt-get install -y tmux
-      - name: Install poetry via pipx
-        run: pipx install poetry
-
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python-version }}
-          cache: "poetry"
-
-      - name: Comment on PR if 'eval-this' label is present
-        if: github.event_name == 'pull_request' && github.event.label.name == 'eval-this'
-        uses: KeisukeYamashita/create-comment@v1
-        with:
-          unique: false
-          comment: |
-            Hi! I started running the evaluation on your PR. You will receive a comment with the results shortly.
-
-      - name: Install Python dependencies using Poetry
-        run: poetry install
-
-      - name: Configure config.toml for evaluation
-        env:
-          DEEPSEEK_API_KEY: ${{ secrets.DEEPSEEK_LLM_API_KEY }}
-        run: |
-          echo "[llm.eval]" > config.toml
-          echo "model = \"deepseek/deepseek-chat\"" >> config.toml
-          echo "api_key = \"$DEEPSEEK_API_KEY\"" >> config.toml
-          echo "temperature = 0.0" >> config.toml
-
-      - name: Run SWE-Bench evaluation
-        env:
-          ALLHANDS_API_KEY: ${{ secrets.ALLHANDS_EVAL_RUNTIME_API_KEY }}
-          RUNTIME: remote
-          SANDBOX_REMOTE_RUNTIME_API_URL: https://runtime.eval.all-hands.dev
-          EVAL_DOCKER_IMAGE_PREFIX: us-central1-docker.pkg.dev/evaluation-092424/swe-bench-images
-
-        run: |
-          poetry run ./evaluation/benchmarks/swe_bench/scripts/run_infer.sh llm.eval HEAD CodeActAgent 300 30 $N_PROCESSES "princeton-nlp/SWE-bench_Lite" test
-          OUTPUT_FOLDER=$(find evaluation/evaluation_outputs/outputs/princeton-nlp__SWE-bench_Lite-test/CodeActAgent -name "deepseek-chat_maxiter_50_N_*-no-hint-run_1" -type d | head -n 1)
-          echo "OUTPUT_FOLDER for SWE-bench evaluation: $OUTPUT_FOLDER"
-          poetry run ./evaluation/benchmarks/swe_bench/scripts/eval_infer_remote.sh $OUTPUT_FOLDER/output.jsonl $N_PROCESSES "princeton-nlp/SWE-bench_Lite" test
-
-          poetry run ./evaluation/benchmarks/swe_bench/scripts/eval/summarize_outputs.py $OUTPUT_FOLDER/output.jsonl > summarize_outputs.log 2>&1
-          echo "SWEBENCH_REPORT<<EOF" >> $GITHUB_ENV
-          cat summarize_outputs.log >> $GITHUB_ENV
-          echo "EOF" >> $GITHUB_ENV
-
-      - name: Create tar.gz of evaluation outputs
-        run: |
-          TIMESTAMP=$(date +'%y-%m-%d-%H-%M')
-          tar -czvf evaluation_outputs_${TIMESTAMP}.tar.gz evaluation/evaluation_outputs/outputs
-
-      - name: Upload evaluation results as artifact
-        uses: actions/upload-artifact@v4
-        id: upload_results_artifact
-        with:
-          name: evaluation-outputs
-          path: evaluation_outputs_*.tar.gz
-
-      - name: Get artifact URL
-        run: echo "ARTIFACT_URL=${{ steps.upload_results_artifact.outputs.artifact-url }}" >> $GITHUB_ENV
-
-      - name: Authenticate to Google Cloud
-        uses: 'google-github-actions/auth@v2'
-        with:
-          credentials_json: ${{ secrets.GCP_RESEARCH_OBJECT_CREATOR_SA_KEY }}
-
-      - name: Set timestamp and trigger reason
-        run: |
-          echo "TIMESTAMP=$(date +'%Y-%m-%d-%H-%M')" >> $GITHUB_ENV
-          if [[ "${{ github.event_name }}" == "pull_request" ]]; then
-            echo "TRIGGER_REASON=pr-${{ github.event.pull_request.number }}" >> $GITHUB_ENV
-          elif [[ "${{ github.event_name }}" == "schedule" ]]; then
-            echo "TRIGGER_REASON=schedule" >> $GITHUB_ENV
-          else
-            echo "TRIGGER_REASON=manual-${{ github.event.inputs.reason }}" >> $GITHUB_ENV
-          fi
-
-      - name: Upload evaluation results to Google Cloud Storage
-        uses: 'google-github-actions/upload-cloud-storage@v2'
-        with:
-          path: 'evaluation/evaluation_outputs/outputs'
-          destination: 'openhands-oss-eval-results/${{ env.TIMESTAMP }}-${{ env.TRIGGER_REASON }}'
-
-      - name: Comment with evaluation results and artifact link
-        id: create_comment
-        uses: KeisukeYamashita/create-comment@v1
-        with:
-          number: ${{ github.event_name == 'pull_request' && github.event.pull_request.number || 4504 }}
-          unique: false
-          comment: |
-              Trigger by: ${{ github.event_name == 'pull_request' && format('Pull Request (eval-this label on PR #{0})', github.event.pull_request.number) || github.event_name == 'schedule' && 'Daily Schedule' || format('Manual Trigger: {0}', github.event.inputs.reason) }}
-              Commit: ${{ github.sha }}
-              **SWE-Bench Evaluation Report**
-              ${{ env.SWEBENCH_REPORT }}
-              ---
-              You can download the full evaluation outputs [here](${{ env.ARTIFACT_URL }}).
-
-      - name: Post to a Slack channel
-        id: slack
-        uses: slackapi/slack-github-action@v2.0.0
-        with:
-          channel-id: 'C07SVQSCR6F'
-          slack-message: "*Evaluation Trigger:* ${{ github.event_name == 'pull_request' && format('Pull Request (eval-this label on PR #{0})', github.event.pull_request.number) || github.event_name == 'schedule' && 'Daily Schedule' || format('Manual Trigger: {0}', github.event.inputs.reason) }}\n\nLink to summary: [here](https://github.com/${{ github.repository }}/issues/${{ github.event_name == 'pull_request' && github.event.pull_request.number || 4504 }}#issuecomment-${{ steps.create_comment.outputs.comment-id }})"
-        env:
-          SLACK_BOT_TOKEN: ${{ secrets.EVAL_NOTIF_SLACK_BOT_TOKEN }}

Development.md

@@ -100,7 +100,7 @@ poetry run pytest ./tests/unit/test_*.py
 To reduce build time (e.g., if no changes were made to the client-runtime component), you can use an existing Docker container image by
 setting the SANDBOX_RUNTIME_CONTAINER_IMAGE environment variable to the desired Docker image.
 
-Example: `export SANDBOX_RUNTIME_CONTAINER_IMAGE=ghcr.io/all-hands-ai/runtime:0.28-nikolaik`
+Example: `export SANDBOX_RUNTIME_CONTAINER_IMAGE=ghcr.io/all-hands-ai/runtime:0.29-nikolaik`
 
 ## Develop inside Docker container
 

README.md

@@ -12,7 +12,7 @@
   <a href="https://codecov.io/github/All-Hands-AI/OpenHands?branch=main"><img alt="CodeCov" src="https://img.shields.io/codecov/c/github/All-Hands-AI/OpenHands?style=for-the-badge&color=blue"></a>
   <a href="https://github.com/All-Hands-AI/OpenHands/blob/main/LICENSE"><img src="https://img.shields.io/github/license/All-Hands-AI/OpenHands?style=for-the-badge&color=blue" alt="MIT License"></a>
   <br/>
-  <a href="https://join.slack.com/t/openhands-ai/shared_invite/zt-2ypg5jweb-d~6hObZDbXi_HEL8PDrbHg"><img src="https://img.shields.io/badge/Slack-Join%20Us-red?logo=slack&logoColor=white&style=for-the-badge" alt="Join our Slack community"></a>
+  <a href="https://join.slack.com/t/openhands-ai/shared_invite/zt-2ngejmfw6-9gW4APWOC9XUp1n~SiQ6iw"><img src="https://img.shields.io/badge/Slack-Join%20Us-red?logo=slack&logoColor=white&style=for-the-badge" alt="Join our Slack community"></a>
   <a href="https://discord.gg/ESHStjSjD4"><img src="https://img.shields.io/badge/Discord-Join%20Us-purple?logo=discord&logoColor=white&style=for-the-badge" alt="Join our Discord community"></a>
   <a href="https://github.com/All-Hands-AI/OpenHands/blob/main/CREDITS.md"><img src="https://img.shields.io/badge/Project-Credits-blue?style=for-the-badge&color=FFE165&logo=github&logoColor=white" alt="Credits"></a>
   <br/>
@@ -43,17 +43,17 @@ See the [Running OpenHands](https://docs.all-hands.dev/modules/usage/installatio
 system requirements and more information.
 
 ```bash
-docker pull docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik
+docker pull docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik
 
 docker run -it --rm --pull=always \
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -e LOG_ALL_EVENTS=true \
     -v /var/run/docker.sock:/var/run/docker.sock \
     -v ~/.openhands-state:/.openhands-state \
     -p 3000:3000 \
     --add-host host.docker.internal:host-gateway \
     --name openhands-app \
-    docker.all-hands.dev/all-hands-ai/openhands:0.28
+    docker.all-hands.dev/all-hands-ai/openhands:0.29
 ```
 
 You'll find OpenHands running at [http://localhost:3000](http://localhost:3000)!
@@ -96,7 +96,7 @@ troubleshooting resources, and advanced configuration options.
 OpenHands is a community-driven project, and we welcome contributions from everyone. We do most of our communication
 through Slack, so this is the best place to start, but we also are happy to have you contact us on Discord or Github:
 
-- [Join our Slack workspace](https://join.slack.com/t/openhands-ai/shared_invite/zt-2ypg5jweb-d~6hObZDbXi_HEL8PDrbHg) - Here we talk about research, architecture, and future development.
+- [Join our Slack workspace](https://join.slack.com/t/openhands-ai/shared_invite/zt-2ngejmfw6-9gW4APWOC9XUp1n~SiQ6iw) - Here we talk about research, architecture, and future development.
 - [Join our Discord server](https://discord.gg/ESHStjSjD4) - This is a community-run server for general discussion, questions, and feedback.
 - [Read or post Github Issues](https://github.com/All-Hands-AI/OpenHands/issues) - Check out the issues we're working on, or add your own ideas.
 

containers/dev/compose.yml

@@ -11,7 +11,7 @@ services:
       - BACKEND_HOST=${BACKEND_HOST:-"0.0.0.0"}
       - SANDBOX_API_HOSTNAME=host.docker.internal
       #
-      - SANDBOX_RUNTIME_CONTAINER_IMAGE=${SANDBOX_RUNTIME_CONTAINER_IMAGE:-ghcr.io/all-hands-ai/runtime:0.28-nikolaik}
+      - SANDBOX_RUNTIME_CONTAINER_IMAGE=${SANDBOX_RUNTIME_CONTAINER_IMAGE:-ghcr.io/all-hands-ai/runtime:0.29-nikolaik}
       - SANDBOX_USER_ID=${SANDBOX_USER_ID:-1234}
       - WORKSPACE_MOUNT_PATH=${WORKSPACE_BASE:-$PWD/workspace}
     ports:

docker-compose.yml

@@ -7,7 +7,7 @@ services:
     image: openhands:latest
     container_name: openhands-app-${DATE:-}
     environment:
-      - SANDBOX_RUNTIME_CONTAINER_IMAGE=${SANDBOX_RUNTIME_CONTAINER_IMAGE:-docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik}
+      - SANDBOX_RUNTIME_CONTAINER_IMAGE=${SANDBOX_RUNTIME_CONTAINER_IMAGE:-docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik}
       #- SANDBOX_USER_ID=${SANDBOX_USER_ID:-1234} # enable this only if you want a specific non-root sandbox user but you will have to manually adjust permissions of openhands-state for this user
       - WORKSPACE_MOUNT_PATH=${WORKSPACE_BASE:-$PWD/workspace}
     ports:

docs/docusaurus.config.ts

@@ -75,12 +75,6 @@ const config: Config = {
           position: "left",
           label: "User Guides",
         },
-        {
-          type: "docSidebar",
-          sidebarId: "apiSidebar",
-          position: "left",
-          label: "Python API",
-        },
         {
           type: 'localeDropdown',
           position: 'left',

docs/i18n/fr/docusaurus-plugin-content-docs/current/usage/how-to/cli-mode.md

@@ -52,7 +52,7 @@ LLM_API_KEY="sk_test_12345"
 ```bash
 docker run -it \
     --pull=always \
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -e SANDBOX_USER_ID=$(id -u) \
     -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
     -e LLM_API_KEY=$LLM_API_KEY \
@@ -61,7 +61,7 @@ docker run -it \
     -v /var/run/docker.sock:/var/run/docker.sock \
     --add-host host.docker.internal:host-gateway \
     --name openhands-app-$(date +%Y%m%d%H%M%S) \
-    docker.all-hands.dev/all-hands-ai/openhands:0.28 \
+    docker.all-hands.dev/all-hands-ai/openhands:0.29 \
     python -m openhands.core.cli
 ```
 

docs/i18n/fr/docusaurus-plugin-content-docs/current/usage/how-to/headless-mode.md

@@ -46,7 +46,7 @@ LLM_API_KEY="sk_test_12345"
 ```bash
 docker run -it \
     --pull=always \
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -e SANDBOX_USER_ID=$(id -u) \
     -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
     -e LLM_API_KEY=$LLM_API_KEY \
@@ -56,6 +56,6 @@ docker run -it \
     -v /var/run/docker.sock:/var/run/docker.sock \
     --add-host host.docker.internal:host-gateway \
     --name openhands-app-$(date +%Y%m%d%H%M%S) \
-    docker.all-hands.dev/all-hands-ai/openhands:0.28 \
+    docker.all-hands.dev/all-hands-ai/openhands:0.29 \
     python -m openhands.core.main -t "write a bash script that prints hi" --no-auto-continue
 ```

docs/i18n/fr/docusaurus-plugin-content-docs/current/usage/installation.mdx

@@ -13,16 +13,16 @@
 La façon la plus simple d'exécuter OpenHands est avec Docker.
 
 ```bash
-docker pull docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik
+docker pull docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik
 
 docker run -it --rm --pull=always \
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -e LOG_ALL_EVENTS=true \
     -v /var/run/docker.sock:/var/run/docker.sock \
     -p 3000:3000 \
     --add-host host.docker.internal:host-gateway \
     --name openhands-app \
-    docker.all-hands.dev/all-hands-ai/openhands:0.28
+    docker.all-hands.dev/all-hands-ai/openhands:0.29
 ```
 
 Vous pouvez également exécuter OpenHands en mode [headless scriptable](https://docs.all-hands.dev/modules/usage/how-to/headless-mode), en tant que [CLI interactive](https://docs.all-hands.dev/modules/usage/how-to/cli-mode), ou en utilisant l'[Action GitHub OpenHands](https://docs.all-hands.dev/modules/usage/how-to/github-action).

docs/i18n/fr/docusaurus-plugin-content-docs/current/usage/intro.mdx

@@ -42,7 +42,7 @@ Explorez le code source d'OpenHands sur [GitHub](https://github.com/All-Hands-AI
   />
 </a>
 <br></br>
-<a href="https://join.slack.com/t/openhands-ai/shared_invite/zt-2ypg5jweb-d~6hObZDbXi_HEL8PDrbHg">
+<a href="https://join.slack.com/t/openhands-ai/shared_invite/zt-2ngejmfw6-9gW4APWOC9XUp1n~SiQ6iw">
   <img
     src="https://img.shields.io/badge/Slack-Join%20Us-red?logo=slack&logoColor=white&style=for-the-badge"
     alt="Join our Slack community"

docs/i18n/fr/docusaurus-plugin-content-docs/current/usage/runtimes.md

@@ -13,7 +13,7 @@ C'est le Runtime par défaut qui est utilisé lorsque vous démarrez OpenHands.
 
 ```
 docker run # ...
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -v /var/run/docker.sock:/var/run/docker.sock \
     # ...
 ```

docs/i18n/zh-Hans/docusaurus-plugin-content-docs/current/usage/how-to/cli-mode.md

@@ -50,7 +50,7 @@ LLM_API_KEY="sk_test_12345"
 ```bash
 docker run -it \
     --pull=always \
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -e SANDBOX_USER_ID=$(id -u) \
     -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
     -e LLM_API_KEY=$LLM_API_KEY \
@@ -59,7 +59,7 @@ docker run -it \
     -v /var/run/docker.sock:/var/run/docker.sock \
     --add-host host.docker.internal:host-gateway \
     --name openhands-app-$(date +%Y%m%d%H%M%S) \
-    docker.all-hands.dev/all-hands-ai/openhands:0.28 \
+    docker.all-hands.dev/all-hands-ai/openhands:0.29 \
     python -m openhands.core.cli
 ```
 

docs/i18n/zh-Hans/docusaurus-plugin-content-docs/current/usage/how-to/headless-mode.md

@@ -47,7 +47,7 @@ LLM_API_KEY="sk_test_12345"
 ```bash
 docker run -it \
     --pull=always \
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -e SANDBOX_USER_ID=$(id -u) \
     -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
     -e LLM_API_KEY=$LLM_API_KEY \
@@ -57,6 +57,6 @@ docker run -it \
     -v /var/run/docker.sock:/var/run/docker.sock \
     --add-host host.docker.internal:host-gateway \
     --name openhands-app-$(date +%Y%m%d%H%M%S) \
-    docker.all-hands.dev/all-hands-ai/openhands:0.28 \
+    docker.all-hands.dev/all-hands-ai/openhands:0.29 \
     python -m openhands.core.main -t "write a bash script that prints hi" --no-auto-continue
 ```

docs/i18n/zh-Hans/docusaurus-plugin-content-docs/current/usage/installation.mdx

@@ -11,16 +11,16 @@
 在 Docker 中运行 OpenHands 是最简单的方式。
 
 ```bash
-docker pull docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik
+docker pull docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik
 
 docker run -it --rm --pull=always \
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -e LOG_ALL_EVENTS=true \
     -v /var/run/docker.sock:/var/run/docker.sock \
     -p 3000:3000 \
     --add-host host.docker.internal:host-gateway \
     --name openhands-app \
-    docker.all-hands.dev/all-hands-ai/openhands:0.28
+    docker.all-hands.dev/all-hands-ai/openhands:0.29
 ```
 
 你也可以在可脚本化的[无头模式](https://docs.all-hands.dev/modules/usage/how-to/headless-mode)下运行 OpenHands，作为[交互式 CLI](https://docs.all-hands.dev/modules/usage/how-to/cli-mode)，或使用 [OpenHands GitHub Action](https://docs.all-hands.dev/modules/usage/how-to/github-action)。

docs/i18n/zh-Hans/docusaurus-plugin-content-docs/current/usage/intro.mdx

@@ -42,7 +42,7 @@ OpenHands 是一个**自主 AI 软件工程师**，能够执行复杂的工程
   />
 </a>
 <br></br>
-<a href="https://join.slack.com/t/openhands-ai/shared_invite/zt-2ypg5jweb-d~6hObZDbXi_HEL8PDrbHg">
+<a href="https://join.slack.com/t/openhands-ai/shared_invite/zt-2ngejmfw6-9gW4APWOC9XUp1n~SiQ6iw">
   <img
     src="https://img.shields.io/badge/Slack-Join%20Us-red?logo=slack&logoColor=white&style=for-the-badge"
     alt="Join our Slack community"

docs/i18n/zh-Hans/docusaurus-plugin-content-docs/current/usage/runtimes.md

@@ -11,7 +11,7 @@
 
 ```
 docker run # ...
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -v /var/run/docker.sock:/var/run/docker.sock \
     # ...
 ```

docs/modules/usage/cloud/cloud-github-resolver.md

@@ -5,7 +5,7 @@ The GitHub Resolver automates code fixes and provides intelligent assistance for
 ## Setup
 
 The Cloud GitHub Resolver is available automatically when you
-[grant OpenHands Cloud repository access](./openhands-cloud.md#adding-repositories).
+[grant OpenHands Cloud repository access](./openhands-cloud.md#adding-repository-access).
 
 ## Usage
 

docs/modules/usage/how-to/cli-mode.md

@@ -35,7 +35,7 @@ To run OpenHands in CLI mode with Docker:
 ```bash
 docker run -it \
     --pull=always \
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -e SANDBOX_USER_ID=$(id -u) \
     -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
     -e LLM_API_KEY=$LLM_API_KEY \
@@ -45,7 +45,7 @@ docker run -it \
     -v ~/.openhands-state:/.openhands-state \
     --add-host host.docker.internal:host-gateway \
     --name openhands-app-$(date +%Y%m%d%H%M%S) \
-    docker.all-hands.dev/all-hands-ai/openhands:0.28 \
+    docker.all-hands.dev/all-hands-ai/openhands:0.29 \
     python -m openhands.core.cli
 ```
 

docs/modules/usage/how-to/headless-mode.md

@@ -32,7 +32,7 @@ To run OpenHands in Headless mode with Docker:
 ```bash
 docker run -it \
     --pull=always \
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -e SANDBOX_USER_ID=$(id -u) \
     -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
     -e LLM_API_KEY=$LLM_API_KEY \
@@ -43,7 +43,7 @@ docker run -it \
     -v ~/.openhands-state:/.openhands-state \
     --add-host host.docker.internal:host-gateway \
     --name openhands-app-$(date +%Y%m%d%H%M%S) \
-    docker.all-hands.dev/all-hands-ai/openhands:0.28 \
+    docker.all-hands.dev/all-hands-ai/openhands:0.29 \
     python -m openhands.core.main -t "write a bash script that prints hi"
 ```
 

docs/modules/usage/installation.mdx

@@ -58,17 +58,17 @@ A system with a modern processor and a minimum of **4GB RAM** is recommended to
 The easiest way to run OpenHands is in Docker.
 
 ```bash
-docker pull docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik
+docker pull docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik
 
 docker run -it --rm --pull=always \
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -e LOG_ALL_EVENTS=true \
     -v /var/run/docker.sock:/var/run/docker.sock \
     -v ~/.openhands-state:/.openhands-state \
     -p 3000:3000 \
     --add-host host.docker.internal:host-gateway \
     --name openhands-app \
-    docker.all-hands.dev/all-hands-ai/openhands:0.28
+    docker.all-hands.dev/all-hands-ai/openhands:0.29
 ```
 
 You'll find OpenHands running at http://localhost:3000!

docs/modules/usage/runtimes.md

@@ -10,13 +10,16 @@ We also support "remote" runtimes, which are typically managed by third-parties.
 They can make setup a bit simpler and more scalable, especially
 if you're running many OpenHands conversations in parallel (e.g. to do evaluation).
 
+Additionally, we provide a "local" runtime that runs directly on your machine without Docker,
+which can be useful in controlled environments like CI pipelines.
+
 ## Docker Runtime
 This is the default Runtime that's used when you start OpenHands. You might notice
 some flags being passed to `docker run` that make this possible:
 
 ```
 docker run # ...
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.28-nikolaik \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.29-nikolaik \
     -v /var/run/docker.sock:/var/run/docker.sock \
     # ...
 ```
@@ -56,21 +59,12 @@ any files that are mounted into its workspace.
 This setup can cause some issues with file permissions (hence the `SANDBOX_USER_ID` variable)
 but seems to work well on most systems.
 
-## All Hands Runtime
-The All Hands Runtime is currently in beta. You can request access by joining
-the #remote-runtime-limited-beta channel on Slack ([see the README](https://github.com/All-Hands-AI/OpenHands?tab=readme-ov-file#-how-to-join-the-community) for an invite).
+## OpenHands Remote Runtime
 
-To use the All Hands Runtime, set the following environment variables when
-starting OpenHands:
+OpenHands Remote Runtime is currently in beta (read [here](https://runtime.all-hands.dev/) for more details), it allows you to launch runtimes in parallel in the cloud.
+Fill out [this form](https://docs.google.com/forms/d/e/1FAIpQLSckVz_JFwg2_mOxNZjCtr7aoBFI2Mwdan3f75J_TrdMS1JV2g/viewform) to apply if you want to try this out!
 
-```bash
-docker run # ...
-    -e RUNTIME=remote \
-    -e SANDBOX_REMOTE_RUNTIME_API_URL="https://runtime.app.all-hands.dev" \
-    -e SANDBOX_API_KEY="your-all-hands-api-key" \
-    -e SANDBOX_KEEP_RUNTIME_ALIVE="true" \
-    # ...
-```
+NOTE: This runtime is specifically designed for agent evaluation purposes only through [OpenHands evaluation harness](https://github.com/All-Hands-AI/OpenHands/tree/main/evaluation). It should not be used to launch production OpenHands applications.
 
 ## Modal Runtime
 Our partners at [Modal](https://modal.com/) have also provided a runtime for OpenHands.
@@ -117,3 +111,66 @@ bash -i <(curl -sL https://get.daytona.io/openhands)
 Once executed, OpenHands should be running locally and ready for use.
 
 For more details and manual initialization, view the entire [README.md](https://github.com/All-Hands-AI/OpenHands/blob/main/openhands/runtime/impl/daytona/README.md)
+
+## Local Runtime
+
+The Local Runtime allows the OpenHands agent to execute actions directly on your local machine without using Docker. This runtime is primarily intended for controlled environments like CI pipelines or testing scenarios where Docker is not available.
+
+:::caution
+**Security Warning**: The Local Runtime runs without any sandbox isolation. The agent can directly access and modify files on your machine. Only use this runtime in controlled environments or when you fully understand the security implications.
+:::
+
+### Prerequisites
+
+Before using the Local Runtime, ensure you have the following dependencies installed:
+
+1. You have followed the [Development setup instructions](https://github.com/All-Hands-AI/OpenHands/blob/main/Development.md).
+2. tmux is available on your system.
+
+### Configuration
+
+To use the Local Runtime, besides required configurations like the model, API key, you'll need to set the following options via environment variables or the [config.toml file](https://github.com/All-Hands-AI/OpenHands/blob/main/config.template.toml) when starting OpenHands:
+
+- Via environment variables:
+
+```bash
+# Required
+export RUNTIME=local
+
+# Optional but recommended
+export WORKSPACE_BASE=/path/to/your/workspace
+```
+
+- Via `config.toml`:
+
+```toml
+[core]
+runtime = "local"
+workspace_base = "/path/to/your/workspace"
+```
+
+If `WORKSPACE_BASE` is not set, the runtime will create a temporary directory for the agent to work in.
+
+### Example Usage
+
+Here's an example of how to start OpenHands with the Local Runtime in Headless Mode:
+
+```bash
+# Set the runtime type to local
+export RUNTIME=local
+
+# Optionally set a workspace directory
+export WORKSPACE_BASE=/path/to/your/project
+
+# Start OpenHands
+poetry run python -m openhands.core.main -t "write a bash script that prints hi"
+```
+
+### Use Cases
+
+The Local Runtime is particularly useful for:
+
+- CI/CD pipelines where Docker is not available.
+- Testing and development of OpenHands itself.
+- Environments where container usage is restricted.
+- Scenarios where direct file system access is required.

docs/src/components/CustomFooter.tsx

@@ -8,7 +8,7 @@ function CustomFooter() {
     <footer className="custom-footer">
       <div className="footer-content">
         <div className="footer-icons">
-          <a href="https://join.slack.com/t/openhands-ai/shared_invite/zt-2ypg5jweb-d~6hObZDbXi_HEL8PDrbHg" target="_blank" rel="noopener noreferrer">
+          <a href="https://join.slack.com/t/openhands-ai/shared_invite/zt-2ngejmfw6-9gW4APWOC9XUp1n~SiQ6iw" target="_blank" rel="noopener noreferrer">
             <FaSlack />
           </a>
           <a href="https://discord.gg/ESHStjSjD4" target="_blank" rel="noopener noreferrer">

docs/src/components/HomepageHeader/HomepageHeader.tsx

@@ -46,7 +46,7 @@ export function HomepageHeader() {
           <a href="https://codecov.io/github/All-Hands-AI/OpenHands?branch=main"><img alt="CodeCov" src="https://img.shields.io/codecov/c/github/All-Hands-AI/OpenHands?style=for-the-badge&color=blue" /></a>
           <a href="https://github.com/All-Hands-AI/OpenHands/blob/main/LICENSE"><img src="https://img.shields.io/github/license/All-Hands-AI/OpenHands?style=for-the-badge&color=blue" alt="MIT License" /></a>
           <br/>
-          <a href="https://join.slack.com/t/openhands-ai/shared_invite/zt-2ypg5jweb-d~6hObZDbXi_HEL8PDrbHg"><img src="https://img.shields.io/badge/Slack-Join%20Us-red?logo=slack&logoColor=white&style=for-the-badge" alt="Join our Slack community" /></a>
+          <a href="https://join.slack.com/t/openhands-ai/shared_invite/zt-2ngejmfw6-9gW4APWOC9XUp1n~SiQ6iw"><img src="https://img.shields.io/badge/Slack-Join%20Us-red?logo=slack&logoColor=white&style=for-the-badge" alt="Join our Slack community" /></a>
           <a href="https://discord.gg/ESHStjSjD4"><img src="https://img.shields.io/badge/Discord-Join%20Us-purple?logo=discord&logoColor=white&style=for-the-badge" alt="Join our Discord community" /></a>
           <a href="https://github.com/All-Hands-AI/OpenHands/blob/main/CREDITS.md"><img src="https://img.shields.io/badge/Project-Credits-blue?style=for-the-badge&color=FFE165&logo=github&logoColor=white" alt="Credits" /></a>
           <br/>

evaluation/benchmarks/swe_bench/README.md

@@ -18,6 +18,20 @@ Please follow instruction [here](../../README.md#setup) to setup your local deve
 
 ## Run Inference (Rollout) on SWE-Bench Instances: Generate Patch from Problem Statement
 
+> [!NOTE]
+> **Iterative Evaluation Protocol**
+>
+> We have an iterative approach for more stable and reproducible results:
+> - For each instance, we attempt to generate a solution up to 3 times
+> - Each attempt continues until either:
+>   1. The agent successfully produces a patch with `AgentFinishAction`, or
+>   2. The attempt reaches the maximum iteration limit
+> - If an attempt fails, we retry with a fresh attempt (up to the 3-attempt maximum)
+> - If your LLM config has temperature=0, we will automatically use temperature=0.1 for the 2nd and 3rd attempts
+>
+> To enable this iterative protocol, set `export ITERATIVE_EVAL_MODE=true`
+
+
 ### Running Locally with Docker
 
 Make sure your Docker daemon is running, and you have ample disk space (at least 200-500GB, depends on the SWE-Bench set you are running on) for the instance-level docker image.
@@ -45,7 +59,7 @@ to `CodeActAgent`.
 default, the script evaluates the entire SWE-bench_Lite test set (300 issues). Note:
 in order to use `eval_limit`, you must also set `agent`.
 - `max_iter`, e.g. `20`, is the maximum number of iterations for the agent to run. By
-default, it is set to 30.
+default, it is set to 60.
 - `num_workers`, e.g. `3`, is the number of parallel workers to run the evaluation. By
 default, it is set to 1.
 - `dataset`, a huggingface dataset name. e.g. `princeton-nlp/SWE-bench`, `princeton-nlp/SWE-bench_Lite`, or `princeton-nlp/SWE-bench_Verified`, specifies which dataset to evaluate on.

evaluation/benchmarks/swe_bench/eval_infer.py

@@ -1,9 +1,12 @@
+import copy
 import json
 import os
 import subprocess
 import tempfile
 import time
+from dataclasses import dataclass
 from functools import partial
+from typing import Callable
 
 import pandas as pd
 from tqdm import tqdm
@@ -91,12 +94,22 @@ def get_config(metadata: EvalMetadata, instance: pd.Series) -> AppConfig:
     return config
 
 
+@dataclass
+class ConditionalImports:
+    """We instantiate the values in this dataclass differently if we're evaluating SWE-bench or SWE-Gym."""
+
+    get_eval_report: Callable
+    APPLY_PATCH_FAIL: str
+    APPLY_PATCH_PASS: str
+
+
 def process_instance(
     instance: pd.Series,
     metadata: EvalMetadata,
     reset_logger: bool = True,
     log_dir: str | None = None,
     runtime_failure_count: int = 0,
+    conditional_imports: ConditionalImports | None = None,
 ) -> EvalOutput:
     """
     Evaluate agent performance on a SWE-bench problem instance.
@@ -108,9 +121,18 @@ def process_instance(
         log_dir (str | None, default=None): Path to directory where log files will be written. Must
         be provided if `reset_logger` is set.
 
