agentengine-sdk-python 0.8.1__py3-none-any.whl → 0.8.2__py3-none-any.whl

ksadk/context_engine/contributors.py

@@ -0,0 +1,344 @@
+"""ContextContributor —— 受控动态 Context 扩展点（方案 §8.7 / §17.2）。
+
+动态上下文（Git、规则文件、Memory Recall、Skill manifest、附件、控制面 policy）必须通过
+Contributor 统一进入，不能由 Runner/Hook/业务代码直接拼接到 Prompt（ADR-012）。Contributor
+只负责产生候选 ``ContextItem``，Planner 拥有是否进入请求的最终决策；外部内容不能借此提升
+权限（trust_level 不高于注册配置）。
+
+首批内置 Contributor（方案 §8.7）：WorkspaceRules / Git / MemoryRecall / SkillManifest /
+Attachment / ControlPlanePolicy。P0/P2 接入顺序见方案 §5.5/§5.8：先 shadow 收集来源与状态，
+P2 再逐个接管真实来源。
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+from dataclasses import dataclass, field
+from typing import Any, Literal, Sequence
+
+from ksadk.context_engine.models import ContextItem, ContextTrustLevel
+
+logger = logging.getLogger(__name__)
+
+ContributorFailureMode = Literal["skip", "warn", "fail"]
+ContributorCacheability = Literal["stable", "turn", "none"]
+
+
+@dataclass(frozen=True)
+class ContributorCapabilities:
+    """Contributor 的能力与约束声明（方案 §8.7）。
+
+    ``trust_level`` 不得高于注册配置（方案 §19）；外部 Hook/MCP 一律 ``untrusted``，不能生成
+    ``platform_safety``、不能改变 ownership、不能绕过 approval。
+    """
+
+    contributor_id: str
+    trust_level: ContextTrustLevel
+    max_tokens: int
+    timeout_ms: int
+    cacheability: ContributorCacheability
+    failure_mode: ContributorFailureMode
+
+
+@dataclass(frozen=True)
+class ContextContributionRequest:
+    """单次 Contributor 贡献请求。"""
+
+    user_input: str
+    session_id: str
+    invocation_id: str
+    workspace_root: str = ""
+    user_id: str = ""
+    agent_id: str = ""
+    metadata: dict[str, Any] = field(default_factory=dict)
+
+
+class ContextContributor:
+    """Contributor 基类。子类实现 ``contribute`` 产生候选 ContextItem。
+
+    返回的 ContextItem 一律按 ``capabilities.trust_level`` 标记信任级别；Planner 仍拥有是否
+    进入请求的最终决策。Contributor 不得自行决定高优先级或 required。
+    """
+
+    capabilities: ContributorCapabilities
+
+    async def contribute(self, request: ContextContributionRequest) -> list[ContextItem]:  # noqa: B027
+        return []
+
+    def id(self) -> str:
+        return self.capabilities.contributor_id
+
+
+def _make_item(
+    *,
+    contributor_id: str,
+    trust_level: ContextTrustLevel,
+    kind: str,
+    content: Any,
+    tokens: int,
+    source: str,
+    score: float | None = None,
+    metadata: dict[str, Any] | None = None,
+) -> ContextItem:
+    return ContextItem(
+        item_id=f"contrib:{contributor_id}:{kind}",
+        kind=kind,  # type: ignore[arg-type]
+        content=content,
+        source=source,
+        trust_level=trust_level,
+        priority=0,
+        estimated_tokens=tokens,
+        required=False,  # Contributor 不得自行声明 required（方案 §8.7）
+        droppable=True,
+        truncatable=True,
+        score=score,
+        metadata={"contributor_id": contributor_id, **(metadata or {})},
+    )
+
+
+# ---- 首批内置 Contributor（接口 + 默认实现，shadow 优先）----
+
+
+class WorkspaceRulesContributor(ContextContributor):
+    """工作区规则文件（AGENTS.md / CLAUDE.md） Contributor（方案 §7.6 / §8.7）。
+
+    复用 ``ksadk.prompts.sources.discover_instruction_files`` 的确定性发现逻辑；``trust_level``
+    固定 ``developer``，不高于平台安全。默认 ``cacheability=turn``（规则文件按 turn 读一次）。
+    """
+
+    def __init__(
+        self,
+        *,
+        max_tokens: int = 12000,
+        timeout_ms: int = 3000,
+        failure_mode: ContributorFailureMode = "skip",
+    ) -> None:
+        self.capabilities = ContributorCapabilities(
+            contributor_id="workspace_rules",
+            trust_level="developer",
+            max_tokens=max_tokens,
+            timeout_ms=timeout_ms,
+            cacheability="turn",
+            failure_mode=failure_mode,
+        )
+
+    async def contribute(self, request: ContextContributionRequest) -> list[ContextItem]:
+        from ksadk.prompts.sources import discover_instruction_files
+
+        sections = discover_instruction_files(request.workspace_root or None)
+        if not sections:
+            return []
+        items: list[ContextItem] = []
+        for index, section in enumerate(sections):
+            items.append(
+                _make_item(
+                    contributor_id=self.capabilities.contributor_id,
+                    trust_level=self.capabilities.trust_level,
+                    kind="resource_manifest",
+                    content=section.content,
