Merge branch 'main' into rb/github-patch

This reverts commit 79492b6551.

"b/frontend/__tests__/utils/extractModelAndProvider.test.ts\""

@@ -1,5 +0,0 @@
-expect(extractModelAndProvider("claude-3-5-sonnet-20241022")).toEqual({
-      provider: "anthropic",
-      model: "claude-3-5-sonnet-20241022",
-      separator: "/",
-    });

"b/frontend/__tests__/utils/organizeModelsAndProviders.test.ts\""

@@ -1,65 +0,0 @@
-import { expect, test } from "vitest";
-import { organizeModelsAndProviders } from "../../src/utils/organizeModelsAndProviders";
-
-test("organizeModelsAndProviders", () => {
-  const models = [
-    "azure/ada",
-    "azure/gpt-35-turbo",
-    "azure/gpt-3-turbo",
-    "azure/standard/1024-x-1024/dall-e-2",
-    "vertex_ai_beta/chat-bison",
-    "vertex_ai_beta/chat-bison-32k",
-    "sagemaker/meta-textgeneration-llama-2-13b",
-    "cohere.command-r-v1:0",
-    "cloudflare/@cf/mistral/mistral-7b-instruct-v0.1",
-    "gpt-4o",
-    "together-ai-21.1b-41b",
-    "gpt-4o-mini",
-    "claude-3-5-sonnet-20241022",
-    "claude-3-haiku-20240307",
-    "claude-2",
-    "claude-2.1",
-    "anthropic.unsafe-claude-2.1",
-  ];
-
-  const object = organizeModelsAndProviders(models);
-
-  expect(object).toEqual({
-    azure: {
-      separator: "/",
-      models: [
-        "ada",
-        "gpt-35-turbo",
-        "gpt-3-turbo",
-        "standard/1024-x-1024/dall-e-2",
-      ],
-    },
-    vertex_ai_beta: {
-      separator: "/",
-      models: ["chat-bison", "chat-bison-32k"],
-    },
-    sagemaker: { separator: "/", models: ["meta-textgeneration-llama-2-13b"] },
-    cohere: { separator: ".", models: ["command-r-v1:0"] },
-    cloudflare: {
-      separator: "/",
-      models: ["@cf/mistral/mistral-7b-instruct-v0.1"],
-    },
-    openai: {
-      separator: "/",
-      models: ["gpt-4o", "gpt-4o-mini"],
-    },
-    anthropic: {
-      separator: "/",
-      models: [
-        "claude-3-5-sonnet-20241022",
-        "claude-3-haiku-20240307",
-        "claude-2",
-        "claude-2.1",
-      ],
-    },
-    other: {
-      separator: "",
-      models: ["together-ai-21.1b-41b"],
-    },
-  });
-});

"b/frontend/src/utils/verified-models.ts\""

@@ -1,29 +0,0 @@
-// Here are the list of verified models and providers that we know work well with OpenHands.
-export const VERIFIED_PROVIDERS = ["openai", "azure", "anthropic"];
-export const VERIFIED_MODELS = ["gpt-4o", "claude-3-5-sonnet-20241022"];
-
-// LiteLLM does not return OpenAI models with the provider, so we list them here to set them ourselves for consistency
-// (e.g., they return `gpt-4o` instead of `openai/gpt-4o`)
-export const VERIFIED_OPENAI_MODELS = [
-  "gpt-4o",
-  "gpt-4o-mini",
-  "gpt-4-turbo",
-  "gpt-4",
-  "gpt-4-32k",
-  "o1-mini",
-  "o1-preview",
-];
-
-// LiteLLM does not return the compatible Anthropic models with the provider, so we list them here to set them ourselves
-// (e.g., they return `claude-3-5-sonnet-20241022` instead of `anthropic/claude-3-5-sonnet-20241022`)
-export const VERIFIED_ANTHROPIC_MODELS = [
-  "claude-2",
-  "claude-2.1",
-  "claude-3-5-sonnet-20241022",
-  "claude-3-5-sonnet-20240620",
-  "claude-3-haiku-20240307",
-  "claude-3-opus-20240229",
-  "claude-3-sonnet-20240229",
-  "claude-instant-1",
-  "claude-instant-1.2",
-];

.github/ISSUE_TEMPLATE/bug_template.yml

@@ -31,6 +31,8 @@ body:
       options:
         - Docker command in README
         - Development workflow
+        - app.all-hands.dev
+        - Other
       default: 0
 
   - type: input

.github/workflows/ghcr-build.yml

@@ -1,5 +1,5 @@
 # Workflow that builds, tests and then pushes the OpenHands and runtime docker images to the ghcr.io repository
-name: Build, Test and Publish RT Image
+name: Docker
 
 # Always run on "main"
 # Always run on tags
@@ -399,3 +399,49 @@ jobs:
         run: |
           echo "Some runtime tests failed or were cancelled"
           exit 1
+  update_pr_description:
+    name: Update PR Description
+    if: github.event_name == 'pull_request' && !github.event.pull_request.head.repo.fork && github.actor != 'dependabot[bot]'
+    needs: [ghcr_build_runtime]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Get short SHA
+        id: short_sha
+        run: echo "SHORT_SHA=$(echo ${{ github.event.pull_request.head.sha }} | cut -c1-7)" >> $GITHUB_OUTPUT
+
+      - name: Update PR Description
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          REPO: ${{ github.repository }}
+          SHORT_SHA: ${{ steps.short_sha.outputs.SHORT_SHA }}
+        run: |
+          echo "updating PR description"
+          DOCKER_RUN_COMMAND="docker run -it --rm \
+            -p 3000:3000 \
+            -v /var/run/docker.sock:/var/run/docker.sock \
+            --add-host host.docker.internal:host-gateway \
+            -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:$SHORT_SHA-nikolaik \
+            --name openhands-app-$SHORT_SHA \
+            docker.all-hands.dev/all-hands-ai/openhands:$SHORT_SHA"
+
+          PR_BODY=$(gh pr view $PR_NUMBER --json body --jq .body)
+
+          if echo "$PR_BODY" | grep -q "To run this PR locally, use the following command:"; then
+            UPDATED_PR_BODY=$(echo "${PR_BODY}" | sed -E "s|docker run -it --rm.*|$DOCKER_RUN_COMMAND|")
+          else
+            UPDATED_PR_BODY="${PR_BODY}
+
+          ---
+
+          To run this PR locally, use the following command:
+          \`\`\`
+          $DOCKER_RUN_COMMAND
+          \`\`\`"
+          fi
+
+          echo "updated body: $UPDATED_PR_BODY"
+          gh pr edit $PR_NUMBER --body "$UPDATED_PR_BODY"

.github/workflows/openhands-resolver.yml

@@ -3,11 +3,13 @@ name: Resolve Issues with OpenHands
 on:
   issues:
     types: [labeled]
+  pull_request:
+    types: [labeled]
 
 jobs:
   call-openhands-resolver:
     uses: All-Hands-AI/openhands-resolver/.github/workflows/openhands-resolver.yml@main
     if: github.event.label.name == 'fix-me'
     with:
-      issue_number: ${{ github.event.issue.number }}
+      max_iterations: 50
     secrets: inherit

.gitignore

@@ -174,6 +174,7 @@ evaluation/bird/data
 evaluation/gaia/data
 evaluation/gorilla/data
 evaluation/toolqa/data
+evaluation/scienceagentbench/benchmark
 
 # frontend
 

Development.md

@@ -100,7 +100,7 @@ poetry run pytest ./tests/unit/test_*.py
 ### 9. Use existing Docker image
 To reduce build time (e.g., if no changes were made to the client-runtime component), you can use an existing Docker container image. Follow these steps:
 1. Set the SANDBOX_RUNTIME_CONTAINER_IMAGE environment variable to the desired Docker image.
-2. Example: export SANDBOX_RUNTIME_CONTAINER_IMAGE=ghcr.io/all-hands-ai/runtime:0.9-nikolaik
+2. Example: export SANDBOX_RUNTIME_CONTAINER_IMAGE=ghcr.io/all-hands-ai/runtime:0.13-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/opendevin/shared_invite/zt-2oikve2hu-UDxHeo8nsE69y6T7yFX_BA"><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-2tom0er4l-JeNUGHt_AxpEfIBstbLPiw"><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/>
@@ -33,37 +33,35 @@ Learn more at [docs.all-hands.dev](https://docs.all-hands.dev), or jump to the [
 
 ## ⚡ Quick Start
 
-The easiest way to run OpenHands is in Docker. You can change `WORKSPACE_BASE` below to
-point OpenHands to existing code that you'd like to modify.
-
+The easiest way to run OpenHands is in Docker.
 See the [Installation](https://docs.all-hands.dev/modules/usage/installation) guide for
 system requirements and more information.
 
 ```bash
-export WORKSPACE_BASE=$(pwd)/workspace
-
-docker pull ghcr.io/all-hands-ai/runtime:0.11-nikolaik
+docker pull docker.all-hands.dev/all-hands-ai/runtime:0.13-nikolaik
 
 docker run -it --pull=always \
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=ghcr.io/all-hands-ai/runtime:0.11-nikolaik \
-    -e SANDBOX_USER_ID=$(id -u) \
-    -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
-    -v $WORKSPACE_BASE:/opt/workspace_base \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.13-nikolaik \
     -v /var/run/docker.sock:/var/run/docker.sock \
     -p 3000:3000 \
+    -e LOG_ALL_EVENTS=true \
     --add-host host.docker.internal:host-gateway \
-    --name openhands-app-$(date +%Y%m%d%H%M%S) \
-    ghcr.io/all-hands-ai/openhands:0.11
+    --name openhands-app \
+    docker.all-hands.dev/all-hands-ai/openhands:0.13
 ```
 
 You'll find OpenHands running at [http://localhost:3000](http://localhost:3000)!
 
-You'll need a model provider and API key. One option that works well: [Claude 3.5 Sonnet](https://www.anthropic.com/api), but you have [many options](https://docs.all-hands.dev/modules/usage/llms).
+Finally, you'll need a model provider and API key.
+[Anthropic's Claude 3.5 Sonnet](https://www.anthropic.com/api) (`anthropic/claude-3-5-sonnet-20241022`)
+works best, but you have [many options](https://docs.all-hands.dev/modules/usage/llms).
 
 ---
 
-You can also run OpenHands in a scriptable [headless mode](https://docs.all-hands.dev/modules/usage/how-to/headless-mode),
-or as an [interactive CLI](https://docs.all-hands.dev/modules/usage/how-to/cli-mode).
+You can also [connect OpenHands to your local filesystem](https://docs.all-hands.dev/modules/usage/runtimes),
+run OpenHands in a scriptable [headless mode](https://docs.all-hands.dev/modules/usage/how-to/headless-mode),
+interact with it via a [friendly CLI](https://docs.all-hands.dev/modules/usage/how-to/cli-mode),
+or run it on tagged issues with [a github action](https://github.com/All-Hands-AI/OpenHands-resolver).
 
 Visit [Installation](https://docs.all-hands.dev/modules/usage/installation) for more information and setup instructions.
 
@@ -96,7 +94,7 @@ For details, please check [CONTRIBUTING.md](./CONTRIBUTING.md).
 Whether you're a developer, a researcher, or simply enthusiastic about OpenHands, we'd love to have you in our community.
 Let's make software engineering better together!
 
-- [Slack workspace](https://join.slack.com/t/opendevin/shared_invite/zt-2oikve2hu-UDxHeo8nsE69y6T7yFX_BA) - Here we talk about research, architecture, and future development.
+- [Slack workspace](https://join.slack.com/t/openhands-ai/shared_invite/zt-2tom0er4l-JeNUGHt_AxpEfIBstbLPiw) - Here we talk about research, architecture, and future development.
 - [Discord server](https://discord.gg/ESHStjSjD4) - This is a community-run server for general discussion, questions, and feedback.
 
 ## 📈 Progress

compose.yml

@@ -7,7 +7,7 @@ services:
     image: openhands:latest
     container_name: openhands-app-${DATE:-}
     environment:
-      - SANDBOX_RUNTIME_CONTAINER_IMAGE=${SANDBOX_RUNTIME_CONTAINER_IMAGE:-ghcr.io/all-hands-ai/runtime:0.9-nikolaik}
+      - SANDBOX_RUNTIME_CONTAINER_IMAGE=${SANDBOX_RUNTIME_CONTAINER_IMAGE:-ghcr.io/all-hands-ai/runtime:0.13-nikolaik}
       - SANDBOX_USER_ID=${SANDBOX_USER_ID:-1234}
       - WORKSPACE_MOUNT_PATH=${WORKSPACE_BASE:-$PWD/workspace}
     ports:

config.template.toml

@@ -32,7 +32,8 @@ workspace_base = "./workspace"
 # Enable saving and restoring the session when run from CLI
 #enable_cli_session = false
 
-# Path to store trajectories
+# Path to store trajectories, can be a folder or a file
+# If it's a folder, the session id will be used as the file name
 #trajectories_path="./trajectories"
 
 # File store path

containers/app/Dockerfile

@@ -41,6 +41,7 @@ ENV SANDBOX_LOCAL_RUNTIME_URL=http://host.docker.internal
 ENV USE_HOST_NETWORK=false
 ENV WORKSPACE_BASE=/opt/workspace_base
 ENV OPENHANDS_BUILD_VERSION=$OPENHANDS_BUILD_VERSION
+ENV SANDBOX_USER_ID=0
 RUN mkdir -p $WORKSPACE_BASE
 
 RUN apt-get update -y \

containers/app/entrypoint.sh

@@ -18,6 +18,11 @@ if [ -z "$SANDBOX_USER_ID" ]; then
   exit 1
 fi
 
+if [ -z "$WORKSPACE_MOUNT_PATH" ]; then
+  # This is set to /opt/workspace in the Dockerfile. But if the user isn't mounting, we want to unset it so that OpenHands doesn't mount at all
+  unset WORKSPACE_BASE
+fi
+
 if [[ "$SANDBOX_USER_ID" -eq 0 ]]; then
   echo "Running OpenHands as root"
   export RUN_AS_OPENHANDS=false

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:0.9-nikolaik}
+      - SANDBOX_RUNTIME_CONTAINER_IMAGE=${SANDBOX_RUNTIME_CONTAINER_IMAGE:-ghcr.io/all-hands-ai/runtime:0.13-nikolaik}
       - SANDBOX_USER_ID=${SANDBOX_USER_ID:-1234}
       - WORKSPACE_MOUNT_PATH=${WORKSPACE_BASE:-$PWD/workspace}
     ports:

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

@@ -14,4 +14,97 @@ Pour démarrer une session OpenHands interactive via la ligne de commande, suive
 
 2. Exécutez la commande suivante :
 
+```bash
+poetry run python -m openhands.core.cli
+```
 
+Cette commande démarrera une session interactive où vous pourrez saisir des tâches et recevoir des réponses d'OpenHands.
+
+Vous devrez vous assurer de définir votre modèle, votre clé API et d'autres paramètres via des variables d'environnement
+[ou le fichier `config.toml`](https://github.com/All-Hands-AI/OpenHands/blob/main/config.template.toml).
+
+
+## Avec Docker
+
+Pour exécuter OpenHands en mode CLI avec Docker, suivez ces étapes :
+
+1. Définissez `WORKSPACE_BASE` sur le répertoire que vous souhaitez qu'OpenHands modifie :
+
+```bash
+WORKSPACE_BASE=$(pwd)/workspace
+```
+
+2. Définissez `LLM_MODEL` sur le modèle que vous souhaitez utiliser :
+
+```bash
+LLM_MODEL="anthropic/claude-3-5-sonnet-20241022"
+```
+
+3. Définissez `LLM_API_KEY` sur votre clé API :
+
+```bash
+LLM_API_KEY="sk_test_12345"
+```
+
+4. Exécutez la commande Docker suivante :
+
+```bash
+docker run -it \
+    --pull=always \
+    -e SANDBOX_USER_ID=$(id -u) \
+    -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
+    -e LLM_API_KEY=$LLM_API_KEY \
+    -e LLM_MODEL=$LLM_MODEL \
+    -v $WORKSPACE_BASE:/opt/workspace_base \
+    -v /var/run/docker.sock:/var/run/docker.sock \
+    --add-host host.docker.internal:host-gateway \
+    --name openhands-app-$(date +%Y%m%d%H%M%S) \
+    ghcr.io/all-hands-ai/openhands:0.11 \
+    python -m openhands.core.cli
+```
+
+Cette commande démarrera une session interactive dans Docker où vous pourrez saisir des tâches et recevoir des réponses d'OpenHands.
+
+## Exemples de commandes CLI et de sorties attendues
+
+Voici quelques exemples de commandes CLI et leurs sorties attendues :
+
+### Exemple 1 : Tâche simple
+
+```bash
+Comment puis-je vous aider ? >> Écrivez un script Python qui affiche "Hello, World!"
+```
+
+Sortie attendue :
+
+```bash
+🤖 Bien sûr ! Voici un script Python qui affiche "Hello, World!" :
+
+❯ print("Hello, World!")
+```
+
+### Exemple 2 : Commande Bash
+
+```bash
+Comment puis-je vous aider ? >> Créez un répertoire nommé "test_dir"
+```
+
+Sortie attendue :
+
+```bash
+🤖 Création d'un répertoire nommé "test_dir" :
+
+❯ mkdir test_dir
+```
+
+### Exemple 3 : Gestion des erreurs
+
+```bash
+Comment puis-je vous aider ? >> Supprimez un fichier inexistant
+```
+
+Sortie attendue :
+
+```bash
+🤖 Une erreur s'est produite. Veuillez réessayer.
+```

docs/i18n/fr/docusaurus-plugin-content-docs/current/usage/how-to/evaluation-harness.md

@@ -1,3 +1,5 @@
+
+
 # Évaluation
 
 Ce guide fournit un aperçu de la façon d'intégrer votre propre benchmark d'évaluation dans le framework OpenHands.
@@ -9,4 +11,271 @@ OpenHands en mode développement utilise `config.toml` pour garder une trace de
 
 Voici un exemple de fichier de configuration que vous pouvez utiliser pour définir et utiliser plusieurs LLMs :
 
+```toml
+[llm]
+# IMPORTANT : ajoutez votre clé API ici et définissez le modèle que vous souhaitez évaluer
+model = "claude-3-5-sonnet-20241022"
 
+api_key = "sk-XXX"
+
+[llm.eval_gpt4_1106_preview_llm]
+model = "gpt-4-1106-preview"
+api_key = "XXX"
+temperature = 0.0
+
+[llm.eval_some_openai_compatible_model_llm]
+model = "openai/MODEL_NAME"
+base_url = "https://OPENAI_COMPATIBLE_URL/v1"
+api_key = "XXX"
+temperature = 0.0
+```
+
+
+## Comment utiliser OpenHands en ligne de commande
+
+OpenHands peut être exécuté depuis la ligne de commande en utilisant le format suivant :
+
+```bash
+poetry run python ./openhands/core/main.py \
+        -i <max_iterations> \
+        -t "<task_description>" \
+        -c <agent_class> \
+        -l <llm_config>
+```
+
+Par exemple :
+
+```bash
+poetry run python ./openhands/core/main.py \
+        -i 10 \
+        -t "Écrivez-moi un script bash qui affiche hello world." \
+        -c CodeActAgent \
+        -l llm
+```
+
+Cette commande exécute OpenHands avec :
+- Un maximum de 10 itérations
+- La description de tâche spécifiée
+- En utilisant CodeActAgent
+- Avec la configuration LLM définie dans la section `llm` de votre fichier `config.toml`
+
+## Comment fonctionne OpenHands
+
+Le point d'entrée principal d'OpenHands se trouve dans `openhands/core/main.py`. Voici un flux simplifié de son fonctionnement :
+
+1. Analyse des arguments de ligne de commande et chargement de la configuration
+2. Création d'un environnement d'exécution à l'aide de `create_runtime()`
+3. Initialisation de l'agent spécifié
+4. Exécution du contrôleur à l'aide de `run_controller()`, qui :
+   - Attache l'environnement d'exécution à l'agent
+   - Exécute la tâche de l'agent
+   - Renvoie un état final une fois terminé
+
+La fonction `run_controller()` est le cœur de l'exécution d'OpenHands. Elle gère l'interaction entre l'agent, l'environnement d'exécution et la tâche, en gérant des choses comme la simulation d'entrée utilisateur et le traitement des événements.
+
+
+## Le moyen le plus simple de commencer : Explorer les benchmarks existants
+
+Nous vous encourageons à examiner les différents benchmarks d'évaluation disponibles dans le [répertoire `evaluation/`](https://github.com/All-Hands-AI/OpenHands/blob/main/evaluation) de notre dépôt.
+
+Pour intégrer votre propre benchmark, nous vous suggérons de commencer par celui qui ressemble le plus à vos besoins. Cette approche peut considérablement rationaliser votre processus d'intégration, vous permettant de vous appuyer sur les structures existantes et de les adapter à vos exigences spécifiques.
+
+## Comment créer un workflow d'évaluation
+
+
+Pour créer un workflow d'évaluation pour votre benchmark, suivez ces étapes :
+
+1. Importez les utilitaires OpenHands pertinents :
+   ```python
+    import openhands.agenthub
+    from evaluation.utils.shared import (
+        EvalMetadata,
+        EvalOutput,
+        make_metadata,
+        prepare_dataset,
+        reset_logger_for_multiprocessing,
+        run_evaluation,
+    )
+    from openhands.controller.state.state import State
+    from openhands.core.config import (
+        AppConfig,
+        SandboxConfig,
+        get_llm_config_arg,
+        parse_arguments,
+    )
+    from openhands.core.logger import openhands_logger as logger
+    from openhands.core.main import create_runtime, run_controller
+    from openhands.events.action import CmdRunAction
+    from openhands.events.observation import CmdOutputObservation, ErrorObservation
+    from openhands.runtime.runtime import Runtime
+   ```
+
+2. Créez une configuration :
+   ```python
+   def get_config(instance: pd.Series, metadata: EvalMetadata) -> AppConfig:
+       config = AppConfig(
+           default_agent=metadata.agent_class,
+           runtime='eventstream',
+           max_iterations=metadata.max_iterations,
+           sandbox=SandboxConfig(
+               base_container_image='your_container_image',
+               enable_auto_lint=True,
+               timeout=300,
+           ),
+       )
+       config.set_llm_config(metadata.llm_config)
+       return config
+   ```
+
+3. Initialisez l'environnement d'exécution et configurez l'environnement d'évaluation :
+   ```python
+   def initialize_runtime(runtime: Runtime, instance: pd.Series):
+       # Configurez votre environnement d'évaluation ici
+       # Par exemple, définir des variables d'environnement, préparer des fichiers, etc.
+       pass
+   ```
+
+4. Créez une fonction pour traiter chaque instance :
+   ```python
+   from openhands.utils.async_utils import call_async_from_sync
+   def process_instance(instance: pd.Series, metadata: EvalMetadata) -> EvalOutput:
+       config = get_config(instance, metadata)
+       runtime = create_runtime(config)
+       call_async_from_sync(runtime.connect)
+       initialize_runtime(runtime, instance)
+
+       instruction = get_instruction(instance, metadata)
+
+       state = run_controller(
+           config=config,
+           task_str=instruction,
+           runtime=runtime,
+           fake_user_response_fn=your_user_response_function,
+       )
+
+       # Évaluez les actions de l'agent
+       evaluation_result = await evaluate_agent_actions(runtime, instance)
+
+       return EvalOutput(
+           instance_id=instance.instance_id,
+           instruction=instruction,
+           test_result=evaluation_result,
+           metadata=metadata,
+           history=compatibility_for_eval_history_pairs(state.history),
+           metrics=state.metrics.get() if state.metrics else None,
+           error=state.last_error if state and state.last_error else None,
+       )
+   ```
+
+5. Exécutez l'évaluation :
+   ```python
+   metadata = make_metadata(llm_config, dataset_name, agent_class, max_iterations, eval_note, eval_output_dir)
+   output_file = os.path.join(metadata.eval_output_dir, 'output.jsonl')
+   instances = prepare_dataset(your_dataset, output_file, eval_n_limit)
+
+   await run_evaluation(
+       instances,
+       metadata,
+       output_file,
+       num_workers,
+       process_instance
+   )
+   ```
+
+Ce workflow configure la configuration, initialise l'environnement d'exécution, traite chaque instance en exécutant l'agent et en évaluant ses actions, puis collecte les résultats dans un objet `EvalOutput`. La fonction `run_evaluation` gère la parallélisation et le suivi de la progression.
+
+N'oubliez pas de personnaliser les fonctions `get_instruction`, `your_user_response_function` et `evaluate_agent_actions` en fonction des exigences spécifiques de votre benchmark.
+
+En suivant cette structure, vous pouvez créer un workflow d'évaluation robuste pour votre benchmark dans le framework OpenHands.
+
+
+## Comprendre la `user_response_fn`
+
+La `user_response_fn` est un composant crucial dans le workflow d'évaluation d'OpenHands. Elle simule l'interaction de l'utilisateur avec l'agent, permettant des réponses automatisées pendant le processus d'évaluation. Cette fonction est particulièrement utile lorsque vous souhaitez fournir des réponses cohérentes et prédéfinies aux requêtes ou actions de l'agent.
+
+
+### Workflow et interaction
+
+Le workflow correct pour gérer les actions et la `user_response_fn` est le suivant :
+
+1. L'agent reçoit une tâche et commence à la traiter
+2. L'agent émet une Action
+3. Si l'Action est exécutable (par exemple, CmdRunAction, IPythonRunCellAction) :
+   - Le Runtime traite l'Action
+   - Le Runtime renvoie une Observation
+4. Si l'Action n'est pas exécutable (généralement une MessageAction) :
+   - La `user_response_fn` est appelée
+   - Elle renvoie une réponse utilisateur simulée
+5. L'agent reçoit soit l'Observation, soit la réponse simulée
+6. Les étapes 2 à 5 se répètent jusqu'à ce que la tâche soit terminée ou que le nombre maximum d'itérations soit atteint
+
+Voici une représentation visuelle plus précise :
+
+```
+                 [Agent]
+                    |
+                    v
+               [Émettre une Action]
+                    |
+                    v
+            [L'Action est-elle exécutable ?]
+           /                       \
+         Oui                        Non
+          |                          |
+          v                          v
+     [Runtime]               [user_response_fn]
+          |                          |
+          v                          v
+  [Renvoyer une Observation]    [Réponse simulée]
+           \                        /
+            \                      /
+             v                    v
+           [L'agent reçoit le feedback]
+                    |
+                    v
+         [Continuer ou terminer la tâche]
+```
+
+Dans ce workflow :
+
+- Les actions exécutables (comme l'exécution de commandes ou de code) sont gérées directement par le Runtime
+- Les actions non exécutables (généralement lorsque l'agent veut communiquer ou demander des clarifications) sont gérées par la `user_response_fn`
+- L'agent traite ensuite le feedback, qu'il s'agisse d'une Observation du Runtime ou d'une réponse simulée de la `user_response_fn`
+
+Cette approche permet une gestion automatisée des actions concrètes et des interactions utilisateur simulées, ce qui la rend adaptée aux scénarios d'évaluation où vous souhaitez tester la capacité de l'agent à effectuer des tâches avec une intervention humaine minimale.
+
+### Exemple d'implémentation
+
+Voici un exemple de `user_response_fn` utilisée dans l'évaluation SWE-Bench :
+
+```python
+def codeact_user_response(state: State | None) -> str:
+    msg = (
+        'Veuillez continuer à travailler sur la tâche avec l\'approche que vous jugez appropriée.\n'
+        'Si vous pensez avoir résolu la tâche, veuillez d\'abord envoyer votre réponse à l\'utilisateur via un message, puis <execute_bash> exit </execute_bash>.\n'
+        'IMPORTANT : VOUS NE DEVEZ JAMAIS DEMANDER DE L\'AIDE HUMAINE.\n'
+    )
+
+    if state and state.history:
+        # vérifier si l'agent a essayé de parler à l'utilisateur 3 fois, si oui, faire savoir à l'agent qu'il peut abandonner
+        user_msgs = [
+            event
+            for event in state.history
+            if isinstance(event, MessageAction) and event.source == 'user'
+        ]
+        if len(user_msgs) >= 2:
+            # faire savoir à l'agent qu'il peut abandonner lorsqu'il a essayé 3 fois
+            return (
+                msg
+                + 'Si vous voulez abandonner, exécutez : <execute_bash> exit </execute_bash>.\n'
+            )
+    return msg
+```
+
+Cette fonction fait ce qui suit :
+
+1. Fournit un message standard encourageant l'agent à continuer à travailler
+2. Vérifie combien de fois l'agent a tenté de communiquer avec l'utilisateur
+3. Si l'agent a fait plusieurs tentatives, il lui donne la possibilité d'abandonner
+
+En utilisant cette fonction, vous pouvez garantir un comportement cohérent sur plusieurs exécutions d'évaluation et empêcher l'agent de rester bloqué en attendant une entrée humaine.