+        conditional_imports: A dataclass containing values that are imported differently based on
+        whether we're evaluating SWE-bench or SWE-Gym.
+
     Raises:
         AssertionError: if the `reset_logger` flag is set without a provided log directory.
+
+        AssertionError: if `conditional_imports` is not provided.
     """
+    assert (
+        conditional_imports is not None
+    ), 'conditional_imports must be provided to run process_instance using multiprocessing'
+
     # Setup the logger properly, so you can run multi-processing to parallelize the evaluation
     if reset_logger:
         assert (
@@ -124,7 +146,7 @@ def process_instance(
     config = get_config(metadata, instance)
     instance_id = instance.instance_id
     model_patch = instance['model_patch']
-    test_spec: TestSpec = instance['test_spec']
+    test_spec = instance['test_spec']
     logger.info(f'Starting evaluation for instance {instance_id}.')
 
     if 'test_result' not in instance.keys():
@@ -154,6 +176,11 @@ def process_instance(
         logger.warning(
             f'This is the {runtime_failure_count + 1}th attempt for instance {instance.instance_id}, setting resource factor to {config.sandbox.remote_runtime_resource_factor}'
         )
+    metadata = copy.deepcopy(metadata)
+    metadata.details['runtime_failure_count'] = runtime_failure_count
+    metadata.details['remote_runtime_resource_factor'] = (
+        config.sandbox.remote_runtime_resource_factor
+    )
 
     try:
         runtime = create_runtime(config)
@@ -196,7 +223,9 @@ def process_instance(
         instance['test_result']['apply_patch_output'] = apply_patch_output
 
         if 'APPLY_PATCH_FAIL' in apply_patch_output:
-            logger.info(f'[{instance_id}] {APPLY_PATCH_FAIL}:\n{apply_patch_output}')
+            logger.info(
+                f'[{instance_id}] {conditional_imports.APPLY_PATCH_FAIL}:\n{apply_patch_output}'
+            )
             instance['test_result']['report']['failed_apply_patch'] = True
 
             return EvalOutput(
@@ -205,7 +234,9 @@ def process_instance(
                 metadata=metadata,
             )
         elif 'APPLY_PATCH_PASS' in apply_patch_output:
-            logger.info(f'[{instance_id}] {APPLY_PATCH_PASS}:\n{apply_patch_output}')
+            logger.info(
+                f'[{instance_id}] {conditional_imports.APPLY_PATCH_PASS}:\n{apply_patch_output}'
+            )
 
             # Run eval script in background and save output to log file
             log_file = '/tmp/eval_output.log'
@@ -271,14 +302,20 @@ def process_instance(
                         with open(test_output_path, 'w') as f:
                             f.write(test_output)
                         try:
-                            _report = get_eval_report(
+                            extra_kwargs = {}
+                            if 'SWE-Gym' in metadata.dataset:
+                                # SWE-Gym uses a different version of the package, hence a different eval report argument
+                                extra_kwargs['log_path'] = test_output_path
+                            else:
+                                extra_kwargs['test_log_path'] = test_output_path
+                            _report = conditional_imports.get_eval_report(
                                 test_spec=test_spec,
                                 prediction={
                                     'model_patch': model_patch,
                                     'instance_id': instance_id,
                                 },
-                                test_log_path=test_output_path,
                                 include_tests_status=True,
+                                **extra_kwargs,
                             )
                             report = _report[instance_id]
                             logger.info(
@@ -345,7 +382,6 @@ if __name__ == '__main__':
         )
         from swegym.harness.test_spec import (
             SWEbenchInstance,
-            TestSpec,
             make_test_spec,
         )
         from swegym.harness.utils import load_swebench_dataset
@@ -357,7 +393,6 @@ if __name__ == '__main__':
         )
         from swebench.harness.test_spec.test_spec import (
             SWEbenchInstance,
-            TestSpec,
             make_test_spec,
         )
         from swebench.harness.utils import load_swebench_dataset
@@ -440,12 +475,21 @@ if __name__ == '__main__':
             .decode('utf-8')
             .strip(),  # Current commit
             dataset=args.dataset,  # Dataset name from args
+            details={},
         )
 
     # The evaluation harness constrains the signature of `process_instance_func` but we need to
     # pass extra information. Build a new function object to avoid issues with multiprocessing.
     process_instance_func = partial(
-        process_instance, log_dir=output_file.replace('.jsonl', '.logs')
+        process_instance,
+        log_dir=output_file.replace('.jsonl', '.logs'),
+        # We have to explicitly pass these imports to the process_instance function, otherwise
+        # they won't be available in the multiprocessing context.
+        conditional_imports=ConditionalImports(
+            get_eval_report=get_eval_report,
+            APPLY_PATCH_FAIL=APPLY_PATCH_FAIL,
+            APPLY_PATCH_PASS=APPLY_PATCH_PASS,
+        ),
     )
 
     run_evaluation(

evaluation/benchmarks/swe_bench/resource/mapping.py

@@ -23,7 +23,7 @@ def get_resource_mapping(dataset_name: str) -> dict[str, float]:
     if dataset_name not in _global_resource_mapping:
         file_path = os.path.join(CUR_DIR, f'{dataset_name}.json')
         if not os.path.exists(file_path):
-            logger.warning(f'Resource mapping for {dataset_name} not found.')
+            logger.info(f'Resource mapping for {dataset_name} not found.')
             return None
 
         with open(file_path, 'r') as f:

evaluation/benchmarks/swe_bench/run_infer.py

@@ -1,4 +1,5 @@
 import asyncio
+import copy
 import json
 import os
 import tempfile
@@ -36,9 +37,10 @@ from openhands.core.config import (
 )
 from openhands.core.logger import openhands_logger as logger
 from openhands.core.main import create_runtime, run_controller
+from openhands.critic import AgentFinishedCritic
 from openhands.events.action import CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation, ErrorObservation
-from openhands.events.serialization.event import event_to_dict
+from openhands.events.serialization.event import event_from_dict, event_to_dict
 from openhands.runtime.base import Runtime
 from openhands.utils.async_utils import call_async_from_sync
 from openhands.utils.shutdown_listener import sleep_if_should_continue
@@ -121,7 +123,9 @@ You SHOULD NEVER attempt to browse the web.
 
 
 # TODO: migrate all swe-bench docker to ghcr.io/openhands
-DEFAULT_DOCKER_IMAGE_PREFIX = os.environ.get('EVAL_DOCKER_IMAGE_PREFIX', 'docker.io/xingyaoww/')
+DEFAULT_DOCKER_IMAGE_PREFIX = os.environ.get(
+    'EVAL_DOCKER_IMAGE_PREFIX', 'docker.io/xingyaoww/'
+)
 logger.info(f'Default docker image prefix: {DEFAULT_DOCKER_IMAGE_PREFIX}')
 
 
@@ -149,7 +153,8 @@ def get_config(
 ) -> AppConfig:
     # We use a different instance image for the each instance of swe-bench eval
     use_official_image = bool(
-        'verified' in metadata.dataset.lower() or 'lite' in metadata.dataset.lower()
+        ('verified' in metadata.dataset.lower() or 'lite' in metadata.dataset.lower())
+        and 'swe-gym' not in metadata.dataset.lower()
     )
     base_container_image = get_instance_docker_image(
         instance['instance_id'], use_official_image
@@ -475,6 +480,13 @@ def process_instance(
         logger.warning(
             f'This is the {runtime_failure_count + 1}th attempt for instance {instance.instance_id}, setting resource factor to {config.sandbox.remote_runtime_resource_factor}'
         )
+
+    metadata = copy.deepcopy(metadata)
+    metadata.details['runtime_failure_count'] = runtime_failure_count
+    metadata.details['remote_runtime_resource_factor'] = (
+        config.sandbox.remote_runtime_resource_factor
+    )
+
     runtime = create_runtime(config)
     call_async_from_sync(runtime.connect)
 
@@ -560,20 +572,6 @@ def filter_dataset(dataset: pd.DataFrame, filter_column: str) -> pd.DataFrame:
     return dataset
 
 
-# A list of instances that are known to be tricky to infer
-# (will cause runtime failure even with resource factor = 8)
-SWEGYM_EXCLUDE_IDS = [
-    'dask__dask-10422',
-    'pandas-dev__pandas-50548',
-    'pandas-dev__pandas-53672',
-    'pandas-dev__pandas-54174',
-    'pandas-dev__pandas-55518',
-    'pandas-dev__pandas-58383',
-    'pydata__xarray-6721',
-    'pytest-dev__pytest-10081',
-    'pytest-dev__pytest-7236',
-]
-
 if __name__ == '__main__':
     parser = get_parser()
     parser.add_argument(
@@ -598,11 +596,20 @@ if __name__ == '__main__':
         f'Loaded dataset {args.dataset} with split {args.split}: {len(swe_bench_tests)} tasks'
     )
     if 'SWE-Gym' in args.dataset:
-        swe_bench_tests = swe_bench_tests[
-            ~swe_bench_tests['instance_id'].isin(SWEGYM_EXCLUDE_IDS)
-        ]
+        with open(
+            os.path.join(
+                os.path.dirname(os.path.abspath(__file__)),
+                'split',
+                'swegym_verified_instances.json',
+            ),
+            'r',
+        ) as f:
+            swegym_verified_instances = json.load(f)
+            swe_bench_tests = swe_bench_tests[
+                swe_bench_tests['instance_id'].isin(swegym_verified_instances)
+            ]
         logger.info(
-            f'{len(swe_bench_tests)} tasks left after excluding SWE-Gym excluded tasks'
+            f'{len(swe_bench_tests)} tasks left after filtering for SWE-Gym verified instances'
         )
 
     llm_config = None
@@ -633,20 +640,132 @@ if __name__ == '__main__':
 
     output_file = os.path.join(metadata.eval_output_dir, 'output.jsonl')
     print(f'### OUTPUT FILE: {output_file} ###')
-    instances = prepare_dataset(swe_bench_tests, output_file, args.eval_n_limit)
 
-    if len(instances) > 0 and not isinstance(
-        instances['PASS_TO_PASS'][instances['PASS_TO_PASS'].index[0]], str
-    ):
-        for col in ['PASS_TO_PASS', 'FAIL_TO_PASS']:
-            instances[col] = instances[col].apply(lambda x: str(x))
-
-    run_evaluation(
-        instances,
-        metadata,
-        output_file,
-        args.eval_num_workers,
-        process_instance,
-        timeout_seconds=8 * 60 * 60,  # 8 hour PER instance should be more than enough
-        max_retries=5,
+    # Run evaluation in iterative mode:
+    # If a rollout fails to output AgentFinishAction, we will try again until it succeeds OR total 3 attempts have been made.
+    ITERATIVE_EVAL_MODE = (
+        os.environ.get('ITERATIVE_EVAL_MODE', 'false').lower() == 'true'
     )
+    ITERATIVE_EVAL_MODE_MAX_ATTEMPTS = int(
+        os.environ.get('ITERATIVE_EVAL_MODE_MAX_ATTEMPTS', '3')
+    )
+
+    if not ITERATIVE_EVAL_MODE:
+        # load the dataset
+        instances = prepare_dataset(swe_bench_tests, output_file, args.eval_n_limit)
+        if len(instances) > 0 and not isinstance(
+            instances['PASS_TO_PASS'][instances['PASS_TO_PASS'].index[0]], str
+        ):
+            for col in ['PASS_TO_PASS', 'FAIL_TO_PASS']:
+                instances[col] = instances[col].apply(lambda x: str(x))
+
+        run_evaluation(
+            instances,
+            metadata,
+            output_file,
+            args.eval_num_workers,
+            process_instance,
+            timeout_seconds=8
+            * 60
+            * 60,  # 8 hour PER instance should be more than enough
+            max_retries=5,
+        )
+    else:
+        critic = AgentFinishedCritic()
+
+        def get_cur_output_file_path(attempt: int) -> str:
+            return (
+                f'{output_file.removesuffix(".jsonl")}.critic_attempt_{attempt}.jsonl'
+            )
+
+        eval_ids = None
+        for attempt in range(1, ITERATIVE_EVAL_MODE_MAX_ATTEMPTS + 1):
+            cur_output_file = get_cur_output_file_path(attempt)
+            logger.info(
+                f'Running evaluation with critic {critic.__class__.__name__} for attempt {attempt} of {ITERATIVE_EVAL_MODE_MAX_ATTEMPTS}.'
+            )
+
+            # For deterministic eval, we set temperature to 0.1 for (>1) attempt
+            # so hopefully we get slightly different results
+            if attempt > 1 and metadata.llm_config.temperature == 0:
+                logger.info(
+                    f'Detected temperature is 0 for (>1) attempt {attempt}. Setting temperature to 0.1...'
+                )
+                metadata.llm_config.temperature = 0.1
+
+            # Load instances - at first attempt, we evaluate all instances
+            # On subsequent attempts, we only evaluate the instances that failed the previous attempt determined by critic
+            instances = prepare_dataset(
+                swe_bench_tests, cur_output_file, args.eval_n_limit, eval_ids=eval_ids
+            )
+            if len(instances) > 0 and not isinstance(
+                instances['PASS_TO_PASS'][instances['PASS_TO_PASS'].index[0]], str
+            ):
+                for col in ['PASS_TO_PASS', 'FAIL_TO_PASS']:
+                    instances[col] = instances[col].apply(lambda x: str(x))
+
+            # Run evaluation - but save them to cur_output_file
+            logger.info(
+                f'Evaluating {len(instances)} instances for attempt {attempt}...'
+            )
+            run_evaluation(
+                instances,
+                metadata,
+                cur_output_file,
+                args.eval_num_workers,
+                process_instance,
+                timeout_seconds=8
+                * 60
+                * 60,  # 8 hour PER instance should be more than enough
+                max_retries=5,
+            )
+
+            # When eval is done, we update eval_ids to the instances that failed the current attempt
+            instances_failed = []
+            logger.info(
+                f'Use critic {critic.__class__.__name__} to check {len(instances)} instances for attempt {attempt}...'
+            )
+            with open(cur_output_file, 'r') as f:
+                for line in f:
+                    instance = json.loads(line)
+                    history = [event_from_dict(event) for event in instance['history']]
+                    critic_result = critic.evaluate(history)
+                    if not critic_result.success:
+                        instances_failed.append(instance['instance_id'])
+            logger.info(
+                f'{len(instances_failed)} instances failed the current attempt {attempt}: {instances_failed}'
+            )
+            eval_ids = instances_failed
+
+            # If no instances failed, we break
+            if len(instances_failed) == 0:
+                break
+
+        # Then we should aggregate the results from all attempts into the original output file
+        # and remove the intermediate files
+        logger.info(
+            'Aggregating results from all attempts into the original output file...'
+        )
+        fout = open(output_file, 'w')
+        added_instance_ids = set()
+        for attempt in reversed(range(1, ITERATIVE_EVAL_MODE_MAX_ATTEMPTS + 1)):
+            cur_output_file = get_cur_output_file_path(attempt)
+            if not os.path.exists(cur_output_file):
+                logger.warning(
+                    f'Intermediate output file {cur_output_file} does not exist. Skipping...'
+                )
+                continue
+
+            with open(cur_output_file, 'r') as f:
+                for line in f:
+                    instance = json.loads(line)
+                    if instance['instance_id'] not in added_instance_ids:
+                        fout.write(line)
+                        added_instance_ids.add(instance['instance_id'])
+            logger.info(
+                f'Aggregated instances from {cur_output_file}. Total instances added so far: {len(added_instance_ids)}'
+            )
+        fout.close()
+        logger.info(
+            f'Done! Total {len(added_instance_ids)} instances added to {output_file}'
+        )

evaluation/benchmarks/swe_bench/scripts/eval/download_gold_patch.py

@@ -9,7 +9,7 @@ parser.add_argument(
     '--dataset_name',
     type=str,
     help='Name of the dataset to download',
-    default='princeton-nlp/SWE-bench_Lite',
+    default='princeton-nlp/SWE-bench_Verified',
 )
 parser.add_argument('--split', type=str, help='Split to download', default='test')
 args = parser.parse_args()
@@ -20,7 +20,12 @@ print(
     f'Downloading gold patches from {args.dataset_name} (split: {args.split}) to {output_filepath}'
 )
 patches = [
-    {'instance_id': row['instance_id'], 'model_patch': row['patch']} for row in dataset
+    {
+        'instance_id': row['instance_id'],
+        'model_patch': row['patch'],
+        'model_name_or_path': 'gold',
+    }
+    for row in dataset
 ]
 print(f'{len(patches)} gold patches loaded')
 pd.DataFrame(patches).to_json(output_filepath, lines=True, orient='records')

evaluation/benchmarks/swe_bench/scripts/run_infer.sh

@@ -25,8 +25,8 @@ if [ -z "$AGENT" ]; then
 fi
 
 if [ -z "$MAX_ITER" ]; then
-  echo "MAX_ITER not specified, use default 100"
-  MAX_ITER=100
+  echo "MAX_ITER not specified, use default 60"
+  MAX_ITER=60
 fi
 
 if [ -z "$RUN_WITH_BROWSING" ]; then

evaluation/benchmarks/testgeneval/NOTES.md

@@ -0,0 +1,12 @@
+codamosa_ids = ['pydata__xarray-4750-16496', 'pydata__xarray-3239-16458', 'pydata__xarray-4966-16515', 'pydata__xarray-3302-16459', 'pydata__xarray-5126-16518', 'pydata__xarray-4994-16516', 'pydata__xarray-3905-16478', 'pydata__xarray-4182-16484', 'pydata__xarray-5131-16520', 'pydata__xarray-5662-16532', 'pydata__xarray-3364-16461', 'pydata__xarray-5731-16534', 'pydata__xarray-3239-16457', 'pydata__xarray-7203-16577', 'pydata__xarray-3156-16454', 'pydata__xarray-5126-16519', 'pydata__xarray-5365-16529', 'pydata__xarray-4629-16492', 'pydata__xarray-4248-16486', 'pydata__xarray-4339-16487', 'pydata__xarray-3151-16453', 'pydata__xarray-3114-16452', 'pydata__xarray-5033-16517', 'pydata__xarray-4802-16505', 'pydata__xarray-5455-16530', 'pydata__xarray-6400-16539', 'pydata__xarray-3239-16456', 'pydata__xarray-4419-16488']
+
+pynguin_ids = ['pydata__xarray-6548-16541', 'pydata__xarray-7003-16557', 'pydata__xarray-3114-16452', 'pydata__xarray-4339-16487', 'pydata__xarray-6889-16549', 'pydata__xarray-3239-16458', 'pydata__xarray-3364-16461', 'pydata__xarray-3239-16457', 'pydata__xarray-5365-16529', 'pydata__xarray-5131-16520', 'pydata__xarray-7229-16578', 'pydata__xarray-6461-16540', 'pydata__xarray-4419-16488', 'pydata__xarray-7147-16571', 'pydata__xarray-3151-16453', 'pydata__xarray-4966-16515', 'pydata__xarray-4629-16492', 'pydata__xarray-3239-16456', 'pydata__xarray-7400-16582', 'pydata__xarray-4994-16516', 'pydata__xarray-3302-16459', 'pydata__xarray-6601-16544', 'pydata__xarray-6882-16548', 'pydata__xarray-6135-16535', 'pydata__xarray-7393-16581', 'pydata__xarray-5731-16534', 'pydata__xarray-7203-16577']
+
+ids = ['pydata__xarray-3114-16452', 'pydata__xarray-3151-16453', 'pydata__xarray-3156-16454', 'pydata__xarray-3239-16456', 'pydata__xarray-3239-16457', 'pydata__xarray-3239-16458', 'pydata__xarray-3302-16459', 'pydata__xarray-3364-16461', 'pydata__xarray-3677-16471', 'pydata__xarray-3905-16478', 'pydata__xarray-4182-16484', 'pydata__xarray-4248-16486', 'pydata__xarray-4339-16487', 'pydata__xarray-4419-16488', 'pydata__xarray-4629-16492', 'pydata__xarray-4750-16496', 'pydata__xarray-4802-16505', 'pydata__xarray-4966-16515', 'pydata__xarray-4994-16516', 'pydata__xarray-5033-16517', 'pydata__xarray-5126-16518', 'pydata__xarray-5126-16519', 'pydata__xarray-5131-16520', 'pydata__xarray-5365-16529', 'pydata__xarray-5455-16530', 'pydata__xarray-5662-16532', 'pydata__xarray-5731-16534', 'pydata__xarray-6135-16535', 'pydata__xarray-6135-16536', 'pydata__xarray-6386-16537', 'pydata__xarray-6394-16538', 'pydata__xarray-6400-16539', 'pydata__xarray-6461-16540', 'pydata__xarray-6548-16541', 'pydata__xarray-6599-16543', 'pydata__xarray-6601-16544', 'pydata__xarray-6882-16548', 'pydata__xarray-6889-16549', 'pydata__xarray-7003-16557', 'pydata__xarray-7147-16571', 'pydata__xarray-7150-16572', 'pydata__xarray-7203-16577', 'pydata__xarray-7229-16578', 'pydata__xarray-7393-16581', 'pydata__xarray-7400-16582']
+
+
+Command eval (our approach):
+poetry run ./evaluation/benchmarks/testgeneval/scripts/eval_infer_remote.sh evaluation/evaluation_outputs/outputs/kjain14__testgeneval-test/CodeActAgent/gpt-4o_maxiter_25_N_v0.20.0-no-hint-run_1/output.jsonl 10 kjain14/testgeneval test true
+
+Command run (our approach):
+./evaluation/benchmarks/testgeneval/scripts/run_infer.sh llm.eval_gpt HEAD CodeActAgent -1 25 10 kjain14/testgeneval test 1 ../TestGenEval/results/testgeneval/preds/gpt-4o-2024-08-06__testgeneval__0.2__test.jsonl

evaluation/benchmarks/testgeneval/README.md

@@ -0,0 +1,80 @@
+# TestGenEval Benchmark Evaluation
+
+This folder contains the evaluation harness for the TestGenEval benchmark, which is based on the original TestGenEval benchmark ([paper](https://arxiv.org/abs/2410.00752)). TestGenEval is designed to evaluate the ability of language models to generate unit tests for given Python functions.
+
+## Setup Environment and LLM Configuration
+
+1. Follow the instructions [here](../../README.md#setup) to set up your local development environment and configure your LLM.
+
+2. Install the TestGenEval dependencies:
+```bash
+poetry install --with testgeneval
+```
+
+## Run Inference
+
+To generate tests using your model, run the following command:
+
+```bash
+./evaluation/benchmarks/testgeneval/scripts/run_infer.sh [model_config] [git-version] [agent] [eval_limit] [max_iter] [num_workers] [dataset] [dataset_split]
+
+# Example
+./evaluation/benchmarks/testgeneval/scripts/run_infer.sh llm.eval_gpt4_1106_preview HEAD CodeActAgent 100 30 1 kjain14/testgenevallite test
+```
+
+Parameters:
+- `model_config`: The config group name for your LLM settings (e.g., `eval_gpt4_1106_preview`)
+- `git-version`: The git commit hash or release tag of OpenHands to evaluate (e.g., `HEAD` or `0.6.2`)
+- `agent`: The name of the agent for benchmarks (default: `CodeActAgent`)
+- `eval_limit`: Limit the evaluation to the first N instances (optional)
+- `max_iter`: Maximum number of iterations for the agent to run (default: 30)
+- `num_workers`: Number of parallel workers for evaluation (default: 1)
+- `dataset`: HuggingFace dataset name (default: `kjain14/testgenevallite`)
+- `dataset_split`: Dataset split to use (default: `test`)
+
+After running the inference, you will obtain an `output.jsonl` file (by default saved to `evaluation/evaluation_outputs`).
+
+## Evaluate Generated Tests
+
+To evaluate the generated tests, use the `eval_infer.sh` script:
+
+```bash
+./evaluation/benchmarks/testgeneval/scripts/eval_infer.sh $YOUR_OUTPUT_JSONL [instance_id] [dataset_name] [split] [num_workers] [skip_mutation]
+
+# Example
+./evaluation/benchmarks/testgeneval/scripts/eval_infer.sh evaluation/evaluation_outputs/outputs/kjain14__testgenevallite-test/CodeActAgent/gpt-4-1106-preview_maxiter_50_N_v1.0/output.jsonl
+```
+
+Optional arguments:
+- `instance_id`: Evaluate a single instance (optional)
+- `dataset_name`: Name of the dataset to use (default: `kjain14/testgenevallite`)
+- `split`: Dataset split to use (default: `test`)
+- `num_workers`: Number of workers for running docker (default: 1)
+- `skip_mutation`: Skip mutation testing (enter `true` if desired)
+
+The evaluation results will be saved to `evaluation/evaluation_outputs/outputs/kjain14__testgenevallite-test/CodeActAgent/gpt-4-1106-preview_maxiter_50_N_v1.0/` with `output.testgeneval.jsonl` containing the metrics.
+
+## Metrics
+
+The TestGenEval benchmark evaluates generated tests based on the following metrics:
+
+1. Correctness: Measures if the generated tests are syntactically correct and run without errors.
+2. Coverage: Assesses the code coverage achieved by the generated tests.
+3. Mutation Score: Evaluates the effectiveness of the tests in detecting intentionally introduced bugs (mutations).
+4. Readability: Analyzes the readability of the generated tests using various metrics.
+
+## Submit Your Evaluation Results
+
+To contribute your evaluation results:
+
+1. Fork [our HuggingFace evaluation outputs](https://huggingface.co/spaces/OpenHands/evaluation).
+2. Add your results to the forked repository.
+3. Submit a Pull Request with your evaluation results following the guide [here](https://huggingface.co/docs/hub/en/repositories-pull-requests-discussions#pull-requests-and-discussions).
+
+## Additional Resources
+
+- [TestGenEval Paper](https://arxiv.org/abs/2410.00752)
+- [OpenHands Documentation](https://github.com/All-Hands-AI/OpenHands)
+- [HuggingFace Datasets](https://huggingface.co/datasets)
+
+For any questions or issues, please open an issue in the [OpenHands repository](https://github.com/All-Hands-AI/OpenHands/issues).

evaluation/benchmarks/testgeneval/compute_readability.py

@@ -0,0 +1,347 @@
+import math
+
+
+def total_byte_entropy_stats(python_code):
+    # Count the occurrence of each byte (character for simplicity)
+    byte_counts = {}
+    for byte in python_code.encode('utf-8'):
+        byte_counts[byte] = byte_counts.get(byte, 0) + 1
+
+    total_bytes = sum(byte_counts.values())
+    entropy = -sum(
+        (count / total_bytes) * math.log2(count / total_bytes)
+        for count in byte_counts.values()
+    )
+
+    return {'total_byte_entropy': entropy}
+
+
+def average_nulls_stats(tree, num_lines):
+    total_nulls = 0
+    nulls_per_line = {}  # Dictionary to count nulls per line
+
+    def traverse(node):
+        nonlocal total_nulls
+        if node.type == 'null_literal':
+            total_nulls += 1
+            line_number = node.start_point[0]  # Get line number
+            if line_number in nulls_per_line:
+                nulls_per_line[line_number] += 1
+            else:
+                nulls_per_line[line_number] = 1
+        for child in node.children:
+            traverse(child)
+
+    traverse(tree.root_node)
+
+    # Calculate average nulls per line
+    avg_nulls = total_nulls / num_lines if num_lines > 0 else 0
+
+    # Calculate max nulls on any line
+    max_nulls_on_any_line = max(nulls_per_line.values()) if nulls_per_line else 0
+
+    return {
+        'avg_nulls': avg_nulls,
+        'total_nulls': total_nulls,
+        'max_nulls': max_nulls_on_any_line,
+        'has_nulls': 1 if total_nulls > 0 else 0,
+    }
+
+
+def arithmetic_operations_stats(tree, num_lines):
+    # Dictionary to hold counts of each arithmetic operation
+    op_counts = {'+': 0, '-': 0, '*': 0, '/': 0, '%': 0}
+    total_ops = 0
+
+    # Function to traverse the AST and update operation counts
+    def traverse(node):
+        nonlocal total_ops
+        if node.type == 'binary_expression' or node.type == 'update_expression':
+            for child in node.children:
+                if child.type == 'operator':
+                    op = child.text.decode('utf8')
+                    if op in op_counts:
+                        op_counts[op] += 1
+                        total_ops += 1
+        else:
+            for child in node.children:
+                traverse(child)
+
+    traverse(tree.root_node)
+
+    return {
+        'total_arithmetic_operations': total_ops,
+        'avg_arithmetic_operations': total_ops / num_lines,
+    }
+
+
+def numbers_floats_stats(tree, num_lines):
+    total_numbers = 0
+    total_floats = 0
+
+    def traverse(node):
+        nonlocal total_numbers, total_floats
+        if node.type in ['integer_literal', 'decimal_literal']:
+            total_numbers += 1
+            if (
+                '.' in node.text.decode('utf8')
+                or 'e' in node.text.decode('utf8').lower()
+            ):
+                total_floats += 1
+        for child in node.children:
+            traverse(child)
+
+    traverse(tree.root_node)
+    return {'total_numbers': total_numbers, 'total_floats': total_floats}
+
+
+def code_stats(python_code):
+    lines = python_code.strip().split('\n')
+    total_line_length = sum(len(line) for line in lines)
+    max_line_length = max(len(line) for line in lines)
+    return {
+        'total_line_length': total_line_length,
+        'max_line_length': max_line_length,
+        'avg_characters': total_line_length / len(lines),
+    }
+
+
+def assertions_stats(tree, num_lines):
+    total_assertions = 0
+
+    def traverse(node):
+        nonlocal total_assertions
+        if node.type == 'assert_statement':
+            total_assertions += 1
+        for child in node.children:
+            traverse(child)
+
+    traverse(tree.root_node)
+    return {
+        'total_assertions': total_assertions,
+        'total_has_assertions': 1 if total_assertions > 0 else 0,
+    }
+
+
+def class_instances_stats(tree, num_lines):
+    total_class_instances = 0
+
+    def traverse(node):
+        nonlocal total_class_instances
+        if node.type == 'object_creation_expression':
+            total_class_instances += 1
+        for child in node.children:
+            traverse(child)
+
+    traverse(tree.root_node)
+    return {'total_class_instances': total_class_instances}
+
+
+def has_execeptions(tree, num_lines):
+    total_has_exceptions = 0
+
+    def traverse(node):
+        nonlocal total_has_exceptions
+        if node.type == 'try_statement':
+            total_has_exceptions += 1
+        for child in node.children:
+            traverse(child)
+
+    traverse(tree.root_node)
+    return {'total_has_exceptions': 1 if total_has_exceptions > 0 else 0}
+
+
+def distinct_methods_stats(tree, num_lines):
+    method_names = set()
+    total_nodes = 0
+
+    def traverse(node):
+        nonlocal total_nodes
+        if node.type == 'method_declaration':
+            for child in node.children:
+                if child.type == 'identifier':
+                    method_names.add(child.text.decode('utf8'))
+                    break
+        total_nodes += 1
+        for child in node.children:
+            traverse(child)
+
+    traverse(tree.root_node)
+    total_distinct_methods = len(method_names)
+    total_method_ratio = (
+        total_distinct_methods / (total_nodes - total_distinct_methods)
+        if total_nodes > total_distinct_methods
+        else 0
+    )
+
+    return {
+        'total_distinct_methods': total_distinct_methods,
+        'total_method_ratio': total_method_ratio,
+    }
+
+
+def loops_stats(tree, num_lines):
+    """
+    Calculate the average number of loops.
+    """
+    total_loops = 0
+
+    def traverse(node):
+        nonlocal total_loops
+        if node.type in ['for_statement', 'while_statement', 'do_statement']:
+            total_loops += 1
+        for child in node.children:
+            traverse(child)
+
+    traverse(tree.root_node)
+    avg_loops = total_loops / num_lines
+    return {'avg_loops': avg_loops}
+
+
+def branches_stats(tree, num_lines):
+    """
+    Calculate the average number of branches (conditional statements).
+    """
+    total_branches = 0
+
+    def traverse(node):
+        nonlocal total_branches
+        if node.type in ['if_statement', 'switch_statement']:
+            total_branches += 1
+        for child in node.children:
+            traverse(child)
+
+    traverse(tree.root_node)
+    # Assuming each branch is its own, this might need refinement based on definition
+    avg_branches = total_branches / num_lines
+    return {'avg_branches': avg_branches}
+
+
+def string_stats(tree, num_lines):
+    string_literals = []
+
+    # Function to traverse the AST and collect string literals
+    def traverse(node):
+        if node.type == 'string_literal':
+            # Extracting the string literal, excluding the quotation marks
+            literal_text = node.text.decode('utf8')[1:-1]
+            string_literals.append(literal_text)
+        for child in node.children:
+            traverse(child)
+
+    traverse(tree.root_node)
+
+    # Calculate the average string length
+    total_length = sum(len(s) for s in string_literals)
+    avg_length = total_length / num_lines
+    return {'avg_str_length': avg_length}
+
+
+def identifier_stats(tree, num_lines):
+    root_node = tree.root_node
+    identifier_counts = {}  # Dictionary to count occurrences of each identifier
+    total_nodes = 0  # Counter for all nodes
+
+    # Function to recursively count identifiers and all nodes, gathering their stats
+    def count(node):
+        nonlocal identifier_counts, total_nodes
+        iden_count = 0
+        max_length = 0
+        total_nodes += 1  # Increment total nodes for every node visited
+        if node.type == 'identifier':
+            identifier = node.text.decode('utf8')  # Assuming UTF-8 encoding
+            iden_count += 1
+            identifier_counts[identifier] = identifier_counts.get(identifier, 0) + 1
+            iden_length = len(identifier)
+            if iden_length > max_length:
+                max_length = iden_length
+        for child in node.children:
+            child_count, child_max_length = count(child)
+            iden_count += child_count
+            if child_max_length > max_length:
+                max_length = child_max_length
+        return iden_count, max_length
+
+    total_identifiers, max_identifier_length = count(root_node)
+    total_unique_identifiers = len(identifier_counts)
+    total_identifier_length = sum(len(k) * v for k, v in identifier_counts.items())
+    avg_identifier_length = total_identifier_length / num_lines
+
+    # Calculate the identifier ratio as total identifiers over total nodes
+    identifier_ratio = total_identifiers / total_nodes if total_nodes > 0 else 0
+
+    return {
+        'total_identifiers': total_identifiers,
+        'total_identifier_length': total_identifier_length,
+        'max_identifier_length': max_identifier_length,
+        'avg_identifier_length': avg_identifier_length,
+        'total_unique_identifiers': total_unique_identifiers,
+        'total_identifier_ratio': identifier_ratio,  # Include the new ratio in the returned dictionary
+        'total_nodes': total_nodes,  # Include total node count for reference or further calculations
+    }
+
+
+def compute_regression(results):
+    components = {
+        'total_line_length': -0.0001,
+        'max_line_length': -0.0021,
+        'total_identifiers': 0.0076,
+        'total_identifier_length': -0.0004,
+        'max_identifier_length': -0.0067,
+        'avg_identifier_length': -0.005,
+        'avg_arithmetic_operations': 0.0225,
+        'avg_branches': 0.9886,
+        'avg_loops': 0.1572,
+        'total_assertions': 0.0119,
+        'total_has_assertions': -0.0147,
+        'avg_characters': 0.1242,
+        'total_class_instances': -0.043,
+        'total_distinct_methods': -0.0127,
+        'avg_str_length': 0.0026,
+        'total_has_exceptions': 0.1206,
+        'total_unique_identifiers': -0.019,
+        'max_nulls': -0.0712,
+        'total_numbers': -0.0078,
+        'avg_nulls': 0.1444,
+        'total_identifier_ratio': 0.334,
+        'total_method_ratio': 0.0406,
+        'total_floats': -0.0174,
+        'total_byte_entropy': -0.3917,
+    }
+    test_score = 0
+
+    for component in components:
+        test_score += components[component] * results[component]
+
+    test_score += 5.7501
+    return test_score
+
+
+def compute_readability(python_code):
+    # Create parser and set up language
+    import tree_sitter_python
+    from tree_sitter import Language, Parser
+
+    parser = Parser(Language(tree_sitter_python.language()))
+
+    results = code_stats(python_code)
+
+    num_lines = len(python_code.strip().split('\n'))
+    results.update(total_byte_entropy_stats(python_code))
+
+    tree = parser.parse(bytes(python_code, 'utf8'))
+
+    results.update(identifier_stats(tree, num_lines))
+    results.update(loops_stats(tree, num_lines))
+    results.update(branches_stats(tree, num_lines))
+    results.update(distinct_methods_stats(tree, num_lines))
+    results.update(has_execeptions(tree, num_lines))
+    results.update(class_instances_stats(tree, num_lines))
+    results.update(assertions_stats(tree, num_lines))
+    results.update(numbers_floats_stats(tree, num_lines))
+    results.update(average_nulls_stats(tree, num_lines))
+    results.update(arithmetic_operations_stats(tree, num_lines))
+    results.update(string_stats(tree, num_lines))
+
+    score = compute_regression(results)
+    return score