+                    tokens=int(section.metadata.get("tokens", 0)) or 1,
+                    source=section.source,
+                    metadata={"path": section.metadata.get("path"), "kind": "rule_file"},
+                )
+            )
+        return items
+
+
+class MemoryRecallContributor(ContextContributor):
+    """Memory Recall Contributor（方案 §8.7 / §10.6）。
+
+    调用 ``MemoryCoordinator.recall`` 召回长期记忆，失败返回空（不污染模型输入，方案 §10.8）。
+    返回的 ContextItem 一律 ``untrusted``，不能覆盖 PromptSection（方案 §8.1 / §19）。
+    """
+
+    def __init__(
+        self,
+        coordinator: Any,
+        *,
+        max_tokens: int = 4000,
+        timeout_ms: int = 3000,
+        top_k: int = 8,
+        min_score: float = 0.45,
+    ) -> None:
+        self._coordinator = coordinator
+        self._top_k = top_k
+        self._min_score = min_score
+        self.capabilities = ContributorCapabilities(
+            contributor_id="memory_recall",
+            trust_level="untrusted",
+            max_tokens=max_tokens,
+            timeout_ms=timeout_ms,
+            cacheability="turn",
+            failure_mode="skip",
+        )
+
+    async def contribute(self, request: ContextContributionRequest) -> list[ContextItem]:
+        from ksadk.memory.coordinator import (
+            agent_user_scope_id,
+            build_search_request,
+            recall_to_context_item,
+        )
+
+        req = build_search_request(
+            query=request.user_input,
+            user_id=agent_user_scope_id(
+                agent_id=request.agent_id,
+                user_id=request.user_id,
+            ),
+            top_k=self._top_k,
+            max_tokens=self.capabilities.max_tokens,
+            min_score=self._min_score,
+        )
+        result = self._coordinator.recall(req)
+        ctx = recall_to_context_item(result)
+        if ctx is None:
+            return []
+        from ksadk.context_engine.tokenizer import get_default_token_counter
+
+        tokens = get_default_token_counter().count_text(ctx["formatted_text"])
+        return [
+            _make_item(
+                contributor_id=self.capabilities.contributor_id,
+                trust_level=self.capabilities.trust_level,
+                kind="recalled_memory",
+                content=ctx["formatted_text"],
+                tokens=tokens,
+                source="memory_provider",
+                score=None,
+                metadata={"recall_count": ctx.get("recall_count", 0), "status": result.status},
+            )
+        ]
+
+
+class SkillManifestContributor(ContextContributor):
+    """Skill manifest Contributor（方案 §7.7 / §8.7）：只暴露 name/desc/version，不进正文。"""
+
+    def __init__(
+        self,
+        manifests: Sequence[dict[str, Any]] | None = None,
+        *,
+        max_tokens: int = 8000,
+        timeout_ms: int = 3000,
+    ) -> None:
+        self._manifests = list(manifests or [])
+        self.capabilities = ContributorCapabilities(
+            contributor_id="skill_manifest",
+            trust_level="resource",
+            max_tokens=max_tokens,
+            timeout_ms=timeout_ms,
+            cacheability="stable",
+            failure_mode="skip",
+        )
+
+    def set_manifests(self, manifests: Sequence[dict[str, Any]]) -> None:
+        self._manifests = list(manifests)
+
+    async def contribute(self, request: ContextContributionRequest) -> list[ContextItem]:
+        if not self._manifests:
+            return []
+        import json
+
+        text = json.dumps(self._manifests, ensure_ascii=False)
+        from ksadk.context_engine.tokenizer import get_default_token_counter
+
+        return [
+            _make_item(
+                contributor_id=self.capabilities.contributor_id,
+                trust_level=self.capabilities.trust_level,
+                kind="resource_manifest",
+                content=text,
+                tokens=get_default_token_counter().count_text(text),
+                source="skill_manifest",
+                metadata={"skill_count": len(self._manifests)},
+            )
+        ]
+
+
+# ---- 并发执行与约束（方案 §8.7 / §17.2）----
+
+
+@dataclass(frozen=True)
+class ContributionResult:
+    """一批 Contributor 的执行结果。"""
+
+    items: list[ContextItem]
+    status: dict[str, str]  # contributor_id → "ok" / "timeout" / "error" / "skipped"
+    warnings: tuple[str, ...]