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

@@ -1,3 +1,5 @@
+
+
 # Mode sans interface
 
 Vous pouvez exécuter OpenHands avec une seule commande, sans démarrer l'application web.
@@ -11,4 +13,46 @@ Pour exécuter OpenHands en mode sans interface avec Python,
 [suivez les instructions de configuration de développement](https://github.com/All-Hands-AI/OpenHands/blob/main/Development.md),
 puis exécutez :
 
+```bash
+poetry run python -m openhands.core.main -t "write a bash script that prints hi"
+```
 
+Vous devrez vous assurer de définir votre modèle, votre clé API et d'autres paramètres via des variables d'environnement
+[ou le fichier `config.toml`](https://github.com/All-Hands-AI/OpenHands/blob/main/config.template.toml).
+
+## Avec Docker
+
+1. Définissez `WORKSPACE_BASE` sur le répertoire que vous voulez qu'OpenHands modifie :
+
+```bash
+WORKSPACE_BASE=$(pwd)/workspace
+```
+
+2. Définissez `LLM_MODEL` sur le modèle que vous voulez utiliser :
+
+```bash
+LLM_MODEL="anthropic/claude-3-5-sonnet-20241022"
+```
+
+3. Définissez `LLM_API_KEY` sur votre clé API :
+
+```bash
+LLM_API_KEY="sk_test_12345"
+```
+
+4. Exécutez la commande Docker suivante :
+
+```bash
+docker run -it \
+    --pull=always \
+    -e SANDBOX_USER_ID=$(id -u) \
+    -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
+    -e LLM_API_KEY=$LLM_API_KEY \
+    -e LLM_MODEL=$LLM_MODEL \
+    -v $WORKSPACE_BASE:/opt/workspace_base \
+    -v /var/run/docker.sock:/var/run/docker.sock \
+    --add-host host.docker.internal:host-gateway \
+    --name openhands-app-$(date +%Y%m%d%H%M%S) \
+    ghcr.io/all-hands-ai/openhands:0.11 \
+    python -m openhands.core.main -t "write a bash script that prints hi"
+```

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

@@ -14,4 +14,96 @@ OpenHands 可以在交互式命令行模式下运行,允许用户通过命令行
 
 2. 运行以下命令:
 
+```bash
+poetry run python -m openhands.core.cli
+```
 
+该命令将启动一个交互式会话,你可以在其中输入任务并接收来自 OpenHands 的响应。
+
+你需要确保通过环境变量[或 `config.toml` 文件](https://github.com/All-Hands-AI/OpenHands/blob/main/config.template.toml)设置你的模型、API 密钥和其他设置。
+
+
+## 使用 Docker
+
+要在 Docker 中以命令行模式运行 OpenHands,请按照以下步骤操作:
+
+1. 将 `WORKSPACE_BASE` 设置为你希望 OpenHands 编辑的目录:
+
+```bash
+WORKSPACE_BASE=$(pwd)/workspace
+```
+
+2. 将 `LLM_MODEL` 设置为你要使用的模型:
+
+```bash
+LLM_MODEL="anthropic/claude-3-5-sonnet-20241022"
+```
+
+3. 将 `LLM_API_KEY` 设置为你的 API 密钥:
+
+```bash
+LLM_API_KEY="sk_test_12345"
+```
+
+4. 运行以下 Docker 命令:
+
+```bash
+docker run -it \
+    --pull=always \
+    -e SANDBOX_USER_ID=$(id -u) \
+    -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
+    -e LLM_API_KEY=$LLM_API_KEY \
+    -e LLM_MODEL=$LLM_MODEL \
+    -v $WORKSPACE_BASE:/opt/workspace_base \
+    -v /var/run/docker.sock:/var/run/docker.sock \
+    --add-host host.docker.internal:host-gateway \
+    --name openhands-app-$(date +%Y%m%d%H%M%S) \
+    ghcr.io/all-hands-ai/openhands:0.11 \
+    python -m openhands.core.cli
+```
+
+该命令将在 Docker 中启动一个交互式会话,你可以在其中输入任务并接收来自 OpenHands 的响应。
+
+## 命令行命令和预期输出示例
+
+以下是一些命令行命令及其预期输出的示例:
+
+### 示例 1: 简单任务
+
+```bash
+How can I help? >> Write a Python script that prints "Hello, World!"
+```
+
+预期输出:
+
+```bash
+🤖 Sure! Here is a Python script that prints "Hello, World!":
+
+❯ print("Hello, World!")
+```
+
+### 示例 2: Bash 命令
+
+```bash
+How can I help? >> Create a directory named "test_dir"
+```
+
+预期输出:
+
+```bash
+🤖 Creating a directory named "test_dir":
+
+❯ mkdir test_dir
+```
+
+### 示例 3: 错误处理
+
+```bash
+How can I help? >> Delete a non-existent file
+```
+
+预期输出:
+
+```bash
+🤖 An error occurred. Please try again.
+```

docs/i18n/zh-Hans/docusaurus-plugin-content-docs/current/usage/how-to/evaluation-harness.md

@@ -9,4 +9,270 @@
 
 以下是一个示例配置文件，您可以使用它来定义和使用多个 LLM：
 
+```toml
+[llm]
+# 重要：在此处添加您的 API 密钥，并将模型设置为您要评估的模型
+model = "claude-3-5-sonnet-20241022"
+api_key = "sk-XXX"
 
+[llm.eval_gpt4_1106_preview_llm]
+model = "gpt-4-1106-preview"
+api_key = "XXX"
+temperature = 0.0
+
+[llm.eval_some_openai_compatible_model_llm]
+model = "openai/MODEL_NAME"
+base_url = "https://OPENAI_COMPATIBLE_URL/v1"
+api_key = "XXX"
+temperature = 0.0
+```
+
+
+## 如何在命令行中使用 OpenHands
+
+可以使用以下格式从命令行运行 OpenHands：
+
+```bash
+poetry run python ./openhands/core/main.py \
+        -i <max_iterations> \
+        -t "<task_description>" \
+        -c <agent_class> \
+        -l <llm_config>
+```
+
+例如：
+
+```bash
+poetry run python ./openhands/core/main.py \
+        -i 10 \
+        -t "Write me a bash script that prints hello world." \
+        -c CodeActAgent \
+        -l llm
+```
+
+此命令使用以下参数运行 OpenHands：
+- 最大迭代次数为 10
+- 指定的任务描述
+- 使用 CodeActAgent
+- 使用 `config.toml` 文件的 `llm` 部分中定义的 LLM 配置
+
+## OpenHands 如何工作
+
+OpenHands 的主要入口点在 `openhands/core/main.py` 中。以下是它工作原理的简化流程：
+
+1. 解析命令行参数并加载配置
+2. 使用 `create_runtime()` 创建运行时环境
+3. 初始化指定的代理
+4. 使用 `run_controller()` 运行控制器，它：
+   - 将运行时附加到代理
+   - 执行代理的任务
+   - 完成后返回最终状态
+
+`run_controller()` 函数是 OpenHands 执行的核心。它管理代理、运行时和任务之间的交互，处理用户输入模拟和事件处理等事项。
+
+
+## 入门最简单的方法：探索现有基准
+
+我们鼓励您查看我们仓库的 [`evaluation/` 目录](https://github.com/All-Hands-AI/OpenHands/blob/main/evaluation)中提供的各种评估基准。
+
+要集成您自己的基准，我们建议从最接近您需求的基准开始。这种方法可以显著简化您的集成过程，允许您在现有结构的基础上进行构建并使其适应您的特定要求。
+
+## 如何创建评估工作流
+
+
+要为您的基准创建评估工作流，请按照以下步骤操作：
+
+1. 导入相关的 OpenHands 实用程序：
+   ```python
+    import openhands.agenthub
+    from evaluation.utils.shared import (
+        EvalMetadata,
+        EvalOutput,
+        make_metadata,
+        prepare_dataset,
+        reset_logger_for_multiprocessing,
+        run_evaluation,
+    )
+    from openhands.controller.state.state import State
+    from openhands.core.config import (
+        AppConfig,
+        SandboxConfig,
+        get_llm_config_arg,
+        parse_arguments,
+    )
+    from openhands.core.logger import openhands_logger as logger
+    from openhands.core.main import create_runtime, run_controller
+    from openhands.events.action import CmdRunAction
+    from openhands.events.observation import CmdOutputObservation, ErrorObservation
+    from openhands.runtime.runtime import Runtime
+   ```
+
+2. 创建配置：
+   ```python
+   def get_config(instance: pd.Series, metadata: EvalMetadata) -> AppConfig:
+       config = AppConfig(
+           default_agent=metadata.agent_class,
+           runtime='eventstream',
+           max_iterations=metadata.max_iterations,
+           sandbox=SandboxConfig(
+               base_container_image='your_container_image',
+               enable_auto_lint=True,
+               timeout=300,
+           ),
+       )
+       config.set_llm_config(metadata.llm_config)
+       return config
+   ```
+
+3. 初始化运行时并设置评估环境：
+   ```python
+   def initialize_runtime(runtime: Runtime, instance: pd.Series):
+       # 在此处设置您的评估环境
+       # 例如，设置环境变量、准备文件等
+       pass
+   ```
+
+4. 创建一个函数来处理每个实例：
+   ```python
+   from openhands.utils.async_utils import call_async_from_sync
+   def process_instance(instance: pd.Series, metadata: EvalMetadata) -> EvalOutput:
+       config = get_config(instance, metadata)
+       runtime = create_runtime(config)
+       call_async_from_sync(runtime.connect)
+       initialize_runtime(runtime, instance)
+
+       instruction = get_instruction(instance, metadata)
+
+       state = run_controller(
+           config=config,
+           task_str=instruction,
+           runtime=runtime,
+           fake_user_response_fn=your_user_response_function,
+       )
+
+       # 评估代理的操作
+       evaluation_result = await evaluate_agent_actions(runtime, instance)
+
+       return EvalOutput(
+           instance_id=instance.instance_id,
+           instruction=instruction,
+           test_result=evaluation_result,
+           metadata=metadata,
+           history=compatibility_for_eval_history_pairs(state.history),
+           metrics=state.metrics.get() if state.metrics else None,
+           error=state.last_error if state and state.last_error else None,
+       )
+   ```
+
+5. 运行评估：
+   ```python
+   metadata = make_metadata(llm_config, dataset_name, agent_class, max_iterations, eval_note, eval_output_dir)
+   output_file = os.path.join(metadata.eval_output_dir, 'output.jsonl')
+   instances = prepare_dataset(your_dataset, output_file, eval_n_limit)
+
+   await run_evaluation(
+       instances,
+       metadata,
+       output_file,
+       num_workers,
+       process_instance
+   )
+   ```
+
+此工作流设置配置，初始化运行时环境，通过运行代理并评估其操作来处理每个实例，然后将结果收集到 `EvalOutput` 对象中。`run_evaluation` 函数处理并行化和进度跟踪。
+
+请记住根据您特定的基准要求自定义 `get_instruction`、`your_user_response_function` 和 `evaluate_agent_actions` 函数。
+
+通过遵循此结构，您可以在 OpenHands 框架内为您的基准创建强大的评估工作流。
+
+
+## 理解 `user_response_fn`
+
+`user_response_fn` 是 OpenHands 评估工作流中的关键组件。它模拟用户与代理的交互，允许在评估过程中自动响应。当您想要为代理的查询或操作提供一致的、预定义的响应时，此函数特别有用。
+
+
+### 工作流和交互
+
+处理操作和 `user_response_fn` 的正确工作流如下：
+
+1. 代理接收任务并开始处理
+2. 代理发出操作
+3. 如果操作可执行（例如 CmdRunAction、IPythonRunCellAction）：
+   - 运行时处理操作
+   - 运行时返回观察结果
+4. 如果操作不可执行（通常是 MessageAction）：
+   - 调用 `user_response_fn`
+   - 它返回模拟的用户响应
+5. 代理接收观察结果或模拟响应
+6. 重复步骤 2-5，直到任务完成或达到最大迭代次数
+
+以下是更准确的可视化表示：
+
+```
+                 [代理]
+                    |
+                    v
+               [发出操作]
+                    |
+                    v
+            [操作是否可执行？]
+           /                       \
+         是                         否
+          |                          |
+          v                          v
+     [运行时]                 [user_response_fn]
+          |                          |
+          v                          v
+  [返回观察结果]           [模拟响应]
+           \                        /
+            \                      /
+             v                    v
+           [代理接收反馈]
+                    |
+                    v
+         [继续或完成任务]
+```
+
+在此工作流中：
+
+- 可执行的操作（如运行命令或执行代码）由运行时直接处理
+- 不可执行的操作（通常是当代理想要通信或寻求澄清时）由 `user_response_fn` 处理
+- 然后，代理处理反馈，无论是来自运行时的观察结果还是来自 `user_response_fn` 的模拟响应
+
+这种方法允许自动处理具体操作和模拟用户交互，使其适用于您想要测试代理在最少人工干预的情况下完成任务的能力的评估场景。
+
+### 示例实现
+
+以下是 SWE-Bench 评估中使用的 `user_response_fn` 示例：
+
+```python
+def codeact_user_response(state: State | None) -> str:
+    msg = (
+        'Please continue working on the task on whatever approach you think is suitable.\n'
+        'If you think you have solved the task, please first send your answer to user through message and then <execute_bash> exit </execute_bash>.\n'
+        'IMPORTANT: YOU SHOULD NEVER ASK FOR HUMAN HELP.\n'
+    )
+
+    if state and state.history:
+        # 检查代理是否已尝试与用户对话 3 次，如果是，让代理知道它可以放弃
+        user_msgs = [
+            event
+            for event in state.history
+            if isinstance(event, MessageAction) and event.source == 'user'
+        ]
+        if len(user_msgs) >= 2:
+            # 让代理知道它在尝试 3 次后可以放弃
+            return (
+                msg
+                + 'If you want to give up, run: <execute_bash> exit </execute_bash>.\n'
+            )
+    return msg
+```
+
+此函数执行以下操作：
+
+1. 提供一条标准消息，鼓励代理继续工作
+2. 检查代理尝试与用户通信的次数
+3. 如果代理已多次尝试，它会提供放弃的选项
+
+通过使用此函数，您可以确保在多次评估运行中保持一致的行为，并防止代理在等待人工输入时陷入困境。

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

@@ -13,4 +13,48 @@
 [请按照开发设置说明](https://github.com/All-Hands-AI/OpenHands/blob/main/Development.md),
 然后运行:
 
+```bash
+poetry run python -m openhands.core.main -t "write a bash script that prints hi"
+```
 
+你需要确保通过环境变量
+[或 `config.toml` 文件](https://github.com/All-Hands-AI/OpenHands/blob/main/config.template.toml)
+设置你的模型、API 密钥和其他设置。
+
+## 使用 Docker
+
+1. 将 `WORKSPACE_BASE` 设置为你希望 OpenHands 编辑的目录:
+
+```bash
+WORKSPACE_BASE=$(pwd)/workspace
+```
+
+2. 将 `LLM_MODEL` 设置为你要使用的模型:
+
+```bash
+LLM_MODEL="anthropic/claude-3-5-sonnet-20241022"
+
+```
+
+3. 将 `LLM_API_KEY` 设置为你的 API 密钥:
+
+```bash
+LLM_API_KEY="sk_test_12345"
+```
+
+4. 运行以下 Docker 命令:
+
+```bash
+docker run -it \
+    --pull=always \
+    -e SANDBOX_USER_ID=$(id -u) \
+    -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
+    -e LLM_API_KEY=$LLM_API_KEY \
+    -e LLM_MODEL=$LLM_MODEL \
+    -v $WORKSPACE_BASE:/opt/workspace_base \
+    -v /var/run/docker.sock:/var/run/docker.sock \
+    --add-host host.docker.internal:host-gateway \
+    --name openhands-app-$(date +%Y%m%d%H%M%S) \
+    ghcr.io/all-hands-ai/openhands:0.11 \
+    python -m openhands.core.main -t "write a bash script that prints hi"
+```

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

@@ -50,6 +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:0.13-nikolaik \
     -e SANDBOX_USER_ID=$(id -u) \
     -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
     -e LLM_API_KEY=$LLM_API_KEY \
@@ -58,7 +59,7 @@ docker run -it \
     -v /var/run/docker.sock:/var/run/docker.sock \
     --add-host host.docker.internal:host-gateway \
     --name openhands-app-$(date +%Y%m%d%H%M%S) \
-    ghcr.io/all-hands-ai/openhands:0.11 \
+    docker.all-hands.dev/all-hands-ai/openhands:0.13 \
     python -m openhands.core.cli
 ```
 
@@ -107,4 +108,3 @@ Expected Output:
 ```bash
 🤖 An error occurred. Please try again.
 ```
-

docs/modules/usage/how-to/evaluation-harness.md

@@ -9,4 +9,270 @@ OpenHands in development mode uses `config.toml` to keep track of most configura
 
 Here's an example configuration file you can use to define and use multiple LLMs:
 
+```toml
+[llm]
+# IMPORTANT: add your API key here, and set the model to the one you want to evaluate
+model = "claude-3-5-sonnet-20241022"
+api_key = "sk-XXX"
 
+[llm.eval_gpt4_1106_preview_llm]
+model = "gpt-4-1106-preview"
+api_key = "XXX"
+temperature = 0.0
+
+[llm.eval_some_openai_compatible_model_llm]
+model = "openai/MODEL_NAME"
+base_url = "https://OPENAI_COMPATIBLE_URL/v1"
+api_key = "XXX"
+temperature = 0.0
+```
+
+
+## How to use OpenHands in the command line
+
+OpenHands can be run from the command line using the following format:
+
+```bash
+poetry run python ./openhands/core/main.py \
+        -i <max_iterations> \
+        -t "<task_description>" \
+        -c <agent_class> \
+        -l <llm_config>
+```
+
+For example:
+
+```bash
+poetry run python ./openhands/core/main.py \
+        -i 10 \
+        -t "Write me a bash script that prints hello world." \
+        -c CodeActAgent \
+        -l llm
+```
+
+This command runs OpenHands with:
+- A maximum of 10 iterations
+- The specified task description
+- Using the CodeActAgent
+- With the LLM configuration defined in the `llm` section of your `config.toml` file
+
+## How does OpenHands work
+
+The main entry point for OpenHands is in `openhands/core/main.py`. Here's a simplified flow of how it works:
+
+1. Parse command-line arguments and load the configuration
+2. Create a runtime environment using `create_runtime()`
+3. Initialize the specified agent
+4. Run the controller using `run_controller()`, which:
+   - Attaches the runtime to the agent
+   - Executes the agent's task
+   - Returns a final state when complete
+
+The `run_controller()` function is the core of OpenHands's execution. It manages the interaction between the agent, the runtime, and the task, handling things like user input simulation and event processing.
+
+
+## Easiest way to get started: Exploring Existing Benchmarks
+
+We encourage you to review the various evaluation benchmarks available in the [`evaluation/` directory](https://github.com/All-Hands-AI/OpenHands/blob/main/evaluation) of our repository.
+
+To integrate your own benchmark, we suggest starting with the one that most closely resembles your needs. This approach can significantly streamline your integration process, allowing you to build upon existing structures and adapt them to your specific requirements.
+
+## How to create an evaluation workflow
+
+
+To create an evaluation workflow for your benchmark, follow these steps:
+
+1. Import relevant OpenHands utilities:
+   ```python
+    import openhands.agenthub
+    from evaluation.utils.shared import (
+        EvalMetadata,
+        EvalOutput,
+        make_metadata,
+        prepare_dataset,
+        reset_logger_for_multiprocessing,
+        run_evaluation,
+    )
+    from openhands.controller.state.state import State
+    from openhands.core.config import (
+        AppConfig,
+        SandboxConfig,
+        get_llm_config_arg,
+        parse_arguments,
+    )
+    from openhands.core.logger import openhands_logger as logger
+    from openhands.core.main import create_runtime, run_controller
+    from openhands.events.action import CmdRunAction
+    from openhands.events.observation import CmdOutputObservation, ErrorObservation
+    from openhands.runtime.runtime import Runtime
+   ```
+
+2. Create a configuration:
+   ```python
+   def get_config(instance: pd.Series, metadata: EvalMetadata) -> AppConfig:
+       config = AppConfig(
+           default_agent=metadata.agent_class,
+           runtime='eventstream',
+           max_iterations=metadata.max_iterations,
+           sandbox=SandboxConfig(
+               base_container_image='your_container_image',
+               enable_auto_lint=True,
+               timeout=300,
+           ),
+       )
+       config.set_llm_config(metadata.llm_config)
+       return config
+   ```
+
+3. Initialize the runtime and set up the evaluation environment:
+   ```python
+   def initialize_runtime(runtime: Runtime, instance: pd.Series):
+       # Set up your evaluation environment here
+       # For example, setting environment variables, preparing files, etc.
+       pass
+   ```
+
+4. Create a function to process each instance:
+   ```python
+   from openhands.utils.async_utils import call_async_from_sync
+   def process_instance(instance: pd.Series, metadata: EvalMetadata) -> EvalOutput:
+       config = get_config(instance, metadata)
+       runtime = create_runtime(config)
+       call_async_from_sync(runtime.connect)
+       initialize_runtime(runtime, instance)
+
+       instruction = get_instruction(instance, metadata)
+
+       state = run_controller(
+           config=config,
+           task_str=instruction,
+           runtime=runtime,
+           fake_user_response_fn=your_user_response_function,
+       )
+
+       # Evaluate the agent's actions
+       evaluation_result = await evaluate_agent_actions(runtime, instance)
+
+       return EvalOutput(
+           instance_id=instance.instance_id,
+           instruction=instruction,
+           test_result=evaluation_result,
+           metadata=metadata,
+           history=compatibility_for_eval_history_pairs(state.history),
+           metrics=state.metrics.get() if state.metrics else None,
+           error=state.last_error if state and state.last_error else None,
+       )
+   ```
+
+5. Run the evaluation:
+   ```python
+   metadata = make_metadata(llm_config, dataset_name, agent_class, max_iterations, eval_note, eval_output_dir)
+   output_file = os.path.join(metadata.eval_output_dir, 'output.jsonl')
+   instances = prepare_dataset(your_dataset, output_file, eval_n_limit)
+
+   await run_evaluation(
+       instances,
+       metadata,
+       output_file,
+       num_workers,
+       process_instance
+   )
+   ```
+
+This workflow sets up the configuration, initializes the runtime environment, processes each instance by running the agent and evaluating its actions, and then collects the results into an `EvalOutput` object. The `run_evaluation` function handles parallelization and progress tracking.
+
+Remember to customize the `get_instruction`, `your_user_response_function`, and `evaluate_agent_actions` functions according to your specific benchmark requirements.
+
+By following this structure, you can create a robust evaluation workflow for your benchmark within the OpenHands framework.
+
+
+## Understanding the `user_response_fn`
+
+The `user_response_fn` is a crucial component in OpenHands's evaluation workflow. It simulates user interaction with the agent, allowing for automated responses during the evaluation process. This function is particularly useful when you want to provide consistent, predefined responses to the agent's queries or actions.
+
+
+### Workflow and Interaction
+
+The correct workflow for handling actions and the `user_response_fn` is as follows:
+
+1. Agent receives a task and starts processing
+2. Agent emits an Action
+3. If the Action is executable (e.g., CmdRunAction, IPythonRunCellAction):
+   - The Runtime processes the Action
+   - Runtime returns an Observation
+4. If the Action is not executable (typically a MessageAction):
+   - The `user_response_fn` is called
+   - It returns a simulated user response
+5. The agent receives either the Observation or the simulated response
+6. Steps 2-5 repeat until the task is completed or max iterations are reached
+
+Here's a more accurate visual representation:
+
+```
+                 [Agent]
+                    |
+                    v
+               [Emit Action]
+                    |
+                    v
+            [Is Action Executable?]
+           /                       \
+         Yes                        No
+          |                          |
+          v                          v
+     [Runtime]               [user_response_fn]
+          |                          |
+          v                          v
+  [Return Observation]    [Simulated Response]
+           \                        /
+            \                      /
+             v                    v
+           [Agent receives feedback]
+                    |
+                    v
+         [Continue or Complete Task]
+```
+
+In this workflow:
+
+- Executable actions (like running commands or executing code) are handled directly by the Runtime
+- Non-executable actions (typically when the agent wants to communicate or ask for clarification) are handled by the `user_response_fn`
+- The agent then processes the feedback, whether it's an Observation from the Runtime or a simulated response from the `user_response_fn`
+
+This approach allows for automated handling of both concrete actions and simulated user interactions, making it suitable for evaluation scenarios where you want to test the agent's ability to complete tasks with minimal human intervention.
+
+### Example Implementation
+
+Here's an example of a `user_response_fn` used in the SWE-Bench evaluation:
+
+```python
+def codeact_user_response(state: State | None) -> str:
+    msg = (
+        'Please continue working on the task on whatever approach you think is suitable.\n'
+        'If you think you have solved the task, please first send your answer to user through message and then <execute_bash> exit </execute_bash>.\n'
+        'IMPORTANT: YOU SHOULD NEVER ASK FOR HUMAN HELP.\n'
+    )
+
+    if state and state.history:
+        # check if the agent has tried to talk to the user 3 times, if so, let the agent know it can give up
+        user_msgs = [
+            event
+            for event in state.history
+            if isinstance(event, MessageAction) and event.source == 'user'
+        ]
+        if len(user_msgs) >= 2:
+            # let the agent know that it can give up when it has tried 3 times
+            return (
+                msg
+                + 'If you want to give up, run: <execute_bash> exit </execute_bash>.\n'
+            )
+    return msg
+```
+
+This function does the following:
+
+1. Provides a standard message encouraging the agent to continue working
+2. Checks how many times the agent has attempted to communicate with the user
+3. If the agent has made multiple attempts, it provides an option to give up
+
+By using this function, you can ensure consistent behavior across multiple evaluation runs and prevent the agent from getting stuck waiting for human input.

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

@@ -19,6 +19,15 @@ OpenHands provides a user-friendly Graphical User Interface (GUI) mode for inter
 3. Enter the corresponding `API Key` for your chosen provider.
 4. Click "Save" to apply the settings.
 
+### GitHub Token Setup
+
+OpenHands automatically exports a `GITHUB_TOKEN` to the shell environment if it is available. This can happen in two ways:
+
+1. Locally (OSS): The user directly inputs their GitHub token.
+2. Online (SaaS): The token is obtained through GitHub OAuth authentication.
+
+When you reach the `/app` route, the app checks if a token is present. If it finds one, it sets it in the environment for the agent to use.
+
 ### Advanced Settings
 
 1. Toggle `Advanced Options` to access additional settings.

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

@@ -11,4 +11,49 @@ To run OpenHands in headless mode with Python,
 [follow the Development setup instructions](https://github.com/All-Hands-AI/OpenHands/blob/main/Development.md),
 and then run:
 
+```bash
+poetry run python -m openhands.core.main -t "write a bash script that prints hi"
+```
 
+You'll need to be sure to set your model, API key, and other settings via environment variables
+[or the `config.toml` file](https://github.com/All-Hands-AI/OpenHands/blob/main/config.template.toml).
+
+## With Docker
+
+1. Set `WORKSPACE_BASE` to the directory you want OpenHands to edit:
+
+```bash
+WORKSPACE_BASE=$(pwd)/workspace
+```
+
+2. Set `LLM_MODEL` to the model you want to use:
+
+```bash
+LLM_MODEL="anthropic/claude-3-5-sonnet-20241022"
+
+```
+
+3. Set `LLM_API_KEY` to your API key:
+
+```bash
+LLM_API_KEY="sk_test_12345"
+```
+
+4. Run the following Docker command:
+
+```bash
+docker run -it \
+    --pull=always \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.13-nikolaik \
+    -e SANDBOX_USER_ID=$(id -u) \
+    -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
+    -e LLM_API_KEY=$LLM_API_KEY \
+    -e LLM_MODEL=$LLM_MODEL \
+    -e LOG_ALL_EVENTS=true \
+    -v $WORKSPACE_BASE:/opt/workspace_base \
+    -v /var/run/docker.sock:/var/run/docker.sock \
+    --add-host host.docker.internal:host-gateway \
+    --name openhands-app-$(date +%Y%m%d%H%M%S) \
+    docker.all-hands.dev/all-hands-ai/openhands:0.13 \
+    python -m openhands.core.main -t "write a bash script that prints hi"
+```

docs/modules/usage/how-to/openshift-example.md

@@ -150,7 +150,7 @@ metadata:
 spec:
   containers:
   - name: openhands-app-2024
-    image: ghcr.io/all-hands-ai/openhands:main
+    image: docker.all-hands.dev/all-hands-ai/openhands:main
     env:
     - name: SANDBOX_USER_ID
       value: "1000"
@@ -164,7 +164,7 @@ spec:
     ports:
     - containerPort: 3000
   - name: openhands-sandbox-2024
-    image: ghcr.io/all-hands-ai/sandbox:main
+    image: docker.all-hands.dev/all-hands-ai/runtime:main
     ports:
     - containerPort: 51963
     command: ["/usr/sbin/sshd", "-D", "-p 51963", "-o", "PermitRootLogin=yes"]
@@ -205,10 +205,10 @@ LAST SEEN   TYPE     REASON                   OBJECT
 9s          Normal   SuccessfulAttachVolume   pod/openhands-app-2024                AttachVolume.Attach succeeded for volume "pvc-2b1d223a-1c8f-4990-8e3d-68061a9ae252"
 9s          Normal   SuccessfulAttachVolume   pod/openhands-app-2024                AttachVolume.Attach succeeded for volume "pvc-31f15b25-faad-4665-a25f-201a530379af"
 6s          Normal   AddedInterface           pod/openhands-app-2024                Add eth0 [10.128.2.48/23] from openshift-sdn
-6s          Normal   Pulled                   pod/openhands-app-2024                Container image "ghcr.io/all-hands-ai/openhands:main" already present on machine
+6s          Normal   Pulled                   pod/openhands-app-2024                Container image "docker.all-hands.dev/all-hands-ai/openhands:main" already present on machine
 6s          Normal   Created                  pod/openhands-app-2024                Created container openhands-app-2024
 6s          Normal   Started                  pod/openhands-app-2024                Started container openhands-app-2024
-6s          Normal   Pulled                   pod/openhands-app-2024                Container image "ghcr.io/all-hands-ai/sandbox:main" already present on machine
+6s          Normal   Pulled                   pod/openhands-app-2024                Container image "docker.all-hands.dev/all-hands-ai/sandbox:main" already present on machine
 5s          Normal   Created                  pod/openhands-app-2024                Created container openhands-sandbox-2024
 5s          Normal   Started                  pod/openhands-app-2024                Started container openhands-sandbox-2024
 83s         Normal   WaitForFirstConsumer     persistentvolumeclaim/workspace-pvc   waiting for first consumer to be created before binding
@@ -334,7 +334,7 @@ spec:
     spec:
       containers:
       - name: openhands-app-2024
-        image: ghcr.io/all-hands-ai/openhands:main
+        image: docker.all-hands.dev/all-hands-ai/openhands:main
         env:
         - name: SANDBOX_USER_ID
           value: "1000"
@@ -356,7 +356,7 @@ spec:
         ports:
         - containerPort: 3000
       - name: openhands-sandbox-2024
-        image: ghcr.io/opendevin/sandbox:main
+        image: docker.all-hands.dev/all-hands-ai/runtime:main
     #    securityContext:
     #      privileged: true  # Add this to allow privileged access
         ports:

docs/modules/usage/installation.mdx

@@ -8,24 +8,19 @@
 
 ## Start the app
 
-The easiest way to run OpenHands is in Docker. You can change `WORKSPACE_BASE` below to point OpenHands to
-existing code that you'd like to modify.
+The easiest way to run OpenHands is in Docker.
 
