feat: persist automation task outcomes

migrations/versions/015_add_run_task_outcome.py

@@ -0,0 +1,25 @@
+"""Add task_outcome column to automation_runs table.
+
+Revision ID: 015
+Revises: 014
+Create Date: 2026-08-12
+"""
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from alembic import op
+
+
+revision: str = "015"
+down_revision: str = "014"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    op.add_column("automation_runs", sa.Column("task_outcome", sa.JSON, nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("automation_runs", "task_outcome")

openhands/automation/dispatcher.py

@@ -244,6 +244,8 @@ async def _fail(error: str, disable: bool = False) -> None:
     env_vars = backend.build_env_vars()
     env_vars["AUTOMATION_CALLBACK_URL"] = callback_url
     env_vars["AUTOMATION_RUN_ID"] = run_id
+    env_vars["AUTOMATION_CALLBACK_SUPPORTS_AGENT_OUTCOME"] = "true"
+
     env_vars["AUTOMATION_USER_ID"] = str(automation.user_id)
     env_vars["AUTOMATION_ORG_ID"] = str(automation.org_id)
     env_vars["AUTOMATION_API_URL"] = settings.resolved_base_url

openhands/automation/execution.py

@@ -519,6 +519,9 @@ async def run_automation(
         env_vars["AUTOMATION_CALLBACK_URL"] = callback_url
     if run_id:
         env_vars["AUTOMATION_RUN_ID"] = run_id
+    if callback_url:
+        env_vars["AUTOMATION_CALLBACK_SUPPORTS_AGENT_OUTCOME"] = "true"
+
     api_url = api_url.rstrip("/")
     sandbox_id: str | None = None
 

openhands/automation/models.py

@@ -164,6 +164,11 @@ class AutomationRun(Base):
     # force-terminated by the watchdog / cancelled so no callback ever fired.
     cost: Mapped[float | None] = mapped_column(Float, nullable=True)
 
+    # Latest structured task outcome reported by the SDK/agent or recovered
+    # from the agent-server when the completion callback is missed.
+    # Uses generic JSON type for cross-database compatibility.
+    task_outcome: Mapped[dict | None] = mapped_column(JSON, nullable=True)
+
     # Pre-computed deadline: started_at + max_duration. Set when transitioning
     # to RUNNING, used by the staleness watchdog for efficient indexed queries.
     timeout_at: Mapped[datetime | None] = mapped_column(

openhands/automation/presets/plugin/sdk_main.py

@@ -86,9 +86,8 @@
 # may be missing here even when the agent-server has a valid session key.
 # We still fall back to SESSION_API_KEY for compatibility with cloud-mode
 # deployments and older agent-server versions that only set the bare name.
-session_key = (
-    os.environ.get("OH_SESSION_API_KEYS_0")
-    or os.environ.get("SESSION_API_KEY", "")
+session_key = os.environ.get("OH_SESSION_API_KEYS_0") or os.environ.get(
+    "SESSION_API_KEY", ""
 )
 model_profile = os.environ.get("AUTOMATION_MODEL") or None
 automation_user_id = os.environ.get("AUTOMATION_USER_ID") or None
@@ -100,10 +99,7 @@
 if IS_LOCAL_MODE:
     # Local mode: AGENT_SERVER_URL required
     print(f"  AGENT_SERVER_URL: {'OK' if agent_server_url else 'MISSING'}")
-    print(
-        f"  OH_SESSION_API_KEYS_0: "
-        f"{'OK' if session_key else 'NONE (may fail auth)'}"
-    )
+    print(f"  OH_SESSION_API_KEYS_0: {'OK' if session_key else 'NONE (may fail auth)'}")
     if not agent_server_url:
         print("FAIL: AGENT_SERVER_URL not set for local mode", file=sys.stderr)
         sys.exit(1)
@@ -449,6 +445,7 @@ def event_callback(event) -> None:
     conversation = Conversation(**conversation_kwargs)
     assert isinstance(conversation, RemoteConversation)
     print(f"  conversation created: {type(conversation).__name__}")
+    print(f"  conversation id: {conversation.id}")
     print(f"  plugins loaded: {len(plugin_sources)}")
     if experiment_tags:
         print(f"  experiment tags: {experiment_tags}")