+
+
+async def run_contributors(
+    contributors: Sequence[ContextContributor],
+    request: ContextContributionRequest,
+    *,
+    default_timeout_ms: int = 3000,
+    default_failure_mode: ContributorFailureMode = "skip",
+) -> ContributionResult:
+    """并发执行 Contributors，各自受超时、预算与 failure policy 约束（方案 §8.7）。
+
+    - 超时 → 该 Contributor 返回空，status=timeout。
+    - 异常 → 按 failure_mode：skip 返空 / warn 返空 + warning / fail 抛给上层。
+    - 返回的 ContextItem 总 token 受各自 ``max_tokens`` 约束（Planner 再做全局预算）。
+    """
+
+    async def _run_one(c: ContextContributor) -> tuple[str, list[ContextItem], str, str | None]:
+        timeout = max(c.capabilities.timeout_ms, 1) / 1000.0
+        try:
+            items = await asyncio.wait_for(c.contribute(request), timeout=timeout)
+        except asyncio.TimeoutError:
+            return c.id(), [], "timeout", f"{c.id()}: timeout"
+        except Exception as exc:  # noqa: BLE001
+            if c.capabilities.failure_mode == "fail":
+                raise
+            return c.id(), [], "error", f"{c.id()}: {exc}"
+        # 单 Contributor 总 token 约束
+        if items and sum(i.estimated_tokens for i in items) > c.capabilities.max_tokens:
+            items = items[: c.capabilities.max_tokens]  # best-effort 限条数
+        return c.id(), items, "ok", None
+
+    results = await asyncio.gather(*[_run_one(c) for c in contributors], return_exceptions=True)
+    all_items: list[ContextItem] = []
+    status: dict[str, str] = {}
+    warnings: list[str] = []
+    for contributor, r in zip(contributors, results):
+        if isinstance(r, Exception):
+            if contributor.capabilities.failure_mode == "fail":
+                raise r
+            warnings.append(f"{contributor.id()}: {r}")
+            status[contributor.id()] = "error"
+            continue
+        cid, items, st, warn = r
+        status[cid] = st
+        all_items.extend(items)
+        if warn:
+            warnings.append(warn)
+    return ContributionResult(items=all_items, status=status, warnings=tuple(warnings))
+
+
+def run_contributors_sync(
+    contributors: Sequence[ContextContributor],
+    request: ContextContributionRequest,
+    **kwargs: Any,
+) -> ContributionResult:
+    """同步入口（无运行中事件循环时用 ``asyncio.run``）。"""
+    try:
+        asyncio.get_running_loop()
+        raise RuntimeError("call run_contributors within a running loop instead")
+    except RuntimeError as exc:
+        if "call run_contributors" in str(exc):
+            raise
+        # 无运行中 loop → asyncio.run 安全
+        return asyncio.run(run_contributors(contributors, request, **kwargs))
+
+
+__all__ = [
+    "ContributionResult",
+    "ContextContributionRequest",
+    "ContextContributor",
+    "ContributorCapabilities",
+    "MemoryRecallContributor",
+    "SkillManifestContributor",
+    "WorkspaceRulesContributor",
+    "run_contributors",
+    "run_contributors_sync",
+]