devcouncil 0.3.0 → 0.4.0

package/src/devcouncil/executors/claude_sdk.py

@@ -0,0 +1,376 @@
+"""In-process Claude executor backed by the Claude Agent SDK.
+
+Unlike :class:`CodingCliExecutor`, which shells out to ``claude -p`` and can only enforce
+scope *after* the subprocess has already written to disk, this executor arbitrates every
+tool call live through the SDK's ``can_use_tool`` callback. The callback is backed by the
+same :class:`HookPolicy` the shell write-gate uses, so an out-of-scope write or command is
+denied *before it happens* — real containment that, unlike the shell PreToolUse gate, does
+not fail-closed on an interactive session (the decision is evaluated against the run's own
+task, not a globally-resolved lease).
+
+The Claude Agent SDK is an OPTIONAL dependency: it is imported lazily so nothing else in
+DevCouncil requires it, and a missing install produces an actionable error rather than an
+import-time crash. Wire it up with ``pip install claude-agent-sdk``.
+"""
+
+from __future__ import annotations
+
+import importlib
+import logging
+from pathlib import Path
+from typing import Any, Coroutine, List, Optional, TypeVar
+
+from devcouncil.domain.requirement import Requirement
+from devcouncil.domain.task import Task
+from devcouncil.execution.executor import ExecutionResult, Executor
+from devcouncil.execution.hook_policy import HookDecision, HookPolicy
+from devcouncil.execution.prompt_builder import PromptBuilder
+from devcouncil.executors.transient_retry import transient_error_in_text
+from devcouncil.telemetry.stages import log_step
+
+logger = logging.getLogger(__name__)
+
+_T = TypeVar("_T")
+
+
+class ClaudeSdkExecutor(Executor):
+    """Run a DevCouncil task through the Claude Agent SDK with lease-aware gating."""
+
+    def __init__(
+        self,
+        project_root: Path | str,
+        *,
+        active_task: Task | None = None,
+        model: str | None = None,
+        advisor_model: str | None = None,
+        permission_mode: str = "default",
+        env: dict[str, str] | None = None,
+    ):
+        # NOTE: default (not acceptEdits). acceptEdits auto-approves file edits, which would
+        # skip the can_use_tool callback entirely and defeat the lease gate; "default" routes
+        # every write/command through our callback so scope is actually enforced.
+        self.project_root = Path(project_root)
+        # The task whose scope authorizes writes/commands. Passed explicitly so the gate is
+        # evaluated against THIS run's task rather than a globally-resolved lease (which is
+        # why this can enforce interactively where the shell PreToolUse gate cannot).
+        self.active_task = active_task
+        self.model = model
+        self.advisor_model = (advisor_model or "").strip() or None
+        self.permission_mode = permission_mode
+        # Extra environment for the SDK's underlying Claude Code process. Lets a run
+        # target an alternative Anthropic-compatible endpoint (ANTHROPIC_BASE_URL /
+        # ANTHROPIC_AUTH_TOKEN, e.g. a local proxy) — same knob the subprocess
+        # executor exposes via per-profile ``env`` overrides.
+        self.env = dict(env) if env else {}
+        self.policy = HookPolicy(project_root=self.project_root)
+        self.last_agent_session_id: str | None = None
+        # (tool_name, reason) for every call the gate denied — surfaced in the result.
+        self.denials: list[tuple[str, str]] = []
+
+    # -- policy ---------------------------------------------------------------------
+
+    def permission_decision(self, tool_name: str, tool_input: dict[str, Any]) -> HookDecision:
+        """Lease-aware allow/deny for a single tool call, reusing DevCouncil's HookPolicy.
+
+        Pure and side-effect free (bar reading repo policy) so it is unit-testable on its
+        own: the same call the SDK would make, evaluated against this run's task."""
