open-swe/agent/server.py
Johannes du Plessis 82852f9eda
feat: tune reviewer for precision — web/wiki tools + recalibrated prompt (#1312)
* feat: tune reviewer for precision — web/wiki tools + recalibrated prompt

Reviewer agent now has web_search, fetch_url, and http_request alongside the
finding tools, so it can verify library semantics and consult the DeepWiki
auto-generated wiki for public repos (https://deepwiki.com/<owner>/<repo>)
before flagging cross-file or architectural concerns.

Prompt rewritten to push precision over recall:
- explicit severity ladder pushing reviews toward bimodal high/low instead of
  defaulting to medium
- ≤200-char description target (gold set averages ~186 chars; we were at ~436)
- mandatory docs / wiki / code lookup before flagging concurrency, security,
  or perf — the three categories that dominated false positives
- "do not flag" list covering compiler/linter-catchable nits, speculative
  claims without a concrete attacker/interleaving/scale, style preferences
  the codebase doesn't share, and test-quality nits on non-test diffs
- smart file-selection guidance for large PRs (deprioritize generated /
  vendored / pure-rename hunks)

Eval config switched to openai:gpt-5.5 + high reasoning effort for the next
benchmark run.

* trim prompt

* subagent prompting

* confidence ratings

* added medium

* enforce confidence threshold

* .

* reviewer: precision-tuned prompt + drop confidence gate

Rewrites the reviewer system prompt around a defensibility bar (anchor +
failure mode + maintainer wouldn't say "not a bug"), an explicit do-not-file
list (style nits, speculation, scope-policing, same-bug fan-out), and a
checklist of 10 bug archetypes drawn from a per-PR audit of the eval golden
set. The audit showed 145 FPs in the last eval split ~28% speculative, ~26%
style-nit, ~31% real-but-unscored (mostly same-archetype fan-out); the new
prompt targets each class directly.

Confidence is still recorded on every finding for post-hoc calibration but
no longer gates publication — the audit showed the gate was a no-op (agent
self-rated 65% of findings "high" regardless), and the prompt's defensibility
bar is the actual discipline. Drops CONFIDENCE_ORDER, CONFIDENCE_THRESHOLD,
the confidence_threshold kwarg on filter_findings_for_publish, the
confidence_filtered score_mode, and the min_confidence kwarg on the eval
target's _extract_comments — all dead once the gate is gone.

Also removes the "informational" severity tier from the Severity enum,
SEVERITY_ORDER, and all validators / tests / docstrings. It was reserved for
FYI observations the dataset never rewards.

* benchmax

* adding google provider

* slight steering

* tuning

* more tuning

* fix

* cleanup

* reducing overfitting

* Add per-repo review style profiles and inject them into the reviewer.

Dashboard users can analyze historical PR review feedback per repository,
edit the resulting style guide, and have it loaded from LangGraph Store at
reviewer runtime (including Martian eval runs) keyed by owner/name.

Co-authored-by: Cursor <cursoragent@cursor.com>

* Fix review style job errors leaking exception details to clients.

Return generic dashboard messages while logging full stack traces server-side.

