open-swe/docs/CUSTOMIZATION.md

# Customization Guide

Open SWE is designed to be forked and customized for your org. The core agent is assembled in a single function — `get_agent()` in `agent/server.py` — where you can swap out the sandbox, model, tools, and triggers.

```python
# agent/server.py — the key lines
model_id = os.environ.get("LLM_MODEL_ID", DEFAULT_LLM_MODEL_ID)
model_kwargs = {"max_tokens": DEFAULT_LLM_MAX_TOKENS}
if model_id == DEFAULT_LLM_MODEL_ID:
    model_kwargs["reasoning"] = DEFAULT_LLM_REASONING

return create_deep_agent(
    model=make_model(model_id, **model_kwargs),
    system_prompt=construct_system_prompt(...),
    tools=[http_request, fetch_url, linear_comment, slack_thread_reply],
    backend=sandbox_backend,
    middleware=[
        ToolErrorMiddleware(),
        check_message_queue_before_model,
        ensure_no_empty_msg,
        notify_step_limit_reached,
    ],
)
```

---

## 1. Sandbox

By default, Open SWE runs each task in a [LangSmith cloud sandbox](https://docs.smith.langchain.com/) — an isolated Linux environment where the agent clones the repo and executes commands. Sandbox creation and connection is handled in `agent/integrations/langsmith.py`.

### Using a custom sandbox snapshot

Build a snapshot in LangSmith (UI or `SandboxClient.create_snapshot`) from your Docker image and point Open SWE at its UUID:

```bash
DEFAULT_SANDBOX_SNAPSHOT_ID="<snapshot-uuid>"                      # Required
DEFAULT_SANDBOX_SNAPSHOT_FS_CAPACITY_BYTES="34359738368"           # Optional, default 32 GiB
DEFAULT_SANDBOX_VCPUS="4"                                          # Optional, default 4
DEFAULT_SANDBOX_MEM_BYTES="16106127360"                            # Optional, default 15 GiB
DEFAULT_SANDBOX_IDLE_TTL_SECONDS="7200"                            # Optional, default 7200 (2 h); 0 disables
DEFAULT_SANDBOX_DELETE_AFTER_STOP_SECONDS="86400"                  # Optional, default 86400 (24 h); 0 disables
REPO_SNAPSHOT_BASE_IMAGE="<registry>/<open-swe-sandbox-image>"      # Optional; required for admin-generated repo snapshot templates
```

This is useful for pre-installing languages, frameworks, or internal tools that your repos depend on — reducing setup time per agent run. The default snapshot includes the GitHub CLI; agents invoke it as `GH_TOKEN=dummy gh <command>` and rely on the LangSmith proxy for the real credentials.

`REPO_SNAPSHOT_BASE_IMAGE` should point to the published Docker image used to create your default Open SWE sandbox snapshot (typically the image built from this repository's `Dockerfile`). The admin **Repository Snapshots** page uses it as the base image when generating per-repo Dockerfile templates. If it is not configured, template generation fails closed instead of suggesting a bare image that would be missing Open SWE's required sandbox tools.

For LangSmith sandboxes, Open SWE configures two GitHub proxy rules whenever a sandbox is created or reattached to a run:

- `github.com` / `*.github.com` receive Basic auth for git-over-HTTPS operations.
- `api.github.com` receives Bearer auth for `gh` and REST API operations.

The proxy token is minted at runtime from the GitHub App installation credentials. Do not store GitHub access tokens as deployment environment variables.