+        call_data = {"tool_name": tool_name, "tool_input": tool_input or {}}
+        return self.policy.evaluate(call_data, self.active_task)
+
+    # -- SDK loading ----------------------------------------------------------------
+
+    @staticmethod
+    def _load_sdk():
+        try:
+            return importlib.import_module("claude_agent_sdk")
+        except ModuleNotFoundError as exc:
+            raise RuntimeError(
+                "The Claude Agent SDK is not installed. Install it with "
+                "`pip install claude-agent-sdk` to use the claude-sdk executor."
+            ) from exc
+
+    @staticmethod
+    def _allow(sdk, tool_input: dict[str, Any]):
+        """Build the SDK's 'allow' permission result, preferring its typed class."""
+        allow_cls = getattr(sdk, "PermissionResultAllow", None)
+        if allow_cls is not None:
+            try:
+                return allow_cls(updated_input=tool_input)
+            except TypeError:
+                return allow_cls()
+        # Documented plain-object contract as a fallback.
+        return {"behavior": "allow", "updatedInput": tool_input}
+
+    @staticmethod
+    def _deny(sdk, reason: str):
+        deny_cls = getattr(sdk, "PermissionResultDeny", None)
+        if deny_cls is not None:
+            try:
+                return deny_cls(message=reason)
+            except TypeError:
+                return deny_cls(reason)
+        return {"behavior": "deny", "message": reason}
+
+    # -- run ------------------------------------------------------------------------
+
+    def run_task(self, task: Task, requirements: List[Requirement]) -> ExecutionResult:
+        # A task can be passed at construction (for the gate) or here; the run argument wins.
+        if task is not None:
+            self.active_task = task
+        log_step(
+            f"executor/claude-sdk: starting task {task.id}",
+            project_root=self.project_root,
+            task_id=task.id,
+        )
+        try:
+            result = self._run_coroutine_from_sync(self._run_async(task, requirements))
+            log_step(
+                f"executor/claude-sdk: finished task {task.id}",
+                project_root=self.project_root,
+                task_id=task.id,
+                success=result.success,
+            )
+            return result
+        except RuntimeError as exc:
+            # Missing SDK or a nested event loop — report cleanly, never crash the caller.
+            logger.error("claude-sdk executor could not run %s: %s", getattr(task, "id", "?"), exc)
+            log_step(
+                f"executor/claude-sdk: finished task {getattr(task, 'id', '?')}",
+                project_root=self.project_root,
+                task_id=getattr(task, "id", None),
+                success=False,
+                error=str(exc),
+            )
+            return ExecutionResult(success=False, message=str(exc))
+
+    @staticmethod
+    def _run_coroutine_from_sync(coro: Coroutine[Any, Any, _T]) -> _T:
+        """Run an async coroutine from sync code, including nested-loop contexts.
+
+        ``asyncio.run()`` raises when a loop is already running (pytest-asyncio,
+        Jupyter). With no running loop we use ``asyncio.run()``; otherwise we run
+        the coroutine on a fresh loop in a worker thread.
+        """
+        import asyncio
+        import concurrent.futures
+
+        try:
+            asyncio.get_running_loop()
+        except RuntimeError:
+            return asyncio.run(coro)
+        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as executor:
+            return executor.submit(asyncio.run, coro).result()
+
+    async def _run_async(self, task: Task, requirements: List[Requirement]) -> ExecutionResult:
+        sdk = self._load_sdk()
+        prompt = PromptBuilder(self.project_root).build_task_prompt(task, requirements)
+        # Parity with the subprocess executor: a repair run must carry the correction
+        # manifest so the model knows what the prior attempt got wrong.
+        from devcouncil.planning.correction_manifest import repair_prompt_prefix
+
+        prefix = repair_prompt_prefix(self.project_root, task.id)
+        if prefix:
+            prompt = f"{prefix}{prompt}"
+
+        # Claude-only advisor: resolve attach, build options (old SDKs may drop
+        # extra_args), then steer only when the advisor flag actually landed.