Co-authored-by: Cursor <cursoragent@cursor.com>

---------

Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-05-20 18:35:00 +00:00

449 lines
17 KiB
Python

"""Main entry point and CLI loop for Open SWE agent."""
# ruff: noqa: E402
# Suppress deprecation warnings from langchain_core (e.g., Pydantic V1 on Python 3.14+)
# ruff: noqa: E402
import logging
import os
import warnings
from typing import Any
logger = logging.getLogger(__name__)
from langgraph.graph.state import RunnableConfig
from langgraph.pregel import Pregel
from langgraph_sdk import get_client
warnings.filterwarnings("ignore", module="langchain_core._api.deprecation")
import asyncio
# Suppress Pydantic v1 compatibility warnings from langchain on Python 3.14+
warnings.filterwarnings("ignore", message=".*Pydantic V1.*", category=UserWarning)
# Now safe to import agent (which imports LangChain modules)
from deepagents import create_deep_agent
from deepagents.backends import LangSmithSandbox
from deepagents.backends.protocol import SandboxBackendProtocol
from langchain.agents.middleware import ModelCallLimitMiddleware
from langsmith.sandbox import SandboxClientError
from .dashboard.agent_overrides import (
load_profile,
normalize_profile_overrides,
resolve_github_login,
)
from .integrations.langsmith import _configure_github_proxy
from .middleware import (
ModelFallbackMiddleware,
SandboxCircuitBreakerMiddleware,
SanitizeToolInputsMiddleware,
SlackAssistantStatusMiddleware,
ToolErrorMiddleware,
check_message_queue_before_model,
ensure_no_empty_msg,
notify_step_limit_reached,
)
from .prompt import construct_system_prompt
from .tools import (
fetch_url,
http_request,
linear_comment,
linear_create_issue,
linear_delete_issue,
linear_get_issue,
linear_get_issue_comments,
linear_list_teams,
linear_update_issue,
request_pr_review,
slack_read_thread_messages,
slack_thread_reply,
web_search,
)
from .utils.auth import resolve_github_token
from .utils.authorship import resolve_triggering_user_identity
from .utils.github_app import get_github_app_installation_token
from .utils.model import (
DEFAULT_LLM_REASONING,
ModelKwargs,
fallback_model_id_for,
make_model,
provider_model_kwargs,
)
from .utils.sandbox import create_sandbox
from .utils.sandbox_paths import aresolve_sandbox_work_dir
client = get_client()
SANDBOX_CREATING = "__creating__"
SANDBOX_CREATION_TIMEOUT = 180
SANDBOX_POLL_INTERVAL = 1.0
from .utils.sandbox_state import (
SANDBOX_BACKENDS,
get_sandbox_id_from_metadata,
set_sandbox_backend,
unwrap_sandbox_backend,
)
async def _start_langsmith_sandbox_if_needed(sandbox_backend: SandboxBackendProtocol) -> None:
"""Start a LangSmith sandbox before operations that require it to be running."""
if os.getenv("SANDBOX_TYPE", "langsmith") != "langsmith":
return
current_backend = unwrap_sandbox_backend(sandbox_backend)
if not isinstance(current_backend, LangSmithSandbox):
return
sandbox = current_backend._sandbox # noqa: SLF001
status = await asyncio.to_thread(sandbox._client.get_sandbox_status, sandbox.name) # noqa: SLF001
status_name = getattr(status, "status", status)
status_name = getattr(status_name, "value", status_name)
status_text = str(status_name or "").lower()
if status_text in {"running", "ready"}:
return
logger.info(
"Starting LangSmith sandbox %s before proxy refresh (status=%s)",
current_backend.id,
status_text or "unknown",
)
await asyncio.to_thread(sandbox.start)
async def _create_sandbox_with_proxy() -> SandboxBackendProtocol:
"""Create a new sandbox with GitHub proxy auth configured.
Uses create_sandbox (generic factory) so non-langsmith providers still work.
For langsmith sandboxes, configures the proxy with the installation token.
"""
sandbox_backend = await asyncio.to_thread(create_sandbox)