evaluation/benchmarks/testgeneval/eval_infer.py

@@ -0,0 +1,629 @@
+import os
+import tempfile
+import time
+from functools import partial
+
+import pandas as pd
+from report_utils import (
+    check_coverage,
+    check_mutation,
+    count_methods,
+    get_lines_of_code,
+)
+
+from evaluation.benchmarks.testgeneval.compute_readability import compute_readability
+from evaluation.benchmarks.testgeneval.constants import (
+    COVERAGE_PREFIX,
+    MUTATION_BUFFER,
+    MUTATION_TEMPLATE,
+    MUTATION_TIMEOUT,
+    TESTS_SUFFIX,
+)
+from evaluation.benchmarks.testgeneval.metrics import (
+    bleu,
+    edit_sim,
+    exact_match,
+    rouge_l,
+)
+from evaluation.benchmarks.testgeneval.pygments_utils import tokenize_code
+from evaluation.benchmarks.testgeneval.run_infer import get_instance_docker_image
+from evaluation.benchmarks.testgeneval.test_filter import filter_tests
+from evaluation.benchmarks.testgeneval.test_spec import (
+    TestGenEvalInstance,
+    TestSpec,
+    make_test_spec,
+)
+from evaluation.benchmarks.testgeneval.utils import load_testgeneval_dataset
+from evaluation.utils.shared import (
+    EvalMetadata,
+    EvalOutput,
+    prepare_dataset,
+    reset_logger_for_multiprocessing,
+    run_evaluation,
+)
+from openhands.core.config import AppConfig, SandboxConfig, get_parser
+from openhands.core.logger import openhands_logger as logger
+from openhands.core.main import create_runtime
+from openhands.events.action import CmdRunAction
+from openhands.events.observation import CmdOutputObservation
+from openhands.utils.async_utils import call_async_from_sync
+
+DOCKER_IMAGE_PREFIX = os.environ.get('EVAL_DOCKER_IMAGE_PREFIX', 'docker.io/kdjain/')
+logger.info(f'Using docker image prefix: {DOCKER_IMAGE_PREFIX}')
+
+
+def get_config(instance: pd.Series) -> AppConfig:
+    base_container_image = get_instance_docker_image(instance['instance_id_swebench'])
+    assert (
+        base_container_image
+    ), f"Invalid container image for instance {instance['instance_id_swebench']}."
+    logger.info(f'Using instance container image: {base_container_image}.')
+    return AppConfig(
+        run_as_openhands=False,
+        runtime=os.environ.get('RUNTIME', 'eventstream'),
+        sandbox=SandboxConfig(
+            base_container_image=base_container_image,
+            use_host_network=False,
+            timeout=1800,
+            api_key=os.environ.get('ALLHANDS_API_KEY'),
+            remote_runtime_api_url=os.environ.get(
+                'SANDBOX_REMOTE_RUNTIME_API_URL', 'http://localhost:8000'
+            ),
+        ),
+        workspace_base=None,
+        workspace_mount_path=None,
+    )
+
+
+def compute_lexical_metrics(pred_suite, gold_suite):
+    pred_loc = get_lines_of_code(pred_suite)
+    gold_loc = get_lines_of_code(gold_suite)
+    pred_methods = count_methods(pred_suite)
+    gold_methods = count_methods(gold_suite)
+    readability_pred = compute_readability(pred_suite)
+    readability_gold = compute_readability(gold_suite)
+
+    preds = tokenize_code(pred_suite)
+    golds = tokenize_code(gold_suite)
+
+    return {
+        'pred_loc': pred_loc,
+        'gold_loc': gold_loc,
+        'pred_readability': readability_pred,
+        'gold_readability': readability_gold,
+        'pred_methods': pred_methods,
+        'gold_methods': gold_methods,
+        'bleu': bleu(preds, golds),
+        'xmatch': exact_match(preds, golds),
+        'edit_sim': edit_sim(preds, golds),
+        'rouge_f': rouge_l(golds, preds)['f'],
+        'rouge_p': rouge_l(golds, preds)['p'],
+        'rouge_r': rouge_l(golds, preds)['r'],
+    }
+
+
+def run_command(runtime, command, timeout=600):
+    action = CmdRunAction(command=command)
+    action.set_hard_timeout(timeout)
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    assert obs.exit_code == 0
+    return obs
+
+
+def run_tests(runtime, instance, test_script, log_file='/tmp/test_output.log'):
+    action = CmdRunAction(command=f'bash {test_script} > {log_file} 2>&1 & echo $!')
+    action.set_hard_timeout(60)
+    obs = runtime.run_action(action)
+
+    assert isinstance(obs, CmdOutputObservation), 'Failed to start test script.'
+    pid = obs.content.split()[-1].strip()
+    logger.info(f'[{instance.instance_id}] Test process started with PID: {pid}')
+
+    start_time = time.time()
+    timeout = 1800
+    while True:
+        elapsed_time = time.time() - start_time
+        if elapsed_time > timeout:
+            logger.info(f'[{instance.instance_id}] Test process timed out.')
+            instance['test_result']['report']['test_timeout'] = True
+            break
+
+        check_action = CmdRunAction(command=f'ps -p {pid} > /dev/null; echo $?')
+        check_obs = runtime.run_action(check_action)
+        if (
+            isinstance(check_obs, CmdOutputObservation)
+            and len(check_obs.content.split()) > 0
+            and check_obs.content.split()[-1].strip() == '1'
+        ):
+            logger.info(f'[{instance.instance_id}] Test process completed.')
+            break
+        time.sleep(30)
+
+    test_action = CmdRunAction(command=f'cat {log_file}')
+    test_action.set_hard_timeout(300)
+    test_obs = runtime.run_action(test_action)
+    assert isinstance(test_obs, CmdOutputObservation), 'Failed to retrieve test output.'
+    return test_obs.exit_code, test_obs.content, elapsed_time
+
+
+def run_mutation_testing(
+    runtime, instance, mutation_script, log_file='/tmp/mutation_output.log'
+):
+    action = CmdRunAction(command=f'bash {mutation_script} > {log_file} 2>&1 & echo $!')
+    action.set_hard_timeout(60)
+    obs = runtime.run_action(action)
+
+    assert isinstance(obs, CmdOutputObservation), 'Failed to start test script.'
+    pid = obs.content.split()[-1].strip()
+    logger.info(f'[{instance.instance_id}] Mutation process started with PID: {pid}')
+
+    start_time = time.time()
+    timeout = 4000
+    while True:
+        elapsed_time = time.time() - start_time
+        if elapsed_time > timeout:
+            logger.info(f'[{instance.instance_id}] Mutation process timed out.')
+            instance['test_result']['report']['mutation_timeout'] = True
+            break
+
+        check_action = CmdRunAction(command=f'ps -p {pid} > /dev/null; echo $?')
+        check_obs = runtime.run_action(check_action)
+        if (
+            isinstance(check_obs, CmdOutputObservation)
+            and len(check_obs.content.split()) > 0
+            and check_obs.content.split()[-1].strip() == '1'
+        ):
+            logger.info(f'[{instance.instance_id}] Mutation process completed.')
+            break
+        time.sleep(30)
+
+    assert isinstance(obs, CmdOutputObservation), 'Failed to run mutation script.'
+    mutation_action = CmdRunAction(command=f'cat {log_file}')
+    mutation_action.set_hard_timeout(300)
+    mutation_obs = runtime.run_action(mutation_action)
+    assert isinstance(
+        mutation_obs, CmdOutputObservation
+    ), 'Failed to retrieve mutation output.'
+    return mutation_obs.exit_code, mutation_obs.content
+
+
+def grade_test_output(
+    test_suite: str, instance: pd.Series, test_output: str, test_spec: TestSpec, runtime
+):
+    """
+    Two-pass test grading with short-circuiting:
+    1. Run all tests to identify passing/failing tests
+    2. If no failing tests, evaluate coverage immediately
+    3. Otherwise, run only passing tests for coverage analysis
+    """
+    unit_test_output, coverage_output = '', ''
+    if TESTS_SUFFIX in test_output:
+        unit_test_output = test_output.split(TESTS_SUFFIX)[0]
+
+    if not unit_test_output:
+        return (
+            False,
+            0,
+            '',
+            '',
+            {
+                'total_tests': 0,
+                'passing_tests': 0,
+                'failing_tests': 0,
+                'any_pass': False,
+                'all_pass': False,
+                'passing_test_names': [],
+                'failing_test_names': [],
+            },
+        )
+
+    logger.info('Calling filter unit tests')
+    filtered_content, passing_tests, failing_tests = filter_tests(
+        test_suite, unit_test_output, test_spec.repo
+    )
+
+    total_tests = len(passing_tests) + len(failing_tests)
+    test_stats = {
+        'total_tests': total_tests,
+        'passing_tests': len(passing_tests),
+        'failing_tests': len(failing_tests),
+        'any_pass': len(passing_tests) > 0,
+        'all_pass': len(failing_tests) == 0 and total_tests > 0,
+        'passing_test_names': passing_tests,
+        'failing_test_names': failing_tests,
+    }
+
+    if not passing_tests:
+        return False, 0, unit_test_output, coverage_output, test_stats
+
+    # If all tests pass, evaluate coverage immediately
+    if not failing_tests:
+        coverage = 0
+        cov_success = False
+        if COVERAGE_PREFIX in test_output:
+            coverage_output = test_output.split(COVERAGE_PREFIX)[1]
+            _, coverage = check_coverage(coverage_output, test_spec.code_file)
+            cov_success = True
+        # test_stats['filtered_suite'] = test_suite
+        return cov_success, coverage, unit_test_output, coverage_output, test_stats
+
+    cov_success = False
+    coverage = 0
+    # Second pass - run coverage on passing tests
+    if filtered_content:
+        with tempfile.TemporaryDirectory() as temp_dir:
+            test_suite_path = os.path.join(temp_dir, 'test_suite.py')
+            with open(test_suite_path, 'w') as f:
+                f.write(filtered_content)
+            runtime.copy_to(test_suite_path, '/tmp')
+
+        run_command(runtime, f'cp /tmp/test_suite.py /testbed/{test_spec.test_file}')
+        _, test_output_second_pass, _ = run_tests(runtime, instance, '/tmp/test.sh')
+
+        coverage, coverage_output, unit_test_output = 0, '', test_output_second_pass
+
+        if COVERAGE_PREFIX in test_output_second_pass:
+            coverage_output = test_output_second_pass.split(COVERAGE_PREFIX)[1]
+            unit_test_output = test_output_second_pass.split(TESTS_SUFFIX)[0]
+            _, coverage = check_coverage(coverage_output, test_spec.code_file)
+            cov_success = True
+
+    # test_stats['filtered_suite'] = filtered_content
+    return cov_success, coverage, unit_test_output, coverage_output, test_stats
+
+
+def process_instance(
+    instance: pd.Series,
+    metadata: EvalMetadata,
+    reset_logger: bool = True,
+    log_dir: str | None = None,
+) -> EvalOutput:
+    """
+    Evaluate agent performance on a TestGenEval problem instance.
+
+    Note that this signature differs from the expected input to `run_evaluation`. Use
+    `functools.partial` to provide optional arguments before passing to the evaluation harness.
+
+    Args:
+        log_dir (str | None, default=None): Path to directory where log files will be written. Must
+        be provided if `reset_logger` is set.
+
+    Raises:
+        AssertionError: if the `reset_logger` flag is set without a provided log directory.
+    """
+    if reset_logger:
+        assert (
+            log_dir is not None
+        ), "Can't reset logger without a provided log directory."
+        os.makedirs(log_dir, exist_ok=True)
+        reset_logger_for_multiprocessing(logger, instance.instance_id, log_dir)
+    else:
+        logger.info(f'Starting evaluation for instance {instance.instance_id}.')
+
+    config = get_config(instance)
+    id = instance.instance_id
+    logger.info(f'Starting evaluation for instance {id}.')
+
+    instance['test_result']['id'] = id
+    instance['test_result']['report'] = {
+        'test_output': '',
+        # 'coverage_output': '',
+        # 'mutation_output': '',
+        'empty_generation': False,
+        'error_eval': False,
+        'all_tests_pass': False,
+        'tests_pass': False,
+        'test_timeout': False,
+        'mutation_timeout': False,
+        'coverage_success': False,
+        'mutation_success': False,
+        'coverage': 0,
+        'mutation_score': 0,
+        'mutation_error_interval': -1,
+        'num_mutants': -1,
+    }
+
+    instance['test_result']['lexical'] = {
+        'pred_loc': -1,
+        'gold_loc': -1,
+        'pred_readability': -1,
+        'gold_readability': -1,
+        'pred_methods': -1,
+        'gold_methods': -1,
+        'bleu': -1,
+        'xmatch': -1,
+        'edit_sim': -1,
+        'rouge_f': -1,
+        'rouge_p': -1,
+        'rouge_r': -1,
+    }
+
+    if instance['test_suite'] == '' or instance['test_suite'] is None:
+        instance['test_result']['report']['empty_generation'] = True
+        return EvalOutput(
+            instance_id=instance.instance_id, test_result=instance['test_result']
+        )
+
+    if not args.skip_lexical:
+        lexical_metrics = compute_lexical_metrics(
+            instance['test_suite'], instance['instance']['test_src']
+        )
+        instance['test_result']['lexical'] = lexical_metrics
+
+    test_suite = instance['test_suite']
+    test_spec: TestSpec = instance['test_spec']
+    runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
+    with tempfile.TemporaryDirectory() as temp_dir:
+        test_suite_path = os.path.join(temp_dir, 'test_suite.py')
+        with open(test_suite_path, 'w') as f:
+            f.write(test_suite)
+        runtime.copy_to(test_suite_path, '/tmp')
+
+        test_script_path = os.path.join(temp_dir, 'test.sh')
+        with open(test_script_path, 'w') as f:
+            f.write(test_spec.test_script)
+        runtime.copy_to(test_script_path, '/tmp')
+
+        mutation_script_path = os.path.join(temp_dir, 'mutation.sh')
+        with open(mutation_script_path, 'w') as f:
+            f.write(test_spec.mutation_script)
+        runtime.copy_to(mutation_script_path, '/tmp')
+
+    try:
+        run_command(runtime, 'chmod +x /tmp/test.sh /tmp/mutation.sh')
+        run_command(runtime, f'cp /tmp/test_suite.py /testbed/{test_spec.test_file}')
+
+        # First pass - run all tests
+        _, test_output, test_time = run_tests(runtime, instance, '/tmp/test.sh')
+
+        # Grade tests with two-pass approach
+        coverage_success, coverage, unit_test_output, coverage_output, test_stats = (
+            grade_test_output(test_suite, instance, test_output, test_spec, runtime)
+        )
+
+        # Update report with test statistics
+        instance['test_result']['report'].update(
+            {
+                'test_output': unit_test_output,
+                # 'coverage_output': coverage_output,
+                'tests_pass': test_stats['any_pass'],  # Changed to use any_pass
+                'all_tests_pass': test_stats['all_pass'],  # Added all_pass metric
+                'coverage_success': coverage_success,
+                'coverage': coverage if coverage_success else 0,
+                'test_stats': test_stats,
+            }
+        )
+
+        # Only run mutation testing if we have passing tests and coverage
+        if (
+            not args.skip_mutation
+            and coverage_success
+            and test_stats['any_pass']
+            and coverage > 0
+        ):
+            mutation_timeout = max(10, 1.5 * test_time)
+            mutation_toml = MUTATION_TEMPLATE.format(
+                test_cmd=test_spec.test_cmd,
+                source_fp=test_spec.code_file,
+                timeout=mutation_timeout,
+            )
+
+            with tempfile.TemporaryDirectory() as temp_dir:
+                mutation_toml_path = os.path.join(temp_dir, 'mutation.toml')
+                with open(mutation_toml_path, 'w') as f:
+                    f.write(mutation_toml)
+                runtime.copy_to(mutation_toml_path, '/tmp')
+
+            run_command(runtime, 'cp /tmp/mutation.toml /testbed/mutation.toml')
+
+            mutation_code, mutation_output = run_mutation_testing(
+                runtime, instance, '/tmp/mutation.sh'
+            )
+            # instance['test_result']['report']['mutation_output'] = mutation_output
+            if mutation_output and mutation_code == 0:
+                (
+                    mutation_success,
+                    num_mutants,
+                    mutation_score,
+                    mutation_confidence_interval,
+                ) = check_mutation(mutation_output)
+                instance['test_result']['report']['num_mutants'] = num_mutants
+                instance['test_result']['report']['mutation_success'] = mutation_success
+                instance['test_result']['report']['mutation_score'] = mutation_score
+                instance['test_result']['report']['mutation_error_interval'] = (
+                    mutation_confidence_interval
+                )
+
+        return EvalOutput(
+            instance_id=instance.instance_id, test_result=instance['test_result']
+        )
+    except Exception as e:
+        logger.error(f'Error processing instance {instance.instance_id}: {e}')
+        raise RuntimeError(
+            instance.instance_id,
+            'Unexpected output...',
+            logger,
+        )
+
+    finally:
+        runtime.close()
+
+
+def count_and_log_fields(evaluated_predictions, fields, key):
+    """
+    Count and log the sum of specified fields in the evaluated predictions,
+    ignoring fields with a value of -1. If all values for a field are -1,
+    return -1.
+
+    :param evaluated_predictions: DataFrame containing evaluation results
+    :param fields: List of field names to count
+    :param key: Key to access the field values ('report' or 'lexical')
+    """
+
+    def count_field(row, field):
+        value = row['test_result'][key][field]
+        return (
+            value if value != -1 else None
+        )  # Ignore -1 fields by treating them as None
+
+    for field in fields:
+        # Extract the valid values for the field, ignoring -1
+        valid_values = evaluated_predictions.apply(
+            count_field, args=(field,), axis=1
+        ).dropna()
+
+        if valid_values.empty:  # If all values are -1
+            logger.info(f'# {field}: -1 (All values are -1)')
+        else:
+            count = valid_values.sum()  # Sum of valid values
+            length = len(valid_values)  # Count of valid entries
+            logger.info(f'# {field}: {length}. ({count / length:.2f})')
+
+
+if __name__ == '__main__':
+    parser = get_parser()
+    parser.add_argument(
+        '--input-file', type=str, required=True, help='Path to input predictions file'
+    )
+    parser.add_argument(
+        '--dataset',
+        type=str,
+        default='kjain14/testgeneval',
+        help='Dataset to evaluate on',
+    )
+    parser.add_argument(
+        '--split', type=str, default='test', help='Split to evaluate on'
+    )
+    parser.add_argument(
+        '--skip_mutation', action='store_true', help='Skip mutation testing'
+    )
+    parser.add_argument(
+        '--skip_lexical', action='store_true', help='Skip lexical metrics'
+    )
+    parser.add_argument(
+        '--mutation_timeout',
+        type=int,
+        default=MUTATION_TIMEOUT,
+        help='Mutation timeout',
+    )
+    parser.add_argument(
+        '--mutation_buffer',
+        type=int,
+        default=MUTATION_BUFFER,
+        help='Mutation buffer',
+    )
+    args, _ = parser.parse_known_args()
+
+    dataset: list[TestGenEvalInstance] = load_testgeneval_dataset(
+        args.dataset, args.split
+    )
+
+    logger.info(
+        f'Loaded dataset {args.dataset} with split {args.split} to run inference on.'
+    )
+
+    # Load predictions
+    assert args.input_file.endswith('.jsonl'), 'Input file must be a jsonl file.'
+    predictions = pd.read_json(args.input_file, lines=True)
+    assert (
+        'instance_id' in predictions.columns
+    ), 'Input file must contain instance_id column.'
+
+    if 'test_suite' not in predictions.columns and (
+        'test_result' in predictions.columns
+        and 'test_suite' in predictions['test_result'].iloc(0)
+    ):
+        raise ValueError(
+            'Input file must contain test_suite column OR test_result column with test_suite field.'
+        )
+
+    if 'instance_id_swebench' not in predictions.columns:
+        predictions['instance_id_swebench'] = predictions['instance'].apply(
+            lambda x: x['instance_id_swebench']
+        )
+
+    if 'instance_id' not in predictions.columns and (
+        'instance_id' in predictions['instance'].iloc(0)
+    ):
+        raise ValueError(
+            'Input file must contain id column OR instance column with id field.'
+        )
+
+    if 'instance_id' not in predictions.columns:
+        predictions['instance_id'] = predictions['instance'].apply(
+            lambda x: x['instance_id']
+        )
+
+    if 'test_suite' not in predictions.columns:
+        predictions['test_suite'] = predictions['test_result'].apply(
+            lambda x: x['test_suite']
+        )
+
+    assert len(predictions['instance_id'].unique()) == len(
+        predictions
+    ), 'instance_id column must be unique.'
+
+    assert {'instance_id_swebench', 'test_suite', 'instance_id'}.issubset(
+        set(predictions.columns)
+    ), 'Input file must contain id, instance_id and test_suite columns.'
+
+    predictions['test_spec'] = predictions['instance'].apply(
+        lambda x: make_test_spec(x, args.mutation_timeout, args.mutation_buffer)
+    )
+
+    output_file = args.input_file.replace('.jsonl', '.testgeneval.jsonl')
+    instances = prepare_dataset(predictions, output_file, args.eval_n_limit)
+
+    # If possible, load the relevant metadata to avoid issues with `run_evaluation`.
+    metadata: EvalMetadata | None = None
+    metadata_filepath = os.path.join(os.path.dirname(args.input_file), 'metadata.json')
+    if os.path.exists(metadata_filepath):
+        with open(metadata_filepath, 'r') as metadata_file:
+            data = metadata_file.read()
+            metadata = EvalMetadata.model_validate_json(data)
+
+    # The evaluation harness constrains the signature of `process_instance_func` but we need to
+    # pass extra information. Build a new function object to avoid issues with multiprocessing.
+    process_instance_func = partial(
+        process_instance, log_dir=output_file.replace('.jsonl', '.logs')
+    )
+
+    run_evaluation(
+        instances,
+        metadata=None,
+        output_file=output_file,
+        num_workers=args.eval_num_workers,
+        process_instance_func=process_instance_func,
+    )
+
+    # Load evaluated predictions & print number of resolved predictions
+    evaluated_predictions = pd.read_json(output_file, lines=True)
+    report_fields = [
+        'coverage',
+        'mutation_score',
+        'tests_pass',
+        'all_tests_pass',
+        'empty_generation',
+        'coverage_success',
+        'test_timeout',
+        'error_eval',
+    ]
+    lexical_fields = [
+        'pred_loc',
+        'gold_loc',
+        'pred_methods',
+        'gold_methods',
+        'bleu',
+        'xmatch',
+        'edit_sim',
+        'rouge_f',
+        'rouge_p',
+        'rouge_r',
+    ]
+
+    # Log report and lexical fields
+    count_and_log_fields(evaluated_predictions, report_fields, key='report')
+    count_and_log_fields(evaluated_predictions, lexical_fields, key='lexical')

evaluation/benchmarks/testgeneval/log_parsers.py

@@ -0,0 +1,291 @@
+import re
+
+from evaluation.benchmarks.testgeneval.constants import TestStatus
+
+
+def parse_log_pytest(log: str) -> dict[str, str]:
+    """
+    Parser for test logs generated with PyTest framework
+
+    Args:
+        log (str): log content
+    Returns:
+        dict: test case to test status mapping
+    """
+    test_status_map = {}
+    for line in log.split('\n'):
+        if any([line.startswith(x.value) for x in TestStatus]):
+            # Additional parsing for FAILED status
+            if line.startswith(TestStatus.FAILED.value):
+                line = line.replace(' - ', ' ')
+            test_case = line.split()
+            if len(test_case) <= 1:
+                continue
+            test_status_map[test_case[1]] = test_case[0]
+    return test_status_map
+
+
+def parse_log_pytest_options(log: str) -> dict[str, str]:
+    """
+    Parser for test logs generated with PyTest framework with options
+
+    Args:
+        log (str): log content
+    Returns:
+        dict: test case to test status mapping
+    """
+    option_pattern = re.compile(r'(.*?)\[(.*)\]')
+    test_status_map = {}
+    for line in log.split('\n'):
+        if any([line.startswith(x.value) for x in TestStatus]):
+            # Additional parsing for FAILED status
+            if line.startswith(TestStatus.FAILED.value):
+                line = line.replace(' - ', ' ')
+            test_case = line.split()
+            if len(test_case) <= 1:
+                continue
+            has_option = option_pattern.search(test_case[1])
+            if has_option:
+                main, option = has_option.groups()
+                if (
+                    option.startswith('/')
+                    and not option.startswith('//')
+                    and '*' not in option
+                ):
+                    option = '/' + option.split('/')[-1]
+                test_name = f'{main}[{option}]'
+            else:
+                test_name = test_case[1]
+            test_status_map[test_name] = test_case[0]
+    return test_status_map
+
+
+def parse_log_django(log: str) -> dict[str, str]:
+    """
+    Parser for test logs generated with Django tester framework
+
+    Args:
+        log (str): log content
+    Returns:
+        dict: test case to test status mapping
+    """
+    test_status_map = {}
+    lines = log.split('\n')
+
+    prev_test = None
+    for line in lines:
+        line = line.strip()
+
+        # This isn't ideal but the test output spans multiple lines
+        if '--version is equivalent to version' in line:
+            test_status_map['--version is equivalent to version'] = (
+                TestStatus.PASSED.value
+            )
+
+        # Log it in case of error
+        if ' ... ' in line:
+            prev_test = line.split(' ... ')[0]
+
+        pass_suffixes = (' ... ok', ' ... OK', ' ...  OK')
+        for suffix in pass_suffixes:
+            if line.endswith(suffix):
+                # TODO: Temporary, exclusive fix for django__django-7188
+                # The proper fix should involve somehow getting the test results to
+                # print on a separate line, rather than the same line
+                if line.strip().startswith(
+                    'Applying sites.0002_alter_domain_unique...test_no_migrations'
+                ):
+                    line = line.split('...', 1)[-1].strip()
+                test = line.rsplit(suffix, 1)[0]
+                test_status_map[test] = TestStatus.PASSED.value
+                break
+        if ' ... skipped' in line:
+            test = line.split(' ... skipped')[0]
+            test_status_map[test] = TestStatus.SKIPPED.value
+        if line.endswith(' ... FAIL'):
+            test = line.split(' ... FAIL')[0]
+            test_status_map[test] = TestStatus.FAILED.value
+        if line.startswith('FAIL:'):
+            test = line.split()[1].strip()
+            test_status_map[test] = TestStatus.FAILED.value
+        if line.endswith(' ... ERROR'):
+            test = line.split(' ... ERROR')[0]
+            test_status_map[test] = TestStatus.ERROR.value
+        if line.startswith('ERROR:'):
+            test = line.split()[1].strip()
+            test_status_map[test] = TestStatus.ERROR.value
+
+        if line.lstrip().startswith('ok') and prev_test is not None:
+            # It means the test passed, but there's some additional output (including new lines)
+            # between "..." and "ok" message
+            test = prev_test
+            test_status_map[test] = TestStatus.PASSED.value
+
+    # TODO: This is very brittle, we should do better
+    # There's a bug in the django logger, such that sometimes a test output near the end gets
+    # interrupted by a particular long multiline print statement.
+    # We have observed this in one of 3 forms:
+    # - "{test_name} ... Testing against Django installed in {*} silenced.\nok"
+    # - "{test_name} ... Internal Server Error: \/(.*)\/\nok"
+    # - "{test_name} ... System check identified no issues (0 silenced).\nok"
+    patterns = [
+        r'^(.*?)\s\.\.\.\sTesting\ against\ Django\ installed\ in\ ((?s:.*?))\ silenced\)\.\nok$',
+        r'^(.*?)\s\.\.\.\sInternal\ Server\ Error:\ \/(.*)\/\nok$',
+        r'^(.*?)\s\.\.\.\sSystem check identified no issues \(0 silenced\)\nok$',
+    ]
+    for pattern in patterns:
+        for match in re.finditer(pattern, log, re.MULTILINE):
+            test_name = match.group(1)
+            test_status_map[test_name] = TestStatus.PASSED.value
+    return test_status_map
+
+
+def parse_log_pytest_v2(log: str) -> dict[str, str]:
+    """
+    Parser for test logs generated with PyTest framework (Later Version)
+
+    Args:
+        log (str): log content
+    Returns:
+        dict: test case to test status mapping
+    """
+    test_status_map = {}
+    escapes = ''.join([chr(char) for char in range(1, 32)])
+    for line in log.split('\n'):
+        line = re.sub(r'\[(\d+)m', '', line)
+        translator = str.maketrans('', '', escapes)
+        line = line.translate(translator)
+        if any([line.startswith(x.value) for x in TestStatus]):
+            if line.startswith(TestStatus.FAILED.value):
+                line = line.replace(' - ', ' ')
+            test_case = line.split()
+            if len(test_case) >= 2:
+                test_status_map[test_case[1]] = test_case[0]
+        # Support older pytest versions by checking if the line ends with the test status
+        elif any([line.endswith(x.value) for x in TestStatus]):
+            test_case = line.split()
+            if len(test_case) >= 2:
+                test_status_map[test_case[0]] = test_case[1]
+    return test_status_map
+
+
+def parse_log_seaborn(log: str) -> dict[str, str]:
+    """
+    Parser for test logs generated with seaborn testing framework
+
+    Args:
+        log (str): log content
+    Returns:
+        dict: test case to test status mapping
+    """
+    test_status_map = {}
+    for line in log.split('\n'):
+        if line.startswith(TestStatus.FAILED.value):
+            test_case = line.split()[1]
+            test_status_map[test_case] = TestStatus.FAILED.value
+        elif f' {TestStatus.PASSED.value} ' in line:
+            parts = line.split()
+            if parts[1] == TestStatus.PASSED.value:
+                test_case = parts[0]
+                test_status_map[test_case] = TestStatus.PASSED.value
+        elif line.startswith(TestStatus.PASSED.value):
+            parts = line.split()
+            test_case = parts[1]
+            test_status_map[test_case] = TestStatus.PASSED.value
+    return test_status_map
+
+
+def parse_log_sympy(log: str) -> dict[str, str]:
+    """
+    Parser for test logs generated with Sympy framework
+
+    Args:
+        log (str): log content
+    Returns:
+        dict: test case to test status mapping
+    """
+    test_status_map = {}
+    pattern = r'(_*) (.*)\.py:(.*) (_*)'
+    matches = re.findall(pattern, log)
+    for match in matches:
+        test_case = f'{match[1]}.py:{match[2]}'
+        test_status_map[test_case] = TestStatus.FAILED.value
+    for line in log.split('\n'):
+        line = line.strip()
+        if line.startswith('test_'):
+            if line.endswith('[FAIL]') or line.endswith('[OK]'):
+                line = line[: line.rfind('[')]
+                line = line.strip()
+            if line.endswith(' E'):
+                test = line.split()[0]
+                test_status_map[test] = TestStatus.ERROR.value
+            if line.endswith(' F'):
+                test = line.split()[0]
+                test_status_map[test] = TestStatus.FAILED.value
+            if line.endswith(' ok'):
+                test = line.split()[0]
+                test_status_map[test] = TestStatus.PASSED.value
+    return test_status_map
+
+
+def parse_log_matplotlib(log: str) -> dict[str, str]:
+    """
+    Parser for test logs generated with PyTest framework
+
+    Args:
+        log (str): log content
+    Returns:
+        dict: test case to test status mapping
+    """
+    test_status_map = {}
+    for line in log.split('\n'):
+        line = line.replace('MouseButton.LEFT', '1')
+        line = line.replace('MouseButton.RIGHT', '3')
+        if any([line.startswith(x.value) for x in TestStatus]):
+            # Additional parsing for FAILED status
+            if line.startswith(TestStatus.FAILED.value):
+                line = line.replace(' - ', ' ')
+            test_case = line.split()
+            if len(test_case) <= 1:
+                continue
+            test_status_map[test_case[1]] = test_case[0]
+    return test_status_map
+
+
+parse_log_astroid = parse_log_pytest
+parse_log_flask = parse_log_pytest
+parse_log_marshmallow = parse_log_pytest
+parse_log_pvlib = parse_log_pytest
+parse_log_pyvista = parse_log_pytest
+parse_log_sqlfluff = parse_log_pytest
+parse_log_xarray = parse_log_pytest
+
+parse_log_pydicom = parse_log_pytest_options
+parse_log_requests = parse_log_pytest_options
+parse_log_pylint = parse_log_pytest_options
+
+parse_log_astropy = parse_log_pytest_v2
+parse_log_scikit = parse_log_pytest_v2
+parse_log_sphinx = parse_log_pytest_v2
+
+
+MAP_REPO_TO_PARSER = {
+    'astropy/astropy': parse_log_astropy,
+    'django/django': parse_log_django,
+    'marshmallow-code/marshmallow': parse_log_marshmallow,
+    'matplotlib/matplotlib': parse_log_matplotlib,
+    'mwaskom/seaborn': parse_log_seaborn,
+    'pallets/flask': parse_log_flask,
+    'psf/requests': parse_log_requests,
+    'pvlib/pvlib-python': parse_log_pvlib,
+    'pydata/xarray': parse_log_xarray,
+    'pydicom/pydicom': parse_log_pydicom,
+    'pylint-dev/astroid': parse_log_astroid,
+    'pylint-dev/pylint': parse_log_pylint,
+    'pytest-dev/pytest': parse_log_pytest,
+    'pyvista/pyvista': parse_log_pyvista,
+    'scikit-learn/scikit-learn': parse_log_scikit,
+    'sqlfluff/sqlfluff': parse_log_sqlfluff,
+    'sphinx-doc/sphinx': parse_log_sphinx,
+    'sympy/sympy': parse_log_sympy,
+}