### Using a different sandbox provider

Set the `SANDBOX_TYPE` environment variable to switch providers. Each provider has a corresponding integration file in `agent/integrations/` and a factory function registered in `agent/utils/sandbox.py`:

| `SANDBOX_TYPE` | Integration file | Required env vars |
|---|---|---|
| `langsmith` (default) | `agent/integrations/langsmith.py` | `LANGSMITH_API_KEY_PROD`, `SANDBOX_TYPE="langsmith"` |
| `daytona` | `agent/integrations/daytona.py` | `DAYTONA_API_KEY`, `SANDBOX_TYPE="daytona"`, optional `DAYTONA_SANDBOX_SNAPSHOT` |
| `runloop` | `agent/integrations/runloop.py` | `RUNLOOP_API_KEY`, `SANDBOX_TYPE="runloop"` |
| `e2b` | `agent/integrations/e2b.py` | `E2B_API_KEY`, `SANDBOX_TYPE="e2b"`, optional `E2B_TEMPLATE` |
| `modal` | `agent/integrations/modal.py` | Modal credentials, `SANDBOX_TYPE="modal"` |
| `local` | `agent/integrations/local.py` | None (no isolation — development only), `SANDBOX_TYPE="local"` |

> **Warning**: `local` runs commands directly on your host with no sandboxing. Only use for local development with human-in-the-loop enabled.

For `langsmith`, sandboxes default to the same LangSmith credentials as tracing. To run sandboxes against a **different** LangSmith workspace, set `SANDBOX_LANGSMITH_API_KEY` (falls back to `LANGSMITH_API_KEY` / `LANGSMITH_API_KEY_PROD`) and optionally `SANDBOX_LANGSMITH_ENDPOINT` (falls back to `LANGSMITH_ENDPOINT`). These apply to sandbox create/connect/delete, the GitHub proxy config, and repo snapshot builds — the `DEFAULT_SANDBOX_SNAPSHOT_ID` must exist in whichever workspace these credentials point at.

### Adding a new sandbox provider

1. **Create an integration file** at `agent/integrations/my_provider.py` with a factory function matching this signature:

```python
def create_my_provider_sandbox(sandbox_id: str | None = None):
    """Create or reconnect to a sandbox.

    Args:
        sandbox_id: Optional existing sandbox ID to reconnect to.
            If None, creates a new sandbox.

    Returns:
        An object implementing SandboxBackendProtocol.
    """
    ...
```

2. **Register it** in `agent/utils/sandbox.py` by adding it to `SANDBOX_FACTORIES`:

```python
SANDBOX_FACTORIES = {
    ...
    "my_provider": ("agent.integrations.my_provider", "create_my_provider_sandbox"),
}
```

The factory must return an object implementing `SandboxBackendProtocol` from `deepagents`. See the existing integration files for reference.

### Building a custom sandbox provider

If none of the built-in providers fit, you can build your own. The agent accepts any backend that implements `SandboxBackendProtocol` from `deepagents`. The protocol requires:

- **File operations**: `ls()`, `read()`, `write()`, `edit()`, `glob()`, `grep()`
- **Shell execution**: `execute(command, timeout=None) -> ExecuteResponse`
- **Identity**: `id` property returning a unique sandbox identifier

The easiest approach is to extend `BaseSandbox` from `deepagents.backends.sandbox` — it implements all file operations by delegating to `execute()`, so you only need to implement the shell execution layer:

```python
from deepagents.backends.sandbox import BaseSandbox
from deepagents.backends.protocol import ExecuteResponse

class MySandbox(BaseSandbox):
    def __init__(self, connection):
        self._conn = connection

    @property
    def id(self) -> str:
        return self._conn.id

    def execute(self, command: str, *, timeout: int | None = None) -> ExecuteResponse:
        result = self._conn.run(command, timeout=timeout or 300)
        return ExecuteResponse(
            output=result.stdout + result.stderr,
            exit_code=result.exit_code,
            truncated=False,
        )
```

See `deepagents.backends.LangSmithSandbox` and `agent/integrations/langsmith.py` for a full reference implementation.

---

## 2. Model

The model is configured in the `get_agent()` function in `agent/server.py`. By default it uses `openai:gpt-5.5` with medium reasoning effort, but you can override the model with the `LLM_MODEL_ID` environment variable:

```bash
# Set the model via environment variable (uses provider:model format)
LLM_MODEL_ID="anthropic:claude-sonnet-5"
```

If `LLM_MODEL_ID` is not set, the default model (`openai:gpt-5.5`) is used.