 ```bash
-export WORKSPACE_BASE=$(pwd)/workspace
+docker pull docker.all-hands.dev/all-hands-ai/runtime:0.13-nikolaik
 
-docker pull ghcr.io/all-hands-ai/runtime:0.11-nikolaik
-
-docker run -it --pull=always \
-    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=ghcr.io/all-hands-ai/runtime:0.11-nikolaik \
-    -e SANDBOX_USER_ID=$(id -u) \
-    -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
-    -v $WORKSPACE_BASE:/opt/workspace_base \
+docker run -it --rm --pull=always \
+    -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.13-nikolaik \
     -v /var/run/docker.sock:/var/run/docker.sock \
     -p 3000:3000 \
+    -e LOG_ALL_EVENTS=true \
     --add-host host.docker.internal:host-gateway \
-    --name openhands-app-$(date +%Y%m%d%H%M%S) \
-    ghcr.io/all-hands-ai/openhands:0.11
+    --name openhands-app \
+    docker.all-hands.dev/all-hands-ai/openhands:0.13
 ```
 
 You can also run OpenHands in a scriptable [headless mode](https://docs.all-hands.dev/modules/usage/how-to/headless-mode), as an [interactive CLI](https://docs.all-hands.dev/modules/usage/how-to/cli-mode), or using the [OpenHands GitHub Action](https://docs.all-hands.dev/modules/usage/how-to/github-action).
@@ -34,9 +29,6 @@ You can also run OpenHands in a scriptable [headless mode](https://docs.all-hand
 
 After running the command above, you'll find OpenHands running at [http://localhost:3000](http://localhost:3000).
 
-The agent will have access to the `./workspace` folder to do its work. You can copy existing code here, or change `WORKSPACE_BASE` in the
-command to point to an existing folder.
-
 Upon launching OpenHands, you'll see a settings modal. You **must** select an `LLM Provider` and `LLM Model` and enter a corresponding `API Key`.
 These can be changed at any time by selecting the `Settings` button (gear icon) in the UI.
 
@@ -52,9 +44,9 @@ The `Advanced Options` also allow you to specify a `Base URL` if required.
 ## Versions
 
 The command above pulls the most recent stable release of OpenHands. You have other options as well:
-- For a specific release, use `ghcr.io/all-hands-ai/openhands:$VERSION`, replacing $VERSION with the version number.
+- For a specific release, use `docker.all-hands.dev/all-hands-ai/openhands:$VERSION`, replacing $VERSION with the version number.
 - We use semver, and release major, minor, and patch tags. So `0.9` will automatically point to the latest `0.9.x` release, and `0` will point to the latest `0.x.x` release.
-- For the most up-to-date development version, you can use `ghcr.io/all-hands-ai/openhands:main`. This version is unstable and is recommended for testing or development purposes only.
+- For the most up-to-date development version, you can use `docker.all-hands.dev/all-hands-ai/openhands:main`. This version is unstable and is recommended for testing or development purposes only.
 
 You can choose the tag that best suits your needs based on stability requirements and desired features.
 

docs/modules/usage/llms/llms.md

@@ -4,11 +4,11 @@ OpenHands can connect to any LLM supported by LiteLLM. However, it requires a po
 
 ## Model Recommendations
 
-Based on a recent evaluation of language models for coding tasks (using the SWE-bench dataset), we can provide some recommendations for model selection. The full analysis can be found in [this blog article](https://www.all-hands.dev/blog/evaluation-of-llms-as-coding-agents-on-swe-bench-at-30x-speed).
+Based on our evaluations of language models for coding tasks (using the SWE-bench dataset), we can provide some recommendations for model selection. Some analyses can be found in [this blog article comparing LLMs](https://www.all-hands.dev/blog/evaluation-of-llms-as-coding-agents-on-swe-bench-at-30x-speed) and [this blog article with some more recent results](https://www.all-hands.dev/blog/openhands-codeact-21-an-open-state-of-the-art-software-development-agent).
 
 When choosing a model, consider both the quality of outputs and the associated costs. Here's a summary of the findings:
 
-- Claude 3.5 Sonnet is the best by a fair amount, achieving a 27% resolve rate with the default agent in OpenHands.
+- Claude 3.5 Sonnet is the best by a fair amount, achieving a 53% resolve rate on SWE-Bench Verified with the default agent in OpenHands.
 - GPT-4o lags behind, and o1-mini actually performed somewhat worse than GPT-4o. We went in and analyzed the results a little, and briefly it seemed like o1 was sometimes "overthinking" things, performing extra environment configuration tasks when it could just go ahead and finish the task.
 - Finally, the strongest open models were Llama 3.1 405 B and deepseek-v2.5, and they performed reasonably, even besting some of the closed models.
 

docs/modules/usage/llms/local-llms.md

@@ -35,32 +35,15 @@ Use the instructions [here](../getting-started) to start OpenHands using Docker.
 But when running `docker run`, you'll need to add a few more arguments:
 
 ```bash
---add-host host.docker.internal:host-gateway \
--e LLM_OLLAMA_BASE_URL="http://host.docker.internal:11434" \
-```
-
-LLM_OLLAMA_BASE_URL is optional. If you set it, it will be used to show the available installed models in the UI.
-
-Example:
-
-```bash
-# The directory you want OpenHands to modify. MUST be an absolute path!
-export WORKSPACE_BASE=$(pwd)/workspace
-
-docker run \
-    -it \
-    --pull=always \
+docker run # ...
     --add-host host.docker.internal:host-gateway \
-    -e SANDBOX_USER_ID=$(id -u) \
     -e LLM_OLLAMA_BASE_URL="http://host.docker.internal:11434" \
-    -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
-    -v $WORKSPACE_BASE:/opt/workspace_base \
-    -v /var/run/docker.sock:/var/run/docker.sock \
-    -p 3000:3000 \
-    ghcr.io/all-hands-ai/openhands:main
+    # ...
 ```
 
-You should now be able to connect to `http://localhost:3000/`
+LLM_OLLAMA_BASE_URL is optional. If you set it, it will be used to show
+the available installed models in the UI.
+
 
 ### Configure the Web Application
 
@@ -176,18 +159,11 @@ CUSTOM_LLM_PROVIDER="openai"
 ### Docker
 
 ```bash
-docker run \
-    -it \
-    --pull=always \
-    -e SANDBOX_USER_ID=$(id -u) \
+docker run # ...
     -e LLM_MODEL="openai/lmstudio" \
     -e LLM_BASE_URL="http://host.docker.internal:1234/v1" \
     -e CUSTOM_LLM_PROVIDER="openai" \
-    -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
-    -v $WORKSPACE_BASE:/opt/workspace_base \
-    -v /var/run/docker.sock:/var/run/docker.sock \
-    -p 3000:3000 \
-    ghcr.io/all-hands-ai/openhands:main
+    # ...
 ```
 
 You should now be able to connect to `http://localhost:3000/`

docs/modules/usage/runtimes.md

@@ -0,0 +1,77 @@
+# Runtime Configuration
+
+A Runtime is an environment where the OpenHands agent can edit files and run
+commands.
+
+By default, OpenHands uses a Docker-based runtime, running on your local computer.
+This means you only have to pay for the LLM you're using, and your code is only ever sent to the LLM.
+
+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).
+
+## 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:0.11-nikolaik \
+    -v /var/run/docker.sock:/var/run/docker.sock \
+    # ...
+```
+
+The `SANDBOX_RUNTIME_CONTAINER_IMAGE` from nikolaik is a pre-built runtime image
+that contains our Runtime server, as well as some basic utilities for Python and NodeJS.
+You can also [build your own runtime image](how-to/custom-sandbox-guide).
+
+### Connecting to Your filesystem
+One useful feature here is the ability to connect to your local filesystem.
+
+To mount your filesystem into the runtime, add the following options to
+the `docker run` command:
+
+```bash
+export WORKSPACE_BASE=/path/to/your/code
+
+docker run # ...
+    -e SANDBOX_USER_ID=$(id -u) \
+    -e WORKSPACE_MOUNT_PATH=$WORKSPACE_BASE \
+    -v $WORKSPACE_BASE:/opt/workspace_base \
+    # ...
+```
+
+Be careful! There's nothing stopping the OpenHands agent from deleting or modifying
+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 for an invite).
+
+To use the All Hands Runtime, set the following environment variables when
+starting OpenHands:
+
+```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_REMOTE_RUNTIME_ALIVE="true" \
+    # ...
+```
+
+## Modal Runtime
+Our partners at [Modal](https://modal.com/) have also provided a runtime for OpenHands.
+
+To use the Modal Runtime, create an account, and then [create an API key](https://modal.com/settings)
+
+You'll then need to set the following environment variables when starting OpenHands:
+```bash
+docker run # ...
+    -e RUNTIME=modal \
+    -e MODAL_API_TOKEN_ID="your-id" \
+    -e MODAL_API_TOKEN_SECRET="your-secret" \
+```

docs/modules/usage/troubleshooting/troubleshooting.md

@@ -16,6 +16,7 @@ Check out [Notes for WSL on Windows Users](troubleshooting/windows) for some tro
 * [404 Resource not found](#404-resource-not-found)
 * [`make build` getting stuck on package installations](#make-build-getting-stuck-on-package-installations)
 * [Sessions are not restored](#sessions-are-not-restored)
+* [Connection to host.docker.internal timed out](#connection-to-host-docker-internal-timed-out)
 
 ### Unable to connect to Docker
 
@@ -153,3 +154,27 @@ should stay accepted.
 ```bash
 EXPORT JWT_SECRET=A_CONST_VALUE
 ```
+
+---
+### Connection to host docker internal timed out
+
+**Symptoms**
+
+When you start the server using the docker command from the main [README](https://github.com/All-Hands-AI/OpenHands/README.md), you get a long timeout
+followed by the a stack trace containing messages like:
+
+* `Connection to host.docker.internal timed out. (connect timeout=310)`
+* `Max retries exceeded with url: /alive`
+
+**Details**
+
+If Docker Engine is installed rather than Docker Desktop, the main command will not work as expected.
+Docker Desktop includes easy DNS configuration for connecting processes running in different containers
+which OpenHands makes use of when the main server is running inside a docker container.
+(Further details: https://forums.docker.com/t/difference-between-docker-desktop-and-docker-engine/124612)
+
+**Workarounds**
+
+* [Install Docker Desktop](https://www.docker.com/products/docker-desktop/)
+* Run OpenHands in [Development Mode](https://github.com/All-Hands-AI/OpenHands/blob/main/Development.md),
+  So that the main server is not run inside a container, but still creates dockerized runtime sandboxes.

docs/package.json

@@ -15,10 +15,10 @@
     "typecheck": "tsc"
   },
   "dependencies": {
-    "@docusaurus/core": "^3.5.2",
-    "@docusaurus/plugin-content-pages": "^3.5.2",
-    "@docusaurus/preset-classic": "^3.5.2",
-    "@docusaurus/theme-mermaid": "^3.5.2",
+    "@docusaurus/core": "^3.6.0",
+    "@docusaurus/plugin-content-pages": "^3.6.0",
+    "@docusaurus/preset-classic": "^3.6.0",
+    "@docusaurus/theme-mermaid": "^3.6.0",
     "@mdx-js/react": "^3.1.0",
     "clsx": "^2.0.0",
     "prism-react-renderer": "^2.4.0",
@@ -29,7 +29,7 @@
   },
   "devDependencies": {
     "@docusaurus/module-type-aliases": "^3.5.1",
-    "@docusaurus/tsconfig": "^3.5.2",
+    "@docusaurus/tsconfig": "^3.6.0",
     "@docusaurus/types": "^3.5.1",
     "typescript": "~5.6.3"
   },

docs/sidebars.ts

@@ -90,6 +90,11 @@ const sidebars: SidebarsConfig = {
             },
           ],
         },
+        {
+          type: 'doc',
+          label: 'Runtime Configuration',
+          id: 'usage/runtimes',
+        },
         {
           type: 'doc',
           label: 'Custom Sandbox',

evaluation/EDA/run_infer.py

@@ -8,6 +8,7 @@ from evaluation.EDA.game import Q20Game, Q20GameCelebrity
 from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -34,7 +35,8 @@ def codeact_user_response_eda(state: State) -> str:
 
     # retrieve the latest model message from history
     if state.history:
-        model_guess = state.history.get_last_agent_message()
+        last_agent_message = state.get_last_agent_message()
+        model_guess = last_agent_message.content if last_agent_message else ''
 
     assert game is not None, 'Game is not initialized.'
     msg = game.generate_user_response(model_guess)
@@ -139,7 +141,8 @@ def process_instance(
     if state is None:
         raise ValueError('State should not be None.')
 
-    final_message = state.history.get_last_agent_message()
+    last_agent_message = state.get_last_agent_message()
+    final_message = last_agent_message.content if last_agent_message else ''
 
     logger.info(f'Final message: {final_message} | Ground truth: {instance["text"]}')
     test_result = game.reward()
@@ -148,7 +151,7 @@ def process_instance(
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     # Save the output
     output = EvalOutput(

evaluation/README.md

@@ -84,4 +84,3 @@ all the preprocessing/evaluation/analysis scripts.
 - Raw data and experimental records should not be stored within this repo.
 - For model outputs, they should be stored at [this huggingface space](https://huggingface.co/spaces/OpenHands/evaluation) for visualization.
 - Important data files of manageable size and analysis scripts (e.g., jupyter notebooks) can be directly uploaded to this repo.
-

evaluation/agent_bench/run_infer.py

@@ -16,6 +16,7 @@ from evaluation.agent_bench.helper import (
 from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -242,7 +243,7 @@ def process_instance(
         raw_ans = ''
 
         # retrieve the last agent message or thought
-        for event in state.history.get_events(reverse=True):
+        for event in reversed(state.history):
             if event.source == 'agent':
                 if isinstance(event, AgentFinishAction):
                     raw_ans = event.thought
@@ -271,7 +272,7 @@ def process_instance(
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     metrics = state.metrics.get() if state.metrics else None
 

evaluation/aider_bench/run_infer.py

@@ -15,6 +15,7 @@ from evaluation.aider_bench.helper import (
 from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -250,7 +251,7 @@ def process_instance(
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
     metrics = state.metrics.get() if state.metrics else None
 
     # Save the output

evaluation/biocoder/run_infer.py

@@ -13,6 +13,7 @@ from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
     codeact_user_response,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -299,7 +300,7 @@ def process_instance(
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     test_result['generated'] = test_result['metadata']['1_copy_change_code']
 

evaluation/bird/run_infer.py

@@ -16,6 +16,7 @@ from tqdm import tqdm
 from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -46,7 +47,7 @@ def codeact_user_response(state: State) -> str:
         # check if the agent has tried to talk to the user 3 times, if so, let the agent know it can give up
         user_msgs = [
             event
-            for event in state.history.get_events()
+            for event in state.history
             if isinstance(event, MessageAction) and event.source == 'user'
         ]
         if len(user_msgs) > 2:
@@ -431,7 +432,7 @@ def process_instance(
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     # Save the output
     output = EvalOutput(

evaluation/browsing_delegation/run_infer.py

@@ -9,6 +9,7 @@ from datasets import load_dataset
 from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -89,7 +90,7 @@ def process_instance(
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     # find the last delegate action
     last_delegate_action = None

evaluation/discoverybench/README.md

@@ -0,0 +1,37 @@
+# DiscoveryBench with OpenHands
+
+[DiscoveryBench](https://github.com/allenai/discoverybench/) [(Paper)](https://arxiv.org/abs/2407.01725v1) contains 264 tasks collected across 6 diverse domains, such as biology, economics, and sociology. It incorporates discovery workflows from published papers to approximate the real-world challenges faced by researchers.
+
+<p align="center">
+  <a href="[https://github.com/allenai/discoverybench](https://github.com/allenai/discoverybench)">
+    <img src="https://raw.githubusercontent.com/allenai/discoverybench/refs/heads/main/assets/discoverybench-openhands-teaser.png" width="100%" alt="DiscoveryBench Background" />
+  </a>
+</p>
+
+
+## Setup Environment and LLM Configuration
+
+1. Please follow instructions mentioned [here](https://github.com/openlocus/OpenHands/blob/discoverybench-openhands-integration/evaluation/README.md#setup) to setup OpenHands development environment and LLMs locally
+
+2. Execute the bash script to start DiscoveryBench Evaluation
+
+```
+./evaluation/discoverybench/scripts/run_infer.sh [YOUR MODEL CONFIG]
+```
+Replace `[YOUR MODEL CONFIG]` with any model the model that you have set up in `config.toml`
+
+
+## Run Inference on DiscoveryBench Instances
+
+When the `run_infer.sh` script is started, it will automatically pull the latest DiscoveryBench instances & set up the agent environment. The OpenHands agent is invoked to process the task within this environment, producing a hypothesis. We then evaluate it against the “gold” hypothesis provided by DiscoveryBench. The evaluation result, along with the agent chat history is logged to `output.jsonl` under `evaluation_outputs`.
+
+
+```
+./evaluation/discoverybench/scripts/run_infer.sh [MODEL_CONFIG] [GIT_COMMIT] [AGENT] [EVAL_LIMIT] [NUM_WORKERS]
+```
+
+- `MODEL_CONFIG`: Name of the model you want to evaluate with
+- `GIT_COMMIT`: This should be the git commit hash or release tag for OpenHands, e.g., HEAD or a specific tag like 0.6.2.
+- `AGENT`: Use CoderActAgent, right now it only supports that.
+- `EVAL_LIMIT`: Number of samples to evaluate.
+- `NUM_WORKERS`: Number of workers to parallelize the evaluation process.

evaluation/discoverybench/eval_utils/README.md

@@ -0,0 +1,7 @@
+## DiscoveryBench Evaluation Utils
+
+- **`eval_w_subhypo_gen.py`**: Implements the DiscoveryBench logic for evaluating agent-generated hypotheses.
+- **`lm_utils.py`**: Provides utility functions necessary for the evaluation process.
+- **`openai_helpers.py`**: Includes helper functions for OpenAI-related tasks.
+- **`openai_semantic_gen_prompts.py`**: Contains prompts used for semantic generation.
+- **`response_parser.py`**: Handles the parsing of agent-generated hypotheses.

evaluation/discoverybench/eval_utils/eval_w_subhypo_gen.py

@@ -0,0 +1,538 @@
+import json
+import logging
+
+from openai import OpenAI
+
+from .lm_utils import run_chatgpt_query_multi_turn
+from .openai_helpers import get_response
+
+logging.basicConfig(
+    format='%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
+    datefmt='%m/%d/%Y %H:%M:%S',
+    level=logging.INFO,
+)
+logger = logging.getLogger(__name__)
+
+
+def get_score_from_answer(type, answer):
+    if type == 'context':
+        answer = answer.replace('Answer:', '').strip()
+        if answer.startswith('A)'):
+            return 1.0
+        elif answer.startswith('B)'):
+            return 0.0
+        return -1.0
+
+    elif type == 'var':
+        try:
+            var_json = json.loads(answer)
+            # print(f"var_json:{var_json}")
+            p = 0.0
+            r = 0.0
+            f1 = 0.0
+            if var_json['sizeB']:
+                p = var_json['intersection'] / var_json['sizeB']
+            if var_json['sizeA']:
+                r = var_json['intersection'] / var_json['sizeA']
+            if p > 0.0 and r > 0.0:
+                f1 = (2 * p * r) / (p + r)
+            else:
+                f1 = 0.0
+            eval_rec = {
+                'p': p,
+                'r': r,
+                'f1': f1,
+                'sizeA': var_json['sizeA'],
+                'sizeB': var_json['sizeB'],
+                'intersection': var_json['intersection'],
+                'explanation': var_json['explanation'],
+            }
+            print(f'var_eval: {eval_rec}')
+            return eval_rec
+        except Exception:  # COMMENT: added Exception
+            return {'p': -1.0, 'r': -1.0, 'f1': -1.0}
+    elif type == 'rel':
+        print(answer)
+        rel_json = json.loads(answer)
+        answer_str = rel_json['answer'].strip()
+        if answer_str.startswith('A') or 'very similar' in answer_str:
+            return 1.0
+        elif (
+            answer_str.startswith('B') or 'similar but general than HypoA' in answer_str
+        ):
+            return 0.5
+        elif answer_str.startswith('C') or 'different' in answer_str:
+            return 0.0
+        return -1.0
+    return -1.0
+
+
+def ask_dimension_question(
+    query,
+    gold_hypo,
+    gold_workflow,
+    gen_hypo,
+    gen_workflow,
+    dataset_meta,
+    llm_used,
+    dimension,
+    dataset_type,
+    use_column_metadata=True,
+):
+    dimension_question = ''
+    answer = ''
+    score = 0.0
+    if dimension == 'var':
+        score = {'p': -1.0, 'r': -1.0, 'f1': -1.0}
+    num_tokens = 256
+    num_retries = 1
+    json_response = False
+
+    messages = [
+        {
+            'role': 'system',
+            'content': 'You are an AI assistant that helps evaluate a data-driven hypothesis. You are a helpful assistant who is not talkative. You only respond with the exact answer to a query without additional conversation.',
+        },
+    ]
+    if dimension == 'context':
+        dimension_question = """\
+        Question: Is HypoB defined in the same context as HypoA?
+        (Context refers to assumptions/stratification under which the hypotheses are defined.)
+        Options: A) same   B) different
+        What is your answer?"""
+    elif dimension == 'var':
+        dimension_question = """\
+        Question: For both HypoA and HypoB, what are the different variables found in the hypotheses? \
+        Return your answer as a JSON object in the following format:
+        ```json
+        {{
+        "sizeA": num of variables used in HypoA
+        "sizeB": num of variables used in HypoB
+        "intersection": num of variables common in HypoA and HypoB. Use *fuzzy matching* to determine intersection, accounting for paraphrases or slightly different surface forms
+        "explanation": a short text explanation about the variables
+        }}```
+        Answer:"""
+        num_tokens = 512
+        num_retries = 1
+        json_response = True
+    elif dimension == 'rel':
+        dimension_question = """\
+        Question: Does HypoB exhibit the same relation as HypoA?
+        Compare using following example hierarchy of relationships (based on specificity): \
+        "there exists a relationship" > "positive relationship" > "positive AND (linear OR quadratic)" > "positive AND linear".
+        Options: A) very similar B) similar but general than HypoA C) different
+        Return your answer as a JSON object in the following format:
+        ```json
+        {{
+        "answer": one of the options from A) very similar B) similar but general than HypoA C) different
+        "explanation": a short text explanation about the relationship comparison
+        }}```
+        Answer:"""
+        num_tokens = 512
+        num_retries = 1
+        json_response = True
+
+    datasets_json = prepare_dataset_metadata_json(
+        dataset_meta, dataset_type=dataset_type, use_column_metadata=use_column_metadata
+    )
+
+    dimension_question_str = f"""\
+        You are going to compare two natural-language hypotheses HypoA and HypoB accompanied with optional workflows: WorkflowA for HypoA and WorkflowB for HypoB. \
+        Both the hypotheses answer the natural language query "QUERY" over the dataset(s) described by dataset description(s) and column description(s) below. \
+        Compare HypoA and HypoB in terms of three aspects: Contexts, Variables, and Relations. \
+        E.g., for the hypothesis "From 1995 to 2009, the number of sandhill cranes around the tundra (Indigilka River) surged by an astounding ~10X":
+        * Contexts refer to stratification of the data under which the given hypothesis is True. E.g., "For all women", "From 1995 to 2009".
+        * Variables refer to the set of variables (either dependent or independent) that are mentioned in the hypothesis. E.g., number of sandhill cranes, location.
+        * Relations refer to the form of relation between the variables. E.g., "surged by ~10x".
+
+        Answer following questions for a given pair of hypotheses, HypoA and HypoB, along with an explanation grounded on the QUERY and the DATASET(S).
+
+        Here is the metadata for the task:
+        ```json
+        {{
+        "datasets": {datasets_json},
+        "query": {query},
+        "HypoA": {gold_hypo},
+        "WorkflowA": {gold_workflow},
+        "HypoB": {gen_hypo},
+        "WorkflowB": {gen_workflow}
+        }}
+        ```
+
+        {dimension_question}"""
+
+    messages.append({'role': 'user', 'content': dimension_question_str})
+    for retry in range(num_retries):
+        response = run_chatgpt_query_multi_turn(
+            messages=messages,
+            model_name=llm_used,
+            max_tokens=num_tokens,
+            temperature=0,  # 0 for greedy best decoding
+            json_response=json_response,
+        )
+        if response is not None:  # COMMENT: changed from != to is not
+            break
+
+    if response is not None:  # COMMENT: changed from != to is not
+        answer = response.choices[0].message.content.strip()
+        score = get_score_from_answer(type=dimension, answer=answer)
+
+    return dimension_question, answer, score
+
+
+def prepare_dataset_metadata_json(dataset_meta, dataset_type, use_column_metadata=True):
+    if dataset_meta is None:  # COMMENT: changed from == to is None
+        return [
+            {
+                'dataset_description': '',
+                'columns': [],
+            }
+        ]
+    datasets_json = []
+    if dataset_type == 'real':
+        for d in dataset_meta['datasets']:
+            datasets_json.append(
+                {
+                    'dataset_description': d['description'],
+                    'columns': [
+                        {'name': col['name'], 'description': col['description']}
+                        for col in d['columns']['raw']
+                    ]
+                    if use_column_metadata
+                    else [],
+                }
+            )
+    else:
+        for d in dataset_meta['datasets']:
+            datasets_json.append(
+                {
+                    'dataset_description': d['description'],
+                    'columns': [
+                        {'name': col['name'], 'description': col['description']}
+                        for col in d['columns']
+                    ]
+                    if use_column_metadata
+                    else [],
+                }
+            )
+    return datasets_json
+
+
+def get_sub_hypotheses(
+    query,
+    hypo,
+    workflow,
+    dataset_meta,
+    llm_used,
+    dataset_type,
+    use_column_metadata=True,
+):
+    client = OpenAI()
+    extraction_prompt = """\
+        Given a set of dataset columns, a ground-truth hypothesis, and the analysis workflow used, your task is to extract three dimensions that define the hypothesis: Context, Variables, and Relations. \
+        Here are the definitions for these dimensions:
+        - Contexts: Boundary conditions that limit the scope of a hypothesis. E.g., “for men over \
+        the age of 30”, “in Asia and Europe”. If the context applies to the full dataset, then extract the context from the dataset_descrption.
+        - Variables: Known concepts that interact in a meaningful way under a given context to \
+        produce the hypothesis. E.g., gender, age, income, or "None" if there is no interacting variable.
+        - Relations: Interactions between a given set of variables under a given context to produce \
+        the hypothesis. E.g., “quadratic relationship”, “inversely proportional”, piecewise conditionals, \
+        or "None" if there is no interacting relationship.
+        Make sure to only use the information present in the hypothesis and the workflow. Do not add any new information. \
+        For each dimension, be specific, and do not omit any important details.
+
+        Here is the metadata for the task:
+        ```json
+        {
+        "datasets": %s,
+        "hypothesis": "%s",
+        "workflow": "%s"
+        }
+        ```
+
+        Return your answer as a JSON object in the following format:
+        ```json
+        {
+        "sub_hypo": [
+            {
+                "text": the hypothesis in natural language,
+                "context": a short text description of the context of the hypothesis,
+                "variables": a list of columns involved in the hypothesis,
+                "relations": a short text description of the relationship between the variables of the hypothesis
+            },
+            ...
+        ]
+        }```
+        """
+    datasets_json = prepare_dataset_metadata_json(
+        dataset_meta, dataset_type, use_column_metadata=use_column_metadata
+    )
+    _prompt = extraction_prompt % (datasets_json, hypo, workflow)
+    sub_hypo_json = get_response(client, _prompt, model=llm_used, max_retry=1)
+
+    if sub_hypo_json is not None:  # COMMENT: changed from != to is not
+        # print(f"full hypothesis: {hypo}")
+        print(f'sub_hypo_json: {sub_hypo_json}')
+    else:
+        sub_hypo_json = {
+            'sub_hypo': [],
+        }
+
+    sub_hypo_json['full_hypo'] = hypo
+
+    return sub_hypo_json
+
+
+def match_context_with_gpt(
+    gold_hyp, gold_context, pred_hyp, pred_context, model='gpt-3.5-turbo'
+):
+    prompt = f"""\
+        Given a gold hypothesis, a gold context, a predicted hypothesis, and a predicted context, your task is \
+        to determine if the predicted context semantically matches the ground-truth context. \
+        Here is the definition for Context: Boundary conditions that limit the scope of a sub-hypothesis. E.g., “for men over the age of 30”, “in Asia and Europe”. If the context applies to the full dataset, then the context is derived from the dataset_descrption. \
+        Here is the definition for Context: Boundary conditions that limit the scope of a sub-hypothesis. E.g., “for men over the age of 30”, “in Asia and Europe”. If the context applies to the full dataset, then the context is derived from the dataset_descrption. \
+        If the predicted context matches the gold context, return true, otherwise return false.
+        If both gold and predicted hypotheses are defined over the context of the full dataset, then also return true.
+        If both gold and predicted hypotheses are defined over the context of the full dataset, then also return true.
+
+        Here is the metadata for the task:
+        ```json
+        {{
+            "gold_hypothesis": "{gold_hyp}",
+            "gold_context": "{gold_context}",
+            "predicted_hypothesis": "{pred_hyp}",
+            "predicted_context": "{pred_context}"
+        }}
+        ```
+
+        Return your answer as a JSON object in the following format:
+        ```json
+        {{
+            "match": true or false
+        }}
+        ```"""
+
+    client = OpenAI()
+    output = get_response(client, prompt, model=model)
+    return output.get('match', False)
+
+
+def is_matching_context(gold_hyp, gold_context, pred_hyp, pred_context, llm_used):
+    if gold_context == pred_context:
+        return True
+    if 'None' in [gold_context, pred_context]:
+        return False
+    return match_context_with_gpt(
+        gold_hyp, gold_context, pred_hyp, pred_context, model=llm_used
+    )
+
+
+def run_eval_gold_vs_gen_NL_subhypo(
+    query,
+    gold_hypo,
+    gold_workflow,
+    gen_hypo,
+    gen_workflow,
+    dataset_meta,
+    llm_used,
+    context_score,
+    dataset_type,
+    use_column_metadata=True,
+):
+    # GPT-4 based evaluation to evaluate generated hypothesis in terms of context, variables, relation
+
+    eval_rec = {
+        'query': query,
+        'HypoA': gold_hypo,
+        'WorkflowA': gold_workflow,
+        'HypoB': gen_hypo,
+        'WorkflowB': gen_workflow,
+    }
+
+    for dimension in ['var', 'rel']:
+        question, answer, score = ask_dimension_question(
+            query,
+            gold_hypo,
+            gold_workflow,
+            gen_hypo,
+            gen_workflow,
+            dataset_meta,
+            llm_used,
+            dimension=dimension,
+            dataset_type=dataset_type,
+            use_column_metadata=use_column_metadata,
+        )
+
+        eval_rec[dimension] = {'question': question, 'answer': answer, 'score': score}
+
+    eval_rec['context'] = context_score
+    eval_rec['accuracy_score'] = (
+        1.0
+        * eval_rec['context']['score']
+        * eval_rec['var']['score']['f1']
+        * eval_rec['rel']['score']
+    )
+
+    return eval_rec
+
+
+def run_eval_gold_vs_gen_NL_hypo_workflow(
+    query,
+    gold_hypo,
+    gold_workflow,
+    gen_hypo,
+    gen_workflow,
+    dataset_meta,
+    llm_used,
+    dataset_type,
+    use_column_metadata=True,
+):
+    # Input: Dataset Metadata, Query, Gold {Hg, Wg}, Predicted {Hp, Wp}
+    # Output: eval_rec json includes final_score
+
+    # Procedure:
+    # Dataset Metadata, Query, Gold {Hg, Wg}, Pred {Hg, Wg}
+    # Gold: [Hg1, Hg2] (compute on the fly) Hg1 is a NL form of subhypothesis
+    # Predicted: [Hp1, Hp2] (compute on the fly)
+
+    # Compute Intersection: [(Hg_i, Hp_j), …]  # tuples of (gold,pred) that matched with context (do this w/o explicit extraction)
+    # # filter so that a gold context and a predicted context are only attached to one tuple
+    # Compute recall_context (programmatically)
+
+    # r_v_list = []
+    # For (Hg_i, Hp_j) in the intersection:
+    #             With Hg_i, Hp_j in NL, ask GPT4 → #variables and #intersection and a paragraph explanation and programmatically calculate f1_v
+    # Hg_i, Hp_j in NL, ask GPT4 → matching score (0, 0.5 or 1) : A) very similar B) similar but general than HypoA C) different + explanation
+    # 	r_v_list ← f1_v * score_r
+    # accuracy_score = mean(r_v_list)
+    # score =   [ recall_context * mean over predicted context(context_score * var_score *rel_score )]
+
+    # recall_context = 1.0  # COMMENT: never used
+    eval_rec = {
+        'query': query,
+        'HypoA': gold_hypo,
+        'WorkflowA': gold_workflow,
+        'HypoB': gen_hypo,
+        'WorkflowB': gen_workflow,
+    }
+
+    gold_sub_hypo_json = get_sub_hypotheses(
+        query=query,
+        hypo=gold_hypo,
+        workflow=gold_workflow,
+        dataset_meta=dataset_meta,
+        llm_used=llm_used,
+        dataset_type=dataset_type,
+        use_column_metadata=use_column_metadata,
+    )
+    if len(gold_sub_hypo_json['sub_hypo']) == 0:
+        gold_sub_hypo_json['sub_hypo'] = [
+            {
+                'text': gold_hypo,
+                'context': 'None',
+                'variables': [],
+                'relations': '',
+                'explanation': 'unable to segment',
+            }
+        ]
+    print(f'gold_sub_hypo_json: {gold_sub_hypo_json}')
+
+    gen_sub_hypo_json = get_sub_hypotheses(
+        query=query,
+        hypo=gen_hypo,
+        workflow=gen_workflow,
+        dataset_meta=dataset_meta,
+        llm_used=llm_used,
+        dataset_type=dataset_type,
+        use_column_metadata=use_column_metadata,
+    )
+    if len(gen_sub_hypo_json['sub_hypo']) == 0:
+        gen_sub_hypo_json['sub_hypo'] = [
+            {
+                'text': gen_hypo,
+                'context': 'None',
+                'variables': [],
+                'relations': '',
+                'explanation': 'unable to segment',
+            }
+        ]
+    print(f'gen_sub_hypo_json: {gen_sub_hypo_json}')
+
+    eval_rec['gold_sub_hypo'] = gold_sub_hypo_json
+    eval_rec['gen_sub_hypo'] = gen_sub_hypo_json
+
+    gold_subh_covered = []
+    gen_subh_to_gold_subh = dict()
+    gen_gold_subh_to_context = dict()
+
+    for p_id, gen_subh in enumerate(gen_sub_hypo_json['sub_hypo']):
+        gen_subh_to_gold_subh[p_id] = -1
+
+        for g_id, gold_subh in enumerate(gold_sub_hypo_json['sub_hypo']):
+            if g_id in gold_subh_covered:
+                continue
+
+            # match context
+            context_bool = is_matching_context(
+                gold_subh['text'],
+                gold_subh.get('context', ''),
+                gen_subh['text'],
+                gen_subh.get('context', ''),
+                llm_used,
+            )
+            if context_bool:
+                context_score = 1.0
+            else:
+                context_score = 0.0
+
+            if context_score == 1.0:  # match only when context_score = 1.0
+                gen_subh_to_gold_subh[p_id] = g_id
+                gold_subh_covered.append(g_id)
+                gen_gold_subh_to_context[f'P{p_id}||G{g_id}'] = {
+                    'question': f"""Comapring: GoldH: {gold_subh["text"]}, GoldC: {gold_subh['context']}\nGenH: {gen_subh['text']}, GenC: {gen_subh['context']}""",
+                    'answer': context_bool,
+                    'score': context_score,
+                }
+                break
+
+    print(f'gen_subh_to_gold_subh: {gen_subh_to_gold_subh}')
+    eval_rec['gen_subh_to_gold_subh'] = gen_subh_to_gold_subh
+    eval_rec['gold_subh_covered'] = gold_subh_covered
+    matched_gold_gen_subh_evals = dict()
+    sum_accuracy_score = 0.0
+    for p_id, g_id in gen_subh_to_gold_subh.items():
+        if g_id >= 0:
+            key = f'P{p_id}||G{g_id}'
+            context_score = gen_gold_subh_to_context[key]
+            subh_eval_rec = run_eval_gold_vs_gen_NL_subhypo(
+                query,
+                gold_hypo,
+                gold_workflow,
+                gen_hypo,
+                gen_workflow,
+                dataset_meta,
+                llm_used,
+                context_score,
+                dataset_type=dataset_type,
+                use_column_metadata=use_column_metadata,
+            )
+            sum_accuracy_score += subh_eval_rec['accuracy_score']
+            matched_gold_gen_subh_evals[key] = subh_eval_rec
+
+    eval_rec['matched_gold_gen_subh_evals'] = matched_gold_gen_subh_evals
+    eval_rec['recall_context'] = (
+        len(gold_subh_covered) / len(gold_sub_hypo_json['sub_hypo'])
+        if len(gold_sub_hypo_json['sub_hypo'])
+        else 0.0
+    )
+    mean_accuracy_score = (
+        sum_accuracy_score / len(gen_subh_to_gold_subh)
+        if len(gen_subh_to_gold_subh)
+        else 0.0
+    )
+    eval_rec['mean_accuracy_score'] = mean_accuracy_score
+    final_score = eval_rec['recall_context'] * mean_accuracy_score
+    eval_rec['final_score'] = final_score
+    print(f'eval_rec: {json.dumps(eval_rec, indent=2)}')
+
+    return eval_rec