evaluation/benchmarks/testgeneval/metrics.py

@@ -0,0 +1,309 @@
+import sys
+from typing import Callable, Dict, List, Optional, Sequence, TypeVar, Union
+
+import nltk
+import numpy as np
+from fuzzywuzzy import fuzz
+from rouge import Rouge
+
+# increase recursion depth to ensure ROUGE can be calculated for long sentences
+if sys.getrecursionlimit() < 10_000:
+    sys.setrecursionlimit(10_000)
+
+
+def bleu(gold: List[str], pred: List[str]) -> float:
+    """
+    Calculate BLEU score, using smoothing method 2 with auto reweighting, in the range of 0~100.
+
+    :param gold: list of gold tokens
+    :param pred: list of predicted tokens
+    :return: BLEU score
+    """
+    if len(pred) == 0 or len(gold) == 0:
+        return 0.0
+    return 100.0 * nltk.translate.bleu_score.sentence_bleu(
+        [gold],
+        pred,
+        smoothing_function=nltk.translate.bleu_score.SmoothingFunction().method2,
+        auto_reweigh=True,
+    )
+
+
+def batch_bleu(golds: List[List[str]], preds: List[List[str]]) -> List[float]:
+    """
+    Calculate BLEU score for a batch of sentences.
+
+    :param golds: list of gold sentences
+    :param preds: list of predicted sentences
+    :return: list of BLEU scores
+    """
+    if len(golds) != len(preds):
+        raise ValueError('golds and preds must have the same length')
+    return [bleu(gold, pred) for gold, pred in zip(golds, preds)]
+
+
+def corpus_bleu(golds: List[List[str]], preds: List[List[str]]) -> float:
+    """
+    Calculate corpus-level BLEU score for a batch of sentences.
+
+    :param golds: list of gold sentences
+    :param preds: list of predicted sentences
+    :return: corpus-level BLEU score
+    """
+    if len(golds) != len(preds):
+        raise ValueError('golds and preds must have the same length')
+    return 100.0 * nltk.translate.bleu_score.corpus_bleu(
+        [[gold] for gold in golds],
+        preds,
+        smoothing_function=nltk.translate.bleu_score.SmoothingFunction().method2,
+        auto_reweigh=True,
+    )
+
+
+def edit_sim(
+    gold: Union[str, List[str]], pred: Union[str, List[str]], sep: str = ' '
+) -> float:
+    """
+    Calculate char-level edit similarity, in the range of 0~100.
+
+    :param gold: gold sentence or list of gold tokens
+    :param pred: predicted sentence or list of predicted tokens
+    :param sep: separator between tokens
+    :return: char-level edit similarity
+    """
+    if len(pred) == 0 or len(gold) == 0:
+        return 0.0
+    if isinstance(gold, list):
+        gold = sep.join(gold)
+    if isinstance(pred, list):
+        pred = sep.join(pred)
+    return fuzz.ratio(gold, pred)
+
+
+def batch_edit_sim(
+    golds: List[Union[str, List[str]]],
+    preds: List[Union[str, List[str]]],
+    sep: str = ' ',
+) -> List[float]:
+    """
+    Calculate char-level edit similarity for a batch of sentences.
+
+    :param golds: list of gold sentences
+    :param preds: list of predicted sentences
+    :param sep: separator between tokens
+    :return: list of char-level edit similarity
+    """
+    if len(golds) != len(preds):
+        raise ValueError('golds and preds must have the same length')
+    return [edit_sim(gold, pred, sep) for gold, pred in zip(golds, preds)]
+
+
+T = TypeVar('T')
+
+
+def exact_match(gold: T, pred: T) -> float:
+    """
+    Calculate exact match accuracy, in the range of {0, 100}.
+
+    :param gold: gold sentence or list of gold tokens
+    :param pred: predicted sentence or list of predicted tokens
+    :return: exact match accuracy
+    """
+    if len(pred) == 0 or len(gold) == 0:
+        return 0.0
+    return 100.0 if gold == pred else 0.0
+
+
+def batch_exact_match(golds: List[T], preds: List[T]) -> List[float]:
+    """
+    Calculate exact match accuracy for a batch of sentences.
+
+    :param golds: list of gold sentences
+    :param preds: list of predicted sentences
+    :return: list of exact match accuracy
+    """
+    if len(golds) != len(preds):
+        raise ValueError('golds and preds must have the same length')
+    return [exact_match(gold, pred) for gold, pred in zip(golds, preds)]
+
+
+def rouge_l(
+    gold: Union[str, List[str]], pred: Union[str, List[str]], sep: str = ' '
+) -> Dict[str, float]:
+    """
+    Calculate ROUGE-L F1, precision, and recall scores, in the range of 0~100.
+
+    :param gold: gold sentence or list of gold tokens
+    :param pred: predicted sentence or list of predicted tokens
+    :return: {"p": precision, "r": recall, "f": F1}
+    """
+    if len(pred) == 0 or len(gold) == 0:
+        return {'p': 0.0, 'r': 0.0, 'f': 0.0}
+    if isinstance(gold, list):
+        gold = sep.join(gold)
+    if isinstance(pred, list):
+        pred = sep.join(pred)
+    try:
+        rouge = Rouge()
+        scores = rouge.get_scores(hyps=pred, refs=gold, avg=True)
+        return {x: scores['rouge-l'][x] * 100.0 for x in ['p', 'r', 'f']}
+    except ValueError:
+        return {'p': 0.0, 'r': 0.0, 'f': 0.0}
+
+
+def batch_rouge_l(
+    golds: List[Union[str, List[str]]],
+    preds: List[Union[str, List[str]]],
+    sep: str = ' ',
+) -> Dict[str, List[float]]:
+    """
+    Calculate ROUGE-L F1, precision, and recall scores for a batch of sentences.
+
+    :param golds: list of gold sentences
+    :param preds: list of predicted sentences
+    :param sep: separator between tokens
+    :return: list of {"p": precision, "r": recall, "f": F1}
+    """
+    if len(golds) != len(preds):
+        raise ValueError('golds and preds must have the same length')
+    scores = [rouge_l(gold, pred, sep) for gold, pred in zip(golds, preds)]
+    return {x: [score[x] for score in scores] for x in ['p', 'r', 'f']}
+
+
+def accuracy(
+    gold: List[str],
+    pred: List[str],
+    ignore: Optional[Sequence[str]] = None,
+) -> float:
+    """
+    Calculate token-level accuracy, in the range of 0~100.
+    If gold and pred are not the same length, the longer one would be truncated.
+
+    :param gold: list of gold tokens
+    :param pred: list of predicted tokens
+    :param ignore: list of (gold) tokens to ignore
+    :return: accuracy
+    """
+    if len(pred) == 0 or len(gold) == 0:
+        return 0.0
+    if ignore is None:
+        ignore = []
+    i = 0
+    total = 0
+    match = 0
+    while i < len(gold) and i < len(pred):
+        if gold[i] in ignore:
+            i += 1
+            continue
+        total += 1
+        if gold[i] == pred[i]:
+            match += 1
+        i += 1
+
+    if total == 0:
+        return 0.0
+    return 100.0 * match / total
+
+
+def batch_accuracy(
+    golds: List[List[str]],
+    preds: List[List[str]],
+    ignore: Optional[Sequence[str]] = None,
+) -> List[float]:
+    """
+    Calculate token-level accuracy for a batch of sentences.
+
+    :param golds: list of gold sentences
+    :param preds: list of predicted sentences
+    :param ignore: list of (gold) tokens to ignore
+    :return: list of accuracy
+    """
+    if len(golds) != len(preds):
+        raise ValueError('golds and preds must have the same length')
+    return [accuracy(gold, pred, ignore) for gold, pred in zip(golds, preds)]
+
+
+def first_match_to_topk(
+    first_match_list: List[int], k_values: List[int]
+) -> Dict[int, List[float]]:
+    """
+    Calculate top-k accuracy with the first match ranks (1-indexed).
+
+    :param first_match: first match ranks (1-indexed)
+    :param k_values: k values to consider
+    :return: a mapping from k to top-k accuracies (ranging from 0~100)
+    """
+    return {k: [100.0 if x <= k else 0.0 for x in first_match_list] for k in k_values}
+
+
+def pass_at_k(n: int, c: int, k: int) -> float:
+    """
+    Sample pass@k metric according to the Codex paper, but in the scale of 0~100.
+    :param n: total number of samples
+    :param c: number of correct samples
+    :param k: k in pass@$k$
+    """
+    if n < k or (n - c) < k:
+        # fallback to the (1 - (1-p)^k) formula
+        return (1 - (1 - (c / n)) ** k) * 100
+    else:
+        return (1.0 - np.prod(1.0 - k / np.arange(n - c + 1, n + 1)).item()) * 100
+
+
+def self_bleu(samples: List[List[str]]) -> float:
+    """
+    Calculate self-BLEU among the samples.
+    :param samples: the chosen m samples
+    :return: self-BLEU
+    """
+    if len(samples) == 0:
+        return 100.0
+
+    scores = []
+    for i in range(len(samples)):
+        scores.append(
+            100.0
+            * nltk.translate.bleu_score.sentence_bleu(
+                [samples[j] for j in range(len(samples)) if j != i],
+                samples[i],
+                smoothing_function=nltk.translate.bleu_score.SmoothingFunction().method2,
+                auto_reweigh=True,
+            )
+        )
+    return np.mean(scores).item()
+
+
+def self_edit_distance(samples: List[Union[str, List[str]]], sep=' ') -> float:
+    """
+    Calculate self-edit-distance among the samples.
+    :param samples: the chosen m samples
+    :param sep: the separator between tokens
+    :return: self-edit-distance
+    """
+    if len(samples) == 0:
+        return 0.0
+
+    scores = []
+    for i in range(len(samples)):
+        sample_i = samples[i]
+        if not isinstance(sample_i, str):
+            sample_i = sep.join(sample_i)
+        for j in range(len(samples)):
+            if i == j:
+                continue
+            sample_j = samples[j]
+            if not isinstance(sample_j, str):
+                sample_j = sep.join(sample_j)
+
+            scores.append(100 - fuzz.ratio(sample_i, sample_j))
+    return np.mean(scores).item()
+
+
+QUALITY_METRICS: Dict[str, Callable[[List[str], List[str]], float]] = {
+    'bleu': bleu,
+    'xmatch': exact_match,
+    'edit-sim': edit_sim,
+    'rouge-f': lambda g, p: rouge_l(g, p)['f'],
+    'rouge-p': lambda g, p: rouge_l(g, p)['p'],
+    'rouge-r': lambda g, p: rouge_l(g, p)['r'],
+}

evaluation/benchmarks/testgeneval/prompt.py

@@ -0,0 +1,114 @@
+CODEACT_TESTGEN_PROMPT_OLD = """Your goal is to generate a high-quality test suite (at least 20+ passing tests) for the code file: {code_file}. Output the test suite at {test_file}\n'
+
+[current directory: /workspace/{workspace_dir_name}]
+
+IMPORTANT: You should ONLY interact with the environment provided to you AND NEVER ASK FOR HUMAN HELP
+
+IMPORTANT: Follow instructions, if you have < 80 tests you should generate more tests rather than trying to fix the ones you have.
+
+IMPORTANT: Code file to test:
+```python
+{code_src}
+```
+
+Here are additional imports that you may need:
+{imports}
+
+Look at code dependencies (NOT {code_file} since you already have contents) and test files you need context for to write a complete test suite.
+
+Aim for 20+ test functions with asserts. Do not hestitate to use the Python interpreter to understand the input output behavior of the code you are testing.
+
+Output your test suite at {test_file}. Each unit test must be a function starting with test_. Include all your test imports and setup before your first test. Do not include a main method to run the tests. Make sure to make it as comprehensive as possible, try to execute all the methods you saw.
+
+When you think you've successfully generated a test suite, run it on for the current project using {coverage_command}.
+
+If you have few tests GENERATE MORE TESTS rather than trying to fix the ones you have (it is possible to filter out failing tests later).
+
+Then run coverage report -m --include {code_file} to see how well your test suite covers the code under test.
+
+When you are trying to improve coverage pick a part of the code that is not covered (indicated by lines on coverage report), examine the code and then
+try to generate a test for it. Feel free to use a code interpreter to understand the input output behavior. ONLY add tests
+not remove them.
+
+If you are unable to see passing and failing tests, FIX YOUR IMPORTS to use the same style as other test files.
+
+You should NOT modify any existing test case files. You SHOULD add new test in a NEW file to reproduce the issue.
+
+You should NEVER use web browsing or any other web-based tools.
+
+You should NEVER install new packages, use existing packages only.
+
+You should ALWAYS use the default Python interpreter available in the <execute_bash> environment to run code related to the provided issue and/or repository.
+
+You should ALWAYS use local imports DO NOT import the general library.
+
+When you think you have a fully adequate test suite, please run the following command: <execute_bash> exit </execute_bash>.
+"""
+
+CODEACT_TESTGEN_PROMPT = """
+Your goal is to generate a comprehensive, **broad-coverage** test suite for the code below, ensuring you test as many lines and branches as possible on the first attempt.
+
+Place your test suite in a new file named {test_file}.
+
+IMPORTANT REQUIREMENTS:
+1. **No external help or resources**—use only the snippet below.
+2. **Focus on breadth over depth**: cover all major functions, classes, and code paths early to minimize coverage iterations.
+3. Each test function must start with `test_` and use `assert` to verify behavior.
+4. Include only necessary imports (standard library or local).
+5. Do **not** modify existing test files—create a brand new one. No `main()` or other non-test code.
+6. Produce **at least 20 test functions**; if coverage is lacking, add more tests rather than removing or changing existing ones.
+7. Use the following commands to check coverage:
+   <execute_bash> {coverage_command} </execute_bash>
+   <execute_bash> coverage report -m --include {code_file} </execute_bash>
+   If lines remain uncovered, add new tests targeting them specifically.
+8. When you're satisfied with coverage, finalize by running:
+   <execute_bash> exit </execute_bash>
+
+Below is the **complete code snippet** to test:
+
+<START_OF_CODE>
+{code_src}
+<END_OF_CODE>
+
+NOTE: if you are testing django, you must use from django.test import SimpleTestCase and class based tests (i.e. class TestSomething(SimpleTestCase)).
+NOTE: if there is an error executing tests you MUST fix it before exiting. DO NOT install new packages.
+NOTE: if outputting a revised test suite REPLACE {test_file} with the revised suite
+
+**Output the final test suite** (20+ tests) for {test_file} in a single code block, no extra commentary. MAKE SURE you run the tests and ensure you can see which tests passed and failed BEFORE exiting.
+"""
+
+CODEACT_TESTGEN_PROMPT_ITERATE = """
+Your goal is to improve the test suite at {test_file} to achieve **broad-coverage** of the code below.
+
+First run the test suite.
+
+If no tests run, then remove {test_file} and create {test_file} with a new suite.
+
+Otherwise, improve it aiming to improve code coverage.
+
+IMPORTANT REQUIREMENTS:
+1. Use the following commands to check coverage (RUN THIS FIRST):
+   <execute_bash> {coverage_command} </execute_bash>
+   <execute_bash> coverage report -m --include {code_file} </execute_bash>
+   If lines remain uncovered, add new tests targeting them specifically.
+2. **No external help or resources**—use only the snippet below.
+3. **Focus on breadth over depth**: cover all major functions, classes, and code paths early to minimize coverage iterations.
+4. Each test function must use `assert` to verify behavior.
+5. Include only necessary imports (standard library or local).
+6. Do **not** modify other test files in the repository. No `main()` or other non-test code.
+7. Produce **at least 20 test functions**; if coverage is lacking, add more tests rather than removing or changing existing ones.
+8. When you're satisfied with coverage, finalize by running:
+   <execute_bash> exit </execute_bash>
+
+Below is the **complete code snippet** to test:
+
+<START_OF_CODE>
+{code_src}
+<END_OF_CODE>
+
+NOTE: if you are testing django, you must use from django.test import SimpleTestCase and class based tests (i.e. class TestSomething(SimpleTestCase)).
+NOTE: if there is an error executing tests you MUST fix it before exiting. DO NOT install new packages.
+NOTE: if outputting a revised test suite REPLACE {test_file} with the revised suite
+
+**Output the final test suite** (20+ tests) for {test_file} in a single code block, no extra commentary. MAKE SURE you run the tests and ensure you can see which tests passed and failed BEFORE exiting.
+"""

evaluation/benchmarks/testgeneval/pygments_utils.py

@@ -0,0 +1,42 @@
+import re
+
+from pygments.lexers.python import PythonLexer
+
+
+def tokenize_code(code):
+    lexer = PythonLexer()
+    tokens = process_pygments_tokens(lexer.get_tokens(code))
+    return tokens
+
+
+def process_pygments_tokens(tokens):
+    new_tokens = []
+
+    for token in tokens:
+        if (
+            str(token[0]) == 'Token.Text'
+            and re.match(r'\s+', token[1])
+            or str(token[0]) == 'Token.Text.Whitespace'
+        ):
+            continue
+        new_tokens.append(token[1])
+
+    new_tokens_final = []
+    i = 0
+    while i < len(new_tokens) - 2:
+        if (
+            new_tokens[i] == '"'
+            and new_tokens[i + 1] == 'STR'
+            and new_tokens[i + 2] == '"'
+        ):
+            new_tokens_final.append('"STR"')
+            i = i + 3
+        else:
+            new_tokens_final.append(new_tokens[i])
+            i = i + 1
+
+    for i in range(len(new_tokens) - 2, len(new_tokens)):
+        if i >= 0:
+            new_tokens_final.append(new_tokens[i])
+
+    return new_tokens_final

evaluation/benchmarks/testgeneval/report_utils.py

@@ -0,0 +1,58 @@
+import json
+import re
+
+
+def check_coverage(coverage_output, code_file):
+    json_cov = json.loads(coverage_output)
+    if code_file in json_cov['files'].keys():
+        file_data = json_cov['files'][code_file]
+        return True, file_data['summary']['percent_covered']
+
+    return False, 0
+
+
+def check_mutation(mutation_output):
+    if 'total jobs: ' in mutation_output:
+        num_mutants = int(mutation_output.split('total jobs: ')[1].split('\n')[0])
+        final_conf = mutation_output.split('\n')[-1]
+        if len(final_conf.strip().split(' ')) == 3:
+            low, val, high = final_conf.split(' ')
+            low = float(low)
+            val = float(val)
+            high = float(high)
+
+            confidence_range = high - val
+            mutation_score = 100 - val
+
+            return True, num_mutants, mutation_score, confidence_range
+
+    return False, -1, 0, -1
+
+
+def count_methods(code_str):
+    """
+    Counts the number of methods/functions in a given string of code.
+
+    Args:
+    code_str (str): A string containing code.
+
+    Returns:
+    int: The number of methods/functions found.
+    """
+    # Regular expression to find Python function definitions
+    pattern = r'\bdef\b\s+\w+\s*\('
+    matches = re.findall(pattern, code_str)
+    return len(matches)
+
+
+def get_lines_of_code(code_str):
+    """
+    Extracts lines of code from a given string.
+
+    Args:
+    code_str (str): A string containing code.
+
+    Returns:
+    list: A list of lines of code.
+    """
+    return len(code_str.strip().split('\n'))

evaluation/benchmarks/testgeneval/run_infer.py

@@ -0,0 +1,578 @@
+import asyncio
+import json
+import os
+import tempfile
+import time
+import traceback
+from typing import Any
+
+import numpy as np
+import pandas as pd
+import toml
+from datasets import load_dataset
+
+import openhands.agenthub
+from evaluation.benchmarks.testgeneval.constants import MAP_REPO_VERSION_TO_SPECS
+from evaluation.benchmarks.testgeneval.prompt import (
+    CODEACT_TESTGEN_PROMPT,
+    CODEACT_TESTGEN_PROMPT_ITERATE,
+)
+from evaluation.benchmarks.testgeneval.utils import get_test_directives
+from evaluation.utils.shared import (
+    EvalException,
+    EvalMetadata,
+    EvalOutput,
+    assert_and_raise,
+    codeact_user_response,
+    get_metrics,
+    is_fatal_evaluation_error,
+    make_metadata,
+    prepare_dataset,
+    reset_logger_for_multiprocessing,
+    run_evaluation,
+    update_llm_config_for_completions_logging,
+)
+from openhands.controller.state.state import State
+from openhands.core.config import (
+    AgentConfig,
+    AppConfig,
+    SandboxConfig,
+    get_llm_config_arg,
+    get_parser,
+)
+from openhands.core.logger import openhands_logger as logger
+from openhands.core.main import create_runtime, run_controller
+from openhands.events.action import CmdRunAction, MessageAction
+from openhands.events.observation import CmdOutputObservation, ErrorObservation
+from openhands.events.serialization.event import event_to_dict
+from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
+
+RUN_WITH_BROWSING = os.environ.get('RUN_WITH_BROWSING', 'false').lower() == 'true'
+
+AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
+    'CodeActAgent': codeact_user_response,
+}
+
+
+def _preprocess_instance(d):
+    for key, value in d.items():
+        if isinstance(value, np.ndarray):
+            d[key] = value.tolist()
+    return d
+
+
+def _get_swebench_workspace_dir_name(instance: pd.Series) -> str:
+    return f'{instance.repo}__{instance.version}'.replace('/', '__')
+
+
+def get_instruction(instance: pd.Series, metadata: EvalMetadata):
+    # workspace_dir_name = _get_swebench_workspace_dir_name(instance)
+    # Prepare instruction
+    coverage_command = ' '.join(
+        [
+            MAP_REPO_VERSION_TO_SPECS[instance['repo']][instance['version']][
+                'test_cmd'
+            ],
+            *get_test_directives(instance),
+        ]
+    )
+
+    # Testing general agents
+    prompt_to_use = (
+        CODEACT_TESTGEN_PROMPT_ITERATE
+        if instance['full_pred'] is not None
+        else CODEACT_TESTGEN_PROMPT
+    )
+    instruction = prompt_to_use.format(
+        code_file=os.path.join('/testbed', instance.code_file),
+        test_file=os.path.join('/testbed', instance.test_file),
+        coverage_command=coverage_command,
+        code_src=instance['code_src'],
+        imports='\n'.join(instance.local_imports),
+        workspace_dir_name=_get_swebench_workspace_dir_name(instance),
+    )
+
+    if RUN_WITH_BROWSING:
+        instruction += (
+            '<IMPORTANT!>\n'
+            'You SHOULD NEVER attempt to browse the web. '
+            '</IMPORTANT!>\n'
+        )
+
+    return instruction
+
+
+# TODO: migrate all swe-bench docker to ghcr.io/openhands
+DOCKER_IMAGE_PREFIX = os.environ.get('EVAL_DOCKER_IMAGE_PREFIX', 'docker.io/kdjain/')
+logger.info(f'Using docker image prefix: {DOCKER_IMAGE_PREFIX}')
+
+
+def get_instance_docker_image(instance_id: str) -> str:
+    image_name = 'sweb.eval.x86_64.' + instance_id
+    image_name = image_name.replace(
+        '__', '_s_'
+    )  # to comply with docker image naming convention
+    return DOCKER_IMAGE_PREFIX.rstrip('/') + '/' + image_name
+
+
+def get_config(
+    instance: pd.Series,
+    metadata: EvalMetadata,
+) -> AppConfig:
+    # We use a different instance image for the each instance of TestGenEval
+    base_container_image = get_instance_docker_image(instance['instance_id_swebench'])
+    logger.info(
+        f'Using instance container image: {base_container_image}. '
+        f'Please make sure this image exists. '
+        f'Submit an issue on https://github.com/All-Hands-AI/OpenHands if you run into any issues.'
+    )
+
+    config = AppConfig(
+        default_agent=metadata.agent_class,
+        run_as_openhands=False,
+        max_iterations=metadata.max_iterations,
+        runtime=os.environ.get('RUNTIME', 'eventstream'),
+        sandbox=SandboxConfig(
+            base_container_image=base_container_image,
+            enable_auto_lint=True,
+            use_host_network=False,
+            # large enough timeout, since some testcases take very long to run
+            timeout=300,
+            # Add platform to the sandbox config to solve issue 4401
+            platform='linux/amd64',
+            api_key=os.environ.get('ALLHANDS_API_KEY', None),
+            remote_runtime_api_url=os.environ.get(
+                'SANDBOX_REMOTE_RUNTIME_API_URL', 'http://localhost:8000'
+            ),
+            keep_runtime_alive=False,
+            remote_runtime_init_timeout=3600,
+        ),
+        # do not mount workspace
+        workspace_base=None,
+        workspace_mount_path=None,
+    )
+    config.set_llm_config(
+        update_llm_config_for_completions_logging(
+            metadata.llm_config, metadata.eval_output_dir, instance['id']
+        )
+    )
+    agent_config = AgentConfig(
+        codeact_enable_jupyter=False,
+        codeact_enable_browsing=RUN_WITH_BROWSING,
+        codeact_enable_llm_editor=False,
+        condenser=metadata.condenser_config,
+        enable_prompt_extensions=False,
+    )
+    config.set_agent_config(agent_config)
+    return config
+
+
+def initialize_runtime(
+    runtime: Runtime,
+    instance: pd.Series,  # this argument is not required
+):
+    """Initialize the runtime for the agent.
+
+    This function is called before the runtime is used to run the agent.
+    """
+    logger.info('-' * 30)
+    logger.info('BEGIN Runtime Initialization Fn')
+    logger.info('-' * 30)
+    workspace_dir_name = _get_swebench_workspace_dir_name(instance)
+    obs: CmdOutputObservation
+
+    instance['instance_id'] = instance['instance_id_swebench']
+
+    # Set instance id
+    action = CmdRunAction(
+        command=f"""echo 'export SWE_INSTANCE_ID={instance['instance_id_swebench']}' >> ~/.bashrc && echo 'export PIP_CACHE_DIR=~/.cache/pip' >> ~/.bashrc && echo "alias git='git --no-pager'" >> ~/.bashrc"""
+    )
+    action.set_hard_timeout(600)
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    assert_and_raise(
+        obs.exit_code == 0, f'Failed to export SWE_INSTANCE_ID: {str(obs)}'
+    )
+
+    action = CmdRunAction(command="""export USER=$(whoami); echo USER=${USER} """)
+    action.set_hard_timeout(600)
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    assert_and_raise(obs.exit_code == 0, f'Failed to export USER: {str(obs)}')
+
+    # inject the init script
+    script_dir = os.path.dirname(__file__)
+
+    # inject the instance info
+    action = CmdRunAction(command='mkdir -p /swe_util/eval_data/instances')
+    action.set_hard_timeout(600)
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    assert_and_raise(
+        obs.exit_code == 0,
+        f'Failed to create /swe_util/eval_data/instances: {str(obs)}',
+    )
+
+    swe_instance_json_name = 'swe-bench-instance.json'
+    swe_prediction = 'test_suite.py'
+    with tempfile.TemporaryDirectory() as temp_dir:
+        # Construct the full path for the desired file name within the temporary directory
+        temp_file_path = os.path.join(temp_dir, swe_instance_json_name)
+        # Write to the file with the desired name within the temporary directory
+        with open(temp_file_path, 'w') as f:
+            if not isinstance(instance, dict):
+                preprocessed_instance = _preprocess_instance(instance.to_dict())
+                json.dump([preprocessed_instance], f)
+            else:
+                preprocessed_instance = _preprocess_instance(instance)
+                json.dump([preprocessed_instance], f)
+
+        # Copy the file to the desired location
+        runtime.copy_to(temp_file_path, '/swe_util/eval_data/instances/')
+
+        if instance['full_pred'] is not None:
+            temp_file_path_pred = os.path.join(temp_dir, swe_prediction)
+            with open(temp_file_path_pred, 'w') as f:
+                f.write(instance['full_pred'])
+
+            runtime.copy_to(temp_file_path_pred, '/tmp')
+
+            # Copy the file to the desired location
+            action = CmdRunAction(
+                command=f"cp /tmp/test_suite.py /testbed/{instance['test_file']}"
+            )
+            action.set_hard_timeout(600)
+            logger.info(action, extra={'msg_type': 'ACTION'})
+            obs = runtime.run_action(action)
+            logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+            assert_and_raise(
+                obs.exit_code == 0, f'Failed to copy test file: {str(obs)}'
+            )
+
+            action = CmdRunAction(
+                command='git -C /testbed add . && git -C /testbed commit -m "Add test file"'
+            )
+            action.set_hard_timeout(600)
+            logger.info(action, extra={'msg_type': 'ACTION'})
+            obs = runtime.run_action(action)
+            logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+            assert_and_raise(obs.exit_code == 0, f'Failed to cat ~/.bashrc: {str(obs)}')
+
+    # inject the instance swe entry
+    runtime.copy_to(
+        str(os.path.join(script_dir, 'scripts/setup/instance_swe_entry.sh')),
+        '/swe_util/',
+    )
+    action = CmdRunAction(command='cat ~/.bashrc')
+    action.set_hard_timeout(600)
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    assert_and_raise(obs.exit_code == 0, f'Failed to cat ~/.bashrc: {str(obs)}')
+
+    action = CmdRunAction(command='source ~/.bashrc')
+    action.set_hard_timeout(600)
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    if isinstance(obs, ErrorObservation):
+        logger.error(f'Failed to source ~/.bashrc: {str(obs)}')
+    assert_and_raise(obs.exit_code == 0, f'Failed to source ~/.bashrc: {str(obs)}')
+
+    action = CmdRunAction(command='source /swe_util/instance_swe_entry.sh')
+    action.set_hard_timeout(600)
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    assert_and_raise(
+        obs.exit_code == 0,
+        f'Failed to source /swe_util/instance_swe_entry.sh: {str(obs)}',
+    )
+
+    action = CmdRunAction(command=f'cd /workspace/{workspace_dir_name}')
+    action.set_hard_timeout(600)
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    assert_and_raise(
+        obs.exit_code == 0,
+        f'Failed to cd to /workspace/{workspace_dir_name}: {str(obs)}',
+    )
+
+    action = CmdRunAction(command='git reset --hard')
+    action.set_hard_timeout(600)
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    assert_and_raise(obs.exit_code == 0, f'Failed to git reset --hard: {str(obs)}')
+
+    action = CmdRunAction(
+        command='for remote_name in $(git remote); do git remote remove "${remote_name}"; done'
+    )
+    action.set_hard_timeout(600)
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    assert_and_raise(obs.exit_code == 0, f'Failed to remove git remotes: {str(obs)}')
+
+    logger.info('-' * 30)
+    logger.info('END Runtime Initialization Fn')
+    logger.info('-' * 30)
+
+
+def complete_runtime(
+    runtime: Runtime,
+    instance: pd.Series,  # this argument is not required, but it is used to get the workspace_dir_name
+) -> dict[str, Any]:
+    """Complete the runtime for the agent.
+
+    This function is called before the runtime is used to run the agent.
+    If you need to do something in the sandbox to get the correctness metric after
+    the agent has run, modify this function.
+    """
+    try:
+        logger.info('-' * 30)
+        logger.info('BEGIN Runtime Completion Fn')
+        logger.info('-' * 30)
+        obs: CmdOutputObservation
+        workspace_dir_name = _get_swebench_workspace_dir_name(instance)
+
+        action = CmdRunAction(command=f'cd /workspace/{workspace_dir_name}')
+        action.set_hard_timeout(600)
+        logger.info(action, extra={'msg_type': 'ACTION'})
+        obs = runtime.run_action(action)
+        logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+        assert_and_raise(
+            obs.exit_code == 0,
+            f'Failed to cd to /workspace/{workspace_dir_name}: {str(obs)}',
+        )
+
+        action = CmdRunAction(command=f'cat {instance.test_file}')
+        action.set_hard_timeout(600)
+        logger.info(action, extra={'msg_type': 'ACTION'})
+        obs = runtime.run_action(action)
+        logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+        assert_and_raise(
+            obs.exit_code == 0,
+            f'Failed to find file: {instance.test_file} in /workspace/{workspace_dir_name}',
+        )
+
+        test_suite = obs.content.strip()
+    except Exception:
+        # Print stack trace
+        print('Skipping, exception in complete_runtime')
+        print(traceback.format_exc())
+        test_suite = instance['full_pred'] if instance['full_pred'] is not None else ''
+
+    # action = CmdRunAction(command='git add -A')
+    # action.set_hard_timeout(600)
+    # logger.info(action, extra={'msg_type': 'ACTION'})
+    # obs = runtime.run_action(action)
+    # logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    # assert_and_raise(obs.exit_code == 0, f'Failed to git add -A: {str(obs)}')
+
+    logger.info('-' * 30)
+    logger.info('END Runtime Completion Fn')
+    logger.info('-' * 30)
+    return {
+        'test_suite': test_suite,
+    }
+
+
+def process_instance(
+    instance: pd.Series,
+    metadata: EvalMetadata,
+    reset_logger: bool = True,
+) -> EvalOutput:
+    config = get_config(instance, metadata)
+    start_time = time.time()  # Track start time
+
+    # Setup the logger properly, so you can run multi-processing to parallelize the evaluation
+    if reset_logger:
+        log_dir = os.path.join(metadata.eval_output_dir, 'infer_logs')
+        reset_logger_for_multiprocessing(logger, instance.id, log_dir)
+    else:
+        logger.info(f'Starting evaluation for instance {instance.id}.')
+
+    runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
+
+    try:
+        initialize_runtime(runtime, instance)
+
+        instruction = get_instruction(instance, metadata)
+
+        # Here's how you can run the agent (similar to the `main` function) and get the final task state
+        state: State | None = asyncio.run(
+            run_controller(
+                config=config,
+                initial_user_action=MessageAction(content=instruction),
+                runtime=runtime,
+                fake_user_response_fn=AGENT_CLS_TO_FAKE_USER_RESPONSE_FN[
+                    metadata.agent_class
+                ],
+            )
+        )
+
+        # if fatal error, throw EvalError to trigger re-run
+        if is_fatal_evaluation_error(state.last_error):
+            raise EvalException('Fatal error detected: ' + state.last_error)
+
+        # ======= THIS IS SWE-Bench specific =======
+        return_val = complete_runtime(runtime, instance)
+        test_suite = return_val['test_suite']
+        logger.info(
+            f'Got test suite for instance {instance.instance_id}:\n--------\n{test_suite}\n--------'
+        )
+    finally:
+        runtime.close()
+
+    end_time = time.time()
+    elapsed_time = end_time - start_time
+    logger.info(
+        f'Evaluation for instance {instance.instance_id} took {elapsed_time:.2f} seconds.'
+    )
+
+    # ==========================================
+
+    # ======= Attempt to evaluate the agent's edits =======
+    # we use eval_infer.sh to evaluate the agent's edits, not here
+    # because the agent may alter the environment / testcases
+    test_result = {
+        'test_suite': test_suite,
+        'elapsed_time': elapsed_time,
+    }
+
+    # If you are working on some simpler benchmark that only evaluates the final model output (e.g., in a MessageAction)
+    # You can simply get the LAST `MessageAction` from the returned `state.history` and parse it for evaluation.
+    if state is None:
+        raise ValueError('State should not be None.')
+
+    histories = [event_to_dict(event) for event in state.history]
+    metrics = get_metrics(state)
+
+    # Save the output
+    output = EvalOutput(
+        instance_id=instance.id,
+        instruction=instruction,
+        instance=_preprocess_instance(instance.to_dict()),  # SWE Bench specific
+        test_result=test_result,
+        metadata=metadata,
+        history=histories,
+        metrics=metrics,
+        error=state.last_error if state and state.last_error else None,
+    )
+    # print(output)
+    return output
+
+
+def prepare_dataset_pre(dataset: pd.DataFrame, filter_column: str) -> pd.DataFrame:
+    file_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'config.toml')
+    if os.path.exists(file_path):
+        with open(file_path, 'r') as file:
+            data = toml.load(file)
+            if 'selected_ids' in data:
+                selected_ids = data['selected_ids']
+                logger.info(
+                    f'Filtering {len(selected_ids)} tasks from "selected_ids"...'
+                )
+                subset = dataset[dataset[filter_column].isin(selected_ids)]
+                logger.info(f'Retained {subset.shape[0]} tasks after filtering')
+
+                subset['instance_id_swebench'] = subset['instance_id']
+                subset['instance_id'] = subset['id']
+                return subset
+
+    dataset['instance_id_swebench'] = dataset['instance_id']
+    dataset['instance_id'] = dataset['id']
+    return dataset
+
+
+if __name__ == '__main__':
+    parser = get_parser()
+    parser.add_argument(
+        '--dataset',
+        type=str,
+        default='kjain/testgenevallite',
+        help='data set to evaluate on, either full-test or lite-test',
+    )
+    parser.add_argument(
+        '--split',
+        type=str,
+        default='test',
+        help='split to evaluate on',
+    )
+    parser.add_argument(
+        '--testfile_start',
+        action='store_true',
+        help='Whether to start from the 0 shot test file',
+    )
+
+    parser.add_argument(
+        '--zero_shot_path',
+        type=str,
+        help='Path to the zero shot test file predictions',
+    )
+    args, _ = parser.parse_known_args()
+
+    if args.testfile_start and not args.zero_shot_path:
+        raise ValueError(
+            'If you want to start from the 0 shot test file, you must provide the path to the zero shot test file predictions'
+        )
+
+    preds_map = {}
+    if args.testfile_start:
+        with open(args.zero_shot_path, 'r') as f:
+            for line in f:
+                pred = json.loads(line)
+                preds_map[pred['id']] = pred['preds']['full'][0]
+
+    # NOTE: It is preferable to load datasets from huggingface datasets and perform post-processing
+    # so we don't need to manage file uploading to OpenHands's repo
+    dataset = load_dataset(args.dataset, split=args.split)
+    logger.info(f'Loaded dataset {args.dataset} with split {args.split}')
+    testgeneval_filepairs = prepare_dataset_pre(dataset.to_pandas(), 'id')
+
+    llm_config = None
+    if args.llm_config:
+        llm_config = get_llm_config_arg(args.llm_config)
+        llm_config.log_completions = True
+        # modify_params must be False for evaluation purpose, for reproducibility and accurancy of results
+        llm_config.modify_params = False
+
+    if llm_config is None:
+        raise ValueError(f'Could not find LLM config: --llm_config {args.llm_config}')
+
+    details = {}
+    _agent_cls = openhands.agenthub.Agent.get_cls(args.agent_cls)
+
+    dataset_descrption = (
+        args.dataset.replace('/', '__') + '-' + args.split.replace('/', '__')
+    )
+    metadata = make_metadata(
+        llm_config,
+        dataset_descrption,
+        args.agent_cls,
+        args.max_iterations,
+        args.eval_note,
+        args.eval_output_dir,
+        details=details,
+    )
+
+    output_file = os.path.join(metadata.eval_output_dir, 'output.jsonl')
+    instances = prepare_dataset(testgeneval_filepairs, output_file, args.eval_n_limit)
+
+    if not instances.empty:
+        instances['full_pred'] = (
+            instances['instance_id']
+            .map(preds_map)
+            .apply(lambda x: x if pd.notna(x) else None)
+        )
+
+        run_evaluation(
+            instances, metadata, output_file, args.eval_num_workers, process_instance
+        )