sandbox_type = os.getenv("SANDBOX_TYPE", "langsmith")
if sandbox_type == "langsmith":
installation_token = await get_github_app_installation_token()
if not installation_token:
msg = "Cannot configure proxy: GitHub App installation token is unavailable"
logger.error(msg)
raise ValueError(msg)
await _start_langsmith_sandbox_if_needed(sandbox_backend)
await asyncio.to_thread(_configure_github_proxy, sandbox_backend.id, installation_token)
return sandbox_backend
async def _refresh_github_proxy(
sandbox_backend: SandboxBackendProtocol,
) -> None:
"""Refresh GitHub proxy credentials for reused LangSmith sandboxes."""
if os.getenv("SANDBOX_TYPE", "langsmith") != "langsmith":
return
installation_token = await get_github_app_installation_token()
if not installation_token:
logger.warning(
"Skipping GitHub proxy refresh for sandbox %s: installation token unavailable",
sandbox_backend.id,
)
return
current_backend = unwrap_sandbox_backend(sandbox_backend)
await _start_langsmith_sandbox_if_needed(current_backend)
await asyncio.to_thread(_configure_github_proxy, current_backend.id, installation_token)
async def _refresh_github_proxy_or_recreate(
sandbox_backend: SandboxBackendProtocol,
thread_id: str,
) -> SandboxBackendProtocol:
"""Refresh proxy credentials, recreating stale LangSmith sandboxes on failure."""
try:
await _refresh_github_proxy(sandbox_backend)
except Exception: # noqa: BLE001
logger.warning(
"Failed to refresh GitHub proxy for sandbox %s on thread %s, recreating sandbox",
sandbox_backend.id,
thread_id,
exc_info=True,
)
return await _recreate_sandbox(thread_id)
return sandbox_backend
async def _configure_git_identity(sandbox_backend: SandboxBackendProtocol) -> None:
await asyncio.to_thread(
sandbox_backend.execute,
"git config --global user.name 'open-swe[bot]' && "
"git config --global user.email 'open-swe@users.noreply.github.com'",
)
async def _recreate_sandbox(thread_id: str) -> SandboxBackendProtocol:
"""Recreate a sandbox after a connection failure.
Sets the SANDBOX_CREATING sentinel and creates a fresh sandbox
(with proxy auth configured), swapping the per-thread proxy target.
The agent is responsible for cloning repos via tools.
"""
await client.threads.update(
thread_id=thread_id,
metadata={"sandbox_id": SANDBOX_CREATING},
)
try:
sandbox_backend = set_sandbox_backend(thread_id, await _create_sandbox_with_proxy())
except Exception:
logger.exception("Failed to recreate sandbox after connection failure")
await client.threads.update(thread_id=thread_id, metadata={"sandbox_id": None})
raise
return sandbox_backend
async def check_or_recreate_sandbox(
sandbox_backend: SandboxBackendProtocol, thread_id: str
) -> SandboxBackendProtocol:
"""Check if a cached sandbox is reachable; recreate it if not.
Pings the sandbox with a lightweight command. If the sandbox is
unreachable (SandboxClientError), it is torn down and a fresh one
is created via _recreate_sandbox.
Returns the original backend if healthy, or a new one if recreated.
"""
try:
await asyncio.to_thread(sandbox_backend.execute, "echo ok")
except SandboxClientError:
logger.warning(
"Cached sandbox is no longer reachable for thread %s, recreating",
thread_id,
)
sandbox_backend = await _recreate_sandbox(thread_id)
return sandbox_backend
async def _wait_for_sandbox_id(thread_id: str) -> str:
"""Wait for sandbox_id to be set in thread metadata.
Polls thread metadata until sandbox_id is set to a real value
(not the creating sentinel).
Raises:
TimeoutError: If sandbox creation takes too long
"""
elapsed = 0.0
while elapsed < SANDBOX_CREATION_TIMEOUT:
sandbox_id = await get_sandbox_id_from_metadata(thread_id)
if sandbox_id is not None and sandbox_id != SANDBOX_CREATING:
return sandbox_id