evaluation/discoverybench/eval_utils/lm_utils.py

@@ -0,0 +1,64 @@
+import os
+import sys
+import time
+
+from openai import OpenAI
+from tenacity import (
+    retry,
+    stop_after_attempt,  # type: ignore
+    wait_random_exponential,  # type: ignore
+)
+
+if sys.version_info >= (3, 8):
+    from typing import Literal
+else:
+    from typing_extensions import Literal
+
+
+Model = Literal['gpt-4', 'gpt-3.5-turbo', 'text-davinci-003']
+
+OpenAI.api_key = os.getenv('OPENAI_API_KEY')
+OPENAI_GEN_HYP = {
+    'temperature': 0,
+    'max_tokens': 250,
+    'top_p': 1.0,
+    'frequency_penalty': 0,
+    'presence_penalty': 0,
+}
+
+
+@retry(wait=wait_random_exponential(min=1, max=60), stop=stop_after_attempt(6))
+def run_chatgpt_query_multi_turn(
+    messages,
+    model_name='gpt-4-turbo',  # pass "gpt4" for more recent model output
+    max_tokens=256,
+    temperature=0.0,
+    json_response=False,
+):
+    response = None
+    num_retries = 3
+    retry = 0
+    while retry < num_retries:
+        retry += 1
+        try:
+            client = OpenAI()
+
+            if json_response:
+                response = client.chat.completions.create(
+                    model=model_name,
+                    response_format={'type': 'json_object'},
+                    messages=messages,
+                    **OPENAI_GEN_HYP,
+                )
+            else:
+                response = client.chat.completions.create(
+                    model=model_name, messages=messages, **OPENAI_GEN_HYP
+                )
+            break
+
+        except Exception as e:
+            print(e)
+            print('GPT error. Retrying in 2 seconds...')
+            time.sleep(2)
+
+    return response

evaluation/discoverybench/eval_utils/openai_helpers.py

@@ -0,0 +1,190 @@
+import json
+
+
+def OPENAI_TOPIC_GEN_MESSAGES(n=10):
+    return [
+        {
+            'role': 'system',
+            'content': 'You are a helpful assistant who is not talkative. You only respond with the exact answer to a query without additional conversation.',
+        },
+        {
+            'role': 'user',
+            'content': f'Given `n`, come up with a list of `n` distinct topics and their descriptions. The topics can be absolutely anything. Be as creative as possible. Return your answer as a JSON object. \n\nFor example, for `n`=3, a valid answer might be:\n```json\n{{"topics": [\n  {{"id": 1, "topic": "cooking", "description": "Related to recipes, ingredients, chefs, etc."}},\n  {{"id": 2, "topic": "sports", "description": "Related to players, stadiums, trophies, etc."}},\n  {{"id": 3, "topic": "antiquing", "description": "Related to unique items, history, etc."}}\n]}}```\n\nNow, give me a list for `n`={n}. Remember, pick diverse topics from everything possible. No consecutive topics should be broadly similar. Directly respond with the answer JSON object.',
+        },
+    ]
+
+
+OPENAI_GEN_HYP = {
+    'temperature': 1.0,
+    'max_tokens': 4096,
+    'top_p': 1.0,
+    'frequency_penalty': 0,
+    'presence_penalty': 0,
+}
+
+
+def OPENAI_SEMANTICS_GEN_MESSAGES(dependent, relationship, domain, domain_desc):
+    return [
+        {
+            'role': 'system',
+            'content': 'You are a helpful assistant who is not talkative. You only respond with the exact answer to a query without additional conversation.',
+        },
+        {
+            'role': 'user',
+            'content': f'Given the true relationship in a dataset and a given domain, your task is to come up with an interpretation of some real-world concepts that the relationship could be modeling from the provided domain. It\'s okay to be wrong, but suggest something reasonable. Try as much as possible to make sure that the TARGET is actually derivable from the other variables. Give your answer as a JSON object. Here\'s an example:\n\nRelationship for x2 = "(96.4 * x1 ** 3) + (88.72 * x5 ** 2) + (81.96 * x6 ** -2) + (28.13 * x3)  + (97.0) + (0 * x4)"\nDomain="Sales"\nDomain description="Related to product distribution, revenues, marketing, etc."\n\nBased on this, the following real-world concepts might be applicable:\n```json\n{{\n  "dependent": "x2",\n  "relationship": "(96.4 * x1 ** 3) + (88.72 * x5 ** 2) + (81.96 * x6 ** -2) + (28.13 * x3)  + (97.0) + (0 * x4)",\n  "domain": "Sales",\n  "trends": {{\n    "x1": "Positive, cubic factor",\n    "x2": "TARGET",\n    "x3": "Positive, linear factor",\n    "x4": "No relation",\n    "x5": "Positive quadratic factor",\n    "x6": "Positive, inverse quadratic factor"\n  }},\n  "interpretation": {{\n    "x2": {{"description": "Volume of product sales by area", "name": "sales_area", "is_target": true}},\n    "x1": {{"description": "Population by area", "name": "pop_area"}},\n    "x3": {{"description": "Advertising spending", "name": "ad_spend"}},\n    "x4": {{"description": "Gender ratio of marketing team", "name": "gdr_ratio_mkt_team"}},\n    "x5": {{"description": "Intensity of marketing campaign", "name": "mkt_intensity"}}\n  }},\n    "x6": {{"description": "Distance to distribution center", "name": "dist_to_distr_ctr"}}\n}}```\n\nHere\'s a new test question:\nRelationship for {dependent} = "{relationship}"\nDomain = "{domain}"\nDomain description="{domain_desc}"\n\nRespond only with the answer JSON. Make sure that you do not forget to include the TARGET variable in the interpretation object.',
+        },
+    ]
+
+
+def OPENAI_SEMANTICS_GEN_W_MAP_MESSAGES(
+    dependent, relationship, domain, domain_desc, mapping
+):
+    return [
+        {
+            'role': 'system',
+            'content': 'You are a helpful assistant who is not talkative. You only respond with the exact answer to a query without additional conversation.',
+        },
+        {
+            'role': 'user',
+            'content': f'Given a partial mapping from variables to real-world concepts and a true relationship in a dataset, your task is to come up with an interpretation of real-world concepts for the variables without any assigned mapping (those starting with x). Suggest something reasonable. The dependent variable must be derivable only from the other variables in the dependent relationship. Give your answer as a JSON object. Here\'s an example:\n\nExample partial mapping and relationship:\n```json\n{{\n  "domain": "Sales",\n  "domain_description": "Related to product distribution, revenues, marketing, etc.",\n  "variable_mapping": {{\n    "x1": {{"description": "Population by area", "name": "pop_area"}},\n    "x2": {{"description": "Volume of product sales by area", "name": "sales_area"}},\n    "x4": {{"description": "Gender ratio of marketing team", "name": "gdr_ratio_mkt_team"}},\n    "x6": {{"description": "Distance to distribution center", "name": "dist_to_distr_ctr"}}\n  }},\n  "dependent_variable": "sales_area",\n  "dependent_relationship": "(96.4 * pop_area ** 3) + (88.72 * x5 ** 2) + (81.96 * dist_to_distr_ctr ** -2) + (28.13 * x3)  + (97.0)"\n}}```\nBased on this, an example answer would be:\n```json\n{{\n  "dependent_variable": "sales_area",\n  "missing_mapping": ["x3", "x5"],\n  "trends": {{\n    "x3": "Positive, linear factor",\n    "x5": "Positive quadratic factor"\n  }},\n  "interpretation": {{\n    "x3": {{"description": "Advertising spending", "name": "ad_spend"}},\n    "x5": {{"description": "Intensity of marketing campaign", "name": "mkt_intensity"}}\n  }}\n}}```\n\nHere\'s a new test question:\n```json\n{{\n  "domain": "{domain}",\n  "domain_description": "{domain_desc}",\n  "variable_mapping": {json.dumps(mapping, indent=2)},\n  "dependent_variable": "{dependent}",\n  "dependent_relationship": "{relationship}"\n}}```\nRespond only with the answer JSON.',
+        },
+    ]
+
+
+def OPENAI_SEMANTICS_GEN_SUMMARY_MESSAGES(dataset):
+    return [
+        {
+            'role': 'system',
+            'content': 'You are a helpful assistant who is not talkative. You only respond with the exact answer to a query without additional conversation.',
+        },
+        {
+            'role': 'user',
+            'content': f'Given the following descriptions of the columns of a dataset, your task is to come up with a natural language overview of the dataset, which should include (1) what the dataset is about, (2) how the data was collected, (3) when the data was collected, and (3) for what purpose the data was collected. Be specific and creative.\n\nExample dataset:\n```json\n{{  \n  "dataset": {{                                                                                                                                                                                       \n    "x6": {{"description": "Ancient artifact significance score", "name": "artifact_significance_score", "is_target": true}},\n    "x1": {{"description": "Distance to ancient city center", "name": "dist_to_ancient_city_ctr"}},\n    "x2": {{"description": "Quantity of discovered relics", "name": "relic_discovery_qty"}},\n    "x3": {{"description": "Years since last archaeological expedition", "name": "years_since_exp"}},\n    "x4": {{"description": "Number of artifacts in excavation site", "name": "artifact_qty"}},\n    "x5": {{"description": "Soil fertility coefficient", "name": "soil_fertility_coef"}},\n    "x7": {{"description": "Distance to ancient burial grounds", "name": "dist_to_burial_grounds"}},\n    "x8": {{"description": "Population estimate of ancient civilization", "name": "ancient_civilization_pop_estimate"}},\n    "x9": {{"description": "Temperature variation in excavation region", "name": "temp_variation"}}\n  }}\n}}```\nExample description:\nThis dataset is about archaeological explorations and findings linked to ancient civilizations. The data was collected in the form of field metrics during various archaeological expeditions during the late mid-20th century. The purpose of the data collection is to evaluate the significance of ancient artifacts discovered during excavations.\n\nHere is a new test dataset.\n{json.dumps(dataset, indent=2)}\nProvide only the description.',
+        },
+    ]
+
+
+def OPENAI_GEN_HYPO_MESSAGES(dataset):
+    return [
+        {
+            'role': 'system',
+            'content': 'You are a helpful assistant who is not talkative. You only respond with the exact answer to a query without additional conversation.',
+        },
+        {
+            'role': 'user',
+            'content': f'Given a dataset with its descriptions and the true functional relationship between its variables, your task is to generate 3 levels of hypotheses for the stated relationship in plain English. The three levels are "broad", "medium" and "narrow". Make sure that the hypotheses sound natural. *Only include concepts for variables that are present in the provided functional relationship.* Give your answer as a JSON.\n\nFor example, an example dataset might be the following:\n```json\n{{\n  "domain": "cybersecurity",\n  "summary": "This dataset is about measuring cybersecurity threats in a system. The data was collected by monitoring various cybersecurity metrics in a network environment. The purpose of the data collection is to assess and predict potential cybersecurity risks and vulnerabilities.",\n  "variables": [\n    {{\n      "description": "Level of cybersecurity threat",\n      "name": "cybersecurity_threat",\n      "is_target": true\n    }},\n    {{\n      "description": "Number of failed login attempts",\n      "name": "failed_login_attempts"\n    }},\n    {{\n      "description": "Amount of encrypted data",\n      "name": "encrypted_data"\n    }},\n    {{\n      "description": "Frequency of software updates",\n      "name": "software_updates"\n    }},\n    {{\n      "description": "Number of antivirus software installed",\n      "name": "antivirus_software"\n    }},\n    {{\n      "description": "Quality of firewall protection",\n      "name": "firewall_quality"\n    }}\n  ],\n  "relationship": {{\n    "dependent": "cybersecurity_threat",\n    "relation": "-53.5*encrypted_data**2 - 53.85*failed_login_attempts**2 + 67.75*firewall_quality - 92.16 - 36.68/software_updates**3"\n  }}\n}}```\nGiven this dataset, the following is a valid answer:\n```json\n{{\n  "broad": {{\n    "instruction": "Be vague. Only indicate which concepts might be related but not how they are related",\n    "hypothesis": "Threat to cybersecurity is influenced by several factors including the amount of encrypted data, the number of failed login attempts, the quality of the firewall, as well as how often the software is updated."\n  }},\n  "medium": {{\n    "instruction": "Be slightly more specific. For each factor, indicate carefully whether it positively or negatively affects the relationship, but do not indicate what the exponent is.",\n    "hypothesis": "Cybersecurity threat tends to decrease with the amount of data encryption, the number of failed login attempts, as well as the frequency of software updates to some extent, while improvement in the firewall quality has a positive effect."\n  }},\n  "narrow": {{\n    "instruction": "Be specific. Communicate the concepts, whether there is a positive or negative effect (be careful), and the meaning of the exponent",\n    "hypothesis": "The threat to cybersecurity interacts in a complex manner with various factors. As the amount of encrypted data increases, there is a quadratic decrease in threat. Similarly for the number of failed login attempts, there is a negative quadratic relationship. The quality of the firewall protection on the other hand demonstrates a positive and linear relationship. Finally, the frequency of software updates has an inverse cubic relationship to the threat."\n  }},\n}}\n```\n\nBased on this, provide an answer for the following test dataset:\n```json\n{dataset}```\nRespond only with a JSON.',
+        },
+    ]
+
+
+def create_prompt(usr_msg):
+    return [
+        {
+            'role': 'system',
+            'content': 'You are a helpful assistant who is not talkative. You only respond with the exact answer to a query without additional conversation.',
+        },
+        {'role': 'user', 'content': usr_msg},
+    ]
+
+
+def get_response(client, prompt, max_retry=5, model='gpt-3.5-turbo', verbose=False):
+    n_try = 0
+    while n_try < max_retry:
+        response = client.chat.completions.create(
+            model=model, messages=create_prompt(prompt), **OPENAI_GEN_HYP
+        )
+
+        # COMMENT: changed from
+        # response.choices[0].message.content.strip().strip('```json').strip('```')
+        content = response.choices[0].message.content
+        cleaned_content = content.split('```json')[1].split('```')[0].strip()
+        output = cleaned_content
+        try:
+            response_json = json.loads(output)
+            return response_json
+        except ValueError:
+            if verbose:
+                print(f'Bad JSON output:\n\n{output}')
+            n_try += 1
+            if n_try < max_retry:
+                if verbose:
+                    print('Retrying...')
+            else:
+                if verbose:
+                    print('Retry limit reached')
+    return None
+
+
+def get_code_fix(
+    client, code, error, max_retry=5, model='gpt-3.5-turbo', verbose=False
+):
+    prompt = f"""\
+Given the following code snippet and error message, provide a single-line fix for the error. \
+Note that the code is going to be executed using python `eval`. \
+The code should be executable and should not produce the error message. Be as specific as possible.
+
+Here's the code and the error:
+{{
+    "code": "{code}",
+    "error": "{error}"
+}}
+
+Return only a JSON object with the fixed code in the following format:
+```json
+{{
+    "fixed_code": "..."
+}}"""
+    response = get_response(
+        client, prompt, max_retry=max_retry, model=model, verbose=verbose
+    )
+    return response
+
+
+def get_new_hypothesis(
+    client, target, old, expr, cols, model='gpt-3.5-turbo', verbose=False
+):
+    prompt = f"""\
+Given a target column from a dataset, a pandas expression to derive the column from existing columns, a list of \
+existing columns, and a previously written hypothesis text, carefully check if the hypothesis text is consistent with \
+the pandas expression or not. If it is consistent, simply return the hypothesis as it is. If it is not consistent, \
+provide a new natural language hypothesis that is consistent with the pandas expression using only the provided \
+information. Be specific.
+
+Here's the information:
+```json
+{{
+    "target_column": "{target}",
+    "pandas_expression": "{expr}",
+    "existing_columns": {json.dumps(cols, indent=4)}
+    "old_hypothesis": "{old}",
+}}```
+
+Give your answer as a new JSON with the following format:
+```json
+{{
+    "hypothesis": "..."
+}}"""
+    response = get_response(client, prompt, model=model, verbose=verbose)
+    return response
+
+
+def replace_variable(client, expr, old, new, model='gpt-3.5-turbo', verbose=False):
+    prompt = f"""\
+Given a pandas "expression", replace mentions of the "old" column with its "new" value such that the resultant \
+expression is equivalent to the original expression.
+
+Here's the information:
+```json
+{{
+    "expression": "{expr}",
+    "old": "{old}",
+    "new": "{new}"
+}}```
+
+Give your answer as a new JSON with the following format:
+```json
+{{
+    "new_expression": "..."
+}}"""
+    response = get_response(client, prompt, model=model, verbose=verbose)
+    return response