`max_tokens` is a maximum completion/output token budget, not the model's total context window. For OpenAI reasoning models, this budget can include both internal reasoning tokens and final response tokens.

### Switching models

Use the `provider:model` format:

```python
# Anthropic
model=make_model("anthropic:claude-sonnet-5", temperature=0, max_tokens=16_000)

# OpenAI (uses Responses API by default)
model=make_model("openai:gpt-5.5", max_tokens=128_000, reasoning={"effort": "medium"})

# Google
model=make_model("google_genai:gemini-2.5-pro", temperature=0, max_tokens=16_000)
```

The `make_model()` helper in `agent/utils/model.py` wraps `langchain.chat_models.init_chat_model`. For OpenAI models, it automatically enables the Responses API. For full control, pass a pre-configured model instance directly:

```python
from langchain_anthropic import ChatAnthropic

model = ChatAnthropic(model_name="claude-sonnet-5", temperature=0, max_tokens=16_000)

return create_deep_agent(
    model=model,
    ...
)
```

### Using different models per context

You can route to different models based on task complexity, repo, or trigger source:

```python
async def get_agent(config: RunnableConfig) -> Pregel:
    source = config["configurable"].get("source")
    
    if source == "slack":
        # Faster model for Slack Q&A
        model = make_model("anthropic:claude-sonnet-5", temperature=0, max_tokens=16_000)
    else:
        # Full model for code changes from Linear
        model = make_model("openai:gpt-5.5", max_tokens=128_000, reasoning={"effort": "medium"})
    
    return create_deep_agent(model=model, ...)
```

### Routing through the LangSmith LLM Gateway

Model calls can be proxied through the [LangSmith LLM Gateway](https://docs.langchain.com/langsmith/llm-gateway) (private beta) instead of hitting providers directly. The gateway authenticates with a **LangSmith API key** that has the `gateway:invoke` permission and resolves the real provider key from workspace Provider Secrets, so no provider API keys are needed at runtime — and it adds central spend limits, PII/secrets redaction, and tracing. Your org must have the gateway enabled with Provider Secrets configured.

Routing is opt-in and off by default. Enable it either way:

| Env var | Default | Purpose |
|---|---|---|
| `LANGSMITH_GATEWAY_ENABLED` | `false` | Deployment-level default for gateway routing. |
| `LANGSMITH_GATEWAY_API_KEY` | unset | Optional dedicated LangSmith key for Gateway calls. Prefer this in LangGraph Cloud if the platform-provided `LANGSMITH_API_KEY` lacks `gateway:invoke`. Falls back to `LANGSMITH_API_KEY_PROD`, then `LANGSMITH_API_KEY`. |
| `LANGSMITH_GATEWAY_BASE_URL` | `https://gateway.smith.langchain.com` | Override for a regional or self-hosted gateway host. |
| `LANGSMITH_GATEWAY_OPENAI_USE_RESPONSES` | `true` | Use the OpenAI Responses API through the gateway. Set to `false` only to force Chat Completions for OpenAI models. |

The admin panel (**Admin → LLM Gateway**) exposes a per-workspace toggle stored in team settings; when set it overrides the `LANGSMITH_GATEWAY_ENABLED` env default (a `None`/unset team value inherits the env default).

Routing is applied centrally in `make_model` (`agent/utils/model.py`), which resolves the effective on/off and delegates URL/key wiring to `agent/utils/gateway.py`. **OpenAI, Anthropic, Fireworks, and Google Gemini** are routed (their LangChain integrations accept `base_url` + `api_key`); Google Vertex (service-account auth) and any other provider call the provider directly with a logged warning.

**Caveat — OpenAI endpoint:** open-swe uses the OpenAI Responses API by default because OpenAI reasoning models with function tools reject `reasoning_effort` on Chat Completions. Direct OpenAI calls use a `wss://` base URL; gateway-routed OpenAI uses the HTTPS gateway base URL with Responses enabled. Set `LANGSMITH_GATEWAY_OPENAI_USE_RESPONSES=false` only if you need to force Chat Completions. Anthropic and Fireworks are unaffected.

