Add goal-based validation gate to AgenticPhase tasks

ddev/src/ddev/ai/agent/build.py

@@ -6,15 +6,14 @@
 
 from collections.abc import Callable
 from pathlib import Path
-from typing import TYPE_CHECKING, Any
+from typing import Any
 
 from ddev.ai.agent.anthropic_client import AnthropicAgent
 from ddev.ai.agent.base import BaseAgent
+from ddev.ai.phases.config import AgentConfig
+from ddev.ai.phases.goal import GOAL_REVIEWER_SYSTEM_PROMPT
 from ddev.ai.tools.fs.file_registry import FileRegistry
-from ddev.ai.tools.registry import ToolRegistry
-
-if TYPE_CHECKING:
-    from ddev.ai.phases.config import AgentConfig
+from ddev.ai.tools.registry import ToolRegistry, filter_read_only
 
 SubagentBuilder = Callable[
     [str, str, list[str]],  # (system_prompt, owner_id, tool_names)
@@ -24,6 +23,10 @@
     [str, str, SubagentBuilder | None, Path | None],  # system_prompt, owner_id, subagent_builder, log_dir
     tuple[BaseAgent[Any], ToolRegistry],
 ]
+GoalAgentBuilder = Callable[
+    [str],  # owner_id
+    tuple[BaseAgent[Any], ToolRegistry],
+]
 
 
 def _resolve_client(agent_clients: dict[str, Any], provider: str) -> Any:
@@ -159,3 +162,49 @@ def builder(system_prompt: str, owner_id: str, tool_names: list[str]) -> tuple[B
         )
 
     return builder
+
+
+def build_goal_agent(
+    parent_agent_config: AgentConfig,
+    agent_clients: dict[str, Any],
+    file_registry: FileRegistry,
+    owner_id: str,
+) -> tuple[BaseAgent[Any], ToolRegistry]:
+    """Build the reviewer agent + its ToolRegistry.
+
+    Uses the same provider as the parent agent. Model and max_tokens are left at
+    provider defaults — the parent's overrides are intentionally not forwarded.
+    Tools are filtered to the read-only subset of the parent's tool list.
+    """
+    read_only_tool_names = filter_read_only(parent_agent_config.tools)
+    goal_agent_config = AgentConfig(
+        provider=parent_agent_config.provider,
+        tools=read_only_tool_names,
+    )
+
+    return _build_agent_and_registry(
+        agent_config=goal_agent_config,
+        agent_clients=agent_clients,
+        system_prompt=GOAL_REVIEWER_SYSTEM_PROMPT,
+        owner_id=owner_id,
+        tool_names=read_only_tool_names,
+        file_registry=file_registry,
+    )
+
+
+def make_goal_agent_builder(
+    parent_agent_config: AgentConfig,
+    agent_clients: dict[str, Any],
+    file_registry: FileRegistry,
+) -> GoalAgentBuilder:
+    """Return a closure that builds a (reviewer_agent, reviewer_registry) tuple."""
+
+    def builder(owner_id: str) -> tuple[BaseAgent[Any], ToolRegistry]:
+        return build_goal_agent(
+            parent_agent_config=parent_agent_config,
+            agent_clients=agent_clients,
+            file_registry=file_registry,
+            owner_id=owner_id,
+        )
+
+    return builder

ddev/src/ddev/ai/callbacks/callbacks.py

@@ -72,6 +72,18 @@ class OnPhaseFinishCallback(Protocol):
     async def __call__(self, phase_id: str) -> None: ...
 
 
+class OnBeforeGoalCheckCallback(Protocol):
+    """Called immediately before each reviewer agent run for a task with a goal."""
+
+    async def __call__(self, task_name: str, attempt: int) -> None: ...
+
+
+class OnAfterGoalCheckCallback(Protocol):
+    """Called after each reviewer agent run, with the parsed verdict."""
+
+    async def __call__(self, task_name: str, attempt: int, valid: bool, reason: str) -> None: ...
+
+
 # ---------------------------------------------------------------------------
 # CallbackSet and Callbacks
 # ---------------------------------------------------------------------------
@@ -103,6 +115,8 @@ def __init__(self) -> None:
         self._on_before_agent_send: list[OnBeforeAgentSendCallback] = []
         self._on_phase_start: list[OnPhaseStartCallback] = []
         self._on_phase_finish: list[OnPhaseFinishCallback] = []
+        self._on_before_goal_check: list[OnBeforeGoalCheckCallback] = []
+        self._on_after_goal_check: list[OnAfterGoalCheckCallback] = []
 
     async def _fire(self, handlers: list[Any], *args: Any) -> None:
         for handler in handlers:
@@ -174,6 +188,20 @@ def on_phase_finish(self, func: OnPhaseFinishCallback) -> OnPhaseFinishCallback:
     async def fire_phase_finish(self, phase_id: str) -> None:
         await self._fire(self._on_phase_finish, phase_id)
 
+    def on_before_goal_check(self, func: OnBeforeGoalCheckCallback) -> OnBeforeGoalCheckCallback:
+        self._on_before_goal_check.append(func)
+        return func
+
+    async def fire_before_goal_check(self, task_name: str, attempt: int) -> None:
+        await self._fire(self._on_before_goal_check, task_name, attempt)
+
+    def on_after_goal_check(self, func: OnAfterGoalCheckCallback) -> OnAfterGoalCheckCallback:
+        self._on_after_goal_check.append(func)
+        return func
+
+    async def fire_after_goal_check(self, task_name: str, attempt: int, valid: bool, reason: str) -> None:
+        await self._fire(self._on_after_goal_check, task_name, attempt, valid, reason)
+
 
 class Callbacks:
     """Container of CallbackSet instances. Dispatches each fire_* to all contained sets."""
@@ -216,3 +244,11 @@ async def fire_phase_start(self, phase_id: str) -> None:
     async def fire_phase_finish(self, phase_id: str) -> None:
         for s in self._sets:
             await s.fire_phase_finish(phase_id)
+
+    async def fire_before_goal_check(self, task_name: str, attempt: int) -> None:
+        for s in self._sets:
+            await s.fire_before_goal_check(task_name, attempt)
+
+    async def fire_after_goal_check(self, task_name: str, attempt: int, valid: bool, reason: str) -> None:
+        for s in self._sets:
+            await s.fire_after_goal_check(task_name, attempt, valid, reason)