evaluation/discoverybench/eval_utils/openai_semantic_gen_prompts.py

@@ -0,0 +1,151 @@
+common_hypothesis_features = [
+    '1-2 sentences',
+    'surprising finding',
+    'includes numeric concepts',
+    'includes categorical concepts',
+    'includes binary concepts',
+]
+hypothesis_features = [
+    ['requires within-cluster analysis'],
+    ['requires across-cluster analysis'],
+    ['corresponds to a polynomial relationship of some columns'],
+    ['corresponds to a ratio between some columns'],
+    ['requires temporal analysis'],
+    ['relationship is based on descriptive statistics of some columns'],
+    ['requires concepts based on percentage or percentiles'],
+    ['relationship is only applicable to one cluster in the data and not the others'],
+]
+
+column_features = [
+    [
+        'must have one target column',
+        'must have quantifiable columns',
+        'must have a few categorical columns',
+        'make sure the categorical column values do not contain special characters',
+        'include a few distractor columns',
+    ]
+]
+
+common_pandas_features = [
+    'must be executable using python `eval` to create the target column in variable `df` (pandas dataframe)',
+    "for e.g., df['A']**2 + 3*df['B'] + 9, np.where(df['A'] > 3, 'Yes', 'No'), etc.",
+    'variables in pandas_expression must be from the existing columns listed above',
+    'variables in pandas_expression must NOT contain the target column itself',
+]
+pandas_features = [
+    ['expression is a quadratic polynomial'],
+    ['expression is a cubic polynomial'],
+    ['expression is a ratio of existing columns'],
+    ['expression is derived through logical combination of existing columns'],
+    # workflow
+]
+pandas_features = [common_pandas_features + p for p in pandas_features]
+
+common_derived_features = [
+    '1-2 sentences',
+    'includes numeric concepts',
+    'includes categorical concepts',
+    'includes binary concepts',
+]
+derived_features = [common_derived_features + h for h in hypothesis_features]
+hypothesis_features = [common_hypothesis_features + h for h in hypothesis_features]
+
+PROMPT_HYP = """\
+Given a dataset topic and description, generate an interesting hypothesis based on \
+the provided instructions. Be creative and come up with an unusual finding.
+
+```json
+{
+    "topic": "%s",
+    "description": "%s",
+    "hypothesis_features": %s,
+    "hypothesis": "..."
+}```
+
+Give your answer as a new JSON with the following format:
+```json
+{
+    "hypothesis": "..."
+}
+```"""
+
+PROMPT_COL = """\
+Given a dataset topic, its description, and a true hypothesis that can be determined from it, \
+generate a list of valid columns based on the provided instructions.
+
+```json
+{
+    "topic": "%s",
+    "description": "%s",
+    "hypothesis": "%s",
+    "column_instructions": %s,
+    "columns": [
+        {
+            "col_name": "...",  # should be an "_"-separated string
+            "description": "...",
+            "data_type": "...",  # should be executable using python's `eval` function. E.g., str, float, int, bool
+            "data_range": {...},  # should be either {"min": ..., "max": ...} or {"values": [...]}
+            "is_distractor": true/false,  # boolean indicating whether this is a distractor that could cause confusion during data analysis
+            "is_target": true/false  # boolean indicating whether this is the target variable for the hypothesis; at least one column should be the target
+        },
+        ...
+    ],
+    "pandas_instructions": %s,
+    "pandas_equation_for_hypothesis": {
+        "target_col": "...",
+        "target_col_type": "...",
+        "target_col_range": {...},
+        "independent_cols_in_pandas_expression": [],  # list of column names that will be used to derive the target column
+        "pandas_expression": "..."  # expression to derive df[target_col] using df[ind_col1], df[ind_col2], etc.
+    }
+}```
+
+Give your answer as a new JSON with the "columns" and "pandas_equation_for_hypothesis" keys filled using the following format:
+```json
+{
+    "columns": [...],
+    "pandas_equation_for_hypothesis": {...}
+}
+```"""
+
+PROMPT_DER = """\
+Given a dataset topic, description, a true hypothesis that can be determined from the data, \
+and a target column from the dataset, generate a hypothesis for the target column using new independent columns not present in the existing columns.
+
+```json
+{
+    "topic": "%s",
+    "description": "%s",
+    "hypothesis": "%s",
+    "existing_columns": %s,
+    "target_column": "%s",
+    "new_to_target_instructions": %s,
+    "new_to_target_hypothesis": "...",  # describe a relationship between new columns that explains the target column
+    "new_columns_for_target": [  # do not repeat any of the existing columns in the dataset
+        {
+            "col_name": "...",  # should be an "_"-separated string
+            "description": "...",
+            "data_type": "...",  # should be executable using python's `eval` function. E.g., str, float, int, bool
+            "data_range": {...},  # should be either {"min": ..., "max": ...} or {"values": [...]}
+        },
+        ...
+    ],
+    "pandas_instructions": %s,
+    "pandas_equation_for_new_to_target_hypothesis": {
+        "target_col": "...",
+        "target_col_type": "...",
+        "target_col_range": {...},
+        "independent_cols_in_pandas_expression": [],  # list of column names from new_columns_for_target that will be used to derive target_col
+        "pandas_expression": "..."  # expression to derive df[target_col] using df[ind_col1], df[ind_col2], etc.
+    }
+}```
+
+Give your answer as a new JSON with the "new_to_target_hypothesis", "new_columns_for_target", and \
+"pandas_equation_for_new_to_target_hypothesis" keys filled using the following format:
+```json
+{
+    "new_to_target_hypothesis": "...",
+    "new_columns_for_target": [...],
+    "pandas_equation_for_new_to_target_hypothesis": {...}
+}
+```"""

evaluation/discoverybench/eval_utils/response_parser.py

@@ -0,0 +1,52 @@
+workflow_summary_markers = [
+    'WORKFLOW SUMMARY',
+    'WORKFLOW_SUMMARY',
+    'WORKFLOW-SUMMARY',
+    'Workflow Summary',
+]
+
+final_answer_markers = [
+    'FINAL ANSWER',
+    'FINAL_ANSWER',
+    'FINAL-ANSWER',
+    'Final Answer',
+    'Scientific Hypothesis',
+    'Hypothesis',
+]
+
+next_agent_markers = [
+    'NEXT AGENT',
+    'NEXT-AGENT',
+    'NEXT_AGENT',
+    'FEEDBACK',
+]
+
+
+def extract_between(content, start_markers, end_markers=None):
+    for marker in start_markers:
+        if marker in content:
+            result = content.split(marker, 1)[1]
+            if end_markers:
+                for end_marker in end_markers:
+                    if end_marker in result:
+                        result = result.split(end_marker, 1)[0]
+            return result
+    return ''
+
+
+def extract_gen_hypo_from_logs(content: str):
+    error = ''
+
+    gen_workflow = extract_between(
+        content, workflow_summary_markers, final_answer_markers
+    )
+
+    if not gen_workflow:
+        error += 'No Workflow Summary found in the line. | '
+
+    gen_hypothesis = extract_between(content, final_answer_markers, next_agent_markers)
+
+    if not gen_hypothesis:
+        error += 'No Final Answer in the line.'
+
+    return gen_hypothesis, gen_workflow, error

evaluation/discoverybench/run_infer.py

@@ -0,0 +1,492 @@
+import asyncio
+import json
+import os
+
+import git
+import pandas as pd
+
+from evaluation.discoverybench.eval_utils.eval_w_subhypo_gen import (
+    run_eval_gold_vs_gen_NL_hypo_workflow,
+)
+from evaluation.discoverybench.eval_utils.response_parser import (
+    extract_gen_hypo_from_logs,
+)
+from evaluation.utils.shared import (
+    EvalMetadata,
+    EvalOutput,
+    codeact_user_response,
+    compatibility_for_eval_history_pairs,
+    make_metadata,
+    prepare_dataset,
+    reset_logger_for_multiprocessing,
+    run_evaluation,
+)
+from openhands.controller.state.state import State
+from openhands.core.config import (
+    AgentConfig,
+    AppConfig,
+    SandboxConfig,
+    get_llm_config_arg,
+    parse_arguments,
+)
+from openhands.core.logger import openhands_logger as logger
+from openhands.core.main import create_runtime, run_controller
+from openhands.events.action import AgentFinishAction, CmdRunAction, MessageAction
+from openhands.events.observation import CmdOutputObservation
+from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
+
+EVALUATION_LLM = 'gpt-4-1106-preview'
+
+DATA_FILES = {}
+
+LIBRARIES = [
+    'pandas',
+    'numpy',
+    'scipy',
+    'matplotlib',
+    'seaborn',
+    'scikit-learn',
+    'statsmodels',
+]
+
+AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
+    'CodeActAgent': codeact_user_response,
+}
+
+AGENT_CLS_TO_INST_SUFFIX = {
+    'CodeActAgent': 'When you think you have fixed the issue through code changes, please run the following command: <execute_bash> exit </execute_bash>.\n'
+}
+
+
+def get_config(
+    metadata: EvalMetadata,
+) -> AppConfig:
+    config = AppConfig(
+        default_agent=metadata.agent_class,
+        run_as_openhands=False,
+        runtime='eventstream',
+        max_iterations=metadata.max_iterations,
+        sandbox=SandboxConfig(
+            base_container_image='python:3.12-bookworm',
+            enable_auto_lint=True,
+            use_host_network=False,
+        ),
+        # do not mount workspace
+        workspace_base=None,
+        workspace_mount_path=None,
+    )
+    config.set_llm_config(metadata.llm_config)
+    agent_config = AgentConfig(
+        function_calling=False,
+        codeact_enable_jupyter=True,
+        codeact_enable_browsing_delegate=True,
+    )
+    config.set_agent_config(agent_config)
+    return config
+
+
+def get_dv_query_for_real(
+    datasets, question, domain_knowledge=None, workflow_tags=None
+):
+    """
+    Prepare a structured query for the agent to execute on the specified datasets.
+
+    This function constructs a query by compiling metadata from the provided datasets, along with any relevant domain knowledge and workflow tags.
+
+    Args:
+        datasets: List of datasets
+        question: Query to be answered
+        domain_knowledge: Domain knowledge if any
+        workflow_tags: Workflow tags if any
+
+    Returns:
+        query_to_dv: Query to be run on the dataset
+        dataset_meta: Metadata of the dataset
+    """
+
+    dataset_meta = ''
+    for dataset_metadata in datasets:
+        dataset_meta += 'Dataset name: ' + dataset_metadata['name']
+        dataset_meta += 'Dataset description: ' + dataset_metadata['description']
+        dataset_meta += '\nBrief description of columns: '
+        for col in dataset_metadata['columns']['raw']:
+            dataset_meta += col['name'] + ': ' + col['description'] + ', '
+
+    query_to_dv = dataset_meta
+
+    query_to_dv += f'\nQuery: {question}'
+
+    if domain_knowledge:
+        query_to_dv += (
+            '\nAdditionally, we provide some hints that might be useful to solve the task. Domain Knowledge: \n'
+            + domain_knowledge
+            + '.\n'
+        )
+
+    if workflow_tags:
+        query_to_dv += 'The meta tags are: ' + workflow_tags + '.\n'
+
+    query_to_dv += (
+        'In the final answer, please write down a scientific hypothesis in '
+        'natural language, derived from the provided dataset, clearly stating the '
+        'context of hypothesis (if any), variables chosen (if any) and '
+        'relationship between those variables (if any) including any statistical significance.'
+        'Also generate a summary of the full workflow starting from data loading that led to the final answer as WORKFLOW SUMMARY:'
+    )
+
+    # Run the NL query through datavoyager
+    return query_to_dv, dataset_meta
+
+
+def initialize_runtime(runtime: Runtime, data_files: list[str]):
+    """
+    Initialize the runtime for the agent.
+
+    This function is called before the runtime is used to run the agent.
+    """
+    logger.info(f"{'-' * 50} BEGIN Runtime Initialization Fn {'-' * 50}")
+    obs: CmdOutputObservation
+
+    action = CmdRunAction(command='mkdir -p /workspace')
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    assert obs.exit_code == 0
+
+    action = CmdRunAction(command='cd /workspace')
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    assert obs.exit_code == 0
+
+    for file in data_files:
+        runtime.copy_to(
+            file,
+            '/workspace',
+        )
+
+    for lib in LIBRARIES:
+        action = CmdRunAction(command=f'pip install {lib}')
+        logger.info(action, extra={'msg_type': 'ACTION'})
+        obs = runtime.run_action(action)
+        assert obs.exit_code == 0
+
+    logger.info(f"{'-' * 50} END Runtime Initialization Fn {'-' * 50}")
+
+
+def get_last_agent_finish_action(state: State) -> AgentFinishAction:
+    for event in reversed(state.history):
+        if isinstance(event, AgentFinishAction):
+            return event
+    return None
+
+
+def get_last_message_action(state: State) -> MessageAction:
+    for event in reversed(state.history):
+        if isinstance(event, MessageAction):
+            return event
+    return None
+
+
+def complete_runtime(state: State):
+    last_agent_finish_action = get_last_agent_finish_action(state)
+    last_agent_message_action = get_last_message_action(state)
+
+    if last_agent_finish_action is not None:
+        final_message_1 = last_agent_finish_action.thought
+        gen_hypo_1, gen_workflow_1, error_1 = extract_gen_hypo_from_logs(
+            final_message_1
+        )
+    else:
+        gen_hypo_1, gen_workflow_1, error_1 = '', '', ''
+
+    if last_agent_message_action is not None:
+        final_message_2 = last_agent_message_action.content
+        gen_hypo_2, gen_workflow_2, error_2 = extract_gen_hypo_from_logs(
+            final_message_2
+        )
+    else:
+        gen_hypo_2, gen_workflow_2, error_2 = '', '', ''
+
+    if gen_hypo_1 and gen_hypo_2:
+        test_result = {
+            'gen_hypo': last_agent_finish_action.thought
+            if last_agent_finish_action
+            else last_agent_message_action.content,
+            'gen_workflow': '',
+            'error': '',
+        }
+        return test_result
+
+    test_result = {
+        'gen_hypo': gen_hypo_1 if gen_hypo_1 else gen_hypo_2,
+        'gen_workflow': gen_workflow_1 if gen_workflow_1 else gen_workflow_2,
+        'error': error_1 if error_1 else error_2,
+    }
+
+    return test_result
+
+
+def process_instance(
+    instance: pd.Series,
+    metadata: EvalMetadata,
+    reset_logger: bool = True,
+):
+    """
+    Process and evaluate a single instance of the dataset.
+
+    This function executes the OpenHands agent
+    for a specific instance of the dataset. It retrieves
+    the agent's results and evaluates them against the gold
+    hypothesis.
+
+    Args:
+        instance: A single row of the dataset
+        metadata: Metadata for the evaluation
+        reset_logger: Whether to reset the logger
+
+    Returns:
+        output: EvalOutput object
+    """
+
+    config = get_config(metadata)
+
+    # use a session id for concurrent evaluation
+    sid = 'ID_' + str(instance.instance_id)
+
+    # 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.instance_id, log_dir)
+    else:
+        logger.info(f'Starting evaluation for instance {instance.instance_id}.')
+
+    problem_statement, dataset_metadata = get_dv_query_for_real(
+        datasets=instance.datasets,
+        question=instance.query,
+        domain_knowledge=instance.domain_knowledge,
+        workflow_tags=instance.workflow_tags,
+    )
+
+    # Prepare instruction
+    instruction = (
+        f'You are a discovery agent who can execute a python code only once to answer a query based on one or more datasets. The datasets will be present in the current directory.\n\n'
+        'Environment has been set up for you to start working. You may assume all necessary tools and datasets are installed.\n\n'
+        '# Problem Statement\n'
+        f'{problem_statement}\n\n'
+    )
+    instruction += (
+        'IMPORTANT: You should ONLY interact with the environment provided to you AND NEVER ASK FOR HUMAN HELP.\n'
+        'You should NOT modify any existing test case files. If needed, you can add new test cases in a NEW file to reproduce the issue.\n'
+        'You SHOULD INCLUDE PROPER INDENTATION in your edit commands.\n'
+    )
+    # NOTE: You can actually set slightly different instruction for different agents
+    instruction += AGENT_CLS_TO_INST_SUFFIX[metadata.agent_class]
+
+    # Here's how you can run the agent (similar to the `main` function) and get the final task state
+    runtime = create_runtime(config, sid=sid)
+    call_async_from_sync(runtime.connect)
+    initialize_runtime(runtime, instance.data_files)
+
+    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.get(
+                metadata.agent_class
+            ),
+        )
+    )
+
+    if state is None:
+        raise ValueError('State should not be None.')
+
+    metrics = state.metrics.get() if state.metrics else None
+    test_result = complete_runtime(state)
+
+    # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
+    # for compatibility with the existing output format, we can remake the pairs here
+    # remove when it becomes unnecessary
+    histories = compatibility_for_eval_history_pairs(state.history)
+
+    # DiscoveryBench Evaluation
+    eval_rec = run_eval_gold_vs_gen_NL_hypo_workflow(
+        query=instance.query,
+        gold_hypo=instance.gold_hypo,
+        gold_workflow='',
+        gen_hypo=test_result['gen_hypo'],
+        gen_workflow='',
+        dataset_meta=instance.dataset_metadata,
+        llm_used=EVALUATION_LLM,
+        dataset_type='real',
+    )
+
+    test_result['eval_rec'] = eval_rec
+
+    output = EvalOutput(
+        instance_id=str(instance.instance_id),
+        instruction=instruction,
+        metadata=metadata,
+        history=histories,
+        metrics=metrics,
+        error=state.last_error if state and state.last_error else None,
+        test_result=test_result,
+    )
+
+    return output
+
+
+def update_csv_name(name):
+    name = name.replace('-', '_')
+
+    if 'meta_regression' in name:
+        name = name.replace('meta_regression', 'meta-regression')
+    if 'ML_enabled' in name:
+        name = name.replace('ML_enabled', 'ML-enabled')
+
+    return name
+
+
+def list_csv_files(list_of_datasets):
+    res = []
+    for ele in list_of_datasets:
+        for key, value in ele.items():
+            if key == 'name':
+                csv_file_name = update_csv_name(value)
+                res.append(DATA_FILES[csv_file_name])
+    return res
+
+
+def create_dataset(repo_location: str, split: str = 'test'):
+    """
+    Create a dataset from the discoverybench repository
+    by walking through the repository and extracting metadata
+    from the metadata_{}.json files
+
+    Args:
+        repo_location: Location of the repository
+        split: Split of the dataset to use
+
+    Returns:
+        df: DataFrame containing the dataset instances
+    """
+
+    data_dict = {}
+
+    data_location = os.path.join(repo_location, 'discoverybench', 'real', split)
+    answer_key_location = os.path.join(repo_location, 'eval', 'answer_key_real.csv')
+
+    idx = 0
+
+    for root, dirs, files in os.walk(data_location):
+        for file in files:
+            if file.endswith('.json'):
+                if 'metadata' in file:
+                    metadata = json.load(open(os.path.join(root, file)))
+
+                    dataset = root.split('/')[-1]
+                    metadata_id = file.split('_')[-1].split('.')[0]
+                    domain = metadata.get('domain', '')
+                    domain_knowledge = metadata.get('domain_knowledge', '')
+                    workflow_tags = metadata.get('workflow_tags', '')
+                    datasets = metadata.get('datasets', [])
+                    queries = metadata.get('queries', [])
+                    gold_workflow = metadata.get('workflow')
+
+                    # loop through queries list to get queries
+                    # and each query has qid; add that to dictionary
+                    for query in queries[0]:
+                        qid = query.get('qid', '')
+
+                        data = {
+                            'dataset': dataset,
+                            'metadata_id': metadata_id,
+                            'qid': qid,
+                            'domain': domain,
+                            'domain_knowledge': domain_knowledge,
+                            'workflow_tags': workflow_tags,
+                            'datasets': datasets,
+                            'question_type': query['question_type'],
+                            'query': query['question'],
+                            'gold_workflow': gold_workflow,
+                            'dataset_metadata': metadata,
+                        }
+
+                        data_dict[idx] = data
+                        idx += 1
+
+            if file.endswith('.csv'):
+                DATA_FILES[file] = os.path.join(root, file)
+            if file.endswith('.txt'):
+                DATA_FILES[file] = os.path.join(root, file)
+
+    df = pd.DataFrame.from_dict(data_dict, orient='index')
+
+    df['instance_id'] = df.index
+
+    df['data_files'] = df['datasets'].apply(lambda x: list_csv_files(x))
+
+    answer_key = pd.read_csv(answer_key_location)
+
+    answer_key = answer_key.rename(
+        columns={
+            'metadataid': 'metadata_id',
+            'query_id': 'qid',
+            'gold_hypothesis': 'gold_hypothesis',
+        }
+    )
+
+    df['qid'] = df['qid'].astype(int)
+    df['metadata_id'] = df['metadata_id'].astype(int)
+
+    answer_key['qid'] = answer_key['qid'].astype(int)
+    answer_key['metadata_id'] = answer_key['metadata_id'].astype(int)
+
+    df = pd.merge(df, answer_key, on=['dataset', 'metadata_id', 'qid'], how='left')
+
+    return df
+
+
+if __name__ == '__main__':
+    args = parse_arguments()
+
+    # clone git repositor for csv files
+    repo_url = 'https://github.com/allenai/discoverybench.git'
+    repo_location = 'git-discoverybench-allenai'
+
+    try:
+        git.Repo.clone_from(repo_url, repo_location)
+    except git.exc.GitCommandError:
+        print('Repository already exists')
+
+    dataset = create_dataset(repo_location)
+
+    # check if there is any empty csv_file
+    if dataset['data_files'].isnull().any():
+        raise ValueError('Some csv files are missing.')
+
+    llm_config = None
+    if args.llm_config:
+        llm_config = get_llm_config_arg(args.llm_config)
+    if llm_config is None:
+        raise ValueError(f'Could not find LLM config: --llm_config {args.llm_config}')
+
+    metadata = make_metadata(
+        llm_config,
+        'discoverybench-python',
+        args.agent_cls,
+        args.max_iterations,
+        args.eval_note,
+        args.eval_output_dir,
+    )
+    output_file = os.path.join(metadata.eval_output_dir, 'output.jsonl')
+    instances = prepare_dataset(dataset, output_file, args.eval_n_limit)
+
+    run_evaluation(
+        instances,
+        metadata,
+        output_file,
+        args.eval_num_workers,
+        process_instance,
+    )

evaluation/discoverybench/scripts/run_infer.sh

@@ -0,0 +1,46 @@
+#!/bin/bash
+set -eo pipefail
+
+source "evaluation/utils/version_control.sh"
+
+MODEL_CONFIG=$1
+COMMIT_HASH=$2
+AGENT=$3
+EVAL_LIMIT=$4
+NUM_WORKERS=$5
+
+if [ -z "$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
+
+get_agent_version
+
+echo "AGENT: $AGENT"
+echo "AGENT_VERSION: $AGENT_VERSION"
+echo "MODEL_CONFIG: $MODEL_CONFIG"
+
+COMMAND="poetry run python evaluation/discoverybench/run_infer.py \
+  --agent-cls $AGENT \
+  --llm-config $MODEL_CONFIG \
+  --max-iterations 10 \
+  --max-chars 10000000 \
+  --eval-num-workers $NUM_WORKERS \
+  --eval-note $AGENT_VERSION"
+
+if [ -n "$EVAL_LIMIT" ]; then
+  echo "EVAL_LIMIT: $EVAL_LIMIT"
+  COMMAND="$COMMAND --eval-n-limit $EVAL_LIMIT"
+fi
+
+# Run the command
+eval $COMMAND

evaluation/gaia/run_infer.py

@@ -12,6 +12,7 @@ from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
     codeact_user_response,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -166,7 +167,7 @@ def process_instance(
 
     model_answer_raw = ''
     # get the last message or thought from the agent
-    for event in state.history.get_events(reverse=True):
+    for event in reversed(state.history):
         if event.source == 'agent':
             if isinstance(event, AgentFinishAction):
                 model_answer_raw = event.thought
@@ -203,7 +204,7 @@ def process_instance(
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     # Save the output
     output = EvalOutput(

evaluation/gorilla/run_infer.py

@@ -10,6 +10,7 @@ from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
     codeact_user_response,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -101,7 +102,8 @@ def process_instance(
         raise ValueError('State should not be None.')
 