---

## 3. Tools

Open SWE ships with a small set of custom tools on top of the built-in Deep Agents tools (file operations, shell execution, subagents, todos). GitHub operations are handled by `GH_TOKEN=dummy gh` inside the sandbox.

| Tool | File | Purpose |
|---|---|---|
| `fetch_url` | `agent/tools/fetch_url.py` | Fetch web pages as markdown |
| `http_request` | `agent/tools/http_request.py` | HTTP API calls |
| `linear_comment` | `agent/tools/linear_comment.py` | Post comments on Linear tickets |
| `jira_comment`, `jira_get_issue`, … | `agent/tools/jira_*.py` | Read/comment/create/update Jira issues (`agent/utils/jira.py`) |
| `confluence_get_page`, `confluence_update_page`, … | `agent/tools/confluence_*.py` | Read/write Confluence pages + comments (`agent/utils/confluence.py`) |
| `slack_thread_reply` | `agent/tools/slack_thread_reply.py` | Reply in Slack threads |

### Adding a tool

Create a new file in `agent/tools/`, define a function, and add it to the tools list.

**Example — adding a Datadog search tool:**

```python
# agent/tools/datadog_search.py
import requests
from typing import Any

def datadog_search(query: str, time_range: str = "1h") -> dict[str, Any]:
    """Search Datadog logs for debugging context.

    Args:
        query: Datadog log query string
        time_range: Time range to search (e.g. "1h", "24h", "7d")

    Returns:
        Dictionary with matching log entries
    """
    # Your Datadog API integration here
    ...
```

Then register it in `agent/server.py`:

```python
from .tools import fetch_url, http_request, linear_comment, slack_thread_reply
from .tools.datadog_search import datadog_search

return create_deep_agent(
    ...
    tools=[
        http_request, fetch_url,
        linear_comment, slack_thread_reply,
        datadog_search,  # new tool
    ],
    ...
)
```

The agent will automatically see the tool's name, docstring, and parameter types — the docstring serves as the tool description, so write it clearly.

### Removing tools

If you only use Linear (not Slack), remove `slack_thread_reply` from the tools list and vice versa. If you don't need web fetching, remove `fetch_url`.

### Conditional tools

You can vary the toolset based on the trigger source:

```python
base_tools = [http_request, fetch_url]
source = config["configurable"].get("source")

if source == "linear":
    tools = [*base_tools, linear_comment]
elif source == "slack":
    tools = [*base_tools, slack_thread_reply]
else:
    tools = [*base_tools, linear_comment, slack_thread_reply]

return create_deep_agent(tools=tools, ...)
```

---

## 4. Triggers

Open SWE supports three invocation surfaces: Linear, Slack, and GitHub. Each is implemented as a webhook endpoint in `agent/webapp.py`. You can add, remove, or modify triggers independently.

### Removing a trigger

If you don't use Linear, simply don't configure the Linear webhook and remove the env vars. Same for Slack. The webhook endpoints still exist but won't receive events.

To fully remove a trigger's code, delete the corresponding endpoint from `agent/webapp.py`:

- **Linear**: `linear_webhook()` and `process_linear_issue()`
- **Slack**: `slack_webhook()` and `process_slack_mention()`

### Default repository

Set the default GitHub org and repo used across all triggers (Slack, Linear, GitHub) when no repo is specified:

```bash
DEFAULT_REPO_OWNER="my-org"      # Default GitHub org (used everywhere)
DEFAULT_REPO_NAME="my-repo"      # Default GitHub repo (used everywhere)
```

These are used as the fallback when:
- A Slack message doesn't specify a repo (and no thread metadata exists)
- A Linear issue's team/project isn't in the `LINEAR_TEAM_TO_REPO` mapping
- A user writes `repo:name` without an org prefix — the org defaults to `DEFAULT_REPO_OWNER`