await asyncio.sleep(SANDBOX_POLL_INTERVAL)
elapsed += SANDBOX_POLL_INTERVAL
msg = f"Timeout waiting for sandbox creation for thread {thread_id}"
raise TimeoutError(msg)
def graph_loaded_for_execution(config: RunnableConfig) -> bool:
"""Check if the graph is loaded for actual execution vs introspection."""
return (
config["configurable"].get("__is_for_execution__", False)
if "configurable" in config
else False
)
async def ensure_sandbox_for_thread(thread_id: str) -> SandboxBackendProtocol:
"""Get-or-create a healthy sandbox bound to ``thread_id``.
Implements the four-state lifecycle described in AGENTS.md:
1. Cached in memory → ping; recreate on ``SandboxClientError``.
2. Metadata says ``__creating__`` and no cache → poll until ready.
3. No sandbox at all → create one and persist the id.
4. Metadata has an id but no cache → reconnect; recreate on failure.
For LangSmith sandboxes, also refreshes the GitHub App proxy auth.
Persists the resulting ``sandbox_id`` to thread metadata, and on the
first creation/reconnect for this thread initializes git identity.
"""
sandbox_backend = SANDBOX_BACKENDS.get(thread_id)
sandbox_id = await get_sandbox_id_from_metadata(thread_id)
if sandbox_id == SANDBOX_CREATING and not sandbox_backend:
logger.info("Sandbox creation in progress for thread %s, waiting...", thread_id)
sandbox_id = await _wait_for_sandbox_id(thread_id)
if sandbox_backend:
logger.info("Using cached sandbox backend for thread %s", thread_id)
original_sandbox_id = sandbox_backend.id
sandbox_backend = await check_or_recreate_sandbox(sandbox_backend, thread_id)
if sandbox_backend.id == original_sandbox_id:
sandbox_backend = await _refresh_github_proxy_or_recreate(sandbox_backend, thread_id)
elif sandbox_id is None:
logger.info("Creating new sandbox for thread %s", thread_id)
await client.threads.update(thread_id=thread_id, metadata={"sandbox_id": SANDBOX_CREATING})
try:
sandbox_backend = await _create_sandbox_with_proxy()
logger.info("Sandbox created: %s", sandbox_backend.id)
except Exception:
logger.exception("Failed to create sandbox")
try:
await client.threads.update(thread_id=thread_id, metadata={"sandbox_id": None})
except Exception:
logger.exception("Failed to reset sandbox_id metadata")
raise
else:
logger.info("Connecting to existing sandbox %s", sandbox_id)
created_replacement_sandbox = False
try:
sandbox_backend = await asyncio.to_thread(create_sandbox, sandbox_id)
except Exception:
logger.warning("Failed to connect to existing sandbox %s, creating new one", sandbox_id)
await client.threads.update(
thread_id=thread_id, metadata={"sandbox_id": SANDBOX_CREATING}
)
try:
sandbox_backend = await _create_sandbox_with_proxy()
created_replacement_sandbox = True
except Exception:
logger.exception("Failed to create replacement sandbox")
await client.threads.update(thread_id=thread_id, metadata={"sandbox_id": None})
raise
if not created_replacement_sandbox:
original_sandbox_id = sandbox_backend.id
sandbox_backend = await check_or_recreate_sandbox(sandbox_backend, thread_id)
if sandbox_backend.id == original_sandbox_id:
sandbox_backend = await _refresh_github_proxy_or_recreate(
sandbox_backend, thread_id
)
sandbox_backend = set_sandbox_backend(thread_id, sandbox_backend)
if sandbox_id != sandbox_backend.id:
await client.threads.update(
thread_id=thread_id, metadata={"sandbox_id": sandbox_backend.id}
)
# Re-apply git identity every run: cached/reconnected sandboxes may have
# lost their `--global` config (or had it overwritten), and Vercel preview
# deploys reject commits whose author email can't be resolved to a GitHub
# account.
await _configure_git_identity(sandbox_backend)
return sandbox_backend
DEFAULT_LLM_MODEL_ID = "openai:gpt-5.5"
DEFAULT_LLM_MAX_TOKENS = 64_000