     # retrieve the last message from the agent
-    model_answer_raw = state.history.get_last_agent_message()
+    last_agent_message = state.get_last_agent_message()
+    model_answer_raw = last_agent_message.content if last_agent_message else ''
 
     # attempt to parse model_answer
     ast_eval_fn = instance['ast_eval']
@@ -114,7 +116,7 @@ def process_instance(
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     output = EvalOutput(
         instance_id=instance_id,

evaluation/gpqa/run_infer.py

@@ -28,6 +28,7 @@ from datasets import load_dataset
 from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -244,7 +245,7 @@ Ok now its time to start solving the question. Good luck!
         'C': False,
         'D': False,
     }
-    for event in state.history.get_events(reverse=True):
+    for event in reversed(state.history):
         if (
             isinstance(event, AgentFinishAction)
             and event.source != 'user'
@@ -300,7 +301,7 @@ Ok now its time to start solving the question. Good luck!
         instance_id=str(instance.instance_id),
         instruction=instruction,
         metadata=metadata,
-        history=state.history.compatibility_for_eval_history_pairs(),
+        history=compatibility_for_eval_history_pairs(state.history),
         metrics=metrics,
         error=state.last_error if state and state.last_error else None,
         test_result={

evaluation/humanevalfix/run_infer.py

@@ -21,6 +21,7 @@ from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
     codeact_user_response,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -255,7 +256,7 @@ def process_instance(
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     # Save the output
     output = EvalOutput(

evaluation/integration_tests/run_infer.py

@@ -13,6 +13,7 @@ from evaluation.utils.shared import (
     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 (
@@ -55,18 +56,14 @@ def get_config(
         workspace_base=None,
         workspace_mount_path=None,
     )
-    if metadata.llm_config.log_completions:
-        metadata.llm_config.log_completions_folder = os.path.join(
-            metadata.eval_output_dir, 'llm_completions', instance_id
+    config.set_llm_config(
+        update_llm_config_for_completions_logging(
+            metadata.llm_config, metadata.eval_output_dir, instance_id
         )
-        logger.info(
-            f'Logging LLM completions for instance {instance_id} to '
-            f'{metadata.llm_config.log_completions_folder}'
-        )
-    config.set_llm_config(metadata.llm_config)
+    )
     agent_config = AgentConfig(
         codeact_enable_jupyter=True,
-        codeact_enable_browsing_delegate=True,
+        codeact_enable_browsing=True,
         codeact_enable_llm_editor=False,
     )
     config.set_agent_config(agent_config)
@@ -132,7 +129,7 @@ def process_instance(
     # # result evaluation
     # # =============================================
 
-    histories = [event_to_dict(event) for event in state.history.get_events()]
+    histories = [event_to_dict(event) for event in state.history]
     test_result: TestResult = test_class.verify_result(runtime, histories)
     metrics = state.metrics.get() if state.metrics else None
 

evaluation/integration_tests/tests/t06_github_pr_browsing.py

@@ -0,0 +1,44 @@
+from evaluation.integration_tests.tests.base import BaseIntegrationTest, TestResult
+from openhands.events.action import AgentFinishAction, MessageAction
+from openhands.events.event import Event
+from openhands.events.observation import AgentDelegateObservation
+from openhands.runtime.base import Runtime
+
+
+class Test(BaseIntegrationTest):
+    INSTRUCTION = 'Look at https://github.com/All-Hands-AI/OpenHands/pull/8, and tell me what is happening there and what did @asadm suggest.'
+
+    @classmethod
+    def initialize_runtime(cls, runtime: Runtime) -> None:
+        pass
+
+    @classmethod
+    def verify_result(cls, runtime: Runtime, histories: list[Event]) -> TestResult:
+        # check if the "The answer is OpenHands is all you need!" is in any message
+        message_actions = [
+            event
+            for event in histories
+            if isinstance(
+                event, (MessageAction, AgentFinishAction, AgentDelegateObservation)
+            )
+        ]
+        for event in message_actions:
+            if isinstance(event, AgentDelegateObservation):
+                content = event.content
+            elif isinstance(event, AgentFinishAction):
+                content = event.outputs.get('content', '')
+            elif isinstance(event, MessageAction):
+                content = event.content
+            else:
+                raise ValueError(f'Unknown event type: {type(event)}')
+
+            if (
+                'non-commercial' in content
+                or 'MIT' in content
+                or 'Apache 2.0' in content
+            ):
+                return TestResult(success=True)
+        return TestResult(
+            success=False,
+            reason=f'The answer is not found in any message. Total messages: {len(message_actions)}. Messages: {message_actions}',
+        )

evaluation/logic_reasoning/run_infer.py

@@ -8,6 +8,7 @@ from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
     codeact_user_response,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -225,7 +226,7 @@ def process_instance(
         raise ValueError('State should not be None.')
 
     final_message = ''
-    for event in state.history.get_events(reverse=True):
+    for event in reversed(state.history):
         if isinstance(event, AgentFinishAction):
             final_message = event.thought
             break
@@ -247,7 +248,7 @@ def process_instance(
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     # Save the output
     output = EvalOutput(

evaluation/miniwob/run_infer.py

@@ -10,10 +10,13 @@ import pandas as pd
 from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
+    codeact_user_response,
+    compatibility_for_eval_history_pairs,
     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 (
@@ -29,7 +32,10 @@ from openhands.events.action import (
     CmdRunAction,
     MessageAction,
 )
-from openhands.events.observation import CmdOutputObservation
+from openhands.events.observation import (
+    BrowserOutputObservation,
+    CmdOutputObservation,
+)
 from openhands.runtime.base import Runtime
 from openhands.runtime.browser.browser_env import (
     BROWSER_EVAL_GET_GOAL_ACTION,
@@ -37,7 +43,11 @@ from openhands.runtime.browser.browser_env import (
 )
 from openhands.utils.async_utils import call_async_from_sync
 
-SUPPORTED_AGENT_CLS = {'BrowsingAgent'}
+SUPPORTED_AGENT_CLS = {'BrowsingAgent', 'CodeActAgent'}
+
+AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
+    'CodeActAgent': codeact_user_response,
+}
 
 
 def get_config(
@@ -47,25 +57,32 @@ def get_config(
     config = AppConfig(
         default_agent=metadata.agent_class,
         run_as_openhands=False,
-        runtime='eventstream',
+        runtime=os.environ.get('RUNTIME', 'eventstream'),
         max_iterations=metadata.max_iterations,
         sandbox=SandboxConfig(
             base_container_image='xingyaoww/od-eval-miniwob:v1.0',
             enable_auto_lint=True,
             use_host_network=False,
             browsergym_eval_env=env_id,
+            api_key=os.environ.get('ALLHANDS_API_KEY', None),
+            remote_runtime_api_url=os.environ.get('SANDBOX_REMOTE_RUNTIME_API_URL'),
+            keep_remote_runtime_alive=False,
         ),
         # do not mount workspace
         workspace_base=None,
         workspace_mount_path=None,
     )
-    config.set_llm_config(metadata.llm_config)
+    config.set_llm_config(
+        update_llm_config_for_completions_logging(
+            metadata.llm_config, metadata.eval_output_dir, env_id
+        )
+    )
     return config
 
 
 def initialize_runtime(
     runtime: Runtime,
-) -> str:
+) -> tuple[str, BrowserOutputObservation]:
     """Initialize the runtime for the agent.
 
     This function is called before the runtime is used to run the agent.
@@ -85,8 +102,14 @@ def initialize_runtime(
     logger.info(obs, extra={'msg_type': 'OBSERVATION'})
     goal = obs.content
 
+    # Run noop to get the initial browser observation (e.g., the page URL & content)
+    action = BrowseInteractiveAction(browser_actions='noop(1000)')
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+
     logger.info(f"{'-' * 50} END Runtime Initialization Fn {'-' * 50}")
-    return goal
+    return goal, obs
 
 
 def complete_runtime(
@@ -117,7 +140,7 @@ def process_instance(
     metadata: EvalMetadata,
     reset_logger: bool = True,
 ) -> EvalOutput:
-    env_id = instance.id
+    env_id = instance.instance_id
     config = get_config(metadata, env_id)
 
     # Setup the logger properly, so you can run multi-processing to parallelize the evaluation
@@ -129,7 +152,12 @@ def process_instance(
 
     runtime = create_runtime(config)
     call_async_from_sync(runtime.connect)
-    task_str = initialize_runtime(runtime)
+    task_str, obs = initialize_runtime(runtime)
+
+    task_str += (
+        f'\nInitial browser state (output of `noop(1000)`):\n{obs.get_agent_obs_text()}'
+    )
+
     state: State | None = asyncio.run(
         run_controller(
             config=config,
@@ -137,6 +165,9 @@ def process_instance(
                 content=task_str
             ),  # take output from initialize_runtime
             runtime=runtime,
+            fake_user_response_fn=AGENT_CLS_TO_FAKE_USER_RESPONSE_FN[
+                metadata.agent_class
+            ],
         )
     )
 
@@ -152,19 +183,19 @@ def process_instance(
 
     # Instruction is the first message from the USER
     instruction = ''
-    for event in state.history.get_events():
+    for event in state.history:
         if isinstance(event, MessageAction):
             instruction = event.content
             break
 
     return_val = complete_runtime(runtime)
     logger.info(f'Return value from complete_runtime: {return_val}')
-    reward = max(return_val['rewards'])
+    reward = max(return_val['rewards'], default=0)
 
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     # Save the output
     output = EvalOutput(

evaluation/mint/run_infer.py

@@ -13,6 +13,7 @@ from evaluation.mint.tasks import Task
 from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -28,6 +29,7 @@ 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.events.action import (
+    Action,
     CmdRunAction,
     MessageAction,
 )
@@ -45,7 +47,10 @@ def codeact_user_response_mint(state: State, task: Task, task_config: dict[str,
         task=task,
         task_config=task_config,
     )
-    last_action = state.history.get_last_action()
+    last_action = next(
+        (event for event in reversed(state.history) if isinstance(event, Action)),
+        None,
+    )
     result_state: TaskState = env.step(last_action.message or '')
 
     state.extra_data['task_state'] = result_state
@@ -202,7 +207,7 @@ def process_instance(
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     # Save the output
     output = EvalOutput(

evaluation/ml_bench/run_infer.py

@@ -24,6 +24,7 @@ from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
     codeact_user_response,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -256,7 +257,7 @@ def process_instance(instance: Any, metadata: EvalMetadata, reset_logger: bool =
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     # Save the output
     output = EvalOutput(

evaluation/scienceagentbench/Dockerfile

@@ -0,0 +1,17 @@
+FROM python:3.11-bookworm
+
+
+# For OpenHands agents to explore the dataset directories, please download the full benchmark [here](https://buckeyemailosu-my.sharepoint.com/:u:/g/personal/chen_8336_buckeyemail_osu_edu/EQuA6uJ3CtRHvRfZ2GiN1tYBRVJE4DSUD10MW61fr7HuSQ?e=sCBegG) and unzip it with password `scienceagentbench`.
+# **Please DO NOT redistribute the unzipped data files online.**
+# It will download a benchmark.zip file to the current directory.
+# unzip it and put the benchmark folder under evaluation/scienceagentbench/
+
+RUN mkdir -p /benchmark
+COPY benchmark /benchmark
+
+RUN mkdir -p /workspace
+WORKDIR /workspace
+
+# pushd evaluation/scienceagentbench
+# docker build -t xingyaoww/openhands-eval-scienceagentbench .
+# popd

evaluation/scienceagentbench/Dockerfile.evaluator

@@ -0,0 +1,25 @@
+FROM mambaorg/micromamba:debian12
+
+USER root
+# For https://github.com/OSU-NLP-Group/ScienceAgentBench/tree/main?tab=readme-ov-file#code-generation-with-agents
+
+RUN micromamba create -n sci-agent-eval python=3.10 pip setuptools wheel
+RUN micromamba run -n sci-agent-eval pip install pip-tools
+
+RUN mkdir -p /workspace
+WORKDIR /workspace
+
+RUN apt-get update && apt-get install -y git
+
+RUN git clone https://github.com/OSU-NLP-Group/ScienceAgentBench.git /workspace/
+RUN git checkout 4eddc7db6449a5ade3e37285747c8b208cd54ce7
+
+RUN micromamba create -n sci-agent python=3.10 pip setuptools wheel
+RUN micromamba run -n sci-agent pip install -r requirements.txt
+
+# Replace all occurence of conda with micromamba under the /workspace
+RUN find ./ -type f -exec sed -i 's/conda/micromamba/g' {} \;
+
+# pushd evaluation/scienceagentbench
+# docker build -t xingyaoww/openhands-eval-scienceagentbench-evaluator -f Dockerfile.evaluator .
+# popd

evaluation/scienceagentbench/README.md

@@ -0,0 +1,54 @@
+# ScienceAgentBench Evaluation with OpenHands
+
+This folder contains the evaluation harness for [ScienceAgentBench](https://osu-nlp-group.github.io/ScienceAgentBench/) (paper: https://arxiv.org/abs/2410.05080).
+
+## Setup Environment and LLM Configuration
+
+Please follow instruction [here](../README.md#setup) to setup your local development environment and LLM.
+
+## Setup ScienceAgentBench
+
+To prevent benchmark data contamination, we only provide the annotation sheet on [Huggingface](https://huggingface.co/datasets/osunlp/ScienceAgentBench), which includes all necessary *inputs* to run an agent.
+
+## Run Inference on ScienceAgentBench
+
+```bash
+./evaluation/scienceagentbench/scripts/run_infer.sh [model_config] [git-version] [use_knowledge] [agent] [eval_limit] [max_iter] [num_workers] [dataset] [dataset_split]
+
+# Example
+./evaluation/scienceagentbench/scripts/run_infer.sh llm.eval_gpt4o 0.9.3
+```
+
+where `model_config` is mandatory, and the rest are optional.
+
+- `model_config`, e.g. `eval_gpt4_1106_preview`, is the config group name for your
+LLM settings, as defined in your `config.toml`.
+- `git-version`, e.g. `HEAD`, is the git commit hash of the OpenHands version you would
+like to evaluate. It could also be a release tag like `0.6.2`.
+- `use_knowledge`, e.g. `true`, specifies whether allowing the agent to use expert-provided knowledge as additional input or not. By default, it is set to `false`.
+- `agent`, e.g. `CodeActAgent`, is the name of the agent for benchmarks, defaulting
+to `CodeActAgent`.
+- `eval_limit`, e.g. `10`, limits the evaluation to the first `eval_limit` instances. By
+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.
+- `num_workers`, e.g. `3`, is the number of parallel workers to run the evaluation. By
+default, it is set to 1.
+
+## Evaluate Generated Programs
+
+### Extract Necessary Information from OpenHands Log
+
+After the inference is completed, you may use the following command to extract necessary information from the output log for evaluation:
+
+```bash
+python post_proc.py [log_fname]
+```
+- `log_fname`, e.g. `evaluation/.../output.jsonl`, is the automatically saved trajectory log of an OpenHands agent.
+
+Output will be write to e.g. `evaluation/.../output.converted.jsonl`
+
+### Run evaluation
+
+Please follow the steps [here](https://github.com/OSU-NLP-Group/ScienceAgentBench/tree/main?tab=readme-ov-file#evaluation-of-generated-code) to evaluate the generated programs.

evaluation/scienceagentbench/post_proc.py

@@ -0,0 +1,30 @@
+import json
+from argparse import ArgumentParser
+
+if __name__ == '__main__':
+    parser = ArgumentParser()
+    parser.add_argument(
+        'log_fname',
+        type=str,
+    )
+    args = parser.parse_args()
+
+    fname = args.log_fname
+    out_fname = args.log_fname.replace('.jsonl', '.converted.jsonl')
+
+    log = [json.loads(line) for line in open(fname)]
+
+    simple_log = [
+        json.dumps(
+            {
+                'instance_id': ex['instance_id'],
+                'instruction': ex['instruction'],
+                'test_result': ex['test_result'],
+                'cost': ex['metrics']['accumulated_cost'],
+            }
+        )
+        for ex in log
+    ]
+
+    with open(out_fname, 'w+', encoding='utf-8') as f:
+        f.write('\n'.join(simple_log))

evaluation/scienceagentbench/run_infer.py

@@ -0,0 +1,292 @@
+import asyncio
+import os
+from typing import Any
+
+import pandas as pd
+from datasets import load_dataset
+from tqdm import tqdm
+
+from evaluation.utils.shared import (
+    EvalMetadata,
+    EvalOutput,
+    codeact_user_response,
+    compatibility_for_eval_history_pairs,
+    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 (
+    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
+from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
+
+AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
+    'CodeActAgent': codeact_user_response,
+}
+
+LOCAL_DATASET_PATH = os.path.join(os.path.dirname(__file__), 'benchmark')
+
+
+def format_task_dict(example, use_knowledge):
+    task = {
+        'instance_id': example['instance_id'],
+        'task_inst': example['task_inst'],
+        'dataset_path': '/benchmark/datasets/'
+        + example['dataset_folder_tree'].split('\n')[0][4:],
+        'dataset_folder_tree': example['dataset_folder_tree'],
+        'dataset_preview': example['dataset_preview'],
+        'pred_program_name': 'pred_' + example['gold_program_name'],
+    }
+
+    if use_knowledge:
+        task['task_inst'] += '\n' + str(example['domain_knowledge'])
+
+    return task
+
+
+def get_config(
+    metadata: EvalMetadata,
+    instance_id: str,
+) -> AppConfig:
+    config = AppConfig(
+        default_agent=metadata.agent_class,
+        run_as_openhands=False,
+        runtime=os.environ.get('RUNTIME', 'eventstream'),
+        max_budget_per_task=4,
+        max_iterations=metadata.max_iterations,
+        sandbox=SandboxConfig(
+            base_container_image='docker.io/xingyaoww/openhands-eval-scienceagentbench',
+            enable_auto_lint=True,
+            use_host_network=False,
+            timeout=300,
+            api_key=os.environ.get('ALLHANDS_API_KEY', None),
+            remote_runtime_api_url=os.environ.get('SANDBOX_REMOTE_RUNTIME_API_URL'),
+            keep_remote_runtime_alive=False,
+        ),
+        # 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,
+        )
+    )
+    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(f"{'-' * 50} BEGIN Runtime Initialization Fn {'-' * 50}")
+    obs: CmdOutputObservation
+
+    # Set up workspace directories
+    action = CmdRunAction(command='mkdir -p /workspace/pred_programs')
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    assert obs.exit_code == 0
+
+    action = CmdRunAction(command='mkdir -p /workspace/pred_results')
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+    assert obs.exit_code == 0
+
+    dataset_name = instance['dataset_folder_tree'].split('\n')[0][4:].rstrip('/')
+
+    # Copy the dataset to the workspace
+    dataset_dir = os.path.join(
+        LOCAL_DATASET_PATH,
+        'datasets',
+        dataset_name,
+    )
+    runtime.copy_to(dataset_dir, '/workspace/benchmark/datasets', recursive=True)
+
+    # Check the dataset exists
+    action = CmdRunAction(
+        command='cd /workspace/benchmark/datasets && ls',
+        keep_prompt=False,
+    )
+    obs = runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    assert obs.exit_code == 0
+    assert dataset_name in obs.content
+
+    logger.info(f"{'-' * 50} END Runtime Initialization Fn {'-' * 50}")
+
+
+def complete_runtime(
+    runtime: Runtime,
+    instance: pd.Series,
+) -> 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.
+    """
+    logger.info(f"{'-' * 50} BEGIN Runtime Completion Fn {'-' * 50}")
+    obs: CmdOutputObservation
+
+    test_result = {}
+
+    action = CmdRunAction(command='cd /workspace')
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+
+    assert obs.exit_code == 0
+
+    action = CmdRunAction(
+        command=f'cat pred_programs/{instance.pred_program_name}',
+        keep_prompt=False,
+    )
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = runtime.run_action(action)
+
+    if obs.exit_code == 0:
+        test_result = {'program': obs.content}
+    else:
+        test_result = {'program': 'ERROR'}
+
+    logger.info(f"{'-' * 50} END Runtime Completion Fn {'-' * 50}")
+    return test_result
+
+
+def process_instance(
+    instance: pd.Series,
+    metadata: EvalMetadata,
+    reset_logger: bool = True,
+) -> EvalOutput:
+    instance_id = instance.instance_id.replace('/', '__')
+    config = get_config(metadata, instance_id)
+
+    # Set up 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}.')
+
+    instruction = f"""You are an expert Python programming assistant that helps scientist users to write high-quality code to solve their tasks.
+Given a user request, you are expected to write a complete program that accomplishes the requested task and save any outputs to `/workspace/pred_results/` in the correct format.
+
+Here's the user request you need to work on:
+{instance.task_inst}
+
+You can access the dataset at `{instance.dataset_path}`. Here is the directory structure of the dataset:
+```
+{instance.dataset_folder_tree}
+```
+Here are some helpful previews for the dataset file(s):
+{instance.dataset_preview}
+
+Please save your program as `/workspace/pred_programs/{instance.pred_program_name}`.
+Then, please run the program to check and fix any errors.
+Please do NOT run the program in the background.
+If the program uses some packages that are incompatible, please figure out alternative implementations and do NOT restart the environment.
+
+"""
+
+    runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
+    initialize_runtime(runtime, instance)
+
+    # 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.get(
+                metadata.agent_class
+            ),
+        )
+    )
+
+    # ======= Attempt to evaluate the agent's edits =======
+    test_result = complete_runtime(runtime, instance)
+
+    # 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.')
+    metrics = state.metrics.get() if state.metrics else None
+
+    # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
+    # for compatibility with the existing output format, we can remake the pairs here
+    # remove when it becomes unnecessary
+    histories = compatibility_for_eval_history_pairs(state.history)
+
+    # Save the output
+    output = EvalOutput(
+        instance_id=instance.instance_id,
+        instruction=instruction,
+        metadata=metadata,
+        history=histories,
+        metrics=metrics,
+        error=state.last_error if state and state.last_error else None,
+        test_result=test_result,
+    )
+    return output
+
+
+if __name__ == '__main__':
+    parser = get_parser()
+    parser.add_argument(
+        '--use_knowledge',
+        type=str,
+        default='false',
+        choices=['true', 'false'],
+        help='use expert-provided knowledge or not',
+    )
+    args, _ = parser.parse_known_args()
+
+    sab_dataset = load_dataset('osunlp/ScienceAgentBench', split='validation')
+
+    dataset_processed = []
+    for example in tqdm(sab_dataset):
+        dataset_processed.append(
+            format_task_dict(example, args.use_knowledge == 'true')
+        )
+
+    dataset = pd.DataFrame(dataset_processed)
+
+    llm_config = None
+    if args.llm_config:
+        llm_config = get_llm_config_arg(args.llm_config)
+    if llm_config is None:
+        raise ValueError(f'Could not find LLM config: --llm_config {args.llm_config}')
+
+    metadata = make_metadata(
+        llm_config,
+        'ScienceAgentBench',
+        args.agent_cls,
+        args.max_iterations,
+        args.eval_note,
+        args.eval_output_dir,
+    )
+    output_file = os.path.join(metadata.eval_output_dir, 'output.jsonl')
+    dataset['instance_id'] = dataset['instance_id'].apply(str)
+    instances = prepare_dataset(dataset, output_file, args.eval_n_limit)
+
+    run_evaluation(
+        instances, metadata, output_file, args.eval_num_workers, process_instance
+    )

evaluation/scienceagentbench/scripts/run_infer.sh

@@ -0,0 +1,49 @@
+#!/bin/bash
+set -eo pipefail
+
+source "evaluation/utils/version_control.sh"
+
+MODEL_CONFIG=$1
+COMMIT_HASH=$2
+USE_KNOWLEDGE=$3
+AGENT=$4
+EVAL_LIMIT=$5
+NUM_WORKERS=$6
+
+if [ -z "$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 "$USE_KNOWLEDGE" ]; then
+  echo "Use knowledge not specified, use default False"
+  USE_KNOWLEDGE=false
+fi
+
+get_agent_version
+
+echo "AGENT: $AGENT"
+echo "AGENT_VERSION: $AGENT_VERSION"
+echo "MODEL_CONFIG: $MODEL_CONFIG"
+
+COMMAND="poetry run python evaluation/scienceagentbench/run_infer.py \
+  --agent-cls $AGENT \
+  --llm-config $MODEL_CONFIG \
+  --use_knowledge $USE_KNOWLEDGE \
+  --max-iterations 30 \
+  --eval-num-workers $NUM_WORKERS \
+  --eval-note $AGENT_VERSION" \
+
+if [ -n "$EVAL_LIMIT" ]; then
+  echo "EVAL_LIMIT: $EVAL_LIMIT"
+  COMMAND="$COMMAND --eval-n-limit $EVAL_LIMIT"
+fi
+
+# Run the command
+eval $COMMAND

evaluation/swe_bench/eval_infer.py

@@ -83,6 +83,7 @@ def get_config(instance: pd.Series) -> AppConfig:
             timeout=1800,
             api_key=os.environ.get('ALLHANDS_API_KEY', None),
             remote_runtime_api_url=os.environ.get('SANDBOX_REMOTE_RUNTIME_API_URL'),
+            remote_runtime_init_timeout=1800,
         ),
         # do not mount workspace
         workspace_base=None,
@@ -239,7 +240,7 @@ def process_instance(
                         # Create a directory structure that matches the expected format
                         # NOTE: this is a hack to make the eval report format consistent
                         # with the original SWE-Bench eval script
-                        log_dir = os.path.join(temp_dir, 'logs', instance_id)
+                        log_dir = os.path.join(temp_dir, 'logs', instance_id.lower())
                         os.makedirs(log_dir, exist_ok=True)
                         test_output_path = os.path.join(log_dir, 'test_output.txt')
                         with open(test_output_path, 'w') as f:

evaluation/swe_bench/run_infer.py

@@ -20,6 +20,7 @@ from evaluation.utils.shared import (
     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 (
@@ -40,6 +41,7 @@ from openhands.utils.async_utils import call_async_from_sync
 
 USE_HINT_TEXT = os.environ.get('USE_HINT_TEXT', 'false').lower() == 'true'
 USE_INSTANCE_IMAGE = os.environ.get('USE_INSTANCE_IMAGE', 'false').lower() == 'true'
+RUN_WITH_BROWSING = os.environ.get('RUN_WITH_BROWSING', 'false').lower() == 'true'
 
 AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
     'CodeActAgent': codeact_user_response,
@@ -79,7 +81,7 @@ def get_instruction(instance: pd.Series, metadata: EvalMetadata):
             '</pr_description>\n\n'
             'Can you help me implement the necessary changes to the repository so that the requirements specified in the <pr_description> are met?\n'
             "I've already taken care of all changes to any of the test files described in the <pr_description>. This means you DON'T have to modify the testing logic or any of the tests in any way!\n"
-            'Your task is to make the minimal changes to non-tests files in the /repo directory to ensure the <pr_description> is satisfied.\n'
+            'Your task is to make the minimal changes to non-tests files in the /workspace directory to ensure the <pr_description> is satisfied.\n'
             'Follow these steps to resolve the issue:\n'
             '1. As a first step, it might be a good idea to explore the repo to familiarize yourself with its structure.\n'
             '2. Create a script to reproduce the error and execute it with `python <filename.py>` using the BashTool, to confirm the error\n'
@@ -88,6 +90,13 @@ def get_instruction(instance: pd.Series, metadata: EvalMetadata):
             '5. Think about edgecases and make sure your fix handles them as well\n'
             "Your thinking should be thorough and so it's fine if it's very long.\n"
         )
+
+    if RUN_WITH_BROWSING:
+        instruction += (
+            '<IMPORTANT!>\n'
+            'You SHOULD NEVER attempt to browse the web. '
+            '</IMPORTANT!>\n'
+        )
     return instruction
 
 
@@ -101,7 +110,7 @@ def get_instance_docker_image(instance_id: str) -> str:
     image_name = image_name.replace(
         '__', '_s_'
     )  # to comply with docker image naming convention
-    return DOCKER_IMAGE_PREFIX.rstrip('/') + '/' + image_name
+    return (DOCKER_IMAGE_PREFIX.rstrip('/') + '/' + image_name).lower()
 
 
 def get_config(
@@ -137,23 +146,20 @@ def get_config(
             api_key=os.environ.get('ALLHANDS_API_KEY', None),
             remote_runtime_api_url=os.environ.get('SANDBOX_REMOTE_RUNTIME_API_URL'),
             keep_remote_runtime_alive=False,
+            remote_runtime_init_timeout=1800,
         ),
         # do not mount workspace
         workspace_base=None,
         workspace_mount_path=None,
     )
-    if metadata.llm_config.log_completions:
-        metadata.llm_config.log_completions_folder = os.path.join(
-            metadata.eval_output_dir, 'llm_completions', instance['instance_id']
+    config.set_llm_config(
+        update_llm_config_for_completions_logging(
+            metadata.llm_config, metadata.eval_output_dir, instance['instance_id']
         )
-        logger.info(
-            f'Logging LLM completions for instance {instance["instance_id"]} to '
-            f'{metadata.llm_config.log_completions_folder}'
-        )
-    config.set_llm_config(metadata.llm_config)
+    )
     agent_config = AgentConfig(
         codeact_enable_jupyter=False,
-        codeact_enable_browsing_delegate=False,
+        codeact_enable_browsing=RUN_WITH_BROWSING,
         codeact_enable_llm_editor=False,
     )
     config.set_agent_config(agent_config)
@@ -438,7 +444,8 @@ def process_instance(
     if state is None:
         raise ValueError('State should not be None.')
 
-    histories = [event_to_dict(event) for event in state.history.get_events()]
+    # NOTE: this is NO LONGER the event stream, but an agent history that includes delegate agent's events
+    histories = [event_to_dict(event) for event in state.history]
     metrics = state.metrics.get() if state.metrics else None
 
     # Save the output

evaluation/swe_bench/scripts/run_infer.sh

@@ -34,6 +34,11 @@ if [ -z "$USE_INSTANCE_IMAGE" ]; then
   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"
@@ -47,6 +52,8 @@ 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_agent_version
 
@@ -67,6 +74,10 @@ 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

evaluation/toolqa/run_infer.py

@@ -9,6 +9,7 @@ from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
     codeact_user_response,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -126,7 +127,8 @@ def process_instance(instance: Any, metadata: EvalMetadata, reset_logger: bool =
         raise ValueError('State should not be None.')
 