### Repository extraction from messages

Slack, Linear, Jira, and Confluence all support specifying a target repo directly in the message or comment text. The shared utility `extract_repo_from_text()` in `agent/utils/repo.py` handles parsing these formats:

- `repo:owner/name` — explicit org and repo
- `repo owner/name` — space syntax (same result)
- `repo:name` — repo name only; the org defaults to `DEFAULT_REPO_OWNER`
- `https://github.com/owner/name` — GitHub URL

### Customizing Linear routing

The `LINEAR_TEAM_TO_REPO` dict in `agent/utils/linear_team_repo_map.py` maps Linear teams and projects to GitHub repos:

```python
LINEAR_TEAM_TO_REPO = {
    "Engineering": {
        "projects": {
            "backend": {"owner": "my-org", "name": "backend"},
            "frontend": {"owner": "my-org", "name": "frontend"},
        },
        "default": {"owner": "my-org", "name": "monorepo"},
    },
}
```

Users can also override the team/project mapping on a per-comment basis by including `repo:owner/name` in their `@openswe` comment. This takes priority over the mapping — the mapping is used as a fallback when no repo is specified in the comment. If the team/project isn't found in the mapping either, `DEFAULT_REPO_OWNER`/`DEFAULT_REPO_NAME` is used.

### Customizing Slack routing

Slack repo resolution (`get_slack_repo_config` in `agent/webapp.py`) checks, in order:

1. Repo carried over from the existing Slack thread's metadata.
2. A `repo:owner/name` (or GitHub URL) token in the channel's **topic or purpose** (its "description"). This lets a channel be pinned to a repo without anyone repeating it per-message.
3. The triggering user's dashboard `default_repo`.
4. The team default repo.
5. `SLACK_REPO_OWNER`/`SLACK_REPO_NAME`, falling back to `DEFAULT_REPO_OWNER`/`DEFAULT_REPO_NAME`.

Users can still override per-message with `repo:owner/name` syntax in their Slack message (this is read from the message text by the agent). A shorthand `repo:name` (without the org) is also supported — the org defaults to `DEFAULT_REPO_OWNER`.

Reading the channel topic/purpose requires the bot's Slack token to have the `channels:read` (and `groups:read` for private channels) scope so `conversations.info` succeeds.

### Adding a new trigger

To add a new invocation surface (e.g. Jira, Discord, a custom API):

1. **Add a webhook endpoint** in `agent/webapp.py`:

```python
@app.post("/webhooks/my-trigger")
async def my_trigger_webhook(request: Request, background_tasks: BackgroundTasks):
    # Parse the incoming event
    payload = await request.json()
    
    # Extract task description and repo info
    task_description = payload["description"]
    repo_config = {"owner": "my-org", "name": "my-repo"}
    
    # Create a LangGraph run
    background_tasks.add_task(process_my_trigger, task_description, repo_config)
    return {"status": "accepted"}
```

2. **Create a processing function** that builds the prompt and starts an agent run:

```python
async def process_my_trigger(task_description: str, repo_config: dict):
    thread_id = generate_deterministic_id(task_description)
    langgraph_client = get_client(url=LANGGRAPH_URL)
    
    await langgraph_client.runs.create(
        thread_id,
        "agent",
        input={"messages": [{"role": "user", "content": task_description}]},
        config={"configurable": {
            "repo": repo_config,
            "source": "my-trigger",
            "user_email": "user@example.com",
        }},
        if_not_exists="create",
    )
```

3. **Add a communication tool** (optional) so the agent can report back:

```python
# agent/tools/my_trigger_reply.py
def my_trigger_reply(message: str) -> dict:
    """Post a reply to the triggering service."""
    # Your API call here
    ...
```

The key fields in `config.configurable` are:
- `repo`: `{"owner": "...", "name": "..."}` — which GitHub repo to work on
- `source`: string identifying the trigger (used for auth routing and communication)
- `user_email`: the triggering user's email (for GitHub OAuth resolution)