openhands/automation/presets/prompt/sdk_main.py

@@ -89,9 +89,8 @@
 # may be missing here even when the agent-server has a valid session key.
 # We still fall back to SESSION_API_KEY for compatibility with cloud-mode
 # deployments and older agent-server versions that only set the bare name.
-session_key = (
-    os.environ.get("OH_SESSION_API_KEYS_0")
-    or os.environ.get("SESSION_API_KEY", "")
+session_key = os.environ.get("OH_SESSION_API_KEYS_0") or os.environ.get(
+    "SESSION_API_KEY", ""
 )
 model_profile = os.environ.get("AUTOMATION_MODEL") or None
 automation_user_id = os.environ.get("AUTOMATION_USER_ID") or None
@@ -103,10 +102,7 @@
 if IS_LOCAL_MODE:
     # Local mode: AGENT_SERVER_URL required
     print(f"  AGENT_SERVER_URL: {'OK' if agent_server_url else 'MISSING'}")
-    print(
-        f"  OH_SESSION_API_KEYS_0: "
-        f"{'OK' if session_key else 'NONE (may fail auth)'}"
-    )
+    print(f"  OH_SESSION_API_KEYS_0: {'OK' if session_key else 'NONE (may fail auth)'}")
     if not agent_server_url:
         print("FAIL: AGENT_SERVER_URL not set for local mode", file=sys.stderr)
         sys.exit(1)
@@ -401,6 +397,7 @@ def event_callback(event) -> None:
     conversation = Conversation(**conversation_kwargs)
     assert isinstance(conversation, RemoteConversation)
     print(f"  conversation created: {type(conversation).__name__}")
+    print(f"  conversation id: {conversation.id}")
 
     # Set a descriptive title so automation runs are distinguishable in the
     # conversation panel — otherwise the agent-server's autotitle falls back

openhands/automation/router.py

@@ -424,6 +424,12 @@ async def complete_run(
         values["conversation_id"] = body.conversation_id
     if body.cost is not None:
         values["cost"] = body.cost
+    task_outcome = (
+        body.agent_outcome if body.agent_outcome is not None else body.task_outcome
+    )
+    if task_outcome is not None:
+        values["task_outcome"] = task_outcome
+
     if body.status == "FAILED" and body.error:
         values["error_detail"] = body.error
 

openhands/automation/schemas.py

@@ -653,6 +653,14 @@ class RunCompleteRequest(BaseModel):
     conversation_id: str | None = None
     error: str | None = None
     cost: float | None = None
+    agent_outcome: dict[str, Any] | None = Field(
+        default=None,
+        description="Latest SDK task outcome reported in the completion callback.",
+    )
+    task_outcome: dict[str, Any] | None = Field(
+        default=None,
+        description="Alias accepted for clients that send task_outcome directly.",
+    )
 
 
 class AutomationRunResponse(BaseModel):
@@ -664,6 +672,8 @@ class AutomationRunResponse(BaseModel):
     error_detail: str | None
     conversation_id: str | None
     cost: float | None = None
+    task_outcome: dict[str, Any] | None = None
+
     timeout_at: UtcDatetime | None
     sandbox_id: str | None
     bash_command_id: str | None = None

openhands/automation/utils/agent_server.py

@@ -6,6 +6,8 @@
 """
 
 import logging
+import re
+from typing import Any
 
 import httpx
 from pydantic.dataclasses import dataclass
@@ -15,6 +17,49 @@
 
 logger = logging.getLogger(__name__)
 
+_UUID_RE = (
+    r"[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}"
+)
+_CONVERSATION_URL_RE = re.compile(rf"/conversations/(?P<id>{_UUID_RE})")
+_CONVERSATION_ID_RE = re.compile(
+    rf"conversation(?:\s+id)?\s*:\s*(?P<id>{_UUID_RE})",
+    re.IGNORECASE,
+)
+
+
+def extract_conversation_id(output: str) -> str | None:
+    """Extract a conversation id printed by automation preset scripts."""
+    for pattern in (_CONVERSATION_ID_RE, _CONVERSATION_URL_RE):
+        match = pattern.search(output)
+        if match:
+            return match.group("id")
+    return None
+
+
+async def get_conversation_task_outcome(
+    client: httpx.AsyncClient,
+    agent_url: str,
+    session_key: str,
+    conversation_id: str,
+) -> dict[str, Any] | None:
+    """Fetch the latest task_outcome from an agent-server conversation."""
+    try:
+        resp = await client.get(
+            f"{agent_url.rstrip('/')}/api/conversations/{conversation_id}",
+            headers={"X-Session-API-Key": session_key},
+            timeout=30.0,
+        )
+        resp.raise_for_status()
+        outcome = resp.json().get("task_outcome")
+        return outcome if isinstance(outcome, dict) else None
+    except Exception as e:
+        logger.warning(
+            "Failed to fetch task outcome for conversation %s: %s",
+            conversation_id,
+            e,
+        )
+        return None
+
 
