WIP

evaluation/benchmarks/swe_bench/run_infer.py

@@ -104,6 +104,7 @@ def set_dataset_type(dataset_name: str) -> str:
 
 AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
     'CodeActAgent': codeact_user_response,
+    'OpenCodeAgent': codeact_user_response,
 }
 
 

openhands/agenthub/__init__.py

@@ -8,6 +8,7 @@
     codeact_agent,
     dummy_agent,
     loc_agent,
+    opencode_agent,
     readonly_agent,
     visualbrowsing_agent,
 )
@@ -21,4 +22,5 @@
     'visualbrowsing_agent',
     'readonly_agent',
     'loc_agent',
+    'opencode_agent',
 ]

openhands/agenthub/opencode_agent/README.md

@@ -0,0 +1,212 @@
+# OpenCode Agent
+
+The OpenCode Agent is inspired by Claude Code's tool-based approach to software development assistance.
+
+## Overview
+
+This agent provides a comprehensive set of tools for:
+- **File Operations**: read, write, edit, glob, grep, list_dir
+- **Web Access**: webfetch, websearch
+- **User Interaction**: question tool for clarifying requirements
+- **Task Management**: todo_read, todo_write for tracking progress
+- **Code Execution**: bash commands
+- **Structured Edits**: apply_patch for multi-file atomic changes
+
+## Key Features
+
+### 1. File Operations
+- `read`: Read file contents with line numbers
+- `write`: Create or overwrite files
+- `edit`: Make precise string replacements
+- `glob`: Pattern-based file search
+- `grep`: Content search across files
+- `list_dir`: Directory listing
+
+### 2. Web Access
+- `webfetch`: Fetch web content in text/markdown/html format
+- `websearch`: Search the web for current information
+
+### 3. User Interaction
+- `question`: Ask users multiple-choice or custom questions
+  - Supports single/multiple selection
+  - Helps clarify requirements before implementation
+
+### 4. Task Management
+- `todo_read`: Read current task list
+- `todo_write`: Update task list with new tasks or status changes
+  - Tracks: pending, in_progress, completed, cancelled
+  - Helps organize multi-step work
+
+### 5. Structured Editing
+- `apply_patch`: Apply unified diff patches
+  - Supports add, update, delete, move operations
+  - Atomic multi-file changes
+  - LSP diagnostics after changes
+
+## Tool Descriptions
+
+### Core File Tools
+
+#### read
+Reads files with line numbers. Supports:
+- Absolute paths
+- Offset and limit for large files
+- Image files (jpeg, png, gif, webp)
+
+#### write
+Creates or overwrites files. Automatically creates parent directories.
+
+#### edit
+Makes exact string replacements in files. Requires reading the file first.
+
+#### glob
+Finds files matching glob patterns (e.g., `**/*.py`, `src/**/*.ts`).
+
+#### grep
+Searches file contents with regex support. Output modes:
+- `content`: Show matching lines with context
+- `files_with_matches`: Just file paths
+- `count`: Match counts per file
+
+#### list_dir
+Lists directory contents with optional recursion.
+
+### Web Tools
+
+#### webfetch
+Fetches web content with format conversion:
+- `text`: Plain text extraction
+- `markdown`: HTML to markdown conversion
+- `html`: Raw HTML
+
+#### websearch
+Searches the web with customizable options:
+- `type`: auto/fast/deep search
+- `livecrawl`: fallback/preferred for freshness
+- Returns content optimized for LLMs
+
+### Interaction Tools
+
+#### question
+Asks users structured questions:
+```python
+{
+  "questions": [{
+    "question": "Which library should we use?",
+    "header": "Library",
+    "options": [
+      {"label": "React", "description": "Popular UI library"},
+      {"label": "Vue", "description": "Progressive framework"}
+    ],
+    "multiple": false,
+    "custom": true
+  }]
+}
+```
+
+### Task Management Tools
+
+#### todo_read
+Returns current task list as JSON array.
+
+#### todo_write
+Updates task list:
+```python
+{
+  "todos": [
+    {"id": "1", "title": "Implement feature", "status": "completed"},
+    {"id": "2", "title": "Write tests", "status": "in_progress"}
+  ]
+}
+```
+
+### Patch Tool
+
+#### apply_patch
+Applies unified diff patches:
+```diff
+*** Begin Patch
+--- path/to/file.py
++++ path/to/file.py
+@@ -10,7 +10,7 @@
+ unchanged
+-old line
++new line
+ unchanged
+*** End Patch
+```
+
+Supports:
+- Add files: `+++ new/file.py`
+- Delete files: `+++ /dev/null`
+- Move files: Different `---` and `+++` paths
+- Update files: Standard diffs
+
+## Usage
+
+### Quick Start
+
+To use the OpenCode agent in code:
+
+```python
+from openhands.core.config import AgentConfig
+from openhands.llm.llm_registry import LLMRegistry
+
+config = AgentConfig(agent_name='OpenCodeAgent')
+llm_registry = LLMRegistry()
+agent = OpenCodeAgent(config, llm_registry)
+```
+
+Or via configuration:
+```yaml
+agent:
+  name: OpenCodeAgent
+```
+
+### Evaluation and Benchmarking
+
+To run SWE-bench or other evaluations with OpenCodeAgent:
+
+```bash
+# Using run_infer.sh script
+export AGENT="OpenCodeAgent"
+./evaluation/benchmarks/swe_bench/scripts/run_infer.sh llama3_1 test OpenCodeAgent
+
+# Or call run_infer.py directly
+poetry run python evaluation/benchmarks/swe_bench/run_infer.py \
+  --agent-cls OpenCodeAgent \
+  --llm-config llama3_1 \
+  --max-iterations 30 \
+  --dataset princeton-nlp/SWE-bench_Lite \
+  --split test
+```
+
+**For detailed instructions on:**
+- Dynamic agent selection (OpenCodeAgent vs CodeActAgent)
+- Custom system prompt overrides
+- Configuration options
+- Troubleshooting
+
+**See the comprehensive [Usage Guide](./USAGE_GUIDE.md)**
+
+## Design Principles
+
+1. **Clear Tool Boundaries**: Each tool has a specific purpose
+2. **User Interaction**: Question tool for clarification before action
+3. **Task Tracking**: Todo tools for managing complex workflows
+4. **Atomic Operations**: apply_patch for multi-file changes
+5. **Web Integration**: Fetch and search for current information
+
+## Comparison to CodeActAgent
+
+| Feature | CodeActAgent | OpenCodeAgent |
+|---------|--------------|---------------|
+| File ops | ✓ | ✓ |
+| Web access | Browser tool | webfetch + websearch |
+| User questions | - | ✓ question tool |
+| Task tracking | task_tracker | todo_read/write |
+| Multi-file edits | Individual edits | apply_patch |
+| Jupyter | ✓ | - |
+| Plan mode | ✓ | - |
+
+OpenCodeAgent is optimized for structured workflows with clear tool boundaries and explicit user interaction patterns.

openhands/agenthub/opencode_agent/__init__.py

@@ -0,0 +1,4 @@
+from openhands.agenthub.opencode_agent.opencode_agent import OpenCodeAgent
+from openhands.controller.agent import Agent
+
+Agent.register('OpenCodeAgent', OpenCodeAgent)