---

## 5. System prompt

The system prompt is assembled in `agent/prompt.py` from modular sections. You can customize behavior by editing individual sections:

| Section | What it controls |
|---|---|
| `WORKING_ENV_SECTION` | Sandbox paths and execution constraints |
| `TASK_EXECUTION_SECTION` | Workflow steps (understand → implement → verify → submit) |
| `CODING_STANDARDS_SECTION` | Code style, testing, and quality rules |
| `COMMIT_PR_SECTION` | PR title/body format and commit conventions |
| `CODE_REVIEW_GUIDELINES_SECTION` | How the agent reviews code changes |
| `COMMUNICATION_SECTION` | Formatting and messaging guidelines |

### Default prompt file

Open SWE supports a `default_prompt.md` file for org-level instructions that apply to **every** agent run, regardless of which repository is being worked on. This is the recommended way to set default repository preferences, org conventions, and shared guidelines.

The file is loaded at agent startup and injected into the system prompt between the task overview and repository setup sections.

**Location:** [`agent/resources/default_prompt.md`](../agent/resources/default_prompt.md) for the bundled default.

**Override:** Set the `DEFAULT_PROMPT_PATH` environment variable to use a different file:

```bash
DEFAULT_PROMPT_PATH="/path/to/my-org-prompt.md"
```

**Format:** Write plain markdown. The content is injected as-is under a `### Custom Instructions` heading in the system prompt. Example:

```markdown
# Default Prompt

## Default Repository

When no repository is specified, work on the **my-app** repository under **my-org**.

## Organization Conventions

- Use conventional commits: feat:, fix:, chore:
- Always tag the requesting user when work is complete
```

**Loading order:** Default prompt → System prompt sections → AGENTS.md (per-repo). If the file is missing or empty, it is silently skipped — no error is raised.

**When to use `default_prompt.md` vs `AGENTS.md`:**

| | `default_prompt.md` | `AGENTS.md` |
|---|---|---|
| Scope | All tasks, all repos | Single repository |
| Location | Open SWE project root | Target repo root |
| Use for | Default repo, org conventions | Repo-specific coding standards |

### Using AGENTS.md

Drop an `AGENTS.md` file in the root of any repository to add repo-specific instructions. The agent reads it from the sandbox at startup and appends it to the system prompt. This is the easiest way to encode conventions per-repo without modifying Open SWE's code.

---

## 6. Middleware

Middleware hooks run around the agent loop. Open SWE includes:

| Middleware | Type | Purpose |
|---|---|---|
| `ToolErrorMiddleware` | Tool error handler | Catches and formats tool errors |
| `check_message_queue_before_model` | Before model | Injects follow-up messages that arrived mid-run |
| `ensure_no_empty_msg` | After model | Re-injects a tool call when the model stops without one, so runs don't end prematurely |
| `notify_step_limit_reached` | After agent | Posts a Slack reply when the agent hits the model-call limit |

There is intentionally no after-agent middleware that opens a PR for the agent. The agent is responsible for committing, pushing, opening/updating the draft PR, and replying in the source channel. If you want a deterministic backstop for your fork, add an `@after_agent` hook here.

Add custom middleware by appending to the middleware list in `get_agent()`. See the [LangChain middleware docs](https://python.langchain.com/docs/concepts/agents/#middleware) for the `@before_model` and `@after_agent` decorators.

**Example — adding a CI check after agent completion:**

```python
from langchain.agents.middleware import AgentState, after_agent
from langgraph.runtime import Runtime

@after_agent
async def run_ci_check(state: AgentState, runtime: Runtime):
    """Run CI checks after the agent finishes."""
    # Trigger your CI pipeline here
    ...
```

Then add it to the middleware list:

```python
middleware=[
    ToolErrorMiddleware(),
    check_message_queue_before_model,
    ensure_no_empty_msg,
    notify_step_limit_reached,
    run_ci_check,  # new middleware
],
```