evaluation/benchmarks/testgeneval/scripts/docker/add_testing_dependencies.py

@@ -0,0 +1,128 @@
+import argparse
+import os
+import subprocess
+
+from datasets import load_dataset
+
+
+# Function to run shell commands
+def run_command(command):
+    try:
+        subprocess.run(command, check=True, shell=True)
+    except subprocess.CalledProcessError as e:
+        print(f'An error occurred: {e}')
+
+
+# Function to log in to Docker Hub
+def docker_login():
+    print('Logging into Docker Hub...')
+    run_command('docker login')
+
+
+# Function to generate Dockerfile content based on image type
+def generate_dockerfile_content(
+    base_image, dependencies, datum, patch_path, test_patch_path
+):
+    dockerfile_content = f"""
+FROM {base_image}
+SHELL ["/bin/bash", "-c"]
+RUN source /opt/miniconda3/bin/activate && conda activate testbed && pip install {' '.join(dependencies)}
+COPY {patch_path} /app/patch.diff
+RUN git apply /app/patch.diff
+RUN rm /app/patch.diff
+COPY {test_patch_path} /app/patch.diff
+RUN git apply /app/patch.diff
+RUN git config --global user.email ""
+RUN git config --global user.name "TestGenEval"
+RUN rm /app/patch.diff
+RUN rm {datum['test_file']}
+"""
+
+    # Add specific content based on image type
+    dockerfile_content += 'RUN git add .\nRUN git commit -m "Testing fixes"'
+
+    return dockerfile_content
+
+
+# Function to build, push, and clean up Docker images
+def build_and_push_image(dockerfile_content, image_name):
+    with open('Dockerfile.temp', 'w') as dockerfile:
+        dockerfile.write(dockerfile_content)
+    run_command(f'docker build -f Dockerfile.temp -t {image_name} .')
+    run_command(f'docker push {image_name}')
+    run_command(f'docker rmi {image_name}')
+    os.remove('Dockerfile.temp')
+
+
+# Function to process images with .eval in the name
+def process_images(dataset, original_namespace, new_namespace, start_instance_id):
+    dependencies = ['coverage', 'cosmic-ray']
+
+    found_start = len(start_instance_id) == 0
+    for datum in dataset:
+        if not found_start and datum['instance_id'] == start_instance_id:
+            found_start = True
+        elif found_start:
+            full_image_name = f'{original_namespace}/sweb.eval.x86_64.{datum["instance_id"].replace("__", "_s_")}:latest'
+            print(f'Processing image: {full_image_name}')
+            run_command(f'docker pull {full_image_name}')
+
+            # Save patches and preds_context to regular files
+            patch_file_path = 'patch.diff'
+            test_patch_file_path = 'test_patch.diff'
+
+            with open(patch_file_path, 'w') as patch_file, open(
+                test_patch_file_path, 'w'
+            ) as test_patch_file:
+                patch_file.write(datum['patch'])
+                test_patch_file.write(datum['test_patch'])
+
+            # Define image types and corresponding tags
+            new_image_name = f'{new_namespace}/sweb.eval.x86_64.{datum["instance_id"].replace("__", "_s_")}:latest'
+            dockerfile_content = generate_dockerfile_content(
+                full_image_name,
+                dependencies,
+                datum,
+                patch_file_path,
+                test_patch_file_path,
+            )
+            build_and_push_image(dockerfile_content, new_image_name)
+
+            # Cleanup regular files and images
+            os.remove(patch_file_path)
+            os.remove(test_patch_file_path)
+            run_command(f'docker rmi {full_image_name}')
+            run_command('docker system prune -f')  # Clean up dangling resources
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(
+        description='Process Docker images with .eval in the name.'
+    )
+    parser.add_argument('--dataset', type=str, default='kjain14/testgeneval')
+    parser.add_argument('--split', type=str, default='test')
+    parser.add_argument(
+        '--new_namespace',
+        type=str,
+        default='kdjain',
+        help='The new Docker Hub namespace to push the images',
+    )
+    parser.add_argument(
+        '--original_namespace',
+        type=str,
+        default='xingyaoww',
+        help='The original Docker Hub namespace',
+    )
+    parser.add_argument(
+        '--start_instance_id',
+        type=str,
+        default='',
+        help='The instance_id to start processing from',
+    )
+    args = parser.parse_args()
+    dataset = load_dataset(args.dataset)[args.split]
+
+    docker_login()
+    process_images(
+        dataset, args.original_namespace, args.new_namespace, args.start_instance_id
+    )

evaluation/benchmarks/testgeneval/scripts/docker/all-swebench-lite-instance-images.txt

@@ -0,0 +1,196 @@
+sweb.base.x86_64:latest
+sweb.env.x86_64.088a7e628bda9770f9757b:latest
+sweb.env.x86_64.0d80c7dec81ee2f2f513e2:latest
+sweb.env.x86_64.0f99bce2750f3109957bec:latest
+sweb.env.x86_64.1b3b218535da0abf4469cb:latest
+sweb.env.x86_64.1c1a6945f732f9391228c5:latest
+sweb.env.x86_64.1f92e6d7cef88badc4f744:latest
+sweb.env.x86_64.27dd9791e13f5c857a09f9:latest
+sweb.env.x86_64.297af196949a2a635bce66:latest
+sweb.env.x86_64.2baaea72acc974f6c02079:latest
+sweb.env.x86_64.2e50125951bc69cddd7421:latest
+sweb.env.x86_64.2f217c8b4490bfa0e2ba14:latest
+sweb.env.x86_64.31244378a92e3bcce809ac:latest
+sweb.env.x86_64.428468730904ff6b4232aa:latest
+sweb.env.x86_64.5d1fda9d55d65d8a4e5bdb:latest
+sweb.env.x86_64.6b007979cf533f0f3016e8:latest
+sweb.env.x86_64.7037e8c448a4b8ebfe9b13:latest
+sweb.env.x86_64.71498c7426dbf05599642f:latest
+sweb.env.x86_64.756beac07713d7e8dc1129:latest
+sweb.env.x86_64.78278ae2cf880e395f1337:latest
+sweb.env.x86_64.8f1f7b974f0c57c7aeba39:latest
+sweb.env.x86_64.934a137824256b612e9dc5:latest
+sweb.env.x86_64.a0efca7a0fe6719dbf65c2:latest
+sweb.env.x86_64.a18371b03f944585b4f08c:latest
+sweb.env.x86_64.a33dddf55cdff5d8e23374:latest
+sweb.env.x86_64.aa92880033da20ca313928:latest
+sweb.env.x86_64.b649f0ff62fad147f7f073:latest
+sweb.env.x86_64.b7ce4be3b3c35f68c61248:latest
+sweb.env.x86_64.c70909fdac4897d1c685df:latest
+sweb.env.x86_64.c795f4b88616b8462021ed:latest
+sweb.env.x86_64.cc47cc71483942d0c3a15e:latest
+sweb.env.x86_64.dc5ff4c0e3fe8db5afc4da:latest
+sweb.env.x86_64.e3afd7f04b325a4de4982d:latest
+sweb.env.x86_64.e5bb89bf78258a7d14c34b:latest
+sweb.env.x86_64.e83e37f52c09532c62acfb:latest
+sweb.env.x86_64.efa6065ed5bf204410fd53:latest
+sweb.eval.x86_64.django_s_django-17087:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-10508:latest
+sweb.eval.x86_64.django_s_django-14017:latest
+sweb.eval.x86_64.django_s_django-11422:latest
+sweb.eval.x86_64.sympy_s_sympy-14774:latest
+sweb.eval.x86_64.django_s_django-14915:latest
+sweb.eval.x86_64.sympy_s_sympy-22005:latest
+sweb.eval.x86_64.pytest-dev_s_pytest-5221:latest
+sweb.eval.x86_64.sympy_s_sympy-17022:latest
+sweb.eval.x86_64.django_s_django-15996:latest
+sweb.eval.x86_64.django_s_django-15252:latest
+sweb.eval.x86_64.sympy_s_sympy-21171:latest
+sweb.eval.x86_64.django_s_django-11797:latest
+sweb.eval.x86_64.django_s_django-16046:latest
+sweb.eval.x86_64.django_s_django-11583:latest
+sweb.eval.x86_64.django_s_django-15738:latest
+sweb.eval.x86_64.sympy_s_sympy-21612:latest
+sweb.eval.x86_64.astropy_s_astropy-12907:latest
+sweb.eval.x86_64.django_s_django-11620:latest
+sweb.eval.x86_64.sympy_s_sympy-16792:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-13779:latest
+sweb.eval.x86_64.django_s_django-16041:latest
+sweb.eval.x86_64.sympy_s_sympy-13471:latest
+sweb.eval.x86_64.sympy_s_sympy-20442:latest
+sweb.eval.x86_64.sympy_s_sympy-20049:latest
+sweb.eval.x86_64.django_s_django-14411:latest
+sweb.eval.x86_64.django_s_django-13447:latest
+sweb.eval.x86_64.django_s_django-12856:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-10949:latest
+sweb.eval.x86_64.django_s_django-14787:latest
+sweb.eval.x86_64.django_s_django-11815:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-13584:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-14087:latest
+sweb.eval.x86_64.django_s_django-15388:latest
+sweb.eval.x86_64.django_s_django-11179:latest
+sweb.eval.x86_64.sympy_s_sympy-24102:latest
+sweb.eval.x86_64.sympy_s_sympy-24213:latest
+sweb.eval.x86_64.django_s_django-15781:latest
+sweb.eval.x86_64.pytest-dev_s_pytest-8906:latest
+sweb.eval.x86_64.django_s_django-13710:latest
+sweb.eval.x86_64.django_s_django-13925:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-14092:latest
+sweb.eval.x86_64.pytest-dev_s_pytest-7373:latest
+sweb.eval.x86_64.matplotlib_s_matplotlib-25498:latest
+sweb.eval.x86_64.pytest-dev_s_pytest-5227:latest
+sweb.eval.x86_64.sympy_s_sympy-15678:latest
+sweb.eval.x86_64.django_s_django-13551:latest
+sweb.eval.x86_64.django_s_django-14155:latest
+sweb.eval.x86_64.django_s_django-13933:latest
+sweb.eval.x86_64.sympy_s_sympy-21055:latest
+sweb.eval.x86_64.django_s_django-13660:latest
+sweb.eval.x86_64.django_s_django-16527:latest
+sweb.eval.x86_64.pytest-dev_s_pytest-5692:latest
+sweb.eval.x86_64.mwaskom_s_seaborn-3010:latest
+sweb.eval.x86_64.django_s_django-12700:latest
+sweb.eval.x86_64.sympy_s_sympy-11400:latest
+sweb.eval.x86_64.sympy_s_sympy-23117:latest
+sweb.eval.x86_64.sympy_s_sympy-20639:latest
+sweb.eval.x86_64.sympy_s_sympy-23262:latest
+sweb.eval.x86_64.django_s_django-15498:latest
+sweb.eval.x86_64.django_s_django-12453:latest
+sweb.eval.x86_64.django_s_django-14999:latest
+sweb.eval.x86_64.sympy_s_sympy-13480:latest
+sweb.eval.x86_64.sympy_s_sympy-21847:latest
+sweb.eval.x86_64.sympy_s_sympy-15011:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-25570:latest
+sweb.eval.x86_64.sphinx-doc_s_sphinx-7975:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-14983:latest
+sweb.eval.x86_64.django_s_django-14534:latest
+sweb.eval.x86_64.sympy_s_sympy-14396:latest
+sweb.eval.x86_64.matplotlib_s_matplotlib-25442:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-15535:latest
+sweb.eval.x86_64.sympy_s_sympy-22714:latest
+sweb.eval.x86_64.django_s_django-15789:latest
+sweb.eval.x86_64.sympy_s_sympy-21627:latest
+sweb.eval.x86_64.sympy_s_sympy-24066:latest
+sweb.eval.x86_64.pylint-dev_s_pylint-7993:latest
+sweb.eval.x86_64.django_s_django-14752:latest
+sweb.eval.x86_64.sympy_s_sympy-18835:latest
+sweb.eval.x86_64.django_s_django-17051:latest
+sweb.eval.x86_64.sympy_s_sympy-12171:latest
+sweb.eval.x86_64.pydata_s_xarray-3364:latest
+sweb.eval.x86_64.mwaskom_s_seaborn-3190:latest
+sweb.eval.x86_64.pytest-dev_s_pytest-7168:latest
+sweb.eval.x86_64.django_s_django-12747:latest
+sweb.eval.x86_64.django_s_django-15695:latest
+sweb.eval.x86_64.matplotlib_s_matplotlib-22835:latest
+sweb.eval.x86_64.sympy_s_sympy-12481:latest
+sweb.eval.x86_64.django_s_django-15851:latest
+sweb.eval.x86_64.sympy_s_sympy-14024:latest
+sweb.eval.x86_64.django_s_django-14608:latest
+sweb.eval.x86_64.pytest-dev_s_pytest-9359:latest
+sweb.eval.x86_64.django_s_django-16873:latest
+sweb.eval.x86_64.matplotlib_s_matplotlib-25433:latest
+sweb.eval.x86_64.sympy_s_sympy-13031:latest
+sweb.eval.x86_64.pytest-dev_s_pytest-7432:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-25747:latest
+sweb.eval.x86_64.django_s_django-12286:latest
+sweb.eval.x86_64.django_s_django-11910:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-12471:latest
+sweb.eval.x86_64.pylint-dev_s_pylint-5859:latest
+sweb.eval.x86_64.django_s_django-11133:latest
+sweb.eval.x86_64.astropy_s_astropy-14365:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-13496:latest
+sweb.eval.x86_64.sympy_s_sympy-19487:latest
+sweb.eval.x86_64.sympy_s_sympy-13895:latest
+sweb.eval.x86_64.sympy_s_sympy-15345:latest
+sweb.eval.x86_64.django_s_django-13590:latest
+sweb.eval.x86_64.django_s_django-13757:latest
+sweb.eval.x86_64.django_s_django-16379:latest
+sweb.eval.x86_64.django_s_django-13768:latest
+sweb.eval.x86_64.pytest-dev_s_pytest-8365:latest
+sweb.eval.x86_64.django_s_django-14580:latest
+sweb.eval.x86_64.sympy_s_sympy-20154:latest
+sweb.eval.x86_64.sympy_s_sympy-12419:latest
+sweb.eval.x86_64.django_s_django-12125:latest
+sweb.eval.x86_64.sympy_s_sympy-24152:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-15512:latest
+sweb.eval.x86_64.sympy_s_sympy-18621:latest
+sweb.eval.x86_64.pydata_s_xarray-4248:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-11040:latest
+sweb.eval.x86_64.django_s_django-11099:latest
+sweb.eval.x86_64.django_s_django-16816:latest
+sweb.eval.x86_64.django_s_django-13265:latest
+sweb.eval.x86_64.django_s_django-16139:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-10297:latest
+sweb.eval.x86_64.django_s_django-14016:latest
+sweb.eval.x86_64.pallets_s_flask-5063:latest
+sweb.eval.x86_64.astropy_s_astropy-7746:latest
+sweb.eval.x86_64.matplotlib_s_matplotlib-24265:latest
+sweb.eval.x86_64.django_s_django-13448:latest
+sweb.eval.x86_64.django_s_django-12908:latest
+sweb.eval.x86_64.sphinx-doc_s_sphinx-8627:latest
+sweb.eval.x86_64.sympy_s_sympy-14317:latest
+sweb.eval.x86_64.pytest-dev_s_pytest-6116:latest
+sweb.eval.x86_64.sympy_s_sympy-23191:latest
+sweb.eval.x86_64.pydata_s_xarray-5131:latest
+sweb.eval.x86_64.django_s_django-11019:latest
+sweb.eval.x86_64.matplotlib_s_matplotlib-23913:latest
+sweb.eval.x86_64.django_s_django-15790:latest
+sweb.eval.x86_64.django_s_django-12497:latest
+sweb.eval.x86_64.matplotlib_s_matplotlib-26020:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-25638:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-25500:latest
+sweb.eval.x86_64.sympy_s_sympy-19007:latest
+sweb.eval.x86_64.django_s_django-12308:latest
+sweb.eval.x86_64.pytest-dev_s_pytest-7220:latest
+sweb.eval.x86_64.django_s_django-11848:latest
+sweb.eval.x86_64.django_s_django-15347:latest
+sweb.eval.x86_64.pytest-dev_s_pytest-7490:latest
+sweb.eval.x86_64.sympy_s_sympy-18532:latest
+sweb.eval.x86_64.django_s_django-14997:latest
+sweb.eval.x86_64.sympy_s_sympy-24909:latest
+sweb.eval.x86_64.django_s_django-13220:latest
+sweb.eval.x86_64.sympy_s_sympy-21614:latest
+sweb.eval.x86_64.django_s_django-15902:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-13497:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-13439:latest
+sweb.eval.x86_64.scikit-learn_s_scikit-learn-14894:latest
+sweb.eval.x86_64.django_s_django-12983:latest

evaluation/benchmarks/testgeneval/scripts/docker/compare_txt_files.py

@@ -0,0 +1,31 @@
+def print_diff_ignore_order(file1, file2):
+    with open(file1, 'r') as f1, open(file2, 'r') as f2:
+        file1_lines = set(f1.readlines())
+        file2_lines = set(f2.readlines())
+
+    only_in_file1 = file1_lines - file2_lines
+    only_in_file2 = file2_lines - file1_lines
+
+    if only_in_file1:
+        print(f'Lines in {file1} but not in {file2}:')
+        for line in sorted(only_in_file1):
+            print(f'- {line.strip()}')
+
+    # if only_in_file2:
+    #     print(f"Lines in {file2} but not in {file1}:")
+    #     for line in sorted(only_in_file2):
+    #         print(f"+ {line.strip()}")
+
+    if not only_in_file1 and not only_in_file2:
+        print('The files have the same content (ignoring line order).')
+
+
+if __name__ == '__main__':
+    # Usage
+    lite1 = 'all-swebench-lite-instance-images.txt'  # Replace with the path to your first file
+    lite2 = '../../swe_bench/scripts/docker/all-swebench-lite-instance-images.txt'  # Replace with the path to your second file
+    print_diff_ignore_order(lite1, lite2)
+
+    full1 = 'all-swebench-full-instance-images.txt'  # Replace with the path to your first file
+    full2 = '../../swe_bench/scripts/docker/all-swebench-full-instance-images.txt'  # Replace with the path to your second file
+    print_diff_ignore_order(full1, full2)

evaluation/benchmarks/testgeneval/scripts/docker/delete_all_images.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+# Script will delete all repositories and tags in your Docker Hub account
+set -e
+
+# Set username and password from command-line arguments
+UNAME=$1
+UPASS=$2
+
+# Get token to interact with Docker Hub
+TOKEN=$(curl -s -H "Content-Type: application/json" -X POST -d '{"username": "'${UNAME}'", "password": "'${UPASS}'"}' https://hub.docker.com/v2/users/login/ | jq -r .token)
+
+# Ensure token retrieval was successful
+if [[ -z "$TOKEN" ]]; then
+    echo "Failed to obtain authentication token. Please check your credentials."
+    exit 1
+fi
+
+# Get list of repositories for that user account
+echo "Listing repositories in Docker Hub account '${UNAME}':"
+REPO_LIST=$(curl -s -H "Authorization: JWT ${TOKEN}" "https://hub.docker.com/v2/repositories/${UNAME}/?page_size=10000" | jq -r '.results|.[]|.name')
+if [[ -z "$REPO_LIST" ]]; then
+    echo "No repositories found for user '${UNAME}' or failed to fetch repositories."
+    exit 1
+fi
+
+# Loop through each repository and delete its tags and the repository itself
+for rep in ${REPO_LIST}; do
+    echo "Processing repository: ${UNAME}/${rep}"
+
+    # Get all tags for the repository
+    IMAGES=$(curl -s -H "Authorization: JWT ${TOKEN}" "https://hub.docker.com/v2/repositories/${UNAME}/${rep}/tags/?page_size=100")
+    IMAGE_TAGS=$(echo $IMAGES | jq -r '.results|.[]|.name')
+
+    # Delete each tag
+    for tag in ${IMAGE_TAGS}; do
+        echo "Deleting tag: ${UNAME}/${rep}:${tag}"
+        curl -s -X DELETE -H "Authorization: JWT ${TOKEN}" "https://hub.docker.com/v2/repositories/${UNAME}/${rep}/tags/${tag}/"
+    done
+
+    # Delete the repository itself
+    echo "Deleting repository: ${UNAME}/${rep}"
+    curl -s -X DELETE -H "Authorization: JWT ${TOKEN}" "https://hub.docker.com/v2/repositories/${UNAME}/${rep}/" || {
+        echo "Failed to delete repository '${UNAME}/${rep}'. Please check permissions or API limits."
+    }
+    sleep 1
+done
+
+echo "Script execution completed."

evaluation/benchmarks/testgeneval/scripts/docker/hf_to_image_list.py

@@ -0,0 +1,18 @@
+from datasets import load_dataset
+
+
+def dataset_to_txt(dataset, txt_file, split='test'):
+    with open(txt_file, 'w') as f:
+        for datum in dataset[split]:
+            instance_id = datum['instance_id'].replace('__', '_s_')
+            f.write(f'sweb.eval.x86_64.{instance_id}:latest\n')
+
+
+if __name__ == '__main__':
+    # Load the private dataset
+    dataset = load_dataset('kjain14/testgeneval')
+
+    dataset_lite = load_dataset('kjain14/testgenevallite')
+
+    dataset_to_txt(dataset_lite, 'all-swebench-lite-instance-images.txt', lite=True)
+    dataset_to_txt(dataset, 'all-swebench-full-instance-images.txt')

evaluation/benchmarks/testgeneval/scripts/eval/build_outputs_ablation.py

@@ -0,0 +1,173 @@
+import argparse
+import copy
+import difflib
+import json
+import os
+import traceback
+
+
+def insert_line_in_string(input_string, new_str, insert_line):
+    """
+    Inserts a new line into a string at the specified line number.
+
+    :param input_string: The original string.
+    :param new_str: The string to insert.
+    :param insert_line: The line number at which to insert (1-based index).
+    :return: The modified string.
+    """
+    file_text = input_string.expandtabs()
+    new_str = new_str.expandtabs()
+
+    file_text_lines = file_text.split('\n')
+
+    new_str_lines = new_str.split('\n')
+    new_file_text_lines = (
+        file_text_lines[:insert_line] + new_str_lines + file_text_lines[insert_line:]
+    )
+
+    return '\n'.join(new_file_text_lines)
+
+
+def print_string_diff(original, modified):
+    """
+    Prints the differences between two strings line by line.
+
+    :param original: The original string.
+    :param modified: The modified string.
+    """
+    original_lines = original.splitlines(keepends=True)
+    modified_lines = modified.splitlines(keepends=True)
+
+    diff = difflib.unified_diff(
+        original_lines,
+        modified_lines,
+        fromfile='original',
+        tofile='modified',
+        lineterm='',
+    )
+
+    print(''.join(diff))
+
+
+def parse_json_files(root_dir, output_dir, metadata_objs, preds_objs):
+    final_output = {i: [] for i in range(25)}
+
+    for subdir in sorted(os.listdir(root_dir)):  # Sorting ensures consistent order
+        subdir_path = os.path.join(root_dir, subdir)
+        # subdir_instance = subdir.rsplit('-', 1)[0]
+        metadata = metadata_objs[subdir]
+        orig_test_suite = metadata['test_result']['test_suite']
+
+        if os.path.isdir(subdir_path):  # Check if it's a directory
+            print(f'Processing subdirectory: {subdir}')
+
+            # Now loop through the JSON files in this subdirectory
+            i = 0
+            test_suite = preds_objs[subdir] if subdir in preds_objs else ''
+            for file in sorted(
+                os.listdir(subdir_path)
+            ):  # Sorting ensures consistent order
+                metadata_copy = copy.deepcopy(metadata)
+                if file.endswith('.json'):  # Check for JSON files
+                    file_path = os.path.join(subdir_path, file)
+                    try:
+                        with open(file_path, 'r', encoding='utf-8') as f:
+                            data = json.load(f)  # Load JSON data
+                            try:
+                                tool_calls = data['response']['choices'][0]['message'][
+                                    'tool_calls'
+                                ]
+                                if tool_calls is not None:
+                                    for tool_call in tool_calls:
+                                        tool_call_dict = eval(
+                                            tool_call['function']['arguments']
+                                        )
+
+                                        if (
+                                            tool_call_dict is not None
+                                            and tool_call_dict != {}
+                                        ):
+                                            command = tool_call_dict['command']
+                                            if command == 'create':
+                                                test_suite = tool_call_dict['file_text']
+                                            if (
+                                                command != 'str_replace'
+                                                and command != 'insert'
+                                                and 'coverage' not in command
+                                            ):
+                                                print(command)
+                                            if command == 'insert':
+                                                test_suite_new = insert_line_in_string(
+                                                    test_suite,
+                                                    tool_call_dict['new_str'],
+                                                    tool_call_dict['insert_line'],
+                                                )
+                                                test_suite = test_suite_new
+                                            if command == 'str_replace':
+                                                if (
+                                                    test_suite.count(
+                                                        tool_call_dict['old_str']
+                                                    )
+                                                    == 1
+                                                ):
+                                                    test_suite_new = test_suite.replace(
+                                                        tool_call_dict['old_str'],
+                                                        tool_call_dict['new_str'],
+                                                    )
+                                                else:
+                                                    continue
+                                                test_suite = test_suite_new
+                            except Exception:
+                                print(traceback.format_exc())
+                                continue
+
+                            metadata_copy['test_result']['test_suite'] = test_suite
+                            if i < 25:
+                                final_output[i].append(metadata_copy)
+                                i += 1
+                    except Exception as e:
+                        print(traceback.format_exc())
+                        print(f'  Error loading {file_path}: {e}')
+
+            for j in range(i, 24):
+                final_output[j].append(metadata_copy)
+            metadata_orig = copy.deepcopy(metadata)
+            metadata_orig['test_result']['test_suite'] = orig_test_suite
+            final_output[24].append(metadata_orig)
+
+    for i in range(25):
+        output_file = os.path.join(output_dir, f'output_{i}.jsonl')
+        with open(output_file, 'w') as f:
+            for metadata in final_output[i]:
+                f.write(json.dumps(metadata) + '\n')
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Parse JSON file')
+    parser.add_argument('--root_dir', type=str, help='Root directory', required=True)
+    parser.add_argument(
+        '--output_dir', type=str, help='Output directory', required=True
+    )
+    parser.add_argument(
+        '--starting_preds_file', type=str, help='Starting predictions', default=None
+    )
+    args = parser.parse_args()
+
+    output_file = os.path.join(args.output_dir, 'output.jsonl')
+    metadata_objs = {}
+    with open(output_file, 'r') as f:
+        content = f.readlines()
+        for line in content:
+            metadata = json.loads(line)
+            metadata_objs[metadata['instance_id']] = metadata
+
+    starting_preds_file = args.starting_preds_file
+    preds_objs = {}
+    if starting_preds_file is not None:
+        with open(starting_preds_file, 'r') as f:
+            content = f.readlines()
+            for line in content:
+                pred = json.loads(line)
+                preds_objs[pred['id']] = pred['preds']['full'][0]
+
+    parse_json_files(args.root_dir, args.output_dir, metadata_objs, preds_objs)