+        advisor_extra = self._resolved_advisor_extra_args()
+
+        async def can_use_tool(tool_name: str, tool_input: dict[str, Any], *args: Any, **kwargs: Any):
+            decision = self.permission_decision(tool_name, tool_input or {})
+            if decision.allowed:
+                return self._allow(sdk, tool_input or {})
+            self.denials.append((tool_name, decision.reason))
+            logger.info("claude-sdk gate denied %s: %s", tool_name, decision.reason)
+            return self._deny(sdk, decision.reason)
+
+        options = self._build_options(sdk, can_use_tool, advisor_extra=advisor_extra)
+        if self._options_have_advisor(options, advisor_extra):
+            from devcouncil.executors.advisor_tool import (
+                advisor_steering_text,
+                advisor_user_cost_trim,
+                warn_advisor_preflight,
+            )
+
+            for warning in warn_advisor_preflight(
+                env=self.env or None,
+                main_model=self.model,
+                advisor_model=self.advisor_model,
+            ):
+                logger.warning("%s", warning)
+            nudge = advisor_steering_text(repair=bool(prefix))
+            cost_trim = advisor_user_cost_trim()
+            prompt = f"{nudge}\n\n{cost_trim}\n\n{prompt}"
+        try:
+            return await self._run_async_once(sdk, prompt, options, task)
+        except Exception as exc:  # a mid-run SDK failure is a failed execution, not a crash
+            logger.exception(
+                "claude-sdk executor raised mid-run for %s: %s",
+                getattr(task, "id", "?"),
+                exc,
+            )
+            reason = transient_error_in_text(str(exc))
+            if reason is not None:
+                import time
+
+                logger.warning(
+                    "claude-sdk transient failure for %s (%s); retrying once in 5s",
+                    getattr(task, "id", "?"),
+                    reason,
+                )
+                time.sleep(5.0)
+                try:
+                    return await self._run_async_once(sdk, prompt, options, task)
+                except Exception as retry_exc:
+                    logger.exception(
+                        "claude-sdk retry failed for %s: %s",
+                        getattr(task, "id", "?"),
+                        retry_exc,
+                    )
+                    return ExecutionResult(
+                        success=False,
+                        message=f"Claude Agent SDK run failed after retry: {retry_exc}",
+                    )
+            return ExecutionResult(success=False, message=f"Claude Agent SDK run failed: {exc}")
+
+    async def _run_async_once(self, sdk, prompt: str, options, task: Task) -> ExecutionResult:
+        result_text: Optional[str] = None
+        is_error = False
+        async for message in sdk.query(prompt=prompt, options=options):
+            session_id = self._message_session_id(message)
+            if session_id:
+                self.last_agent_session_id = session_id
+            text, error = self._message_result(message)
+            if text is not None:
+                result_text = text
+            if error:
+                is_error = True
+        if self.denials:
+            denied = "; ".join(f"{name} ({reason})" for name, reason in self.denials)
+            suffix = f" Gate denied {len(self.denials)} out-of-scope call(s): {denied}."
+        else:
+            suffix = ""
+        message = (result_text or "Claude Agent SDK run finished.") + suffix
+        return ExecutionResult(success=not is_error, message=message)
+
+    def _resolved_advisor_extra_args(self) -> dict[str, str] | None:
+        """Build SDK ``extra_args={"advisor": ...}`` when pairing looks safe."""
+        if not self.advisor_model:
+            return None
+        from devcouncil.executors.advisor_tool import decide_advisor_attach
+
+        decision, resolved, reason = decide_advisor_attach(
+            main_model=self.model,
+            advisor_model=self.advisor_model,
+            env=self.env,
+        )
+        if decision != "attach" or not resolved:
+            if reason:
+                logger.warning(
+                    "Skipping SDK advisor %s: %s",
+                    self.advisor_model,
+                    reason,
+                )
+            return None
+        return {"advisor": resolved}
+
+    @staticmethod
+    def _options_have_advisor(options: Any, advisor_extra: dict[str, str] | None) -> bool:
+        """True when advisor was requested and survived into the SDK options object."""