DEFAULT_RECURSION_LIMIT = 9_999
MODEL_CALL_RECURSION_LIMIT = 5_000 # ~half the recursion limit to account for tool calls
def _get_cached_sandbox_backend(thread_id: str) -> SandboxBackendProtocol:
sandbox_backend = SANDBOX_BACKENDS.get(thread_id)
if sandbox_backend is None:
raise RuntimeError(f"No sandbox backend cached for thread {thread_id}")
return sandbox_backend
async def get_agent(config: RunnableConfig) -> Pregel:
"""Get or create an agent with a sandbox for the given thread."""
thread_id = config["configurable"].get("thread_id", None)
config["recursion_limit"] = DEFAULT_RECURSION_LIMIT
if thread_id is None or not graph_loaded_for_execution(config):
logger.info("No thread_id or not for execution, returning agent without sandbox")
return create_deep_agent(
system_prompt="",
tools=[],
).with_config(config)
github_token, new_encrypted, new_expires_at = await resolve_github_token(config, thread_id)
config["metadata"]["github_token_encrypted"] = new_encrypted
config["metadata"]["github_token_expires_at"] = new_expires_at
triggering_user_identity = await asyncio.to_thread(
resolve_triggering_user_identity, config, github_token
)
del github_token
sandbox_backend = await ensure_sandbox_for_thread(thread_id)
linear_issue = config["configurable"].get("linear_issue", {})
linear_project_id = linear_issue.get("linear_project_id", "")
linear_issue_number = linear_issue.get("linear_issue_number", "")
work_dir = await aresolve_sandbox_work_dir(sandbox_backend)
def backend_factory(_runtime: object, _thread_id: str = thread_id) -> SandboxBackendProtocol:
return _get_cached_sandbox_backend(_thread_id)
model_id = os.environ.get("LLM_MODEL_ID", DEFAULT_LLM_MODEL_ID)
profile_effort: str | None = None
profile_login = resolve_github_login(config)
if profile_login:
profile = await load_profile(profile_login)
if profile:
overridden_model, overridden_effort = normalize_profile_overrides(profile)
if overridden_model:
logger.info(
"Applying dashboard profile override for %s: model=%s effort=%s",
profile_login,
overridden_model,
overridden_effort,
)
model_id = overridden_model
profile_effort = overridden_effort
model_kwargs = provider_model_kwargs(
model_id,
profile_effort,
max_tokens=DEFAULT_LLM_MAX_TOKENS,
)
fallback_model_id = os.environ.get("LLM_FALLBACK_MODEL_ID") or fallback_model_id_for(model_id)
fallback_middleware: list[Any] = []
if fallback_model_id and fallback_model_id != model_id:
fallback_kwargs: ModelKwargs = {"max_tokens": DEFAULT_LLM_MAX_TOKENS}
if fallback_model_id.startswith("openai:"):
fallback_kwargs["reasoning"] = DEFAULT_LLM_REASONING
fallback_middleware.append(
ModelFallbackMiddleware(make_model(fallback_model_id, **fallback_kwargs))
)
logger.info("Configured model fallback %s -> %s", model_id, fallback_model_id)
logger.info("Returning agent with sandbox for thread %s", thread_id)
return create_deep_agent(
model=make_model(model_id, **model_kwargs),
system_prompt=construct_system_prompt(
working_dir=work_dir,
linear_project_id=linear_project_id,
linear_issue_number=linear_issue_number,
triggering_user_identity=triggering_user_identity,
),
tools=[
http_request,
fetch_url,
web_search,
linear_comment,
linear_create_issue,
linear_delete_issue,
linear_get_issue,
linear_get_issue_comments,
linear_list_teams,
linear_update_issue,
request_pr_review,
slack_read_thread_messages,
slack_thread_reply,
],
backend=backend_factory,
middleware=[
SanitizeToolInputsMiddleware(),
ModelCallLimitMiddleware(run_limit=MODEL_CALL_RECURSION_LIMIT, exit_behavior="end"),
ToolErrorMiddleware(),
check_message_queue_before_model,
SlackAssistantStatusMiddleware(),
ensure_no_empty_msg,
notify_step_limit_reached,
SandboxCircuitBreakerMiddleware(),
*fallback_middleware,
],
).with_config(config)