evaluation/benchmarks/testgeneval/scripts/eval/compare_outputs.py

@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+import argparse
+
+import pandas as pd
+
+parser = argparse.ArgumentParser(
+    description='Compare two TestGenEval output JSONL files and print the resolved diff'
+)
+parser.add_argument('input_file_1', type=str)
+parser.add_argument('input_file_2', type=str)
+args = parser.parse_args()
+
+df1 = pd.read_json(args.input_file_1, orient='records', lines=True)
+df2 = pd.read_json(args.input_file_2, orient='records', lines=True)
+
+
+# Get the intersection of the ids
+df = pd.merge(df1, df2, on='id', how='inner')
+
+
+def _get_coverage(report):
+    if report is None:
+        return False
+    if isinstance(report, float):
+        return False
+    else:
+        return report.get('test_pass', False)
+
+
+df['test_pass_x'] = df['test_pass_x'].apply(_get_coverage)
+df['test_pass_y'] = df['test_pass_y'].apply(_get_coverage)
+df['diff'] = df.apply(lambda x: x['test_pass_x'] != x['test_pass_y'], axis=1)
+
+df_diff = df[df['diff']].sort_values(
+    by=['test_pass_x', 'test_pass_y'], ascending=[False, False]
+)
+# skip if any of the pass is nan, which means one of the eval is not finished yet
+df_diff = df_diff[df_diff['test_pass_x'].notna() & df_diff['test_pass_y'].notna()]
+
+print(f'X={args.input_file_1}')
+print(f'Y={args.input_file_2}')
+print(f'# diff={df_diff.shape[0]}')
+df_diff = df_diff[['id', 'test_pass_x', 'test_pass_y', 'report_x', 'report_y']]
+
+# x pass but y not
+print('-' * 100)
+df_diff_x_only = df_diff[df_diff['test_pass_x'] & ~df_diff['test_pass_y']].sort_values(
+    by='id'
+)
+print(f'# x pass but y not={df_diff_x_only.shape[0]}')
+print(df_diff_x_only[['id', 'report_x', 'report_y']])
+
+# y pass but x not
+print('-' * 100)
+df_diff_y_only = df_diff[~df_diff['test_pass_x'] & df_diff['test_pass_y']].sort_values(
+    by='id'
+)
+print(f'# y pass but x not={df_diff_y_only.shape[0]}')
+print(df_diff_y_only[['id', 'report_x', 'report_y']])
+# get instance_id from df_diff_y_only
+print('-' * 100)
+print('Instances that x pass but y not:')
+print(df_diff_x_only['id'].tolist())
+
+print('-' * 100)
+print('Instances that y pass but x not:')
+print(df_diff_y_only['id'].tolist())

evaluation/benchmarks/testgeneval/scripts/eval/convert_oh_folder_to_swebench_submission.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+
+FOLDER_PATH=$1
+NEW_FOLDER_PATH=${FOLDER_PATH}.swebench_submission
+mkdir -p $NEW_FOLDER_PATH
+
+# Build all_preds.jsonl
+poetry run python evaluation/testgeneval/scripts/eval/convert_oh_output_to_swe_json.py $FOLDER_PATH/output.jsonl
+mv $FOLDER_PATH/output.swebench.jsonl $NEW_FOLDER_PATH/all_preds.jsonl
+
+# Build trajs/
+mkdir -p $NEW_FOLDER_PATH/trajs
+for instance_dir in $FOLDER_PATH/llm_completions/*/; do
+    instance_id=$(basename "$instance_dir")
+    latest_json=$(ls -t "$instance_dir"/*.json | head -n1)
+    if [ -n "$latest_json" ]; then
+        cat "$latest_json" | jq -r '.messages' > "$NEW_FOLDER_PATH/trajs/$instance_id.json"
+    fi
+done
+
+# Build logs/
+# check if $FOLDER_PATH/eval_outputs exists, if so copy over - else raise error
+if [ -d "$FOLDER_PATH/eval_outputs" ]; then
+    cp -r $FOLDER_PATH/eval_outputs $NEW_FOLDER_PATH/logs
+else
+    echo "Error: $FOLDER_PATH/eval_outputs does not exist. You should run the local docker eval_infer.sh first."
+    exit 1
+fi

evaluation/benchmarks/testgeneval/scripts/eval/convert_oh_output_to_md.py

@@ -0,0 +1,90 @@
+#!/usr/bin/env python3
+"""Convert OpenHands output to a readable markdown format for visualization."""
+
+import argparse
+import json
+import os
+
+import pandas as pd
+from tqdm import tqdm
+
+from openhands.events.serialization import event_from_dict
+
+tqdm.pandas()
+
+parser = argparse.ArgumentParser()
+parser.add_argument('oh_output_file', type=str)
+args = parser.parse_args()
+output_md_folder = args.oh_output_file.replace('.jsonl', '.viz')
+print(f'Converting {args.oh_output_file} to markdown files in {output_md_folder}')
+
+oh_format = pd.read_json(args.oh_output_file, orient='records', lines=True)
+# model name is the folder name of oh_output_file
+model_name = os.path.basename(os.path.dirname(args.oh_output_file))
+
+
+def convert_history_to_str(history):
+    ret = ''
+    separator = '\n\n' + '-' * 100 + '\n'
+
+    for i, event in enumerate(history):
+        if i != 0:
+            ret += separator
+
+        if isinstance(event, list):
+            # "event" is a legacy pair of (action, observation)
+            event_obj = event_from_dict(event[0])
+            ret += f'## {i+1}| {event_obj.__class__.__name__}\n\n'
+            ret += str(event_obj)
+            ret += separator
+
+            event_obj = event_from_dict(event[1])
+            ret += f'## {i+1}| {event_obj.__class__.__name__}\n\n'
+            ret += str(event_obj)
+        else:
+            # "event" is a single event
+            event_obj = event_from_dict(event)
+            ret += f'## {i+1}| {event_obj.__class__.__name__}\n\n'
+            ret += str(event_obj)
+    return ret
+
+
+def write_row_to_md_file(row):
+    if 'test_suite' in row:
+        test_suite = row['test_suite']
+    elif 'test_result' in row and 'test_suite' in row['test_result']:
+        test_suite = row['test_result']['test_suite']
+    else:
+        raise ValueError(f'Row {row} does not have a test_suite')
+
+    if 'report' in row:
+        coverage = row['report'].get('coverage', 0)
+        mutation = row['report'].get('mutation_score', 0)
+    else:
+        coverage = None
+        mutation = None
+
+    id = row['id']
+    filename = f'{id}.md'
+    os.makedirs(output_md_folder, exist_ok=True)
+    filepath = os.path.join(output_md_folder, filename)
+
+    with open(filepath, 'w') as f:
+        f.write(f'# {id} (coverage: {coverage})\n')
+        f.write(f'# {id} (mutation score: {mutation})\n')
+
+        # MetaData
+        f.write('## MetaData\n')
+        f.write('```json\n')
+        f.write(json.dumps(row['metadata'], indent=2))
+        f.write('\n```\n')
+
+        # Trajectory
+        f.write('## History\n')
+        f.write(convert_history_to_str(row['history']))
+
+        f.write('## Test Suite\n')
+        f.write(f'{test_suite}\n')
+
+
+oh_format.progress_apply(write_row_to_md_file, axis=1)

evaluation/benchmarks/testgeneval/scripts/eval/convert_oh_output_to_swe_json.py

@@ -0,0 +1,35 @@
+import argparse
+import os
+
+import pandas as pd
+
+from evaluation.swe_bench.eval_infer import process_git_patch
+
+parser = argparse.ArgumentParser()
+parser.add_argument('oh_output_file', type=str)
+args = parser.parse_args()
+output_filepath = args.oh_output_file.replace('.jsonl', '.swebench.jsonl')
+print(f'Converting {args.oh_output_file} to {output_filepath}')
+
+oh_format = pd.read_json(args.oh_output_file, orient='records', lines=True)
+# model name is the folder name of oh_output_file
+model_name = os.path.basename(os.path.dirname(args.oh_output_file))
+
+
+def convert_row_to_swebench_format(row):
+    if 'git_patch' in row:
+        model_patch = row['git_patch']
+    elif 'test_result' in row and 'git_patch' in row['test_result']:
+        model_patch = row['test_result']['git_patch']
+    else:
+        raise ValueError(f'Row {row} does not have a git_patch')
+
+    return {
+        'instance_id': row['instance_id'],
+        'model_patch': process_git_patch(model_patch),
+        'model_name_or_path': model_name,
+    }
+
+
+swebench_format = oh_format.apply(convert_row_to_swebench_format, axis=1)
+swebench_format.to_json(output_filepath, lines=True, orient='records')

evaluation/benchmarks/testgeneval/scripts/eval/download_gold_test_suites.py

@@ -0,0 +1,28 @@
+import argparse
+
+import pandas as pd
+from datasets import load_dataset
+
+parser = argparse.ArgumentParser()
+parser.add_argument('output_filepath', type=str, help='Path to save the output file')
+parser.add_argument(
+    '--dataset_name',
+    type=str,
+    help='Name of the dataset to download',
+    default='kjain14/testgeneval',
+)
+parser.add_argument('--split', type=str, help='Split to download', default='test')
+args = parser.parse_args()
+
+dataset = load_dataset(args.dataset_name, split=args.split)
+output_filepath = args.output_filepath
+print(
+    f'Downloading gold test suites from {args.dataset_name} (split: {args.split}) to {output_filepath}'
+)
+test_suites = [
+    {'instance_id': row['instance_id'], 'test_suite': row['test_src']}
+    for row in dataset
+]
+print(f'{len(test_suites)} test suites loaded')
+pd.DataFrame(test_suites).to_json(output_filepath, lines=True, orient='records')
+print(f'Test suites saved to {output_filepath}')

evaluation/benchmarks/testgeneval/scripts/eval/summarize_outputs.py

@@ -0,0 +1,120 @@
+#!/usr/bin/env python3
+import argparse
+import json
+from collections import Counter
+
+from openhands.events.serialization import event_from_dict
+from openhands.events.utils import get_pairs_from_events
+
+ERROR_KEYWORDS = [
+    'Agent encountered an error while processing the last action',
+    'APIError',
+    'Action execution failed',
+]
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('output_file', type=str, help='The file to summarize')
+    args = parser.parse_args()
+
+    with open(args.output_file, 'r') as file:
+        lines = file.readlines()
+
+    num_lines = len(lines)
+    num_error_lines = 0
+    num_agent_stuck_in_loop = 0
+
+    coverage = 0
+    mutation_score = 0
+    num_empty_suite = 0
+
+    error_counter = Counter()
+
+    main_agent_cost = []
+    editor_cost = []
+    num_turns = []
+
+    for line in lines:
+        _d = json.loads(line)
+
+        # Cost
+        costs = _d['metrics'].get('costs', [])
+        _cur_main_agent_cost = 0
+        _cur_editor_cost = 0
+        for cost in costs:
+            if isinstance(cost, float):
+                # backward compatible
+                _cur_main_agent_cost += cost
+            else:
+                if 'draft_editor' in cost['model']:
+                    _cur_editor_cost += cost['cost']
+                else:
+                    _cur_main_agent_cost += cost['cost']
+
+        main_agent_cost.append(_cur_main_agent_cost)
+        editor_cost.append(_cur_editor_cost)
+
+        # Turn status
+        history = _d.get('history', [])
+        events = [event_from_dict(event) for event in history]
+        pairs = get_pairs_from_events(events)
+        num_turns.append(len(pairs))
+
+        # Suite & resolve status
+        suite = _d.get('test_result', {}).get('test_suite', '')
+        if suite == '':
+            num_empty_suite += 1
+            continue
+
+        report = _d.get('report', {}) or {}
+        coverage += report.get('coverage', 0)
+        mutation_score += report.get('mutation_score', 0)
+
+        # Error
+        error = _d.get('error', None)
+
+        if error is not None and isinstance(error, str):
+            agent_stuck_in_loop = 'Agent got stuck in a loop' in error
+            contains_error = bool(error) and not agent_stuck_in_loop
+            if agent_stuck_in_loop:
+                error_counter['Agent got stuck in a loop'] += 1
+                num_agent_stuck_in_loop += 1
+            elif contains_error:
+                error_counter[error] += 1
+            continue
+
+        for keyword in ERROR_KEYWORDS:
+            if keyword in line:
+                error_counter[keyword] += 1
+                num_error_lines += 1
+                break
+
+    # print the error counter (with percentage)
+    print(f'Average coverage for {num_lines} ({coverage / num_lines * 100:.2f}%)')
+    print(
+        f'Average mutation score for {num_lines} ({mutation_score / num_lines * 100:.2f}%)'
+    )
+
+    print(
+        f'Number of empty suite: {num_empty_suite} / {num_lines} ({num_empty_suite / num_lines * 100:.2f}%)'
+    )
+    print(
+        f'Number of error lines: {num_error_lines} / {num_lines} ({num_error_lines / num_lines * 100:.2f}%)'
+    )
+    print(
+        f'Number of agent stuck in loop: {num_agent_stuck_in_loop} / {num_lines} ({num_agent_stuck_in_loop / num_lines * 100:.2f}%)'
+    )
+    assert len(num_turns) == num_lines
+    assert len(main_agent_cost) == num_lines
+    assert len(editor_cost) == num_lines
+    print('## Statistics')
+    print(f'Avg. num of turns per instance: {sum(num_turns) / num_lines:.2f}')
+    print(f'Avg. agent cost per instance: {sum(main_agent_cost) / num_lines:.2f} USD')
+    print(f'Avg. editor cost per instance: {sum(editor_cost) / num_lines:.2f} USD')
+    print(
+        f'Avg. total cost per instance: {(sum(main_agent_cost) + sum(editor_cost)) / num_lines:.2f} USD'
+    )
+
+    print('## Detailed error breakdown:')
+    for error, count in error_counter.items():
+        print(f'{error}: {count} ({count / num_lines * 100:.2f}%)')

evaluation/benchmarks/testgeneval/scripts/eval_infer.sh

@@ -0,0 +1,53 @@
+#!/bin/bash
+set -eo pipefail
+
+INPUT_FILE=$1
+NUM_WORKERS=$2
+DATASET=$3
+SPLIT=$4
+SKIP_MUTATION=$5
+
+if [ -z "$INPUT_FILE" ]; then
+  echo "INPUT_FILE not specified (should be a path to a jsonl file)"
+  exit 1
+fi
+
+if [ -z "$DATASET" ]; then
+  echo "DATASET not specified, use default kjain14/testgenevallite"
+  DATASET="kjain14/testgenevallite"
+fi
+
+if [ -z "$SPLIT" ]; then
+  echo "SPLIT not specified, use default test"
+  SPLIT="test"
+fi
+
+if [ -z "$NUM_WORKERS" ]; then
+  echo "NUM_WORKERS not specified, use default 1"
+  NUM_WORKERS=1
+fi
+
+echo "... Evaluating on $INPUT_FILE ..."
+
+COMMAND="poetry run python evaluation/benchmarks/testgeneval/eval_infer.py \
+  --eval-num-workers $NUM_WORKERS \
+  --input-file $INPUT_FILE \
+  --dataset $DATASET \
+  --split $SPLIT"
+
+if [ "$SKIP_MUTATION" == "true" ]; then
+  echo "Skipping mutation evaluation"
+  COMMAND="$COMMAND --skip_mutation"
+fi
+
+if [ -n "$EVAL_LIMIT" ]; then
+  echo "EVAL_LIMIT: $EVAL_LIMIT"
+  COMMAND="$COMMAND --eval-n-limit $EVAL_LIMIT"
+fi
+
+echo $COMMAND
+# Run the command
+eval $COMMAND
+
+# update the output with evaluation results
+# poetry run python evaluation/benchmarks/testgeneval/scripts/eval/update_output_with_eval.py $INPUT_FILE

evaluation/benchmarks/testgeneval/scripts/run_infer.sh

@@ -0,0 +1,122 @@
+#!/bin/bash
+set -eo pipefail
+
+source "evaluation/utils/version_control.sh"
+
+MODEL_CONFIG=$1
+COMMIT_HASH=$2
+AGENT=$3
+EVAL_LIMIT=$4
+MAX_ITER=$5
+NUM_WORKERS=$6
+DATASET=$7
+SPLIT=$8
+N_RUNS=$9
+ZERO_SHOT_PATH=${10}  # New argument for zero-shot path
+
+if [ -z "$NUM_WORKERS" ]; then
+  NUM_WORKERS=1
+  echo "Number of workers not specified, use default $NUM_WORKERS"
+fi
+checkout_eval_branch
+
+if [ -z "$AGENT" ]; then
+  echo "Agent not specified, use default CodeActAgent"
+  AGENT="CodeActAgent"
+fi
+
+if [ -z "$MAX_ITER" ]; then
+  echo "MAX_ITER not specified, use default 100"
+  MAX_ITER=100
+fi
+
+if [ -z "$USE_INSTANCE_IMAGE" ]; then
+  echo "USE_INSTANCE_IMAGE not specified, use default true"
+  USE_INSTANCE_IMAGE=true
+fi
+
+if [ -z "$RUN_WITH_BROWSING" ]; then
+  echo "RUN_WITH_BROWSING not specified, use default false"
+  RUN_WITH_BROWSING=false
+fi
+
+
+if [ -z "$DATASET" ]; then
+  echo "DATASET not specified, use default princeton-nlp/SWE-bench_Lite"
+  DATASET="princeton-nlp/SWE-bench_Lite"
+fi
+
+if [ -z "$SPLIT" ]; then
+  echo "SPLIT not specified, use default test"
+  SPLIT="test"
+fi
+
+export USE_INSTANCE_IMAGE=$USE_INSTANCE_IMAGE
+echo "USE_INSTANCE_IMAGE: $USE_INSTANCE_IMAGE"
+export RUN_WITH_BROWSING=$RUN_WITH_BROWSING
+echo "RUN_WITH_BROWSING: $RUN_WITH_BROWSING"
+
+get_openhands_version
+
+echo "AGENT: $AGENT"
+echo "OPENHANDS_VERSION: $OPENHANDS_VERSION"
+echo "MODEL_CONFIG: $MODEL_CONFIG"
+echo "DATASET: $DATASET"
+echo "SPLIT: $SPLIT"
+
+# Default to NOT use Hint
+if [ -z "$USE_HINT_TEXT" ]; then
+  export USE_HINT_TEXT=false
+fi
+echo "USE_HINT_TEXT: $USE_HINT_TEXT"
+EVAL_NOTE="$OPENHANDS_VERSION"
+# if not using Hint, add -no-hint to the eval note
+if [ "$USE_HINT_TEXT" = false ]; then
+  EVAL_NOTE="$EVAL_NOTE-no-hint"
+fi
+
+if [ "$RUN_WITH_BROWSING" = true ]; then
+  EVAL_NOTE="$EVAL_NOTE-with-browsing"
+fi
+
+if [ -n "$EXP_NAME" ]; then
+  EVAL_NOTE="$EVAL_NOTE-$EXP_NAME"
+fi
+
+function run_eval() {
+  local eval_note=$1
+  COMMAND="poetry run python evaluation/benchmarks/testgeneval/run_infer.py \
+    --agent-cls $AGENT \
+    --llm-config $MODEL_CONFIG \
+    --max-iterations $MAX_ITER \
+    --eval-num-workers $NUM_WORKERS \
+    --eval-note $eval_note \
+    --dataset $DATASET \
+    --split $SPLIT"
+
+  if [ -n "$EVAL_LIMIT" ]; then
+    echo "EVAL_LIMIT: $EVAL_LIMIT"
+    COMMAND="$COMMAND --eval-n-limit $EVAL_LIMIT"
+  fi
+
+  if [ -n "$ZERO_SHOT_PATH" ]; then
+    echo "ZERO_SHOT_PATH: $ZERO_SHOT_PATH"
+    COMMAND="$COMMAND --testfile_start --zero_shot_path $ZERO_SHOT_PATH"
+  fi
+
+  eval $COMMAND
+}
+
+unset SANDBOX_ENV_GITHUB_TOKEN # prevent the agent from using the github token to push
+if [ -z "$N_RUNS" ]; then
+  N_RUNS=1
+  echo "N_RUNS not specified, use default $N_RUNS"
+fi
+
+for i in $(seq 1 $N_RUNS); do
+  current_eval_note="$EVAL_NOTE-run_$i"
+  echo "EVAL_NOTE: $current_eval_note"
+  run_eval $current_eval_note
+done
+
+checkout_original_branch

evaluation/benchmarks/testgeneval/scripts/setup/instance_swe_entry.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+
+source ~/.bashrc
+SWEUTIL_DIR=/swe_util
+
+# FIXME: Cannot read SWE_INSTANCE_ID from the environment variable
+# SWE_INSTANCE_ID=django__django-11099
+if [ -z "$SWE_INSTANCE_ID" ]; then
+    echo "Error: SWE_INSTANCE_ID is not set." >&2
+    exit 1
+fi
+
+# Read the swe-bench-test-lite.json file and extract the required item based on instance_id
+item=$(jq --arg INSTANCE_ID "$SWE_INSTANCE_ID" '.[] | select(.instance_id == $INSTANCE_ID)' $SWEUTIL_DIR/eval_data/instances/swe-bench-instance.json)
+
+if [[ -z "$item" ]]; then
+  echo "No item found for the provided instance ID."
+  exit 1
+fi
+
+WORKSPACE_NAME=$(echo "$item" | jq -r '(.repo | tostring) + "__" + (.version | tostring) | gsub("/"; "__")')
+
+echo "WORKSPACE_NAME: $WORKSPACE_NAME"
+
+# Clear the workspace
+if [ -d /workspace ]; then
+    rm -rf /workspace/*
+else
+    mkdir /workspace
+fi
+# Copy repo to workspace
+if [ -d /workspace/$WORKSPACE_NAME ]; then
+    rm -rf /workspace/$WORKSPACE_NAME
+fi
+mkdir -p /workspace
+ln -s /testbed /workspace/$WORKSPACE_NAME
+
+# Activate instance-specific environment
+. /opt/miniconda3/etc/profile.d/conda.sh
+conda activate testbed

evaluation/benchmarks/testgeneval/scripts/setup/prepare_swe_utils.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+
+set -e
+EVAL_WORKSPACE="evaluation/swe_bench/eval_workspace"
+mkdir -p $EVAL_WORKSPACE
+
+# 1. Prepare REPO
+echo "==== Prepare SWE-bench repo ===="
+OH_SWE_BENCH_REPO_PATH="https://github.com/All-Hands-AI/SWE-bench.git"
+OH_SWE_BENCH_REPO_BRANCH="eval"
+git clone -b $OH_SWE_BENCH_REPO_BRANCH $OH_SWE_BENCH_REPO_PATH $EVAL_WORKSPACE/OH-SWE-bench
+
+# 2. Prepare DATA
+echo "==== Prepare SWE-bench data ===="
+EVAL_IMAGE=ghcr.io/all-hands-ai/eval-swe-bench:builder_with_conda
+EVAL_WORKSPACE=$(realpath $EVAL_WORKSPACE)
+chmod +x $EVAL_WORKSPACE/OH-SWE-bench/swebench/harness/prepare_data.sh
+if [ -d $EVAL_WORKSPACE/eval_data ]; then
+    rm -r $EVAL_WORKSPACE/eval_data
+fi
+docker run \
+    -v $EVAL_WORKSPACE:/workspace \
+    -w /workspace \
+    -u $(id -u):$(id -g) \
+    -e HF_DATASETS_CACHE="/tmp" \
+    --rm -it $EVAL_IMAGE \
+    bash -c "cd OH-SWE-bench/swebench/harness && /swe_util/miniforge3/bin/conda run -n swe-bench-eval ./prepare_data.sh && mv eval_data /workspace/"

evaluation/benchmarks/testgeneval/scripts/setup/swe_entry.sh

@@ -0,0 +1,96 @@
+#!/bin/bash
+
+set -e
+
+# assert user name is `root`
+if [ "$USER" != "root" ]; then
+    echo "Error: This script is intended to be run by the 'root' user only." >&2
+    exit 1
+fi
+
+source ~/.bashrc
+
+SWEUTIL_DIR=/swe_util
+
+# Create logs directory
+LOG_DIR=/openhands/logs
+mkdir -p $LOG_DIR && chmod 777 $LOG_DIR
+
+# FIXME: Cannot read SWE_INSTANCE_ID from the environment variable
+# SWE_INSTANCE_ID=django__django-11099
+if [ -z "$SWE_INSTANCE_ID" ]; then
+    echo "Error: SWE_INSTANCE_ID is not set." >&2
+    exit 1
+fi
+
+# Read the swe-bench-test-lite.json file and extract the required item based on instance_id
+item=$(jq --arg INSTANCE_ID "$SWE_INSTANCE_ID" '.[] | select(.instance_id == $INSTANCE_ID)' $SWEUTIL_DIR/eval_data/instances/swe-bench-test-lite.json)
+
+if [[ -z "$item" ]]; then
+  echo "No item found for the provided instance ID."
+  exit 1
+fi
+
+CONDA_ENV_NAME=$(echo "$item" | jq -r '.repo + "__" + .version | gsub("/"; "__")')
+
+echo "CONDA_ENV_NAME: $CONDA_ENV_NAME"
+
+SWE_TASK_DIR=/openhands/swe_tasks
+mkdir -p $SWE_TASK_DIR
+# Dump test_patch to /workspace/test.patch
+echo "$item" | jq -r '.test_patch' > $SWE_TASK_DIR/test.patch
+# Dump patch to /workspace/gold.patch
+echo "$item" | jq -r '.patch' > $SWE_TASK_DIR/gold.patch
+# Dump the item to /workspace/instance.json except for the "test_patch" and "patch" fields
+echo "$item" | jq 'del(.test_patch, .patch)' > $SWE_TASK_DIR/instance.json
+
+# Clear the workspace
+rm -rf /workspace/*
+# Copy repo to workspace
+if [ -d /workspace/$CONDA_ENV_NAME ]; then
+    rm -rf /workspace/$CONDA_ENV_NAME
+fi
+cp -r $SWEUTIL_DIR/eval_data/testbeds/$CONDA_ENV_NAME /workspace
+
+# Reset swe-bench testbed and install the repo
+. $SWEUTIL_DIR/miniforge3/etc/profile.d/conda.sh
+conda config --set changeps1 False
+conda config --append channels conda-forge
+conda activate swe-bench-eval
+
+mkdir -p $SWE_TASK_DIR/reset_testbed_temp
+mkdir -p $SWE_TASK_DIR/reset_testbed_log_dir
+SWE_BENCH_DIR=/swe_util/OH-SWE-bench
+output=$(
+    export PYTHONPATH=$SWE_BENCH_DIR && \
+    cd $SWE_BENCH_DIR && \
+    python swebench/harness/reset_swe_env.py \
+    --swe_bench_tasks $SWEUTIL_DIR/eval_data/instances/swe-bench-test.json \
+    --temp_dir $SWE_TASK_DIR/reset_testbed_temp \
+    --testbed /workspace \
+    --conda_path $SWEUTIL_DIR/miniforge3 \
+    --instance_id $SWE_INSTANCE_ID \
+    --log_dir $SWE_TASK_DIR/reset_testbed_log_dir \
+    --timeout 900 \
+    --verbose
+)
+
+REPO_PATH=$(echo "$output" | awk -F': ' '/repo_path:/ {print $2}')
+TEST_CMD=$(echo "$output" | awk -F': ' '/test_cmd:/ {print $2}')
+echo "Repo Path: $REPO_PATH"
+echo "Test Command: $TEST_CMD"
+
+echo "export SWE_BENCH_DIR=\"$SWE_BENCH_DIR\"" >> ~/.bashrc
+echo "export REPO_PATH=\"$REPO_PATH\"" >> ~/.bashrc
+echo "export TEST_CMD=\"$TEST_CMD\"" >> ~/.bashrc
+
+if [[ "$REPO_PATH" == "None" ]]; then
+    echo "Error: Failed to retrieve repository path. Tests may not have passed or output was not as expected." >&2
+    exit 1
+fi
+
+# Activate instance-specific environment
+. $SWEUTIL_DIR/miniforge3/etc/profile.d/conda.sh
+conda activate $CONDA_ENV_NAME
+
+set +e

evaluation/benchmarks/testgeneval/test_filter.py

@@ -0,0 +1,327 @@
+import ast
+import re
+from typing import List, Tuple
+
+from evaluation.benchmarks.testgeneval.constants import TestStatus
+from evaluation.benchmarks.testgeneval.log_parsers import (
+    MAP_REPO_TO_PARSER,
+    parse_log_pytest,
+)
+
+
+def indent_text(text, indent_level):
+    return '\n'.join(
+        ' ' * indent_level + line if line.strip() else line for line in text.split('\n')
+    )
+
+
+def extract_preamble_classes_and_functions(code):
+    class_pattern = re.compile(
+        r'(?P<decorators>(?:^@[^\r\n]*(?:\r?\n(?:[ \t]+[^\r\n]*|^\)[^\r\n]*)*)*\r?\n)*?)'
+        r'^class\s+([\w]+)(?:\([^)]*\))?:',  # the class line
+        re.MULTILINE,
+    )
+    # Capture methods with or without decorators
+    method_pattern = re.compile(r'(^(\s*@.*\s*)*^\s*def\s+[\w_]+\(.*\):)', re.MULTILINE)
+
+    # Capture functions with or without decorators
+    function_pattern = re.compile(
+        r'(?P<decorators>(?:^@[^\r\n]*(?:\r?\n(?:[ \t]+[^\r\n]*|^\)[^\r\n]*)*)*\r?\n)*?)'
+        r'^def\s+([\w_]+)\(.*\):',  # the function line
+        re.MULTILINE,
+    )
+
+    preamble = ''
+    classes = []
+    test_functions = []
+
+    current_position = 0
+
+    def extract_class_body(code: str, start_index: int) -> Tuple[str, int]:
+        """
+        Extracts the body of a class from the given code starting from the specified index.
+        Returns the class body and the end index of the class body.
+        """
+        if not code or start_index < 0 or start_index >= len(code):
+            raise ValueError('Invalid code or start index')
+
+        # Split the code into lines
+        lines = code[start_index:].split('\n')
+        class_body_lines = []
+
+        # Find the starting indentation level of the class definition
+        class_start_line = lines[0]
+        start_indent = len(class_start_line) - len(class_start_line.lstrip())
+
+        inside_multiline_comment = False
+        end_index = start_index
+        for i, line in enumerate(lines[1:], start=1):
+            stripped_line = line.strip()
+            current_indent = len(line) - len(line.lstrip())
+
+            # Handle multiline comments or docstrings
+            if stripped_line.startswith('"""') or stripped_line.startswith("'''"):
+                if inside_multiline_comment:
+                    inside_multiline_comment = False
+                else:
+                    inside_multiline_comment = True
+
+            if not inside_multiline_comment:
+                # Stop when we reach a line with less indentation than the class definition
+                if current_indent <= start_indent and stripped_line:
+                    break
+
+            # Add lines that are part of the class body
+            class_body_lines.append(line)
+            # Update the end index to the current line end
+            end_index = start_index + len('\n'.join(lines[: i + 1])) + 1
+
+        return code[start_index:end_index], end_index
+
+    while current_position < len(code):
+        class_match = class_pattern.search(code, current_position)
+        method_match = method_pattern.search(code, current_position)
+
+        if class_match and (
+            not method_match or class_match.start() < method_match.start()
+        ):
+            class_name = class_match.group(0)
+            class_body, end_idx = extract_class_body(code, class_match.end())
+            current_position = end_idx
+
+            methods = []
+            class_prefix = class_name
+            set_prefix = False
+            for method_match in method_pattern.finditer(class_body):
+                method_name = method_match.group()
+                method_start = method_match.start()
+                if not set_prefix:
+                    class_prefix = class_name + class_body[:method_start]
+                    set_prefix = True
+                next_method = method_pattern.search(
+                    class_body, method_start + len(method_name)
+                )
+                method_body = (
+                    class_body[method_start : next_method.start()]
+                    if next_method
+                    else class_body[method_start:]
+                )
+                methods.append((method_name, method_body))
+
+            classes.append((class_prefix, methods, class_match.start()))
+
+        elif method_match:
+            function_name = method_match.group(0)
+            start_idx = method_match.start()
+
+            # Extract the current function's indentation level
+            lines = code[start_idx:].split('\n')
+            current_indent = len(lines[0]) - len(lines[0].lstrip())
+
+            next_function = function_pattern.search(
+                code, start_idx + len(function_name)
+            )
+            while next_function and (
+                class_match is None or next_function.start() < class_match.start()
+            ):
+                # Calculate the indentation of the next function
+                next_function_start = next_function.start()
+                next_line = code[next_function_start:].split('\n', 1)[0]
+                next_indent = len(next_line) - len(next_line.lstrip())
+
+                # Check if the next function is top-level
+                if next_indent <= current_indent:
+                    break
+
+                # Continue searching for the next top-level function
+                next_function = function_pattern.search(
+                    code, next_function.start() + len(next_function.group(0))
+                )
+
+            if next_function:
+                next_function_start = next_function.start()
+                if class_match and next_function_start > class_match.start():
+                    next_function_start = class_match.start()
+                function_body = code[start_idx:next_function_start]
+            else:
+                function_body = code[start_idx:]
+
+            test_functions.append((function_body, start_idx))
+            current_position = start_idx + len(function_body)
+
+        else:
+            break
+
+    if classes and test_functions:
+        preamble = code[: min(classes[0][2], test_functions[0][1])]
+    else:
+        preamble = (
+            code[: classes[0][2]]
+            if classes
+            else code[: test_functions[0][1]]
+            if test_functions
+            else code
+        )
+
+    return preamble.strip(), classes, test_functions
+
+
+def filter_passing_tests(
+    test_content: str, test_output: str, repo: str
+) -> Tuple[str, List[str], List[str]]:
+    """
+    Filter tests based on their execution results.
+    Returns:
+        Tuple containing:
+        - Modified test content with only passing tests
+        - List of passing test names
+        - List of failing test names
+    """
+    # Parse test results using appropriate parser
+    parser = MAP_REPO_TO_PARSER.get(repo, parse_log_pytest)
+    test_results = parser(test_output)
+    # Get passing and failing tests
+    passing_tests = []
+    failing_tests = []
+    for test_name, status in test_results.items():
+        if status == TestStatus.PASSED.value:
+            passing_tests.append(test_name)
+        else:
+            failing_tests.append(test_name)
+
+    if not passing_tests:
+        return '', passing_tests, failing_tests
+
+    # Extract test components
+    preamble, classes, functions = extract_preamble_classes_and_functions(test_content)
+
+    # Filter classes to only include passing methods
+    filtered_classes = []
+    for class_name, methods, start_idx in classes:
+        non_fail_methods = []
+        for method_name, method_body in methods:
+            # Extract the base method name for matching
+            method_full_name = (
+                method_name.split('.')[-1].split('(')[0].strip().split(' ')[-1]
+            )
+            # Check if the method name is in failing_tests or if any failing_test is in the method name
+            if not (
+                any(method_full_name in failing_test for failing_test in failing_tests)
+                or any(
+                    failing_test in method_full_name for failing_test in failing_tests
+                )
+            ):
+                non_fail_methods.append((method_name, method_body))
+
+        if non_fail_methods:
+            filtered_classes.append((class_name, non_fail_methods, start_idx))
+
+    # Filter standalone functions
+    filtered_functions = []
+    for func_body, start_idx in functions:
+        func_name = func_body.split('def ')[1].split('(')[0].strip()
+        if any(func_name in failing_test for failing_test in failing_tests) or any(
+            failing_test in func_name for failing_test in failing_tests
+        ):
+            continue
+
+        filtered_functions.append((func_body, start_idx))
+
+    # Reconstruct test content with only passing tests
+    content_parts = [preamble]
+
+    # Add filtered classes
+    for class_name, methods, _ in filtered_classes:
+        class_content = class_name + '\n'
+        for _, method_body in methods:
+            class_content += method_body + '\n'
+        content_parts.append(class_content)
+
+    # Add filtered functions
+    for func_body, _ in filtered_functions:
+        content_parts.append(func_body)
+
+    return '\n\n'.join(content_parts), passing_tests, failing_tests
+
+
+def filter_tests(
+    test_content: str, test_output: str, repo: str
+) -> Tuple[str, List[str], List[str]]:
+    """
+    Filter tests using AST parsing to remove failing test functions from the test file.
+    Non-test functions (e.g. setup or helper methods) and classes (even if all test methods are failing)
+    are preserved.
+
+    If AST processing fails (for example, because the test file cannot be parsed),
+    this function falls back on the existing regex-based filtering (filter_passing_tests).
+
+    Returns:
+        Tuple containing:
+         - Modified test content (as a string) containing only passing tests.
+         - List of passing test names.
+         - List of failing test names.
+    """
+    try:
+        # Attempt to parse the test file using the AST.
+        tree = ast.parse(test_content)
+
+        # Parse test results using the appropriate parser.
+        parser = MAP_REPO_TO_PARSER.get(repo, parse_log_pytest)
+        test_results = parser(test_output)
+        passing_tests = [
+            name
+            for name, status in test_results.items()
+            if status == TestStatus.PASSED.value
+        ]
+        failing_tests = [
+            name
+            for name, status in test_results.items()
+            if status != TestStatus.PASSED.value
+        ]
+
+        # Helper function to decide if a test name should be considered failing.
+        def is_failing(name: str) -> bool:
+            for ft in failing_tests:
+                if name in ft or ft in name:
+                    return True
+            return False
+
+        new_body = []
+        for node in tree.body:
+            # For top-level function definitions, only filter those that look like tests.
+            if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
+                if node.name.startswith('test') and is_failing(node.name):
+                    continue
+                new_body.append(node)
+            # For classes, filter out failing test methods but preserve other methods (e.g. setup).
+            elif isinstance(node, ast.ClassDef):
+                new_class_body = []
+                for subnode in node.body:
+                    if isinstance(subnode, (ast.FunctionDef, ast.AsyncFunctionDef)):
+                        # Only consider filtering if the method is a test.
+                        qualified_name = f'{node.name}.{subnode.name}'
+                        if is_failing(subnode.name) or is_failing(qualified_name):
+                            continue
+                        new_class_body.append(subnode)
+                    else:
+                        new_class_body.append(subnode)
+                # Always include the class even if no test methods remain, as it might contain
+                # setup, teardown, or other necessary logic.
+                if new_class_body:
+                    node.body = new_class_body
+                    new_body.append(node)
+
+            else:
+                new_body.append(node)
+
+        tree.body = new_body
+
+        # Reconstruct the source code from the filtered AST.
+        # (Requires Python 3.9+ for ast.unparse; otherwise an exception will trigger the fallback.)
+        new_test_content = ast.unparse(tree)
+        return new_test_content, passing_tests, failing_tests
+
+    except Exception:
+        print('AST processing failed; falling back on regex-based filtering.')
+        # If AST processing fails for any reason, fall back on the original regex-based filtering.
+        return filter_passing_tests(test_content, test_output, repo)