 @dataclass(frozen=True)
 class BashCommandResult:
@@ -112,6 +157,8 @@ class VerificationResult:
     stdout: str = ""
     stderr: str = ""
     error: str | None = None
+    conversation_id: str | None = None
+    task_outcome: dict[str, Any] | None = None
 
 
 async def verify_run_on_agent_server(
@@ -168,6 +215,12 @@ async def verify_run_on_agent_server(
             )
 
         success = bash_result.exit_code == 0
+        conversation_id = extract_conversation_id(bash_result.stdout)
+        task_outcome = None
+        if conversation_id:
+            task_outcome = await get_conversation_task_outcome(
+                client, agent_url, session_key, conversation_id
+            )
         logger.info(
             "Verified run status: exit_code=%s, success=%s",
             bash_result.exit_code,
@@ -181,4 +234,6 @@ async def verify_run_on_agent_server(
             exit_code=bash_result.exit_code,
             stdout=bash_result.stdout,
             stderr=bash_result.stderr,
+            conversation_id=conversation_id,
+            task_outcome=task_outcome,
         )

openhands/automation/utils/sandbox.py

@@ -14,6 +14,8 @@
 from openhands.automation.utils.agent_server import (
     BashCommandResult,
     VerificationResult,
+    extract_conversation_id,
+    get_conversation_task_outcome,
     get_last_bash_command_result,
 )
 from openhands.automation.utils.log_context import log_extra
@@ -195,6 +197,12 @@ async def verify_run_status(
             )
 
         success = bash_result.exit_code == 0
+        conversation_id = extract_conversation_id(bash_result.stdout)
+        task_outcome = None
+        if conversation_id:
+            task_outcome = await get_conversation_task_outcome(
+                client, agent_url, session_key, conversation_id
+            )
         logger.info(
             "Verified run status: exit_code=%s, success=%s",
             bash_result.exit_code,
@@ -208,4 +216,6 @@ async def verify_run_status(
             exit_code=bash_result.exit_code,
             stdout=bash_result.stdout,
             stderr=bash_result.stderr,
+            conversation_id=conversation_id,
+            task_outcome=task_outcome,
         )

openhands/automation/watchdog.py

@@ -13,6 +13,7 @@
 
 import asyncio
 import logging
+from typing import Any
 
 from sqlalchemy import inspect, select, update
 from sqlalchemy.engine import CursorResult
@@ -117,6 +118,12 @@ async def _verify_and_mark_run(
         return result.rowcount > 0
 
     if verification.verified:
+        verification_values: dict[str, Any] = {}
+        if verification.conversation_id:
+            verification_values["conversation_id"] = verification.conversation_id
+        if verification.task_outcome is not None:
+            verification_values["task_outcome"] = verification.task_outcome
+
         exit_code = verification.exit_code
 
         # exit_code == 0: Command completed successfully, we just missed the callback
@@ -136,6 +143,7 @@ async def _verify_and_mark_run(
                 .values(
                     status=AutomationRunStatus.COMPLETED,
                     completed_at=now,
+                    **verification_values,
                 )
             )
 
@@ -160,6 +168,7 @@ async def _verify_and_mark_run(
                     status=AutomationRunStatus.FAILED,
                     completed_at=now,
                     error_detail=f"Timed out: {error_msg}",
+                    **verification_values,
                 )
             )
 
@@ -187,6 +196,7 @@ async def _verify_and_mark_run(
                     status=AutomationRunStatus.FAILED,
                     completed_at=now,
                     error_detail=error_detail,
+                    **verification_values,
                 )
             )
 

pyproject.toml

@@ -21,8 +21,8 @@ dependencies = [
   "google-cloud-storage>=2.18",
   "httpx>=0.27",
   "jmespath>=1.0",
-  "openhands-sdk==1.40.1",
-  "openhands-workspace==1.40.1",
+  "openhands-sdk @ git+https://github.com/OpenHands/software-agent-sdk.git@be17960b7846fdfc4bd66ce2fb6dd40bd890f6b3#subdirectory=openhands-sdk",
+  "openhands-workspace @ git+https://github.com/OpenHands/software-agent-sdk.git@be17960b7846fdfc4bd66ce2fb6dd40bd890f6b3#subdirectory=openhands-workspace",
   "pg8000>=1.31",
   "prometheus-client>=0.19",
   "pydantic>=2",

tests/test_db.py

@@ -260,6 +260,8 @@ def test_migrations_run_on_sqlite(self, monkeypatch):
             }
             assert "cost" in run_columns
 
+            assert "task_outcome" in run_columns
+
             engine.dispose()
         finally:
             # Clean up