     # retrieve the last message from the agent
-    model_answer_raw = state.history.get_last_agent_message()
+    last_agent_message = state.get_last_agent_message()
+    model_answer_raw = last_agent_message.content if last_agent_message else ''
 
     # attempt to parse model_answer
     correct = eval_answer(str(model_answer_raw), str(answer))
@@ -137,7 +139,7 @@ def process_instance(instance: Any, metadata: EvalMetadata, reset_logger: bool =
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     # Save the output
     output = EvalOutput(

evaluation/utils/shared.py

@@ -18,6 +18,9 @@ from openhands.core.logger import get_console_handler
 from openhands.core.logger import openhands_logger as logger
 from openhands.events.action import Action
 from openhands.events.action.message import MessageAction
+from openhands.events.event import Event
+from openhands.events.serialization.event import event_to_dict
+from openhands.events.utils import get_pairs_from_events
 
 
 class EvalMetadata(BaseModel):
@@ -112,7 +115,14 @@ def codeact_user_response(
     if state.history:
         # check if the last action has an answer, if so, early exit
         if try_parse is not None:
-            last_action = state.history.get_last_action()
+            last_action = next(
+                (
+                    event
+                    for event in reversed(state.history)
+                    if isinstance(event, Action)
+                ),
+                None,
+            )
             ans = try_parse(last_action)
             if ans is not None:
                 return '/exit'
@@ -120,7 +130,7 @@ def codeact_user_response(
         # check if the agent has tried to talk to the user 3 times, if so, let the agent know it can give up
         user_msgs = [
             event
-            for event in state.history.get_events()
+            for event in state.history
             if isinstance(event, MessageAction) and event.source == 'user'
         ]
         if len(user_msgs) >= 2:
@@ -411,3 +421,35 @@ def reset_logger_for_multiprocessing(
     )
     file_handler.setLevel(logging.INFO)
     logger.addHandler(file_handler)
+
+
+def update_llm_config_for_completions_logging(
+    llm_config: LLMConfig,
+    eval_output_dir: str,
+    instance_id: str,
+) -> LLMConfig:
+    """Update the LLM config for logging completions."""
+    if llm_config.log_completions:
+        llm_config.log_completions_folder = os.path.join(
+            eval_output_dir, 'llm_completions', instance_id
+        )
+        logger.info(
+            f'Logging LLM completions for instance {instance_id} to '
+            f'{llm_config.log_completions_folder}'
+        )
+    return llm_config
+
+
+# history is now available as a filtered stream of events, rather than list of pairs of (Action, Observation)
+# we rebuild the pairs here
+# for compatibility with the existing output format in evaluations
+# remove this when it's no longer necessary
+def compatibility_for_eval_history_pairs(
+    history: list[Event],
+) -> list[tuple[dict, dict]]:
+    history_pairs = []
+
+    for action, observation in get_pairs_from_events(history):
+        history_pairs.append((event_to_dict(action), event_to_dict(observation)))
+
+    return history_pairs

evaluation/webarena/run_infer.py

@@ -10,6 +10,7 @@ import pandas as pd
 from evaluation.utils.shared import (
     EvalMetadata,
     EvalOutput,
+    compatibility_for_eval_history_pairs,
     make_metadata,
     prepare_dataset,
     reset_logger_for_multiprocessing,
@@ -166,7 +167,7 @@ def process_instance(
 
     # Instruction is the first message from the USER
     instruction = ''
-    for event in state.history.get_events():
+    for event in state.history:
         if isinstance(event, MessageAction):
             instruction = event.content
             break
@@ -178,7 +179,7 @@ def process_instance(
     # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
     # for compatibility with the existing output format, we can remake the pairs here
     # remove when it becomes unnecessary
-    histories = state.history.compatibility_for_eval_history_pairs()
+    histories = compatibility_for_eval_history_pairs(state.history)
 
     # Save the output
     output = EvalOutput(

frontend/.eslintrc

@@ -84,4 +84,4 @@
       }
     }
   ]
-}
+}

frontend/.gitignore

@@ -1,4 +1,9 @@
 # i18n translation files make by script using `make build`
 public/locales/**/*
 src/i18n/declaration.ts
-.env
+.env
+node_modules/
+/test-results/
+/playwright-report/
+/blob-report/
+/playwright/.cache/

frontend/__tests__/components/chat/chat-input.test.tsx

@@ -1,5 +1,5 @@
 import userEvent from "@testing-library/user-event";
-import { render, screen } from "@testing-library/react";
+import { fireEvent, render, screen } from "@testing-library/react";
 import { describe, afterEach, vi, it, expect } from "vitest";
 import { ChatInput } from "#/components/chat-input";
 
@@ -158,4 +158,46 @@ describe("ChatInput", () => {
     await user.tab();
     expect(onBlurMock).toHaveBeenCalledOnce();
   });
+
+  it("should handle text paste correctly", () => {
+    const onSubmit = vi.fn();
+    const onChange = vi.fn();
+
+    render(<ChatInput onSubmit={onSubmit} onChange={onChange} />);
+
+    const input = screen.getByTestId("chat-input").querySelector("textarea");
+    expect(input).toBeTruthy();
+
+    // Fire paste event with text data
+    fireEvent.paste(input!, {
+      clipboardData: {
+        getData: (type: string) => type === 'text/plain' ? 'test paste' : '',
+        files: []
+      }
+    });
+  });
+
+  it("should handle image paste correctly", () => {
+    const onSubmit = vi.fn();
+    const onImagePaste = vi.fn();
+
+    render(<ChatInput onSubmit={onSubmit} onImagePaste={onImagePaste} />);
+
+    const input = screen.getByTestId("chat-input").querySelector("textarea");
+    expect(input).toBeTruthy();
+
+    // Create a paste event with an image file
+    const file = new File(["dummy content"], "image.png", { type: "image/png" });
+
+    // Fire paste event with image data
+    fireEvent.paste(input!, {
+      clipboardData: {
+        getData: () => '',
+        files: [file]
+      }
+    });
+
+    // Verify image paste was handled
+    expect(onImagePaste).toHaveBeenCalledWith([file]);
+  });
 });

frontend/__tests__/components/chat/chat-interface.test.tsx

@@ -1,19 +1,156 @@
-import { afterEach, describe, expect, it, vi } from "vitest";
-import { render, screen, within } from "@testing-library/react";
+import { afterEach, beforeAll, describe, expect, it, vi } from "vitest";
+import { act, screen, waitFor, within } from "@testing-library/react";
 import userEvent from "@testing-library/user-event";
+import { renderWithProviders } from "test-utils";
 import { ChatInterface } from "#/components/chat-interface";
-import { SocketProvider } from "#/context/socket";
+import { addUserMessage } from "#/state/chatSlice";
+import { SUGGESTIONS } from "#/utils/suggestions";
+import * as ChatSlice from "#/state/chatSlice";
 
 // eslint-disable-next-line @typescript-eslint/no-unused-vars
 const renderChatInterface = (messages: (Message | ErrorMessage)[]) =>
-  render(<ChatInterface />, { wrapper: SocketProvider });
+  renderWithProviders(<ChatInterface />);
+
+describe("Empty state", () => {
+  const { send: sendMock } = vi.hoisted(() => ({
+    send: vi.fn(),
+  }));
+
+  const { useWsClient: useWsClientMock } = vi.hoisted(() => ({
+    useWsClient: vi.fn(() => ({ send: sendMock, runtimeActive: true })),
+  }));
+
+  beforeAll(() => {
+    vi.mock("#/context/socket", async (importActual) => ({
+      ...(await importActual<typeof import("#/context/ws-client-provider")>()),
+      useWsClient: useWsClientMock,
+    }));
+  });
 
-describe.skip("ChatInterface", () => {
   afterEach(() => {
     vi.clearAllMocks();
   });
 
-  it.todo("should render suggestions if empty");
+  it("should render suggestions if empty", () => {
+    const { store } = renderWithProviders(<ChatInterface />, {
+      preloadedState: {
+        chat: { messages: [] },
+      },
+    });
+
+    expect(screen.getByTestId("suggestions")).toBeInTheDocument();
+
+    act(() => {
+      store.dispatch(
+        addUserMessage({
+          content: "Hello",
+          imageUrls: [],
+          timestamp: new Date().toISOString(),
+        }),
+      );
+    });
+
+    expect(screen.queryByTestId("suggestions")).not.toBeInTheDocument();
+  });
+
+  it("should render the default suggestions", () => {
+    renderWithProviders(<ChatInterface />, {
+      preloadedState: {
+        chat: { messages: [] },
+      },
+    });
+
+    const suggestions = screen.getByTestId("suggestions");
+    const repoSuggestions = Object.keys(SUGGESTIONS.repo);
+
+    // check that there are at most 4 suggestions displayed
+    const displayedSuggestions = within(suggestions).getAllByRole("button");
+    expect(displayedSuggestions.length).toBeLessThanOrEqual(4);
+
+    // Check that each displayed suggestion is one of the repo suggestions
+    displayedSuggestions.forEach((suggestion) => {
+      expect(repoSuggestions).toContain(suggestion.textContent);
+    });
+  });
+
+  it.fails(
+    "should load the a user message to the input when selecting",
+    async () => {
+      // this is to test that the message is in the UI before the socket is called
+      useWsClientMock.mockImplementation(() => ({
+        send: sendMock,
+        runtimeActive: false, // mock an inactive runtime setup
+      }));
+      const addUserMessageSpy = vi.spyOn(ChatSlice, "addUserMessage");
+      const user = userEvent.setup();
+      const { store } = renderWithProviders(<ChatInterface />, {
+        preloadedState: {
+          chat: { messages: [] },
+        },
+      });
+
+      const suggestions = screen.getByTestId("suggestions");
+      const displayedSuggestions = within(suggestions).getAllByRole("button");
+      const input = screen.getByTestId("chat-input");
+
+      await user.click(displayedSuggestions[0]);
+
+      // user message loaded to input
+      expect(addUserMessageSpy).not.toHaveBeenCalled();
+      expect(screen.queryByTestId("suggestions")).toBeInTheDocument();
+      expect(store.getState().chat.messages).toHaveLength(0);
+      expect(input).toHaveValue(displayedSuggestions[0].textContent);
+    },
+  );
+
+  it.fails(
+    "should send the message to the socket only if the runtime is active",
+    async () => {
+      useWsClientMock.mockImplementation(() => ({
+        send: sendMock,
+        runtimeActive: false, // mock an inactive runtime setup
+      }));
+      const user = userEvent.setup();
+      const { rerender } = renderWithProviders(<ChatInterface />, {
+        preloadedState: {
+          chat: { messages: [] },
+        },
+      });
+
+      const suggestions = screen.getByTestId("suggestions");
+      const displayedSuggestions = within(suggestions).getAllByRole("button");
+
+      await user.click(displayedSuggestions[0]);
+      expect(sendMock).not.toHaveBeenCalled();
+
+      useWsClientMock.mockImplementation(() => ({
+        send: sendMock,
+        runtimeActive: true, // mock an active runtime setup
+      }));
+      rerender(<ChatInterface />);
+
+      await waitFor(() =>
+        expect(sendMock).toHaveBeenCalledWith(expect.any(String)),
+      );
+    },
+  );
+});
+
+describe.skip("ChatInterface", () => {
+  beforeAll(() => {
+    // mock useScrollToBottom hook
+    vi.mock("#/hooks/useScrollToBottom", () => ({
+      useScrollToBottom: vi.fn(() => ({
+        scrollDomToBottom: vi.fn(),
+        onChatBodyScroll: vi.fn(),
+        hitBottom: vi.fn(),
+      })),
+    }));
+  });
+
+  afterEach(() => {
+    vi.clearAllMocks();
+  });
 
   it("should render messages", () => {
     const messages: Message[] = [
@@ -128,14 +265,14 @@ describe.skip("ChatInterface", () => {
         timestamp: new Date().toISOString(),
       },
       {
-        error: "Woops!",
+        error: true,
+        id: "",
         message: "Something went wrong",
       },
     ];
     renderChatInterface(messages);
 
     const error = screen.getByTestId("error-message");
-    expect(within(error).getByText("Woops!")).toBeInTheDocument();
     expect(within(error).getByText("Something went wrong")).toBeInTheDocument();
   });
 

frontend/__tests__/components/feedback-form.test.tsx

@@ -5,7 +5,6 @@ import { FeedbackForm } from "#/components/feedback-form";
 
 describe("FeedbackForm", () => {
   const user = userEvent.setup();
-  const onSubmitMock = vi.fn();
   const onCloseMock = vi.fn();
 
   afterEach(() => {
@@ -13,7 +12,7 @@ describe("FeedbackForm", () => {
   });
 
   it("should render correctly", () => {
-    render(<FeedbackForm onSubmit={onSubmitMock} onClose={onCloseMock} />);
+    render(<FeedbackForm polarity="positive" onClose={onCloseMock} />);
 
     screen.getByLabelText("Email");
     screen.getByLabelText("Private");
@@ -24,7 +23,7 @@ describe("FeedbackForm", () => {
   });
 
   it("should switch between private and public permissions", async () => {
-    render(<FeedbackForm onSubmit={onSubmitMock} onClose={onCloseMock} />);
+    render(<FeedbackForm polarity="positive" onClose={onCloseMock} />);
     const privateRadio = screen.getByLabelText("Private");
     const publicRadio = screen.getByLabelText("Public");
 
@@ -40,69 +39,11 @@ describe("FeedbackForm", () => {
     expect(publicRadio).not.toBeChecked();
   });
 
-  it("should call onSubmit when the form is submitted", async () => {
-    render(<FeedbackForm onSubmit={onSubmitMock} onClose={onCloseMock} />);
-    const email = screen.getByLabelText("Email");
-
-    await user.type(email, "test@test.test");
-    await user.click(screen.getByRole("button", { name: "Submit" }));
-
-    expect(onSubmitMock).toHaveBeenCalledWith("private", "test@test.test"); // private is the default value
-  });
-
-  it("should not call onSubmit when the email is invalid", async () => {
-    render(<FeedbackForm onSubmit={onSubmitMock} onClose={onCloseMock} />);
-    const email = screen.getByLabelText("Email");
-    const submitButton = screen.getByRole("button", { name: "Submit" });
-
-    await user.click(submitButton);
-
-    expect(onSubmitMock).not.toHaveBeenCalled();
-
-    await user.type(email, "test");
-    await user.click(submitButton);
-
-    expect(onSubmitMock).not.toHaveBeenCalled();
-  });
-
-  it("should submit public permissions when the public radio is checked", async () => {
-    render(<FeedbackForm onSubmit={onSubmitMock} onClose={onCloseMock} />);
-    const email = screen.getByLabelText("Email");
-    const publicRadio = screen.getByLabelText("Public");
-
-    await user.type(email, "test@test.test");
-    await user.click(publicRadio);
-    await user.click(screen.getByRole("button", { name: "Submit" }));
-
-    expect(onSubmitMock).toHaveBeenCalledWith("public", "test@test.test");
-  });
-
   it("should call onClose when the close button is clicked", async () => {
-    render(<FeedbackForm onSubmit={onSubmitMock} onClose={onCloseMock} />);
+    render(<FeedbackForm polarity="positive" onClose={onCloseMock} />);
     await user.click(screen.getByRole("button", { name: "Cancel" }));
 
-    expect(onSubmitMock).not.toHaveBeenCalled();
     expect(onCloseMock).toHaveBeenCalled();
   });
 
-  it("should disable the buttons if isSubmitting is true", () => {
-    const { rerender } = render(
-      <FeedbackForm onSubmit={onSubmitMock} onClose={onCloseMock} />,
-    );
-    const submitButton = screen.getByRole("button", { name: "Submit" });
-    const cancelButton = screen.getByRole("button", { name: "Cancel" });
-
-    expect(submitButton).not.toBeDisabled();
-    expect(cancelButton).not.toBeDisabled();
-
-    rerender(
-      <FeedbackForm
-        onSubmit={onSubmitMock}
-        onClose={onCloseMock}
-        isSubmitting
-      />,
-    );
-    expect(submitButton).toBeDisabled();
-    expect(cancelButton).toBeDisabled();
-  });
 });

frontend/__tests__/components/file-explorer/FileExplorer.test.tsx

@@ -16,13 +16,16 @@ vi.mock("../../services/fileService", async () => ({
 }));
 
 const renderFileExplorerWithRunningAgentState = () =>
-  renderWithProviders(<FileExplorer error={null} />, {
-    preloadedState: {
-      agent: {
-        curAgentState: AgentState.RUNNING,
+  renderWithProviders(
+    <FileExplorer error={null} isOpen onToggle={() => {}} />,
+    {
+      preloadedState: {
+        agent: {
+          curAgentState: AgentState.RUNNING,
+        },
       },
     },
-  });
+  );
 
 describe.skip("FileExplorer", () => {
   afterEach(() => {

frontend/__tests__/components/interactive-chat-box.test.tsx

@@ -25,6 +25,21 @@ describe("InteractiveChatBox", () => {
     within(chatBox).getByTestId("upload-image-input");
   });
 
+  it.fails("should set custom values", () => {
+    render(
+      <InteractiveChatBox
+        onSubmit={onSubmitMock}
+        onStop={onStopMock}
+        value="Hello, world!"
+      />,
+    );
+
+    const chatBox = screen.getByTestId("interactive-chat-box");
+    const chatInput = within(chatBox).getByTestId("chat-input");
+
+    expect(chatInput).toHaveValue("Hello, world!");
+  });
+
   it("should display the image previews when images are uploaded", async () => {
     const user = userEvent.setup();
     render(<InteractiveChatBox onSubmit={onSubmitMock} onStop={onStopMock} />);

frontend/__tests__/components/suggestion-item.test.tsx

@@ -0,0 +1,30 @@
+import { render, screen } from "@testing-library/react";
+import userEvent from "@testing-library/user-event";
+import { afterEach, describe, expect, it, vi } from "vitest";
+import { SuggestionItem } from "#/components/suggestion-item";
+
+describe("SuggestionItem", () => {
+  const suggestionItem = { label: "suggestion1", value: "a long text value" };
+  const onClick = vi.fn();
+
+  afterEach(() => {
+    vi.clearAllMocks();
+  });
+
+  it("should render a suggestion", () => {
+    render(<SuggestionItem suggestion={suggestionItem} onClick={onClick} />);
+
+    expect(screen.getByTestId("suggestion")).toBeInTheDocument();
+    expect(screen.getByText(/suggestion1/i)).toBeInTheDocument();
+  });
+
+  it("should call onClick when clicking a suggestion", async () => {
+    const user = userEvent.setup();
+    render(<SuggestionItem suggestion={suggestionItem} onClick={onClick} />);
+
+    const suggestion = screen.getByTestId("suggestion");
+    await user.click(suggestion);
+
+    expect(onClick).toHaveBeenCalledWith("a long text value");
+  });
+});

frontend/__tests__/components/suggestions.test.tsx

@@ -0,0 +1,60 @@
+import { render, screen } from "@testing-library/react";
+import userEvent from "@testing-library/user-event";
+import { afterEach, describe, expect, it, vi } from "vitest";
+import { Suggestions } from "#/components/suggestions";
+
+describe("Suggestions", () => {
+  const firstSuggestion = {
+    label: "first-suggestion",
+    value: "value-of-first-suggestion",
+  };
+  const secondSuggestion = {
+    label: "second-suggestion",
+    value: "value-of-second-suggestion",
+  };
+  const suggestions = [firstSuggestion, secondSuggestion];
+
+  const onSuggestionClickMock = vi.fn();
+
+  afterEach(() => {
+    vi.clearAllMocks();
+  });
+
+  it("should render suggestions", () => {
+    render(
+      <Suggestions
+        suggestions={suggestions}
+        onSuggestionClick={onSuggestionClickMock}
+      />,
+    );
+
+    expect(screen.getByTestId("suggestions")).toBeInTheDocument();
+    const suggestionElements = screen.getAllByTestId("suggestion");
+
+    expect(suggestionElements).toHaveLength(2);
+    expect(suggestionElements[0]).toHaveTextContent("first-suggestion");
+    expect(suggestionElements[1]).toHaveTextContent("second-suggestion");
+  });
+
+  it("should call onSuggestionClick when clicking a suggestion", async () => {
+    const user = userEvent.setup();
+    render(
+      <Suggestions
+        suggestions={suggestions}
+        onSuggestionClick={onSuggestionClickMock}
+      />,
+    );
+
+    const suggestionElements = screen.getAllByTestId("suggestion");
+
+    await user.click(suggestionElements[0]);
+    expect(onSuggestionClickMock).toHaveBeenCalledWith(
+      "value-of-first-suggestion",
+    );
+
+    await user.click(suggestionElements[1]);
+    expect(onSuggestionClickMock).toHaveBeenCalledWith(
+      "value-of-second-suggestion",
+    );
+  });
+});

frontend/__tests__/hooks/use-terminal.test.tsx

@@ -2,8 +2,9 @@ import { beforeAll, describe, expect, it, vi } from "vitest";
 import { render } from "@testing-library/react";
 import { afterEach } from "node:test";
 import { useTerminal } from "#/hooks/useTerminal";
-import { SocketProvider } from "#/context/socket";
 import { Command } from "#/state/commandSlice";
+import { WsClientProvider } from "#/context/ws-client-provider";
+import { ReactNode } from "react";
 
 interface TestTerminalComponentProps {
   commands: Command[];
@@ -18,6 +19,17 @@ function TestTerminalComponent({
   return <div ref={ref} />;
 }
 
+interface WrapperProps {
+  children: ReactNode;
+}
+
+
+function Wrapper({children}: WrapperProps) {
+  return (
+    <WsClientProvider enabled={true} token="NO_JWT" ghToken="NO_GITHUB" settings={null}>{children}</WsClientProvider>
+  )
+}
+
 describe("useTerminal", () => {
   const mockTerminal = vi.hoisted(() => ({
     loadAddon: vi.fn(),
@@ -50,7 +62,7 @@ describe("useTerminal", () => {
 
   it("should render", () => {
     render(<TestTerminalComponent commands={[]} secrets={[]} />, {
-      wrapper: SocketProvider,
+      wrapper: Wrapper,
     });
   });
 
@@ -61,7 +73,7 @@ describe("useTerminal", () => {
     ];
 
     render(<TestTerminalComponent commands={commands} secrets={[]} />, {
-      wrapper: SocketProvider,
+      wrapper: Wrapper,
     });
 
     expect(mockTerminal.writeln).toHaveBeenNthCalledWith(1, "echo hello");
@@ -85,7 +97,7 @@ describe("useTerminal", () => {
         secrets={[secret, anotherSecret]}
       />,
       {
-        wrapper: SocketProvider,
+        wrapper: Wrapper,
       },
     );
 

frontend/__tests__/initial-query.test.tsx

@@ -0,0 +1,17 @@
+import { describe, it, expect } from "vitest";
+import store from "../src/store";
+import { setInitialQuery, clearInitialQuery } from "../src/state/initial-query-slice";
+
+describe("Initial Query Behavior", () => {
+  it("should clear initial query when clearInitialQuery is dispatched", () => {
+    // Set up initial query in the store
+    store.dispatch(setInitialQuery("test query"));
+    expect(store.getState().initalQuery.initialQuery).toBe("test query");
+
+    // Clear the initial query
+    store.dispatch(clearInitialQuery());
+
+    // Verify initial query is cleared
+    expect(store.getState().initalQuery.initialQuery).toBeNull();
+  });
+});

frontend/__tests__/routes/_oh.app.test.tsx

@@ -0,0 +1,5 @@
+import { describe, it } from "vitest";
+
+describe("App", () => {
+  it.todo("should render");
+});

frontend/__tests__/utils/cache.test.ts

@@ -0,0 +1,53 @@
+import { afterEach } from "node:test";
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import { cache } from "#/utils/cache";
+
+describe("Cache", () => {
+  const testKey = "key";
+  const testData = { message: "Hello, world!" };
+  const testTTL = 1000; // 1 second
+
+  beforeEach(() => {
+    vi.useFakeTimers();
+  });
+
+  afterEach(() => {
+    vi.useRealTimers();
+  });
+
+  it("gets data from memory if not expired", () => {
+    cache.set(testKey, testData, testTTL);
+
+    expect(cache.get(testKey)).toEqual(testData);
+  });
+
+  it("should expire after 5 minutes by default", () => {
+    cache.set(testKey, testData);
+    expect(cache.get(testKey)).not.toBeNull();
+
+    vi.advanceTimersByTime(5 * 60 * 1000 + 1);
+
+    expect(cache.get(testKey)).toBeNull();
+  });
+
+  it("returns null if cached data is expired", () => {
+    cache.set(testKey, testData, testTTL);
+
+    vi.advanceTimersByTime(testTTL + 1);
+    expect(cache.get(testKey)).toBeNull();
+  });
+
+  it("deletes data from memory", () => {
+    cache.set(testKey, testData, testTTL);
+    cache.delete(testKey);
+    expect(cache.get(testKey)).toBeNull();
+  });
+
+  it("clears all data with the app prefix from memory", () => {
+    cache.set(testKey, testData, testTTL);
+    cache.set("anotherKey", { data: "More data" }, testTTL);
+    cache.clearAll();
+    expect(cache.get(testKey)).toBeNull();
+    expect(cache.get("anotherKey")).toBeNull();
+  });
+});

frontend/__tests__/utils/organizeModelsAndProviders.test.ts

@@ -15,7 +15,7 @@ test("organizeModelsAndProviders", () => {
     "gpt-4o",
     "together-ai-21.1b-41b",
     "gpt-4o-mini",
-    "claude-3-5-sonnet-20240620",
+    "anthropic/claude-3-5-sonnet-20241022",
     "claude-3-haiku-20240307",
     "claude-2",
     "claude-2.1",
@@ -51,7 +51,7 @@ test("organizeModelsAndProviders", () => {
     anthropic: {
       separator: "/",
       models: [
-        "claude-3-5-sonnet-20240620",
+        "claude-3-5-sonnet-20241022",
         "claude-3-haiku-20240307",
         "claude-2",
         "claude-2.1",