evaluation/benchmarks/testgeneval/test_spec.py

@@ -0,0 +1,166 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from evaluation.benchmarks.testgeneval.constants import (
+    COVERAGE_PREFIX,
+    KEY_INSTANCE_ID,
+    MAP_REPO_VERSION_TO_SPECS,
+    TESTS_FAILED,
+    TESTS_SUFFIX,
+    UPDATE_TOX,
+    TestGenEvalInstance,
+)
+from evaluation.benchmarks.testgeneval.utils import (
+    get_test_directives,
+)
+
+DIFF_MODIFIED_FILE_REGEX = r'--- a/(.*)'
+
+
+@dataclass
+class TestSpec:
+    """
+    A dataclass that represents a test specification for a single instance of SWE-bench.
+    """
+
+    instance_id: str
+    id: str
+    repo: str
+    version: str
+    test_cmd: str
+    code_file: str
+    test_file: str
+    baseline_covs: dict
+    local_imports: list[str]
+    test_script_list: list[str]
+    mutation_script_list: list[str]
+
+    @property
+    def test_script(self):
+        return (
+            '\n'.join(['#!/bin/bash', 'set -uo pipefail'] + self.test_script_list)
+            + '\n'
+        )
+        # Don't exit early because we need to revert tests at the end
+
+    @property
+    def mutation_script(self):
+        return (
+            '\n'.join(['#!/bin/bash', 'set -uo pipefail'] + self.mutation_script_list)
+            + '\n'
+        )
+        # Don't exit early because we need to revert tests at the end
+
+
+def make_test_setup(specs, env_name, repo_directory, includes_tox=False):
+    eval_commands = []
+
+    if includes_tox:
+        eval_commands.append(UPDATE_TOX)
+
+    eval_commands += [
+        'source /opt/miniconda3/bin/activate',
+        f'conda activate {env_name}',
+        f'cd {repo_directory}',
+    ]
+    if 'eval_commands' in specs:
+        eval_commands += specs['eval_commands']
+    eval_commands += [
+        f'git config --global --add safe.directory {repo_directory}',  # for nonroot user
+        f'cd {repo_directory}',
+        # This is just informational, so we have a record
+        'git status',
+        'git show',
+        'source /opt/miniconda3/bin/activate',
+        f'conda activate {env_name}',
+    ]
+    if 'install' in specs:
+        eval_commands.append(specs['install'])
+
+    if includes_tox:
+        eval_commands.append('add_coverage_tox "tox.ini"')
+
+    eval_commands.append('[ -f ".coveragerc" ] && rm ".coveragerc"')
+    return eval_commands
+
+
+def make_test_script_list(test_cmd, specs, env_name, repo_directory):
+    """
+    Runs the tests.
+    """
+
+    includes_tox = 'tox' in test_cmd
+    eval_commands = make_test_setup(specs, env_name, repo_directory, includes_tox)
+    eval_commands += [
+        f'{test_cmd} || {{ echo "{TESTS_FAILED}\n{TESTS_SUFFIX}\n" && exit 1; }}',
+        f'echo "{TESTS_SUFFIX}"\n',
+        'coverage json -o coverage.json',
+        f'echo "{COVERAGE_PREFIX}"\n',
+        'cat coverage.json',
+    ]
+
+    return eval_commands
+
+
+def make_mutation_script_list(specs, env_name, repo_directory, mutation_timeout):
+    """
+    Runs the tests.
+    """
+
+    eval_commands = make_test_setup(specs, env_name, repo_directory)
+    eval_commands += [
+        'cosmic-ray init mutation.toml mutation.sqlite',
+        f'timeout {mutation_timeout}s cosmic-ray exec mutation.toml mutation.sqlite',
+        'cr-report mutation.sqlite',
+        'cr-rate mutation.sqlite  --estimate --confidence 95.0',
+    ]
+    return eval_commands
+
+
+def make_test_spec(
+    instance: TestGenEvalInstance, mutation_timeout: int, buffer: int
+) -> TestSpec:
+    if isinstance(instance, TestSpec):
+        return instance
+    instance_id = instance[KEY_INSTANCE_ID]
+    id = instance['id']
+    repo = instance['repo']
+    version = instance['version']
+    baseline_covs = instance['baseline_covs']
+    code_file = instance['code_file']
+    test_file = instance['test_file']
+    local_imports = instance['local_imports']
+
+    env_name = 'testbed'
+    repo_directory = f'/{env_name}'
+    specs = MAP_REPO_VERSION_TO_SPECS[repo][version]
+
+    test_cmd = ' '.join(
+        [
+            MAP_REPO_VERSION_TO_SPECS[instance['repo']][instance['version']][
+                'test_cmd'
+            ],
+            *get_test_directives(instance),
+        ]
+    )
+
+    test_script_list = make_test_script_list(test_cmd, specs, env_name, repo_directory)
+
+    mutation_script_list = make_mutation_script_list(
+        specs, env_name, repo_directory, mutation_timeout - buffer
+    )
+
+    return TestSpec(
+        instance_id=instance_id,
+        id=id,
+        repo=repo,
+        test_script_list=test_script_list,
+        test_cmd=test_cmd,
+        local_imports=local_imports,
+        mutation_script_list=mutation_script_list,
+        code_file=code_file,
+        test_file=test_file,
+        baseline_covs=baseline_covs,
+        version=version,
+    )

evaluation/benchmarks/testgeneval/utils.py

@@ -0,0 +1,73 @@
+import json
+from pathlib import Path
+from typing import cast
+
+from datasets import Dataset, load_dataset
+
+from evaluation.benchmarks.testgeneval.constants import (
+    KEY_INSTANCE_ID,
+    TestGenEvalInstance,
+)
+
+
+def get_test_directives(instance: TestGenEvalInstance) -> list:
+    """
+    Get test directives from the test_patch of a task instance
+
+    Args:
+        instance (dict): task instance
+    Returns:
+        directives (list): List of test directives
+    """
+    # For seq2seq code repos, testing command is fixed
+    if instance['repo'] == 'swe-bench/humaneval':
+        return ['test.py']
+
+    # Get test directives from test patch and remove non-test files
+    directives = [f"/testbed/{instance['test_file']}"]
+
+    # For Django tests, remove extension + "tests/" prefix and convert slashes to dots (module referencing)
+    if instance['repo'] == 'django/django':
+        directives = [instance['test_file']]
+        directives_transformed = []
+        for d in directives:
+            d = d[: -len('.py')] if d.endswith('.py') else d
+            d = d[len('tests/') :] if d.startswith('tests/') else d
+            d = d.replace('/', '.')
+            directives_transformed.append(d)
+        directives = directives_transformed
+
+    return directives
+
+
+def load_testgeneval_dataset(
+    name='kjain14/testgeneval', split='test', ids=None
+) -> list[TestGenEvalInstance]:
+    """
+    Load SWE-bench dataset from Hugging Face Datasets or local .json/.jsonl file
+    """
+    # check that all instance IDs are in the dataset
+    if ids:
+        ids = set(ids)
+    # Load from local .json/.jsonl file
+    if name.endswith('.json') or name.endswith('.jsonl'):
+        dataset = json.loads(Path(name).read_text())
+        dataset_ids = {instance[KEY_INSTANCE_ID] for instance in dataset}
+    else:
+        # Load from Hugging Face Datasets
+        if name.lower() in {'testgeneval'}:
+            name = 'kjain14/testgeneval'
+        elif name.lower() in {'testgeneval-lite', 'testgenevallite', 'lite'}:
+            name = 'kjain14/testgenevallite'
+        dataset = cast(Dataset, load_dataset(name, split=split))
+        dataset_ids = {instance['id'] for instance in dataset}
+    if ids:
+        if ids - dataset_ids:
+            raise ValueError(
+                (
+                    "Some instance IDs not found in dataset!"
+                    f"\nMissing IDs:\n{' '.join(ids - dataset_ids)}"
+                )
+            )
+        dataset = [instance for instance in dataset if instance['id'] in ids]
+    return [cast(TestGenEvalInstance, instance) for instance in dataset]

evaluation/integration_tests/run_infer.py

@@ -34,7 +34,6 @@ from openhands.utils.async_utils import call_async_from_sync
 
 FAKE_RESPONSES = {
     'CodeActAgent': fake_user_response,
-    'DelegatorAgent': fake_user_response,
     'VisualBrowsingAgent': fake_user_response,
 }
 

frontend/__tests__/components/features/payment/payment-form.test.tsx

@@ -4,8 +4,10 @@ import userEvent from "@testing-library/user-event";
 import { afterEach, beforeEach, describe, expect, it, test, vi } from "vitest";
 import OpenHands from "#/api/open-hands";
 import { PaymentForm } from "#/components/features/payment/payment-form";
+import * as featureFlags from "#/utils/feature-flags";
 
 describe("PaymentForm", () => {
+  const billingSettingsSpy = vi.spyOn(featureFlags, "BILLING_SETTINGS");
   const getBalanceSpy = vi.spyOn(OpenHands, "getBalance");
   const createCheckoutSessionSpy = vi.spyOn(OpenHands, "createCheckoutSession");
   const getConfigSpy = vi.spyOn(OpenHands, "getConfig");
@@ -26,6 +28,7 @@ describe("PaymentForm", () => {
       GITHUB_CLIENT_ID: "123",
       POSTHOG_CLIENT_KEY: "456",
     });
+    billingSettingsSpy.mockReturnValue(true);
   });
 
   afterEach(() => {

frontend/__tests__/routes/home.test.tsx

@@ -26,33 +26,32 @@ const createAxiosNotFoundErrorObject = () =>
     },
   );
 
+const getSettingsSpy = vi.spyOn(OpenHands, "getSettings");
+
+const RouterStub = createRoutesStub([
+  {
+    // layout route
+    Component: MainApp,
+    path: "/",
+    children: [
+      {
+        // home route
+        Component: Home,
+        path: "/",
+      },
+      {
+        Component: SettingsScreen,
+        path: "/settings",
+      },
+    ],
+  },
+]);
+
+afterEach(() => {
+  vi.clearAllMocks();
+});
+
 describe("Home Screen", () => {
-  const getSettingsSpy = vi.spyOn(OpenHands, "getSettings");
-  const getConfigSpy = vi.spyOn(OpenHands, "getConfig");
-
-  const RouterStub = createRoutesStub([
-    {
-      // layout route
-      Component: MainApp,
-      path: "/",
-      children: [
-        {
-          // home route
-          Component: Home,
-          path: "/",
-        },
-        {
-          Component: SettingsScreen,
-          path: "/settings",
-        },
-      ],
-    },
-  ]);
-
-  afterEach(() => {
-    vi.clearAllMocks();
-  });
-
   it("should render the home screen", () => {
     renderWithProviders(<RouterStub initialEntries={["/"]} />);
   });
@@ -79,57 +78,82 @@ describe("Home Screen", () => {
     const settingsScreen = await screen.findByTestId("settings-screen");
     expect(settingsScreen).toBeInTheDocument();
   });
+});
 
-  describe("Settings 404", () => {
-    it("should open the settings modal if GET /settings fails with a 404", async () => {
-      const error = createAxiosNotFoundErrorObject();
-      getSettingsSpy.mockRejectedValue(error);
+describe("Settings 404", () => {
+  const getConfigSpy = vi.spyOn(OpenHands, "getConfig");
 
-      renderWithProviders(<RouterStub initialEntries={["/"]} />);
+  it("should open the settings modal if GET /settings fails with a 404", async () => {
+    const error = createAxiosNotFoundErrorObject();
+    getSettingsSpy.mockRejectedValue(error);
 
-      const settingsModal = await screen.findByTestId("ai-config-modal");
-      expect(settingsModal).toBeInTheDocument();
-    });
+    renderWithProviders(<RouterStub initialEntries={["/"]} />);
 
-    it("should navigate to the settings screen when clicking the advanced settings button", async () => {
-      const error = createAxiosNotFoundErrorObject();
-      getSettingsSpy.mockRejectedValue(error);
+    const settingsModal = await screen.findByTestId("ai-config-modal");
+    expect(settingsModal).toBeInTheDocument();
+  });
 
-      const user = userEvent.setup();
-      renderWithProviders(<RouterStub initialEntries={["/"]} />);
+  it("should navigate to the settings screen when clicking the advanced settings button", async () => {
+    const error = createAxiosNotFoundErrorObject();
+    getSettingsSpy.mockRejectedValue(error);
 
-      const settingsScreen = screen.queryByTestId("settings-screen");
-      expect(settingsScreen).not.toBeInTheDocument();
+    const user = userEvent.setup();
+    renderWithProviders(<RouterStub initialEntries={["/"]} />);
 
-      const settingsModal = await screen.findByTestId("ai-config-modal");
-      expect(settingsModal).toBeInTheDocument();
+    const settingsScreen = screen.queryByTestId("settings-screen");
+    expect(settingsScreen).not.toBeInTheDocument();
 
-      const advancedSettingsButton = await screen.findByTestId(
-        "advanced-settings-link",
-      );
-      await user.click(advancedSettingsButton);
+    const settingsModal = await screen.findByTestId("ai-config-modal");
+    expect(settingsModal).toBeInTheDocument();
 
-      const settingsScreenAfter = await screen.findByTestId("settings-screen");
-      expect(settingsScreenAfter).toBeInTheDocument();
+    const advancedSettingsButton = await screen.findByTestId(
+      "advanced-settings-link",
+    );
+    await user.click(advancedSettingsButton);
 
-      const settingsModalAfter = screen.queryByTestId("ai-config-modal");
-      expect(settingsModalAfter).not.toBeInTheDocument();
-    });
+    const settingsScreenAfter = await screen.findByTestId("settings-screen");
+    expect(settingsScreenAfter).toBeInTheDocument();
 
-    it("should not open the settings modal if GET /settings fails but is SaaS mode", async () => {
-      // TODO: Remove HIDE_LLM_SETTINGS check once released
-      vi.spyOn(FeatureFlags, "HIDE_LLM_SETTINGS").mockReturnValue(true);
-      // @ts-expect-error - we only need APP_MODE for this test
-      getConfigSpy.mockResolvedValue({ APP_MODE: "saas" });
-      const error = createAxiosNotFoundErrorObject();
-      getSettingsSpy.mockRejectedValue(error);
+    const settingsModalAfter = screen.queryByTestId("ai-config-modal");
+    expect(settingsModalAfter).not.toBeInTheDocument();
+  });
 
-      renderWithProviders(<RouterStub initialEntries={["/"]} />);
+  it("should not open the settings modal if GET /settings fails but is SaaS mode", async () => {
+    // TODO: Remove HIDE_LLM_SETTINGS check once released
+    vi.spyOn(FeatureFlags, "HIDE_LLM_SETTINGS").mockReturnValue(true);
+    // @ts-expect-error - we only need APP_MODE for this test
+    getConfigSpy.mockResolvedValue({ APP_MODE: "saas" });
+    const error = createAxiosNotFoundErrorObject();
+    getSettingsSpy.mockRejectedValue(error);
 
-      // small hack to wait for the modal to not appear
-      await expect(
-        screen.findByTestId("ai-config-modal", {}, { timeout: 1000 }),
-      ).rejects.toThrow();
-    });
+    renderWithProviders(<RouterStub initialEntries={["/"]} />);
+
+    // small hack to wait for the modal to not appear
+    await expect(
+      screen.findByTestId("ai-config-modal", {}, { timeout: 1000 }),
+    ).rejects.toThrow();
+  });
+});
+
+describe("Setup Payment modal", () => {
+  const getConfigSpy = vi.spyOn(OpenHands, "getConfig");
+
+  afterEach(() => {
+    vi.resetAllMocks();
+  });
+
+  it("should only render if SaaS mode and is new user", async () => {
+    // @ts-expect-error - we only need the APP_MODE for this test
+    getConfigSpy.mockResolvedValue({
+      APP_MODE: "saas",
+    });
+    vi.spyOn(FeatureFlags, "BILLING_SETTINGS").mockReturnValue(true);
+    const error = createAxiosNotFoundErrorObject();
+    getSettingsSpy.mockRejectedValue(error);
+
+    renderWithProviders(<RouterStub initialEntries={["/"]} />);
+
+    const setupPaymentModal = await screen.findByTestId("proceed-to-stripe-button");
+    expect(setupPaymentModal).toBeInTheDocument();
   });
 });

frontend/__tests__/routes/settings.test.tsx

@@ -721,7 +721,7 @@ describe("Settings Screen", () => {
       expect(saveSettingsSpy).toHaveBeenCalledWith(
         expect.objectContaining({
           llm_api_key: "", // empty because it's not set previously
-          github_token: undefined,
+          provider_tokens: undefined,
           language: "no",
         }),
       );
@@ -758,7 +758,7 @@ describe("Settings Screen", () => {
 
       expect(saveSettingsSpy).toHaveBeenCalledWith(
         expect.objectContaining({
-          github_token: undefined,
+          provider_tokens: undefined,
           llm_api_key: "", // empty because it's not set previously
           llm_model: "openai/gpt-4o",
         }),
@@ -801,7 +801,7 @@ describe("Settings Screen", () => {
 
       expect(saveSettingsSpy).toHaveBeenCalledWith({
         ...mockCopy,
-        github_token: undefined, // not set
+        provider_tokens: undefined, // not set
         llm_api_key: "", // reset as well
       });
       expect(screen.queryByTestId("reset-modal")).not.toBeInTheDocument();

frontend/__tests__/services/actions.test.ts

@@ -1,7 +1,9 @@
 import { describe, it, expect, vi, beforeEach } from "vitest";
-import { handleStatusMessage } from "#/services/actions";
+import { handleStatusMessage, handleActionMessage } from "#/services/actions";
 import store from "#/store";
 import { trackError } from "#/utils/error-handler";
+import ActionType from "#/types/action-type";
+import { ActionMessage } from "#/types/message";
 
 // Mock dependencies
 vi.mock("#/utils/error-handler", () => ({
@@ -56,4 +58,89 @@ describe("Actions Service", () => {
       }));
     });
   });
+
+  describe("handleActionMessage", () => {
+    it("should use first-person perspective for task completion messages", () => {
+      // Test partial completion
+      const messagePartial: ActionMessage = {
+        id: 1,
+        action: ActionType.FINISH,
+        source: "agent",
+        message: "",
+        timestamp: new Date().toISOString(),
+        args: {
+          final_thought: "",
+          task_completed: "partial",
+          outputs: "",
+          thought: ""
+        }
+      };
+
+      // Mock implementation to capture the message
+      let capturedPartialMessage = "";
+      (store.dispatch as any).mockImplementation((action: any) => {
+        if (action.type === "chat/addAssistantMessage" &&
+            action.payload.includes("believe that the task was **completed partially**")) {
+          capturedPartialMessage = action.payload;
+        }
+      });
+
+      handleActionMessage(messagePartial);
+      expect(capturedPartialMessage).toContain("I believe that the task was **completed partially**");
+
+      // Test not completed
+      const messageNotCompleted: ActionMessage = {
+        id: 2,
+        action: ActionType.FINISH,
+        source: "agent",
+        message: "",
+        timestamp: new Date().toISOString(),
+        args: {
+          final_thought: "",
+          task_completed: "false",
+          outputs: "",
+          thought: ""
+        }
+      };
+
+      // Mock implementation to capture the message
+      let capturedNotCompletedMessage = "";
+      (store.dispatch as any).mockImplementation((action: any) => {
+        if (action.type === "chat/addAssistantMessage" &&
+            action.payload.includes("believe that the task was **not completed**")) {
+          capturedNotCompletedMessage = action.payload;
+        }
+      });
+
+      handleActionMessage(messageNotCompleted);
+      expect(capturedNotCompletedMessage).toContain("I believe that the task was **not completed**");
+
+      // Test completed successfully
+      const messageCompleted: ActionMessage = {
+        id: 3,
+        action: ActionType.FINISH,
+        source: "agent",
+        message: "",
+        timestamp: new Date().toISOString(),
+        args: {
+          final_thought: "",
+          task_completed: "true",
+          outputs: "",
+          thought: ""
+        }
+      };
+
+      // Mock implementation to capture the message
+      let capturedCompletedMessage = "";
+      (store.dispatch as any).mockImplementation((action: any) => {
+        if (action.type === "chat/addAssistantMessage" &&
+            action.payload.includes("believe that the task was **completed successfully**")) {
+          capturedCompletedMessage = action.payload;
+        }
+      });
+
+      handleActionMessage(messageCompleted);
+      expect(capturedCompletedMessage).toContain("I believe that the task was **completed successfully**");
+    });
+  });
 });

frontend/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "openhands-frontend",
-  "version": "0.28.1",
+  "version": "0.29.0",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "openhands-frontend",
-      "version": "0.28.1",
+      "version": "0.29.0",
       "dependencies": {
         "@heroui/react": "2.7.4",
         "@monaco-editor/react": "^4.7.0-rc.0",

frontend/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openhands-frontend",
-  "version": "0.28.1",
+  "version": "0.29.0",
   "private": true,
   "type": "module",
   "engines": {

frontend/public/mockServiceWorker.js

@@ -8,7 +8,7 @@
  * - Please do NOT serve this file on production.
  */
 
-const PACKAGE_VERSION = '2.7.0'
+const PACKAGE_VERSION = '2.7.3'
 const INTEGRITY_CHECKSUM = '00729d72e3b82faf54ca8b9621dbb96f'
 const IS_MOCKED_RESPONSE = Symbol('isMockedResponse')
 const activeClientIds = new Set()

frontend/src/api/open-hands.ts

@@ -281,6 +281,13 @@ class OpenHands {
     return data.redirect_url;
   }
 
+  static async createBillingSessionResponse(): Promise<string> {
+    const { data } = await openHands.post(
+      "/api/billing/create-customer-setup-session",
+    );
+    return data.redirect_url;
+  }
+
   static async getBalance(): Promise<string> {
     const { data } = await openHands.get<{ credits: string }>(
       "/api/billing/credits",

frontend/src/api/open-hands.types.ts

@@ -48,6 +48,7 @@ export interface GetConfigResponse {
   APP_SLUG?: string;
   GITHUB_CLIENT_ID: string;
   POSTHOG_CLIENT_KEY: string;
+  STRIPE_PUBLISHABLE_KEY?: string;
 }
 
 export interface GetVSCodeUrlResponse {

frontend/src/components/features/payment/setup-payment-modal.tsx

@@ -0,0 +1,48 @@
+import { useMutation } from "@tanstack/react-query";
+import { Trans, useTranslation } from "react-i18next";
+import AllHandsLogo from "#/assets/branding/all-hands-logo.svg?react";
+import { ModalBackdrop } from "#/components/shared/modals/modal-backdrop";
+import { ModalBody } from "#/components/shared/modals/modal-body";
+import OpenHands from "#/api/open-hands";
+import { BrandButton } from "../settings/brand-button";
+import { displayErrorToast } from "#/utils/custom-toast-handlers";
+
+export function SetupPaymentModal() {
+  const { t } = useTranslation();
+  const { mutate, isPending } = useMutation({
+    mutationFn: OpenHands.createBillingSessionResponse,
+    onSuccess: (data) => {
+      window.location.href = data;
+    },
+    onError: () => {
+      displayErrorToast(t("BILLING$ERROR_WHILE_CREATING_SESSION"));
+    },
+  });
+
+  return (
+    <ModalBackdrop>
+      <ModalBody className="border border-tertiary">
+        <AllHandsLogo width={68} height={46} />
+        <div className="flex flex-col gap-2 w-full items-center text-center">
+          <h1 className="text-2xl font-bold">{t("BILLING$YOUVE_GOT_50")}</h1>
+          <p>
+            <Trans
+              i18nKey="BILLING$CLAIM_YOUR_50"
+              components={{ b: <strong /> }}
+            />
+          </p>
+        </div>
+        <BrandButton
+          testId="proceed-to-stripe-button"
+          type="submit"
+          variant="primary"
+          className="w-full"
+          isDisabled={isPending}
+          onClick={mutate}
+        >
+          {t("BILLING$PROCEED_TO_STRIPE")}
+        </BrandButton>
+      </ModalBody>
+    </ModalBackdrop>
+  );
+}

frontend/src/components/features/sidebar/sidebar.tsx

@@ -61,7 +61,7 @@ export function Sidebar() {
       displayErrorToast(
         "Something went wrong while fetching settings. Please reload the page.",
       );
-    } else if (settingsError?.status === 404) {
+    } else if (config?.APP_MODE === "oss" && settingsError?.status === 404) {
       setSettingsModalIsOpen(true);
     }
   }, [

frontend/src/hooks/mutation/use-save-settings.ts

@@ -17,7 +17,7 @@ const saveSettingsMutationFn = async (settings: Partial<PostSettings>) => {
       ? ""
       : settings.LLM_API_KEY?.trim() || undefined,
     remote_runtime_resource_factor: settings.REMOTE_RUNTIME_RESOURCE_FACTOR,
-    github_token: settings.github_token,
+    provider_tokens: settings.provider_tokens,
     unset_github_token: settings.unset_github_token,
     enable_default_condenser: settings.ENABLE_DEFAULT_CONDENSER,
     enable_sound_notifications: settings.ENABLE_SOUND_NOTIFICATIONS,

frontend/src/hooks/query/use-balance.ts

@@ -1,6 +1,7 @@
 import { useQuery } from "@tanstack/react-query";
 import { useConfig } from "./use-config";
 import OpenHands from "#/api/open-hands";
+import { BILLING_SETTINGS } from "#/utils/feature-flags";
 
 export const useBalance = () => {
   const { data: config } = useConfig();
@@ -8,6 +9,6 @@ export const useBalance = () => {
   return useQuery({
     queryKey: ["user", "balance"],
     queryFn: OpenHands.getBalance,
-    enabled: config?.APP_MODE === "saas",
+    enabled: config?.APP_MODE === "saas" && BILLING_SETTINGS(),
   });
 };

frontend/src/hooks/query/use-settings.ts

@@ -21,6 +21,8 @@ const getSettingsQueryFn = async () => {
     ENABLE_DEFAULT_CONDENSER: apiSettings.enable_default_condenser,
     ENABLE_SOUND_NOTIFICATIONS: apiSettings.enable_sound_notifications,
     USER_CONSENTS_TO_ANALYTICS: apiSettings.user_consents_to_analytics,
+    PROVIDER_TOKENS: apiSettings.provider_tokens,
+    IS_NEW_USER: false,
   };
 };
 

frontend/src/i18n/declaration.ts

@@ -312,4 +312,9 @@ export enum I18nKey {
   BUTTON$MARK_NOT_HELPFUL = "BUTTON$MARK_NOT_HELPFUL",
   BUTTON$EXPORT_CONVERSATION = "BUTTON$EXPORT_CONVERSATION",
   BILLING$CLICK_TO_TOP_UP = "BILLING$CLICK_TO_TOP_UP",
+  BILLING$YOUVE_GOT_50 = "BILLING$YOUVE_GOT_50",
+  BILLING$ERROR_WHILE_CREATING_SESSION = "BILLING$ERROR_WHILE_CREATING_SESSION",
+  BILLING$CLAIM_YOUR_50 = "BILLING$CLAIM_YOUR_50",
+  BILLING$PROCEED_TO_STRIPE = "BILLING$PROCEED_TO_STRIPE",
+  BILLING$YOURE_IN = "BILLING$YOURE_IN",
 }

frontend/src/i18n/translation.json

@@ -4647,5 +4647,80 @@
         "fr": "Ajouter des fonds à votre compte",
         "tr": "Hesabınıza bakiye ekleyin",
         "de": "Guthaben zu Ihrem Konto hinzufügen"
+    },
+    "BILLING$YOUVE_GOT_50": {
+        "en": "You've got $50 in free OpenHands credits",
+        "ja": "OpenHandsの無料クレジット$50を獲得しました",
+        "zh-CN": "您获得了 $50 的 OpenHands 免费额度",
+        "zh-TW": "您獲得了 $50 的 OpenHands 免費額度",
+        "ko-KR": "OpenHands 무료 크레딧 $50를 받았습니다",
+        "no": "Du har fått $50 i gratis OpenHands-kreditter",
+        "it": "Hai ottenuto $50 in crediti gratuiti OpenHands",
+        "pt": "Você ganhou $50 em créditos gratuitos OpenHands",
+        "es": "Has recibido $50 en créditos gratuitos de OpenHands",
+        "ar": "لديك 50$ من رصيد OpenHands المجاني",
+        "fr": "Vous avez reçu $50 de crédits OpenHands gratuits",
+        "tr": "OpenHands'de $50 ücretsiz kredi kazandınız",
+        "de": "Sie haben $50 in kostenlosen OpenHands-Guthaben erhalten"
+    },
+    "BILLING$ERROR_WHILE_CREATING_SESSION": {
+        "en": "Error occurred while setting up your payment session. Please try again later.",
+        "ja": "お支払いセッションの設定中にエラーが発生しました。後ほど再度お試しください。",
+        "zh-CN": "设置支付会话时发生错误。请稍后再试。",
+        "zh-TW": "設置支付會話時發生錯誤。請稍後再試。",
+        "ko-KR": "결제 세션 설정 중 오류가 발생했습니다. 나중에 다시 시도해 주세요.",
+        "no": "Det oppstod en feil under oppsett av betalingsøkten. Vennligst prøv igjen senere.",
+        "it": "Si è verificato un errore durante la configurazione della sessione di pagamento. Si prega di riprovare più tardi.",
+        "pt": "Ocorreu um erro ao configurar sua sessão de pagamento. Por favor, tente novamente mais tarde.",
+        "es": "Se produjo un error al configurar tu sesión de pago. Por favor, inténtalo de nuevo más tarde.",
+        "ar": "حدث خطأ أثناء إعداد جلسة الدفع الخاصة بك. يرجى المحاولة مرة أخرى لاحقًا.",
+        "fr": "Une erreur s'est produite lors de la configuration de votre session de paiement. Veuillez réessayer plus tard.",
+        "tr": "Ödeme oturumunuz kurulurken bir hata oluştu. Lütfen daha sonra tekrar deneyin.",
+        "de": "Beim Einrichten Ihrer Zahlungssitzung ist ein Fehler aufgetreten. Bitte versuchen Sie es später erneut."
+    },
+    "BILLING$CLAIM_YOUR_50": {
+        "en": "Add a credit card with Stripe to claim your $50. <b>We won't charge you without asking first!</b>",
+        "ja": "Stripeでクレジットカードを追加して$50を獲得。<b>事前の確認なしで請求することはありません！</b>",
+        "zh-CN": "添加 Stripe 信用卡以领取 $50。<b>我们不会在未经您同意的情况下收费！</b>",
+        "zh-TW": "添加 Stripe 信用卡以領取 $50。<b>我們不會在未經您同意的情況下收費！</b>",
+        "ko-KR": "Stripe에 신용카드를 추가하여 $50를 받으세요. <b>사전 동의 없이 요금이 청구되지 않습니다!</b>",
+        "no": "Legg til et kredittkort med Stripe for å få $50. <b>Vi belaster deg ikke uten å spørre først!</b>",
+        "it": "Aggiungi una carta di credito con Stripe per ottenere $50. <b>Non ti addebiteremo nulla senza chiedere prima!</b>",
+        "pt": "Adicione um cartão de crédito com Stripe para receber $50. <b>Não cobraremos sem perguntar primeiro!</b>",
+        "es": "Añade una tarjeta de crédito con Stripe para reclamar tus $50. <b>¡No te cobraremos sin preguntarte primero!</b>",
+        "ar": "أضف بطاقة ائتمان مع Stripe للحصول على 50$. <b>لن نقوم بالخصم دون إذن مسبق!</b>",
+        "fr": "Ajoutez une carte de crédit avec Stripe pour obtenir 50$. <b>Nous ne vous facturerons pas sans vous demander d'abord !</b>",
+        "tr": "50$ almak için Stripe ile kredi kartı ekleyin. <b>Önce sormadan ücret almayacağız!</b>",
+        "de": "Fügen Sie eine Kreditkarte mit Stripe hinzu, um $50 zu erhalten. <b>Wir belasten Sie nicht ohne vorherige Zustimmung!</b>"
+    },
+    "BILLING$PROCEED_TO_STRIPE": {
+        "en": "Add Billing Info",
+        "ja": "請求情報を追加",
+        "zh-CN": "添加账单信息",
+        "zh-TW": "添加帳單資訊",
+        "ko-KR": "결제 정보 추가",
+        "no": "Legg til betalingsinformasjon",
+        "it": "Aggiungi informazioni di fatturazione",
+        "pt": "Adicionar informações de pagamento",
+        "es": "Añadir información de facturación",
+        "ar": "إضافة معلومات الفواتير",
+        "fr": "Ajouter les informations de facturation",
+        "tr": "Fatura Bilgisi Ekle",
+        "de": "Zahlungsinformationen hinzufügen"
+    },
+    "BILLING$YOURE_IN": {
+        "en": "You're in! You can start using your $50 in free credits now.",
+        "ja": "登録完了！$50分の無料クレジットを今すぐご利用いただけます。",
+        "zh-CN": "您已加入！现在可以开始使用$50的免费额度了。",
+        "zh-TW": "您已加入！現在可以開始使用$50的免費額度了。",
+        "ko-KR": "가입 완료! 지금 바로 $50 상당의 무료 크레딧을 사용하실 수 있습니다.",
+        "no": "Du er med! Du kan begynne å bruke dine $50 i gratis kreditter nå.",
+        "it": "Ci sei! Puoi iniziare a utilizzare i tuoi $50 in crediti gratuiti ora.",
+        "pt": "Você está dentro! Você pode começar a usar seus $50 em créditos gratuitos agora.",
+        "es": "¡Ya estás dentro! Puedes empezar a usar tus $50 en créditos gratuitos ahora.",
+        "ar": "أنت معنا! يمكنك البدء في استخدام رصيدك المجاني البالغ 50 دولارًا الآن.",
+        "fr": "C'est fait ! Vous pouvez commencer à utiliser vos 50 $ de crédits gratuits maintenant.",
+        "tr": "Başardın! Şimdi $50 değerindeki ücretsiz kredilerini kullanmaya başlayabilirsin.",
+        "de": "Du bist dabei! Du kannst jetzt deine $50 an kostenlosen Guthaben nutzen."
     }
 }

frontend/src/mocks/billing-handlers.ts

@@ -1,8 +1,4 @@
 import { delay, http, HttpResponse } from "msw";
-import Stripe from "stripe";
-
-const TEST_STRIPE_SECRET_KEY = "";
-const PRICE_ID = "";
 
 export const STRIPE_BILLING_HANDLERS = [
   http.get("/api/billing/credits", async () => {
@@ -10,27 +6,17 @@ export const STRIPE_BILLING_HANDLERS = [
     return HttpResponse.json({ credits: "100" });
   }),
 