+        if not advisor_extra:
+            return False
+        if isinstance(options, dict):
+            extra = options.get("extra_args") or {}
+        else:
+            extra = getattr(options, "extra_args", None) or {}
+        if not isinstance(extra, dict):
+            return False
+        return bool(extra.get("advisor"))
+
+    def _build_options(self, sdk, can_use_tool, *, advisor_extra: dict[str, str] | None = None):
+        options_cls = getattr(sdk, "ClaudeAgentOptions", None)
+        kwargs: dict[str, Any] = {
+            "cwd": str(self.project_root),
+            "permission_mode": self.permission_mode,
+            "can_use_tool": can_use_tool,
+        }
+        if self.model:
+            kwargs["model"] = self.model
+        if self.env:
+            kwargs["env"] = dict(self.env)
+        if advisor_extra:
+            kwargs["extra_args"] = dict(advisor_extra)
+        if options_cls is None:
+            return kwargs
+        # Only pass kwargs the SDK's options object actually accepts, so a version skew in
+        # field names degrades gracefully instead of raising TypeError.
+        try:
+            return options_cls(**kwargs)
+        except TypeError:
+            # Progressively narrower kwarg sets: an SDK too old for ``env`` / ``extra_args``
+            # still gets cwd/permission_mode/model rather than losing all options.
+            obj = None
+            dropped_extra = bool(advisor_extra)
+            for keys in (
+                {"cwd", "permission_mode", "model", "env", "extra_args"},
+                {"cwd", "permission_mode", "model", "env"},
+                {"cwd", "permission_mode", "model"},
+            ):
+                safe = {k: v for k, v in kwargs.items() if k in keys}
+                try:
+                    obj = options_cls(**safe)
+                    if dropped_extra and "extra_args" not in safe:
+                        logger.warning(
+                            "claude-agent-sdk ClaudeAgentOptions rejected extra_args; "
+                            "advisor_model=%r will not be applied. Upgrade claude-agent-sdk.",
+                            self.advisor_model,
+                        )
+                    break
+                except TypeError:
+                    continue
+            if obj is None:
+                obj = options_cls()
+                if dropped_extra:
+                    logger.warning(
+                        "claude-agent-sdk ClaudeAgentOptions could not accept advisor "
+                        "extra_args; advisor_model=%r dropped.",
+                        self.advisor_model,
+                    )
+            if hasattr(obj, "can_use_tool"):
+                obj.can_use_tool = can_use_tool
+            return obj
+
+    @staticmethod
+    def _message_session_id(message: Any) -> str | None:
+        for getter in (lambda: getattr(message, "session_id", None),
+                       lambda: message.get("session_id") if isinstance(message, dict) else None,
+                       lambda: (getattr(message, "data", {}) or {}).get("session_id")):
+            try:
+                value = getter()
+            except Exception:
+                value = None
+            if value:
+                return str(value)
+        return None
+
+    @staticmethod
+    def _message_result(message: Any) -> tuple[str | None, bool]:
+        """Extract final result text and error flag from a terminal result message."""
+        mtype = getattr(message, "type", None)
+        if mtype is None and isinstance(message, dict):
+            mtype = message.get("type")
+        subtype = getattr(message, "subtype", None)
+        if subtype is None and isinstance(message, dict):
+            subtype = message.get("subtype")
+        is_result = mtype == "result" or type(message).__name__ == "ResultMessage"
+        if not is_result:
+            return None, False
+        result = getattr(message, "result", None)
+        if result is None and isinstance(message, dict):
+            result = message.get("result")
+        is_error = getattr(message, "is_error", None)
+        if is_error is None and isinstance(message, dict):
+            is_error = message.get("is_error")
+        if is_error is None:
+            is_error = subtype not in (None, "success")
+        return (str(result) if result is not None else None), bool(is_error)