frontend/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "openhands-frontend",
-  "version": "0.11.0",
+  "version": "0.13.0",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "openhands-frontend",
-      "version": "0.11.0",
+      "version": "0.13.0",
       "dependencies": {
         "@monaco-editor/react": "^4.6.0",
         "@nextui-org/react": "^2.4.8",
@@ -26,6 +26,7 @@
         "isbot": "^5.1.17",
         "jose": "^5.9.4",
         "monaco-editor": "^0.52.0",
+        "posthog-js": "^1.176.0",
         "react": "^18.3.1",
         "react-dom": "^18.3.1",
         "react-highlight": "^0.15.0",
@@ -45,6 +46,7 @@
         "ws": "^8.18.0"
       },
       "devDependencies": {
+        "@playwright/test": "^1.48.2",
         "@remix-run/dev": "^2.11.2",
         "@remix-run/testing": "^2.11.2",
         "@tailwindcss/typography": "^0.5.15",
@@ -61,6 +63,7 @@
         "@typescript-eslint/parser": "^7.18.0",
         "@vitest/coverage-v8": "^1.6.0",
         "autoprefixer": "^10.4.20",
+        "cross-env": "^7.0.3",
         "eslint": "^8.57.0",
         "eslint-config-airbnb": "^19.0.4",
         "eslint-config-airbnb-typescript": "^18.0.0",
@@ -3378,6 +3381,21 @@
         "url": "https://opencollective.com/unts"
       }
     },
+    "node_modules/@playwright/test": {
+      "version": "1.48.2",
+      "resolved": "https://registry.npmjs.org/@playwright/test/-/test-1.48.2.tgz",
+      "integrity": "sha512-54w1xCWfXuax7dz4W2M9uw0gDyh+ti/0K/MxcCUxChFh37kkdxPdfZDw5QBbuPUJHr1CiHJ1hXgSs+GgeQc5Zw==",
+      "dev": true,
+      "dependencies": {
+        "playwright": "1.48.2"
+      },
+      "bin": {
+        "playwright": "cli.js"
+      },
+      "engines": {
+        "node": ">=18"
+      }
+    },
     "node_modules/@polka/url": {
       "version": "1.0.0-next.28",
       "resolved": "https://registry.npmjs.org/@polka/url/-/url-1.0.0-next.28.tgz",
@@ -7864,6 +7882,16 @@
         "node": ">=6.6.0"
       }
     },
+    "node_modules/core-js": {
+      "version": "3.38.1",
+      "resolved": "https://registry.npmjs.org/core-js/-/core-js-3.38.1.tgz",
+      "integrity": "sha512-OP35aUorbU3Zvlx7pjsFdu1rGNnD4pgw/CWoYzRY3t2EzoVT7shKHY1dlAy3f41cGIO7ZDPQimhGFTlEYkG/Hw==",
+      "hasInstallScript": true,
+      "funding": {
+        "type": "opencollective",
+        "url": "https://opencollective.com/core-js"
+      }
+    },
     "node_modules/core-util-is": {
       "version": "1.0.3",
       "resolved": "https://registry.npmjs.org/core-util-is/-/core-util-is-1.0.3.tgz",
@@ -7896,6 +7924,24 @@
         }
       }
     },
+    "node_modules/cross-env": {
+      "version": "7.0.3",
+      "resolved": "https://registry.npmjs.org/cross-env/-/cross-env-7.0.3.tgz",
+      "integrity": "sha512-+/HKd6EgcQCJGh2PSjZuUitQBQynKor4wrFbRg4DtAgS1aWO+gU52xpH7M9ScGgXSYmAVS9bIJ8EzuaGw0oNAw==",
+      "dev": true,
+      "dependencies": {
+        "cross-spawn": "^7.0.1"
+      },
+      "bin": {
+        "cross-env": "src/bin/cross-env.js",
+        "cross-env-shell": "src/bin/cross-env-shell.js"
+      },
+      "engines": {
+        "node": ">=10.14",
+        "npm": ">=6",
+        "yarn": ">=1"
+      }
+    },
     "node_modules/cross-fetch": {
       "version": "4.0.0",
       "resolved": "https://registry.npmjs.org/cross-fetch/-/cross-fetch-4.0.0.tgz",
@@ -9666,6 +9712,11 @@
         "url": "https://github.com/sponsors/wooorm"
       }
     },
+    "node_modules/fflate": {
+      "version": "0.4.8",
+      "resolved": "https://registry.npmjs.org/fflate/-/fflate-0.4.8.tgz",
+      "integrity": "sha512-FJqqoDBR00Mdj9ppamLa/Y7vxm+PRmNWA67N846RvsoYVMKB4q3y/de5PA7gUmRMYK/8CMz2GDZQmCRN1wBcWA=="
+    },
     "node_modules/file-entry-cache": {
       "version": "6.0.1",
       "resolved": "https://registry.npmjs.org/file-entry-cache/-/file-entry-cache-6.0.1.tgz",
@@ -19406,6 +19457,50 @@
         "pathe": "^1.1.2"
       }
     },
+    "node_modules/playwright": {
+      "version": "1.48.2",
+      "resolved": "https://registry.npmjs.org/playwright/-/playwright-1.48.2.tgz",
+      "integrity": "sha512-NjYvYgp4BPmiwfe31j4gHLa3J7bD2WiBz8Lk2RoSsmX38SVIARZ18VYjxLjAcDsAhA+F4iSEXTSGgjua0rrlgQ==",
+      "dev": true,
+      "dependencies": {
+        "playwright-core": "1.48.2"
+      },
+      "bin": {
+        "playwright": "cli.js"
+      },
+      "engines": {
+        "node": ">=18"
+      },
+      "optionalDependencies": {
+        "fsevents": "2.3.2"
+      }
+    },
+    "node_modules/playwright-core": {
+      "version": "1.48.2",
+      "resolved": "https://registry.npmjs.org/playwright-core/-/playwright-core-1.48.2.tgz",
+      "integrity": "sha512-sjjw+qrLFlriJo64du+EK0kJgZzoQPsabGF4lBvsid+3CNIZIYLgnMj9V6JY5VhM2Peh20DJWIVpVljLLnlawA==",
+      "dev": true,
+      "bin": {
+        "playwright-core": "cli.js"
+      },
+      "engines": {
+        "node": ">=18"
+      }
+    },
+    "node_modules/playwright/node_modules/fsevents": {
+      "version": "2.3.2",
+      "resolved": "https://registry.npmjs.org/fsevents/-/fsevents-2.3.2.tgz",
+      "integrity": "sha512-xiqMQR4xAeHTuB9uWm+fFRcIOgKBMiOBP+eXiyT7jsgVCq1bkVygt00oASowB7EdtpOHaaPgKt812P9ab+DDKA==",
+      "dev": true,
+      "hasInstallScript": true,
+      "optional": true,
+      "os": [
+        "darwin"
+      ],
+      "engines": {
+        "node": "^8.16.0 || ^10.6.0 || >=11.0.0"
+      }
+    },
     "node_modules/possible-typed-array-names": {
       "version": "1.0.0",
       "resolved": "https://registry.npmjs.org/possible-typed-array-names/-/possible-typed-array-names-1.0.0.tgz",
@@ -19653,6 +19748,31 @@
       "resolved": "https://registry.npmjs.org/postcss-value-parser/-/postcss-value-parser-4.2.0.tgz",
       "integrity": "sha512-1NNCs6uurfkVbeXG4S8JFT9t19m45ICnif8zWLd5oPSZ50QnwMfK+H3jv408d4jw/7Bttv5axS5IiHoLaVNHeQ=="
     },
+    "node_modules/posthog-js": {
+      "version": "1.176.0",
+      "resolved": "https://registry.npmjs.org/posthog-js/-/posthog-js-1.176.0.tgz",
+      "integrity": "sha512-T5XKNtRzp7q6CGb7Vc7wAI76rWap9fiuDUPxPsyPBPDkreKya91x9RIsSapAVFafwD1AEin1QMczCmt9Le9BWw==",
+      "dependencies": {
+        "core-js": "^3.38.1",
+        "fflate": "^0.4.8",
+        "preact": "^10.19.3",
+        "web-vitals": "^4.2.0"
+      }
+    },
+    "node_modules/posthog-js/node_modules/web-vitals": {
+      "version": "4.2.4",
+      "resolved": "https://registry.npmjs.org/web-vitals/-/web-vitals-4.2.4.tgz",
+      "integrity": "sha512-r4DIlprAGwJ7YM11VZp4R884m0Vmgr6EAKe3P+kO0PPj3Unqyvv59rczf6UiGcb9Z8QxZVcqKNwv/g0WNdWwsw=="
+    },
+    "node_modules/preact": {
+      "version": "10.24.3",
+      "resolved": "https://registry.npmjs.org/preact/-/preact-10.24.3.tgz",
+      "integrity": "sha512-Z2dPnBnMUfyQfSQ+GBdsGa16hz35YmLmtTLhM169uW944hYL6xzTYkJjC07j+Wosz733pMWx0fgON3JNw1jJQA==",
+      "funding": {
+        "type": "opencollective",
+        "url": "https://opencollective.com/preact"
+      }
+    },
     "node_modules/prelude-ls": {
       "version": "1.2.1",
       "resolved": "https://registry.npmjs.org/prelude-ls/-/prelude-ls-1.2.1.tgz",

frontend/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openhands-frontend",
-  "version": "0.11.0",
+  "version": "0.13.0",
   "private": true,
   "type": "module",
   "engines": {
@@ -25,6 +25,7 @@
     "isbot": "^5.1.17",
     "jose": "^5.9.4",
     "monaco-editor": "^0.52.0",
+    "posthog-js": "^1.176.0",
     "react": "^18.3.1",
     "react-dom": "^18.3.1",
     "react-highlight": "^0.15.0",
@@ -44,11 +45,12 @@
     "ws": "^8.18.0"
   },
   "scripts": {
-    "dev": "npm run make-i18n && VITE_MOCK_API=false remix vite:dev",
-    "dev:mock": "npm run make-i18n && VITE_MOCK_API=true remix vite:dev",
+    "dev": "npm run make-i18n && cross-env VITE_MOCK_API=false remix vite:dev",
+    "dev:mock": "npm run make-i18n && cross-env VITE_MOCK_API=true remix vite:dev",
     "build": "npm run make-i18n && tsc && remix vite:build",
     "start": "npx sirv-cli build/ --single",
     "test": "vitest run",
+    "test:e2e": "playwright test",
     "test:coverage": "npm run make-i18n && vitest run --coverage",
     "dev_wsl": "VITE_WATCH_USE_POLLING=true vite",
     "preview": "vite preview",
@@ -70,6 +72,7 @@
     ]
   },
   "devDependencies": {
+    "@playwright/test": "^1.48.2",
     "@remix-run/dev": "^2.11.2",
     "@remix-run/testing": "^2.11.2",
     "@tailwindcss/typography": "^0.5.15",
@@ -86,6 +89,7 @@
     "@typescript-eslint/parser": "^7.18.0",
     "@vitest/coverage-v8": "^1.6.0",
     "autoprefixer": "^10.4.20",
+    "cross-env": "^7.0.3",
     "eslint": "^8.57.0",
     "eslint-config-airbnb": "^19.0.4",
     "eslint-config-airbnb-typescript": "^18.0.0",

frontend/playwright.config.ts

@@ -0,0 +1,79 @@
+import { defineConfig, devices } from "@playwright/test";
+
+/**
+ * Read environment variables from file.
+ * https://github.com/motdotla/dotenv
+ */
+// import dotenv from 'dotenv';
+// import path from 'path';
+// dotenv.config({ path: path.resolve(__dirname, '.env') });
+
+/**
+ * See https://playwright.dev/docs/test-configuration.
+ */
+export default defineConfig({
+  testDir: "./tests",
+  /* Run tests in files in parallel */
+  fullyParallel: true,
+  /* Fail the build on CI if you accidentally left test.only in the source code. */
+  forbidOnly: !!process.env.CI,
+  /* Retry on CI only */
+  retries: process.env.CI ? 2 : 0,
+  /* Opt out of parallel tests on CI. */
+  workers: process.env.CI ? 1 : undefined,
+  /* Reporter to use. See https://playwright.dev/docs/test-reporters */
+  reporter: "html",
+  /* Shared settings for all the projects below. See https://playwright.dev/docs/api/class-testoptions. */
+  use: {
+    /* Base URL to use in actions like `await page.goto('/')`. */
+    baseURL: "http://127.0.0.1:3000",
+
+    /* Collect trace when retrying the failed test. See https://playwright.dev/docs/trace-viewer */
+    trace: "on-first-retry",
+  },
+
+  /* Configure projects for major browsers */
+  projects: [
+    {
+      name: "chromium",
+      use: { ...devices["Desktop Chrome"] },
+    },
+
+    {
+      name: "firefox",
+      use: { ...devices["Desktop Firefox"] },
+    },
+
+    {
+      name: "webkit",
+      use: { ...devices["Desktop Safari"] },
+    },
+
+    /* Test against mobile viewports. */
+    // {
+    //   name: 'Mobile Chrome',
+    //   use: { ...devices['Pixel 5'] },
+    // },
+    // {
+    //   name: 'Mobile Safari',
+    //   use: { ...devices['iPhone 12'] },
+    // },
+
+    /* Test against branded browsers. */
+    // {
+    //   name: 'Microsoft Edge',
+    //   use: { ...devices['Desktop Edge'], channel: 'msedge' },
+    // },
+    // {
+    //   name: 'Google Chrome',
+    //   use: { ...devices['Desktop Chrome'], channel: 'chrome' },
+    // },
+  ],
+
+  /* Run your local dev server before starting the tests */
+  webServer: {
+    command: "npm run dev:mock -- --port 3000",
+    url: "http://127.0.0.1:3000",
+    reuseExistingServer: !process.env.CI,
+  },
+});

frontend/src/api/github.ts

@@ -122,6 +122,9 @@ export const retrieveGitHubUser = async (
       id: data.id,
       login: data.login,
       avatar_url: data.avatar_url,
+      company: data.company,
+      name: data.name,
+      email: data.email,
     };
 
     return user;
@@ -136,33 +139,6 @@ export const retrieveGitHubUser = async (
   return error;
 };
 
-/**
- * Given a GitHub token and a repository name, creates a repository for the authenticated user
- * @param token The GitHub token
- * @param repositoryName Name of the repository to create
- * @param description Description of the repository
- * @param isPrivate Boolean indicating if the repository should be private
- * @returns The created repository or an error response
- */
-export const createGitHubRepository = async (
-  token: string,
-  repositoryName: string,
-  description?: string,
-  isPrivate = true,
-): Promise<GitHubRepository | GitHubErrorReponse> => {
-  const response = await fetch("https://api.github.com/user/repos", {
-    method: "POST",
-    headers: generateGitHubAPIHeaders(token),
-    body: JSON.stringify({
-      name: repositoryName,
-      description,
-      private: isPrivate,
-    }),
-  });
-
-  return response.json();
-};
-
 export const retrieveLatestGitHubCommit = async (
   token: string,
   repository: string,

frontend/src/api/open-hands.ts

@@ -1,4 +1,5 @@
-import { getValidFallbackHost } from "#/utils/get-valid-fallback-host";
+import { request } from "#/services/api";
+import { cache } from "#/utils/cache";
 import {
   SaveFileSuccessResponse,
   FileUploadSuccessResponse,
@@ -9,36 +10,19 @@ import {
   GetConfigResponse,
 } from "./open-hands.types";
 
-/**
- * Generate the base URL of the OpenHands API
- * @returns Base URL of the OpenHands API
- */
-const generateBaseURL = () => {
-  const fallback = getValidFallbackHost();
-  const baseUrl = import.meta.env.VITE_BACKEND_BASE_URL || fallback;
-
-  if (typeof window === "undefined") {
-    return `http://${baseUrl}`;
-  }
-  return `${window.location.protocol}//${baseUrl}`;
-};
-
-/**
- * Class to interact with the OpenHands API
- */
 class OpenHands {
-  /**
-   * Base URL of the OpenHands API
-   */
-  static BASE_URL = generateBaseURL();
-
   /**
    * Retrieve the list of models available
    * @returns List of models available
    */
   static async getModels(): Promise<string[]> {
-    const response = await fetch(`${OpenHands.BASE_URL}/api/options/models`);
-    return response.json();
+    const cachedData = cache.get<string[]>("models");
+    if (cachedData) return cachedData;
+
+    const data = await request("/api/options/models");
+    cache.set("models", data);
+
+    return data;
   }
 
   /**
@@ -46,8 +30,13 @@ class OpenHands {
    * @returns List of agents available
    */
   static async getAgents(): Promise<string[]> {
-    const response = await fetch(`${OpenHands.BASE_URL}/api/options/agents`);
-    return response.json();
+    const cachedData = cache.get<string[]>("agents");
+    if (cachedData) return cachedData;
+
+    const data = await request(`/api/options/agents`);
+    cache.set("agents", data);
+
+    return data;
   }
 
   /**
@@ -55,178 +44,135 @@ class OpenHands {
    * @returns List of security analyzers available
    */
   static async getSecurityAnalyzers(): Promise<string[]> {
-    const response = await fetch(
-      `${OpenHands.BASE_URL}/api/options/security-analyzers`,
-    );
-    return response.json();
+    const cachedData = cache.get<string[]>("agents");
+    if (cachedData) return cachedData;
+
+    const data = await request(`/api/options/security-analyzers`);
+    cache.set("security-analyzers", data);
+
+    return data;
   }
 
   static async getConfig(): Promise<GetConfigResponse> {
-    const response = await fetch("config.json", {
-      headers: {
-        "Cache-Control": "no-cache",
-      },
-    });
-    return response.json();
+    const cachedData = cache.get<GetConfigResponse>("config");
+    if (cachedData) return cachedData;
+
+    const data = await request("/config.json");
+    cache.set("config", data);
+
+    return data;
   }
 
   /**
    * Retrieve the list of files available in the workspace
-   * @param token User token provided by the server
    * @param path Path to list files from
    * @returns List of files available in the given path. If path is not provided, it lists all the files in the workspace
    */
-  static async getFiles(token: string, path?: string): Promise<string[]> {
-    const url = new URL(`${OpenHands.BASE_URL}/api/list-files`);
-    if (path) url.searchParams.append("path", path);
-
-    const response = await fetch(url.toString(), {
-      headers: {
-        Authorization: `Bearer ${token}`,
-      },
-    });
-
-    return response.json();
+  static async getFiles(path?: string): Promise<string[]> {
+    let url = "/api/list-files";
+    if (path) url += `?path=${encodeURIComponent(path)}`;
+    return request(url);
   }
 
   /**
    * Retrieve the content of a file
-   * @param token User token provided by the server
    * @param path Full path of the file to retrieve
    * @returns Content of the file
    */
-  static async getFile(token: string, path: string): Promise<string> {
-    const url = new URL(`${OpenHands.BASE_URL}/api/select-file`);
-    url.searchParams.append("file", path);
-    const response = await fetch(url.toString(), {
-      headers: {
-        Authorization: `Bearer ${token}`,
-      },
-    });
-
-    const data = await response.json();
+  static async getFile(path: string): Promise<string> {
+    const url = `/api/select-file?file=${encodeURIComponent(path)}`;
+    const data = await request(url);
     return data.code;
   }
 
   /**
    * Save the content of a file
-   * @param token User token provided by the server
    * @param path Full path of the file to save
    * @param content Content to save in the file
    * @returns Success message or error message
    */
   static async saveFile(
-    token: string,
     path: string,
     content: string,
   ): Promise<SaveFileSuccessResponse | ErrorResponse> {
-    const response = await fetch(`${OpenHands.BASE_URL}/api/save-file`, {
+    return request(`/api/save-file`, {
       method: "POST",
       body: JSON.stringify({ filePath: path, content }),
       headers: {
-        Authorization: `Bearer ${token}`,
         "Content-Type": "application/json",
       },
     });
-
-    return response.json();
   }
 
   /**
    * Upload a file to the workspace
-   * @param token User token provided by the server
    * @param file File to upload
    * @returns Success message or error message
    */
   static async uploadFiles(
-    token: string,
     file: File[],
   ): Promise<FileUploadSuccessResponse | ErrorResponse> {
     const formData = new FormData();
     file.forEach((f) => formData.append("files", f));
 
-    const response = await fetch(`${OpenHands.BASE_URL}/api/upload-files`, {
+    return request(`/api/upload-files`, {
       method: "POST",
       body: formData,
-      headers: {
-        Authorization: `Bearer ${token}`,
-      },
     });
-
-    return response.json();
   }
 
   /**
    * Get the blob of the workspace zip
-   * @param token User token provided by the server
    * @returns Blob of the workspace zip
    */
-  static async getWorkspaceZip(token: string): Promise<Blob> {
-    const response = await fetch(`${OpenHands.BASE_URL}/api/zip-directory`, {
-      headers: {
-        Authorization: `Bearer ${token}`,
-      },
-    });
-
+  static async getWorkspaceZip(): Promise<Blob> {
+    const response = await request(`/api/zip-directory`, {}, false, true);
     return response.blob();
   }
 
   /**
    * Send feedback to the server
-   * @param token User token provided by the server
    * @param data Feedback data
    * @returns The stored feedback data
    */
-  static async sendFeedback(
-    token: string,
-    data: Feedback,
-  ): Promise<FeedbackResponse> {
-    const response = await fetch(`${OpenHands.BASE_URL}/api/submit-feedback`, {
+  static async submitFeedback(data: Feedback): Promise<FeedbackResponse> {
+    return request(`/api/submit-feedback`, {
       method: "POST",
       body: JSON.stringify(data),
       headers: {
-        Authorization: `Bearer ${token}`,
         "Content-Type": "application/json",
       },
     });
-
-    return response.json();
   }
 
   /**
-   * Get the GitHub access token
    * @param code Code provided by GitHub
    * @returns GitHub access token
    */
   static async getGitHubAccessToken(
     code: string,
   ): Promise<GitHubAccessTokenResponse> {
-    const response = await fetch(`${OpenHands.BASE_URL}/api/github/callback`, {
+    return request(`/api/github/callback`, {
       method: "POST",
       body: JSON.stringify({ code }),
       headers: {
         "Content-Type": "application/json",
       },
     });
-
-    return response.json();
   }
 
   /**
-   * Check if the user is authenticated
-   * @param login The user's GitHub login handle
-   * @returns Whether the user is authenticated
+   * Authenticate with GitHub token
+   * @returns Response with authentication status and user info if successful
    */
-  static async isAuthenticated(login: string): Promise<boolean> {
-    const response = await fetch(`${OpenHands.BASE_URL}/api/authenticate`, {
-      method: "POST",
-      body: JSON.stringify({ login }),
-      headers: {
-        "Content-Type": "application/json",
+  static async authenticate(): Promise<Response> {
+    return request(
+      `/api/authenticate`,
+      {
+        method: "POST",
       },
-    });
-
-    return response.status === 200;
+      true,
+    );
   }
 }
 

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

@@ -27,11 +27,16 @@ export interface GitHubAccessTokenResponse {
   access_token: string;
 }
 
+export interface AuthenticationResponse {
+  message: string;
+  login?: string; // Only present when allow list is enabled
+}
+
 export interface Feedback {
   version: string;
   email: string;
   token: string;
-  feedback: "positive" | "negative";
+  polarity: "positive" | "negative";
   permissions: "public" | "private";
   trajectory: unknown[];
 }

frontend/src/assets/branding/all-hands-logo-spark.svg

@@ -32,4 +32,4 @@
       <rect width="69" height="46" fill="white" transform="translate(0.5)" />
     </clipPath>
   </defs>
-</svg>
+</svg>

frontend/src/assets/branding/github-logo.svg

@@ -2,4 +2,4 @@
   <path
     d="M15.359 21V17.319C15.3974 16.8654 15.3314 16.4095 15.1651 15.9814C14.9989 15.5534 14.7363 15.1631 14.3949 14.8364C17.6154 14.5035 21 13.3716 21 8.17826C20.9997 6.85027 20.4489 5.57321 19.4615 4.61139C19.9291 3.44954 19.896 2.16532 19.3692 1.02548C19.3692 1.02548 18.159 0.692576 15.359 2.43321C13.0082 1.84237 10.5302 1.84237 8.17949 2.43321C5.37949 0.692576 4.16923 1.02548 4.16923 1.02548C3.64244 2.16532 3.60938 3.44954 4.07692 4.61139C3.08218 5.58034 2.53079 6.86895 2.53846 8.2068C2.53846 13.3621 5.92308 14.494 9.14359 14.865C8.80615 15.1883 8.54591 15.574 8.3798 15.9968C8.2137 16.4196 8.14544 16.8701 8.17949 17.319V21M8.17949 18.1465C3.05128 19.5732 3.05128 15.7686 1 15.293L8.17949 18.1465Z"
     stroke="white" stroke-linecap="round" stroke-linejoin="round" />
-</svg>
+</svg>

frontend/src/components/AgentControlBar.tsx

@@ -6,7 +6,7 @@ import PlayIcon from "#/assets/play";
 import { generateAgentStateChangeEvent } from "#/services/agentStateService";
 import { RootState } from "#/store";
 import AgentState from "#/types/AgentState";
-import { useSocket } from "#/context/socket";
+import { useWsClient } from "#/context/ws-client-provider";
 
 const IgnoreTaskStateMap: Record<string, AgentState[]> = {
   [AgentState.PAUSED]: [
@@ -72,7 +72,7 @@ function ActionButton({
 }
 
 function AgentControlBar() {
-  const { send } = useSocket();
+  const { send } = useWsClient();
   const { curAgentState } = useSelector((state: RootState) => state.agent);
 
   const handleAction = (action: AgentState) => {

frontend/src/components/AgentStatusBar.tsx

@@ -1,6 +1,7 @@
 import React, { useEffect } from "react";
 import { useTranslation } from "react-i18next";
 import { useSelector } from "react-redux";
+import toast from "react-hot-toast";
 import { I18nKey } from "#/i18n/declaration";
 import { RootState } from "#/store";
 import AgentState from "#/types/AgentState";
@@ -16,7 +17,7 @@ enum IndicatorColor {
 }
 
 function AgentStatusBar() {
-  const { t } = useTranslation();
+  const { t, i18n } = useTranslation();
   const { curAgentState } = useSelector((state: RootState) => state.agent);
   const { curStatusMessage } = useSelector((state: RootState) => state.status);
 
@@ -94,15 +95,27 @@ function AgentStatusBar() {
   const [statusMessage, setStatusMessage] = React.useState<string>("");
 
   React.useEffect(() => {
-    if (curAgentState === AgentState.LOADING) {
-      const trimmedCustomMessage = curStatusMessage.status.trim();
-      if (trimmedCustomMessage) {
-        setStatusMessage(t(trimmedCustomMessage));
-        return;
+    let message = curStatusMessage.message || "";
+    if (curStatusMessage?.id) {
+      const id = curStatusMessage.id.trim();
+      if (i18n.exists(id)) {
+        message = t(curStatusMessage.id.trim()) || message;
       }
     }
+    if (curStatusMessage?.type === "error") {
+      toast.error(message);
+      return;
+    }
+    if (curAgentState === AgentState.LOADING && message.trim()) {
+      setStatusMessage(message);
+    } else {
+      setStatusMessage(AgentStatusMap[curAgentState].message);
+    }
+  }, [curStatusMessage.id]);
+
+  React.useEffect(() => {
     setStatusMessage(AgentStatusMap[curAgentState].message);
-  }, [curAgentState, curStatusMessage.status]);
+  }, [curAgentState]);
 
   return (
     <div className="flex flex-col items-center">

frontend/src/components/Jupyter.tsx

@@ -25,7 +25,11 @@ function JupyterCell({ cell }: IJupyterCell): JSX.Element {
           className="scrollbar-custom scrollbar-thumb-gray-500 hover:scrollbar-thumb-gray-400 dark:scrollbar-thumb-white/10 dark:hover:scrollbar-thumb-white/20 overflow-auto px-5"
           style={{ padding: 0, marginBottom: 0, fontSize: "0.75rem" }}
         >
-          <SyntaxHighlighter language="python" style={atomOneDark}>
+          <SyntaxHighlighter
+            language="python"
+            style={atomOneDark}
+            wrapLongLines
+          >
             {code}
           </SyntaxHighlighter>
         </pre>
@@ -78,7 +82,11 @@ function JupyterCell({ cell }: IJupyterCell): JSX.Element {
   );
 }
 
-function JupyterEditor(): JSX.Element {
+interface JupyterEditorProps {
+  maxWidth: number;
+}
+
+function JupyterEditor({ maxWidth }: JupyterEditorProps) {
   const { t } = useTranslation();
 
   const { cells } = useSelector((state: RootState) => state.jupyter);
@@ -88,7 +96,7 @@ function JupyterEditor(): JSX.Element {
     useScrollToBottom(jupyterRef);
 
   return (
-    <div className="flex-1">
+    <div className="flex-1" style={{ maxWidth }}>
       <div
         className="overflow-y-auto h-full"
         ref={jupyterRef}

frontend/src/components/analytics-consent-form-modal.tsx

@@ -0,0 +1,42 @@
+import { useFetcher } from "@remix-run/react";
+import { ModalBackdrop } from "./modals/modal-backdrop";
+import ModalBody from "./modals/ModalBody";
+import ModalButton from "./buttons/ModalButton";
+import {
+  BaseModalTitle,
+  BaseModalDescription,
+} from "./modals/confirmation-modals/BaseModal";
+
+export function AnalyticsConsentFormModal() {
+  const fetcher = useFetcher({ key: "set-consent" });
+
+  return (
+    <ModalBackdrop>
+      <fetcher.Form
+        method="POST"
+        action="/set-consent"
+        className="flex flex-col gap-2"
+      >
+        <ModalBody>
+          <BaseModalTitle title="Your Privacy Preferences" />
+          <BaseModalDescription>
+            We use tools to understand how our application is used to improve
+            your experience. You can enable or disable analytics. Your
+            preferences will be stored and can be updated anytime.
+          </BaseModalDescription>
+
+          <label className="flex gap-2 items-center self-start">
+            <input name="analytics" type="checkbox" defaultChecked />
+            Send anonymous usage data
+          </label>
+
+          <ModalButton
+            type="submit"
+            text="Confirm Preferences"
+            className="bg-primary text-white w-full hover:opacity-80"
+          />
+        </ModalBody>
+      </fetcher.Form>
+    </ModalBackdrop>
+  );
+}