-  http.post("/api/billing/create-checkout-session", async ({ request }) => {
+  http.post("/api/billing/create-checkout-session", async () => {
     await delay();
-    const body = await request.json();
+    return HttpResponse.json({
+      redirect_url: "https://stripe.com/some-checkout",
+    });
+  }),
 
-    if (body && typeof body === "object" && body.amount) {
-      const stripe = new Stripe(TEST_STRIPE_SECRET_KEY);
-      const session = await stripe.checkout.sessions.create({
-        line_items: [
-          {
-            price: PRICE_ID,
-            quantity: body.amount,
-          },
-        ],
-        mode: "payment",
-        success_url: "http://localhost:3001/settings/billing/?checkout=success",
-        cancel_url: "http://localhost:3001/settings/billing/?checkout=cancel",
-      });
-
-      if (session.url) return HttpResponse.json({ redirect_url: session.url });
-    }
-
-    return HttpResponse.json({ message: "Invalid request" }, { status: 400 });
+  http.post("/api/billing/create-customer-setup-session", async () => {
+    await delay();
+    return HttpResponse.json({
+      redirect_url: "https://stripe.com/some-customer-setup",
+    });
   }),
 ];

frontend/src/mocks/handlers.ts

@@ -22,12 +22,13 @@ export const MOCK_DEFAULT_USER_SETTINGS: ApiSettings | PostApiSettings = {
   enable_default_condenser: DEFAULT_SETTINGS.ENABLE_DEFAULT_CONDENSER,
   enable_sound_notifications: DEFAULT_SETTINGS.ENABLE_SOUND_NOTIFICATIONS,
   user_consents_to_analytics: DEFAULT_SETTINGS.USER_CONSENTS_TO_ANALYTICS,
+  provider_tokens: DEFAULT_SETTINGS.PROVIDER_TOKENS,
 };
 
 const MOCK_USER_PREFERENCES: {
-  settings: ApiSettings | PostApiSettings;
+  settings: ApiSettings | PostApiSettings | null;
 } = {
-  settings: MOCK_DEFAULT_USER_SETTINGS,
+  settings: null,
 };
 
 const conversations: Conversation[] = [
@@ -174,22 +175,24 @@ export const handlers = [
   ),
   http.get("/api/options/config", () => {
     const mockSaas = import.meta.env.VITE_MOCK_SAAS === "true";
+
     const config: GetConfigResponse = {
       APP_MODE: mockSaas ? "saas" : "oss",
       GITHUB_CLIENT_ID: "fake-github-client-id",
       POSTHOG_CLIENT_KEY: "fake-posthog-client-key",
+      STRIPE_PUBLISHABLE_KEY: "",
     };
 
     return HttpResponse.json(config);
   }),
   http.get("/api/settings", async () => {
     await delay();
-    const settings: ApiSettings = {
-      ...MOCK_USER_PREFERENCES.settings,
-      language: "no",
-    };
-    // @ts-expect-error - mock types
-    if (settings.github_token) settings.github_token_is_set = true;
+    const { settings } = MOCK_USER_PREFERENCES;
+
+    if (!settings) return HttpResponse.json(null, { status: 404 });
+
+    if (Object.keys(settings.provider_tokens).length > 0)
+      settings.github_token_is_set = true;
 
     return HttpResponse.json(settings);
   }),
@@ -201,17 +204,19 @@ export const handlers = [
       if (typeof body === "object") {
         newSettings = { ...body };
         if (newSettings.unset_github_token) {
-          newSettings.github_token = undefined;
+          newSettings.provider_tokens = { github: "", gitlab: "" };
           newSettings.github_token_is_set = false;
           delete newSettings.unset_github_token;
         }
       }
 
-      MOCK_USER_PREFERENCES.settings = {
+      const fullSettings = {
+        ...MOCK_DEFAULT_USER_SETTINGS,
         ...MOCK_USER_PREFERENCES.settings,
         ...newSettings,
       };
 
+      MOCK_USER_PREFERENCES.settings = fullSettings;
       return HttpResponse.json(null, { status: 200 });
     }
 

frontend/src/routes/_oh._index/route.tsx

@@ -24,7 +24,10 @@ function Home() {
   });
 
   return (
-    <div className="bg-base-secondary h-full rounded-xl flex flex-col items-center justify-center relative overflow-y-auto px-2">
+    <div
+      data-testid="home-screen"
+      className="bg-base-secondary h-full rounded-xl flex flex-col items-center justify-center relative overflow-y-auto px-2"
+    >
       <HeroHeading />
       <div className="flex flex-col gap-8 w-full md:w-[600px] items-center">
         <div className="flex flex-col gap-2 w-full">

frontend/src/routes/_oh/route.tsx

@@ -1,5 +1,13 @@
 import React from "react";
-import { useRouteError, isRouteErrorResponse, Outlet } from "react-router";
+import {
+  useRouteError,
+  isRouteErrorResponse,
+  Outlet,
+  useNavigate,
+  useLocation,
+  useSearchParams,
+} from "react-router";
+import { useTranslation } from "react-i18next";
 import i18n from "#/i18n";
 import { useGitHubAuthUrl } from "#/hooks/use-github-auth-url";
 import { useIsAuthed } from "#/hooks/query/use-is-authed";
@@ -10,6 +18,10 @@ import { AnalyticsConsentFormModal } from "#/components/features/analytics/analy
 import { useSettings } from "#/hooks/query/use-settings";
 import { useAuth } from "#/context/auth-context";
 import { useMigrateUserConsent } from "#/hooks/use-migrate-user-consent";
+import { useBalance } from "#/hooks/query/use-balance";
+import { SetupPaymentModal } from "#/components/features/payment/setup-payment-modal";
+import { BILLING_SETTINGS } from "#/utils/feature-flags";
+import { displaySuccessToast } from "#/utils/custom-toast-handlers";
 
 export function ErrorBoundary() {
   const error = useRouteError();
@@ -44,11 +56,14 @@ export function ErrorBoundary() {
 }
 
 export default function MainApp() {
+  const navigate = useNavigate();
+  const { pathname } = useLocation();
+  const [searchParams] = useSearchParams();
   const { githubTokenIsSet } = useAuth();
   const { data: settings } = useSettings();
+  const { error, isFetching } = useBalance();
   const { migrateUserConsent } = useMigrateUserConsent();
-
-  const [consentFormIsOpen, setConsentFormIsOpen] = React.useState(false);
+  const { t } = useTranslation();
 
   const config = useConfig();
   const {
@@ -62,6 +77,8 @@ export default function MainApp() {
     gitHubClientId: config.data?.GITHUB_CLIENT_ID || null,
   });
 
+  const [consentFormIsOpen, setConsentFormIsOpen] = React.useState(false);
+
   React.useEffect(() => {
     if (settings?.LANGUAGE) {
       i18n.changeLanguage(settings.LANGUAGE);
@@ -84,6 +101,17 @@ export default function MainApp() {
     });
   }, []);
 
+  React.useEffect(() => {
+    // Don't allow users to use the app if it 402s
+    if (error?.status === 402 && pathname !== "/") {
+      navigate("/");
+    } else if (!isFetching && searchParams.get("free_credits") === "success") {
+      displaySuccessToast(t("BILLING$YOURE_IN"));
+      searchParams.delete("free_credits");
+      navigate("/");
+    }
+  }, [error?.status, pathname, isFetching]);
+
   const userIsAuthed = !!isAuthed && !authError;
   const renderWaitlistModal =
     !isFetchingAuth && !userIsAuthed && config.data?.APP_MODE === "saas";
@@ -116,6 +144,10 @@ export default function MainApp() {
           }}
         />
       )}
+
+      {BILLING_SETTINGS() &&
+        config.data?.APP_MODE === "saas" &&
+        settings?.IS_NEW_USER && <SetupPaymentModal />}
     </div>
   );
 }

frontend/src/routes/account-settings.tsx

@@ -61,7 +61,10 @@ function AccountSettings() {
     if (isSuccess) {
       return (
         isCustomModel(resources.models, settings.LLM_MODEL) ||
-        hasAdvancedSettingsSet(settings)
+        hasAdvancedSettingsSet({
+          ...settings,
+          PROVIDER_TOKENS: settings.PROVIDER_TOKENS || {},
+        })
       );
     }
 
@@ -128,37 +131,42 @@ function AccountSettings() {
       : llmBaseUrl;
     const finalLlmApiKey = shouldHandleSpecialSaasCase ? undefined : llmApiKey;
 
-    saveSettings(
-      {
-        github_token:
-          formData.get("github-token-input")?.toString() || undefined,
-        LANGUAGE: languageValue,
-        user_consents_to_analytics: userConsentsToAnalytics,
-        ENABLE_DEFAULT_CONDENSER: enableMemoryCondenser,
-        ENABLE_SOUND_NOTIFICATIONS: enableSoundNotifications,
-        LLM_MODEL: finalLlmModel,
-        LLM_BASE_URL: finalLlmBaseUrl,
-        LLM_API_KEY: finalLlmApiKey,
-        AGENT: formData.get("agent-input")?.toString(),
-        SECURITY_ANALYZER:
-          formData.get("security-analyzer-input")?.toString() || "",
-        REMOTE_RUNTIME_RESOURCE_FACTOR:
-          remoteRuntimeResourceFactor ||
-          DEFAULT_SETTINGS.REMOTE_RUNTIME_RESOURCE_FACTOR,
-        CONFIRMATION_MODE: confirmationModeIsEnabled,
+    const githubToken = formData.get("github-token-input")?.toString();
+    const newSettings = {
+      github_token: githubToken,
+      provider_tokens: githubToken
+        ? {
+            github: githubToken,
+            gitlab: "",
+          }
+        : undefined,
+      LANGUAGE: languageValue,
+      user_consents_to_analytics: userConsentsToAnalytics,
+      ENABLE_DEFAULT_CONDENSER: enableMemoryCondenser,
+      ENABLE_SOUND_NOTIFICATIONS: enableSoundNotifications,
+      LLM_MODEL: finalLlmModel,
+      LLM_BASE_URL: finalLlmBaseUrl,
+      LLM_API_KEY: finalLlmApiKey,
+      AGENT: formData.get("agent-input")?.toString(),
+      SECURITY_ANALYZER:
+        formData.get("security-analyzer-input")?.toString() || "",
+      REMOTE_RUNTIME_RESOURCE_FACTOR:
+        remoteRuntimeResourceFactor ||
+        DEFAULT_SETTINGS.REMOTE_RUNTIME_RESOURCE_FACTOR,
+      CONFIRMATION_MODE: confirmationModeIsEnabled,
+    };
+
+    saveSettings(newSettings, {
+      onSuccess: () => {
+        handleCaptureConsent(userConsentsToAnalytics);
+        displaySuccessToast("Settings saved");
+        setLlmConfigMode(isAdvancedSettingsSet ? "advanced" : "basic");
       },
-      {
-        onSuccess: () => {
-          handleCaptureConsent(userConsentsToAnalytics);
-          displaySuccessToast("Settings saved");
-          setLlmConfigMode(isAdvancedSettingsSet ? "advanced" : "basic");
-        },
-        onError: (error) => {
-          const errorMessage = retrieveAxiosErrorMessage(error);
-          displayErrorToast(errorMessage);
-        },
+      onError: (error) => {
+        const errorMessage = retrieveAxiosErrorMessage(error);
+        displayErrorToast(errorMessage);
       },
-    );
+    });
   };
 
   const handleReset = () => {

frontend/src/services/actions.ts

@@ -62,13 +62,12 @@ const messageActions = {
     let successPrediction = "";
     if (message.args.task_completed === "partial") {
       successPrediction =
-        "The agent thinks that the task was **completed partially**.";
+        "I believe that the task was **completed partially**.";
     } else if (message.args.task_completed === "false") {
-      successPrediction =
-        "The agent thinks that the task was **not completed**.";
+      successPrediction = "I believe that the task was **not completed**.";
     } else if (message.args.task_completed === "true") {
       successPrediction =
-        "The agent thinks that the task was **completed successfully**.";
+        "I believe that the task was **completed successfully**.";
     }
     if (successPrediction) {
       // if final_thought is not empty, add a new line before the success prediction

frontend/src/services/settings.ts

@@ -15,6 +15,11 @@ export const DEFAULT_SETTINGS: Settings = {
   ENABLE_DEFAULT_CONDENSER: true,
   ENABLE_SOUND_NOTIFICATIONS: false,
   USER_CONSENTS_TO_ANALYTICS: false,
+  PROVIDER_TOKENS: {
+    github: "",
+    gitlab: "",
+  },
+  IS_NEW_USER: true,
 };
 
 /**

frontend/src/types/core/variances.ts

@@ -21,7 +21,6 @@ export interface InitConfig {
     LLM_MODEL: string;
   };
   token?: string;
-  github_token?: string;
   latest_event_id?: unknown; // Not sure what this is
 }
 

frontend/src/types/settings.ts

@@ -1,3 +1,5 @@
+export type Provider = "github" | "gitlab";
+
 export type Settings = {
   LLM_MODEL: string;
   LLM_BASE_URL: string;
@@ -11,6 +13,8 @@ export type Settings = {
   ENABLE_DEFAULT_CONDENSER: boolean;
   ENABLE_SOUND_NOTIFICATIONS: boolean;
   USER_CONSENTS_TO_ANALYTICS: boolean | null;
+  PROVIDER_TOKENS: Record<Provider, string>;
+  IS_NEW_USER?: boolean;
 };
 
 export type ApiSettings = {
@@ -26,16 +30,17 @@ export type ApiSettings = {
   enable_default_condenser: boolean;
   enable_sound_notifications: boolean;
   user_consents_to_analytics: boolean | null;
+  provider_tokens: Record<Provider, string>;
 };
 
 export type PostSettings = Settings & {
-  github_token: string;
+  provider_tokens: Record<Provider, string>;
   unset_github_token: boolean;
   user_consents_to_analytics: boolean | null;
 };
 
 export type PostApiSettings = ApiSettings & {
-  github_token: string;
+  provider_tokens: Record<Provider, string>;
   unset_github_token: boolean;
   user_consents_to_analytics: boolean | null;
 };

frontend/src/utils/settings-utils.ts

@@ -59,6 +59,18 @@ export const extractSettings = (formData: FormData): Partial<Settings> => {
     ENABLE_DEFAULT_CONDENSER,
   } = extractAdvancedFormData(formData);
 
+  // Extract provider tokens
+  const githubToken = formData.get("github-token")?.toString();
+  const gitlabToken = formData.get("gitlab-token")?.toString();
+  const providerTokens: Record<string, string> = {};
+
+  if (githubToken) {
+    providerTokens.github = githubToken;
+  }
+  if (gitlabToken) {
+    providerTokens.gitlab = gitlabToken;
+  }
+
   return {
     LLM_MODEL: CUSTOM_LLM_MODEL || LLM_MODEL,
     LLM_API_KEY,
@@ -68,5 +80,6 @@ export const extractSettings = (formData: FormData): Partial<Settings> => {
     CONFIRMATION_MODE,
     SECURITY_ANALYZER,
     ENABLE_DEFAULT_CONDENSER,
+    PROVIDER_TOKENS: providerTokens,
   };
 };

frontend/tailwind.config.js

@@ -17,45 +17,7 @@ export default {
         tertiary: "#454545", // gray, used for inputs
         "tertiary-light": "#B7BDC2", // lighter gray, used for borders and placeholder text
         content: "#ECEDEE", // light gray, used mostly for text
-      },
-    },
-    animation: {
-      enter: "toastIn 400ms cubic-bezier(0.21, 1.02, 0.73, 1)",
-      leave: "toastOut 100ms ease-in forwards",
-    },
-    keyframes: {
-      toastIn: {
-        "0%": {
-          opacity: "0",
-          transform: "translateY(-100%) scale(0.8)",
-        },
-        "80%": {
-          opacity: "1",
-          transform: "translateY(0) scale(1.02)",
-        },
-        "100%": {
-          opacity: "1",
-          transform: "translateY(0) scale(1)",
-        },
-      },
-      toastOut: {
-        "0%": {
-          opacity: "1",
-          transform: "translateY(0) scale(1)",
-        },
-        "100%": {
-          opacity: "0",
-          transform: "translateY(-100%) scale(0.9)",
-        },
-      },
-      colors: {
-        primary: "#C9B974", // nice yellow
-        base: "#171717", // dark background (neutral-900)
-        "base-secondary": "#262626", // lighter background (neutral-800); also used for tooltips
-        danger: "#E76A5E",
-        success: "#A5E75E",
-        tertiary: "#454545", // gray, used for inputs
-        "tertiary-light": "#B7BDC2", // lighter gray, used for borders and placeholder text
+        "content-2": "#F9FBFE",
       },
     },
   },

openhands/agenthub/__init__.py

@@ -6,7 +6,6 @@ load_dotenv()
 from openhands.agenthub import (  # noqa: E402
     browsing_agent,
     codeact_agent,
-    delegator_agent,
     dummy_agent,
     visualbrowsing_agent,
 )
@@ -15,7 +14,6 @@ from openhands.controller.agent import Agent  # noqa: E402
 __all__ = [
     'Agent',
     'codeact_agent',
-    'delegator_agent',
     'dummy_agent',
     'browsing_agent',
     'visualbrowsing_agent',

openhands/agenthub/codeact_agent/codeact_agent.py

@@ -1,8 +1,6 @@
-import json
 import os
 from collections import deque
 
-import openhands
 import openhands.agenthub.codeact_agent.function_calling as codeact_function_calling
 from openhands.controller.agent import Agent
 from openhands.controller.state.state import State
@@ -72,23 +70,17 @@ class CodeActAgent(Agent):
             codeact_enable_browsing=self.config.codeact_enable_browsing,
             codeact_enable_jupyter=self.config.codeact_enable_jupyter,
             codeact_enable_llm_editor=self.config.codeact_enable_llm_editor,
+            llm=self.llm,
         )
         logger.debug(
-            f'TOOLS loaded for CodeActAgent: {json.dumps(self.tools, indent=2, ensure_ascii=False).replace("\\n", "\n")}'
+            f"TOOLS loaded for CodeActAgent: {', '.join([tool.get('function').get('name') for tool in self.tools])}"
         )
         self.prompt_manager = PromptManager(
-            microagent_dir=os.path.join(
-                os.path.dirname(os.path.dirname(openhands.__file__)),
-                'microagents',
-            )
-            if self.config.enable_prompt_extensions
-            else None,
             prompt_dir=os.path.join(os.path.dirname(__file__), 'prompts'),
-            disabled_microagents=self.config.disabled_microagents,
         )
 
         # Create a ConversationMemory instance
-        self.conversation_memory = ConversationMemory(self.prompt_manager)
+        self.conversation_memory = ConversationMemory(self.config, self.prompt_manager)
 
         self.condenser = Condenser.from_config(self.config.condenser)
         logger.debug(f'Using condenser: {type(self.condenser)}')
@@ -168,7 +160,7 @@ class CodeActAgent(Agent):
         if not self.prompt_manager:
             raise Exception('Prompt Manager not instantiated.')
 
-        # Use conversation_memory to process events instead of calling events_to_messages directly
+        # Use ConversationMemory to process initial messages
         messages = self.conversation_memory.process_initial_messages(
             with_caching=self.llm.is_caching_prompt_active()
         )
@@ -180,12 +172,12 @@ class CodeActAgent(Agent):
             f'Processing {len(events)} events from a total of {len(state.history)} events'
         )
 
+        # Use ConversationMemory to process events
         messages = self.conversation_memory.process_events(
             condensed_history=events,
             initial_messages=messages,
             max_message_chars=self.llm.config.max_message_chars,
             vision_is_active=self.llm.vision_is_active(),
-            enable_som_visual_browsing=self.config.enable_som_visual_browsing,
         )
 
         messages = self._enhance_messages(messages)
@@ -216,14 +208,7 @@ class CodeActAgent(Agent):
                 # compose the first user message with examples
                 self.prompt_manager.add_examples_to_initial_message(msg)
 
-                # and/or repo/runtime info
-                if self.config.enable_prompt_extensions:
-                    self.prompt_manager.add_info_to_initial_message(msg)
-
-            # enhance the user message with additional context based on keywords matched
-            if msg.role == 'user':
-                self.prompt_manager.enhance_message(msg)
-
+            elif msg.role == 'user':
                 # Add double newline between consecutive user messages
                 if prev_role == 'user' and len(msg.content) > 0:
                     # Find the first TextContent in the message to add newlines

openhands/agenthub/codeact_agent/function_calling.py

@@ -12,13 +12,13 @@ from litellm import (
 
 from openhands.agenthub.codeact_agent.tools import (
     BrowserTool,
-    CmdRunTool,
     FinishTool,
     IPythonTool,
     LLMBasedFileEditTool,
-    StrReplaceEditorTool,
     ThinkTool,
     WebReadTool,
+    create_cmd_run_tool,
+    create_str_replace_editor_tool,
 )
 from openhands.core.exceptions import (
     FunctionCallNotExistsError,
@@ -39,6 +39,7 @@ from openhands.events.action import (
 )
 from openhands.events.event import FileEditSource, FileReadSource
 from openhands.events.tool import ToolCallMetadata
+from openhands.llm import LLM
 
 
 def combine_thought(action: Action, thought: str) -> Action:
@@ -80,7 +81,7 @@ def response_to_actions(response: ModelResponse) -> list[Action]:
             # CmdRunTool (Bash)
             # ================================================
 
-            if tool_call.function.name == CmdRunTool['function']['name']:
+            if tool_call.function.name == create_cmd_run_tool()['function']['name']:
                 if 'command' not in arguments:
                     raise FunctionCallValidationError(
                         f'Missing required argument "command" in tool call {tool_call.function.name}'
@@ -131,7 +132,10 @@ def response_to_actions(response: ModelResponse) -> list[Action]:
                     start=arguments.get('start', 1),
                     end=arguments.get('end', -1),
                 )
-            elif tool_call.function.name == StrReplaceEditorTool['function']['name']:
+            elif (
+                tool_call.function.name
+                == create_str_replace_editor_tool()['function']['name']
+            ):
                 if 'command' not in arguments:
                     raise FunctionCallValidationError(
                         f'Missing required argument "command" in tool call {tool_call.function.name}'
@@ -219,8 +223,22 @@ def get_tools(
     codeact_enable_browsing: bool = False,
     codeact_enable_llm_editor: bool = False,
     codeact_enable_jupyter: bool = False,
+    llm: LLM | None = None,
 ) -> list[ChatCompletionToolParam]:
-    tools = [CmdRunTool, ThinkTool, FinishTool]
+    SIMPLIFIED_TOOL_DESCRIPTION_LLM_SUBSTRS = ['gpt-', 'o3', 'o1']
+
+    use_simplified_tool_desc = False
+    if llm is not None:
+        use_simplified_tool_desc = any(
+            model_substr in llm.config.model
+            for model_substr in SIMPLIFIED_TOOL_DESCRIPTION_LLM_SUBSTRS
+        )
+
+    tools = [
+        create_cmd_run_tool(use_simplified_description=use_simplified_tool_desc),
+        ThinkTool,
+        FinishTool,
+    ]
     if codeact_enable_browsing:
         tools.append(WebReadTool)
         tools.append(BrowserTool)
@@ -229,5 +247,9 @@ def get_tools(
     if codeact_enable_llm_editor:
         tools.append(LLMBasedFileEditTool)
     else:
-        tools.append(StrReplaceEditorTool)
+        tools.append(
+            create_str_replace_editor_tool(
+                use_simplified_description=use_simplified_tool_desc
+            )
+        )
     return tools

openhands/agenthub/codeact_agent/prompts/additional_info.j2

@@ -1,6 +1,6 @@
 {% if repository_info %}
 <REPOSITORY_INFO>
-At the user's request, repository {{ repository_info.repo_name }} has been cloned to directory {{ repository_info.repo_directory }}.
+At the user's request, repository {{ repository_info.repo_name }} has been cloned to the current working directory {{ repository_info.repo_directory }}.
 </REPOSITORY_INFO>
 {% endif %}
 {% if repository_instructions -%}
@@ -20,6 +20,8 @@ When starting a web server, use the corresponding ports. You should also
 set any options to allow iframes and CORS requests, and allow the server to
 be accessed from any host (e.g. 0.0.0.0).
 {% endif %}
+{% if runtime_info.additional_agent_instructions %}
 {{ runtime_info.additional_agent_instructions }}
+{% endif %}
 </RUNTIME_INFORMATION>
 {% endif %}

openhands/agenthub/codeact_agent/prompts/microagent_info.j2

@@ -1,8 +1,8 @@
 {% for agent_info in triggered_agents %}
 <EXTRA_INFO>
-The following information has been included based on a keyword match for "{{ agent_info.trigger_word }}".
+The following information has been included based on a keyword match for "{{ agent_info.trigger }}".
 It may or may not be relevant to the user's request.
 
-{{ agent_info.agent.content }}
+{{ agent_info.content }}
 </EXTRA_INFO>
 {% endfor %}

openhands/agenthub/codeact_agent/tools/__init__.py

@@ -1,19 +1,19 @@
-from .bash import CmdRunTool
+from .bash import create_cmd_run_tool
 from .browser import BrowserTool
 from .finish import FinishTool
 from .ipython import IPythonTool
 from .llm_based_edit import LLMBasedFileEditTool
-from .str_replace_editor import StrReplaceEditorTool
+from .str_replace_editor import create_str_replace_editor_tool
 from .think import ThinkTool
 from .web_read import WebReadTool
 
 __all__ = [
     'BrowserTool',
-    'CmdRunTool',
+    'create_cmd_run_tool',
     'FinishTool',
     'IPythonTool',
     'LLMBasedFileEditTool',
-    'StrReplaceEditorTool',
+    'create_str_replace_editor_tool',
     'WebReadTool',
     'ThinkTool',
 ]

openhands/agenthub/codeact_agent/tools/bash.py

@@ -1,6 +1,6 @@
 from litellm import ChatCompletionToolParam, ChatCompletionToolParamFunctionChunk
 
-_BASH_DESCRIPTION = """Execute a bash command in the terminal within a persistent shell session.
+_DETAILED_BASH_DESCRIPTION = """Execute a bash command in the terminal within a persistent shell session.
 
 ### Command Execution
 * One command at a time: You can only execute one bash command at a time. If you need to run multiple commands sequentially, use `&&` or `;` to chain them together.
@@ -22,25 +22,39 @@ _BASH_DESCRIPTION = """Execute a bash command in the terminal within a persisten
 * Output truncation: If the output exceeds a maximum length, it will be truncated before being returned.
 """
 
-CmdRunTool = ChatCompletionToolParam(
-    type='function',
-    function=ChatCompletionToolParamFunctionChunk(
-        name='execute_bash',
-        description=_BASH_DESCRIPTION,
-        parameters={
-            'type': 'object',
-            'properties': {
-                'command': {
-                    'type': 'string',
-                    'description': 'The bash command to execute. Can be empty string to view additional logs when previous exit code is `-1`. Can be `C-c` (Ctrl+C) to interrupt the currently running process. Note: You can only execute one bash command at a time. If you need to run multiple commands sequentially, you can use `&&` or `;` to chain them together.',
-                },
-                'is_input': {
-                    'type': 'string',
-                    'description': 'If True, the command is an input to the running process. If False, the command is a bash command to be executed in the terminal. Default is False.',
-                    'enum': ['true', 'false'],
+_SIMPLIFIED_BASH_DESCRIPTION = """Execute a bash command in the terminal.
+* Long running commands: For commands that may run indefinitely, it should be run in the background and the output should be redirected to a file, e.g. command = `python3 app.py > server.log 2>&1 &`.
+* Interact with running process: If a bash command returns exit code `-1`, this means the process is not yet finished. By setting `is_input` to `true`, the assistant can interact with the running process and send empty `command` to retrieve any additional logs, or send additional text (set `command` to the text) to STDIN of the running process, or send command like `C-c` (Ctrl+C), `C-d` (Ctrl+D), `C-z` (Ctrl+Z) to interrupt the process.
+* One command at a time: You can only execute one bash command at a time. If you need to run multiple commands sequentially, you can use `&&` or `;` to chain them together."""
+
+
+def create_cmd_run_tool(
+    use_simplified_description: bool = False,
+) -> ChatCompletionToolParam:
+    description = (
+        _SIMPLIFIED_BASH_DESCRIPTION
+        if use_simplified_description
+        else _DETAILED_BASH_DESCRIPTION
+    )
+    return ChatCompletionToolParam(
+        type='function',
+        function=ChatCompletionToolParamFunctionChunk(
+            name='execute_bash',
+            description=description,
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'command': {
+                        'type': 'string',
+                        'description': 'The bash command to execute. Can be empty string to view additional logs when previous exit code is `-1`. Can be `C-c` (Ctrl+C) to interrupt the currently running process. Note: You can only execute one bash command at a time. If you need to run multiple commands sequentially, you can use `&&` or `;` to chain them together.',
+                    },
+                    'is_input': {
+                        'type': 'string',
+                        'description': 'If True, the command is an input to the running process. If False, the command is a bash command to be executed in the terminal. Default is False.',
+                        'enum': ['true', 'false'],
+                    },
                 },
+                'required': ['command'],
             },
-            'required': ['command'],
-        },
-    ),
-)
+        ),
+    )