agentengine-sdk-python 0.8.4__py3-none-any.whl → 0.8.6__py3-none-any.whl

ksadk/harness/context_engine.py

@@ -0,0 +1,388 @@
+"""Harness ContextEngine（plan §8）——复用 PCM 模块的默认上下文引擎。
+
+Release 交付（plan §17）：
+- Context 分层：Stable Prompt 与动态 Context 分开构建、分别算 Hash（§8.1）；
+- ContextEngine 接口：plan / compact / recover（§8.2）；
+- 动态 Token Budget：预算来源优先级链（§8.3，无固定常数）；
+- 主动/紧急压缩：主动阈值 0.72（Spec 默认）提前压缩；紧急压缩在 overflow 后
+  更强压缩且仅重试一次（§8.4）。
+
+复用（master 已合入的 PCM 模块，本分支已移植）：
+- ``ksadk.context_engine.planner.ContextPlanner``——确定性规划、required/group
+  原子性（Tool Call/Result 不拆分）、优先级裁剪；
+- ``ksadk.context_engine.assembler``——plan 投影为 Chat 输入；
+- ``ksadk.context_engine.policies.ContextBudgetPolicy``——分区预算比例。
+
+新增（plan §8.4.1 差距表）：压缩后校验（关键 ID/金额/日期存在性比对）与重注入。
+"""
+
+from __future__ import annotations
+
+import hashlib
+import re
+from dataclasses import dataclass, field
+from typing import Any
+
+from ksadk.context_engine.assembler import AssembledInput, ContextAssembler
+from ksadk.context_engine.models import ContextBudget, ContextItem, ContextPlan
+from ksadk.context_engine.planner import ContextPlanner, build_budget
+from ksadk.context_engine.policies import ContextBudgetPolicy
+from ksadk.context_engine.tokenizer import get_default_token_counter
+from ksadk.harness.spec import HarnessSpec
+from ksadk.harness.state import HarnessState, Message, MessageRole
+
+#: 关键保留字段模式（§8.4：opaque ID、金额、日期、版本、审批编号）。
+_CRITICAL_PATTERN = re.compile(
+    r"""
+    (?P<id>[A-Z]{1,6}-?\d{2,10}(?:-\d{1,6})?) |  # 形如 AP-1024 / INV-2026-0001 的 opaque ID
+    (?P<amount>¥\d{1,3}(?:,\d{3})*(?:\.\d+)?|\d{1,3}(?:,\d{3})+(?:\.\d+)?) |  # 金额
+    (?P<date>20\d{2}[-/年.]\d{1,2}[-/月.]\d{1,2}) |  # 日期
+    (?P<version>v?\d+\.\d+\.\d+)             |  # 版本号
+    (?P<approval>审批号[:：]?\s*[A-Za-z0-9-]+)     # 审批编号
+    """,
+    re.VERBOSE,
+)
+
+
+def extract_critical_facts(text: str) -> set[str]:
+    """抽取压缩前后都必须保留的关键事实（ID/金额/日期/版本/审批号）。"""
+
+    return {m.group(0) for m in _CRITICAL_PATTERN.finditer(text or "")}
+
+
+@dataclass(frozen=True)
+class ContextRequest:
+    """plan() 的输入：一次模型调用的上下文需求。"""
+
+    spec: HarnessSpec
+    state: HarnessState
+    user_input: str
+    #: 模型上下文窗口（来自 Model Profile 等，非固定常数，§8.3）。
+    context_window_tokens: int
+    #: ContextSection 装配面（G9 §12.3.1）产出的具名段（skill/mcp 目录、
+    #: 环境上下文、项目文档等）。投影与装配顺序由
+    #: ``context_sections.sections_to_context_items`` 固化；key/content-hash
+    #: 经 build_manifest 记入 ContextManifest。G9.0 的 skill_catalog/
+    #: mcp_catalog 字段由此承载并退役。
+    section_items: tuple[ContextItem, ...] = ()
+
+
+@dataclass(frozen=True)
+class CompactionRequest:
+    """compact() 的输入。"""
+
+    messages: tuple[Message, ...]
+    #: 压缩触发方式：proactive（阈值前）/ emergency（模型 overflow 后）。
+    trigger: str
+    #: 保留最近 N 条完整消息为"最近尾部"。
+    keep_recent: int = 6
+    #: 摘要文本（调用方生成，引擎不内嵌模型调用）。
+    summary: str | None = None
+    #: 参与压缩的消息数。
+    compacted_count: int = 0
+
+
+@dataclass(frozen=True)
+class ContextCheckpoint:
+    """压缩结果（§8.4：持久化为 Checkpoint，完整 Transcript 不删除）。"""
+
+    checkpoint_id: str
+    trigger: str
+    compacted_until_seq_id: int
+    summary: str
+    retained_critical_facts: tuple[str, ...] = field(default=())
+    dropped_critical_facts: tuple[str, ...] = field(default=())
+    reinjection: str = ""
+
+    def to_payload(self) -> dict[str, Any]:
+        return {
+            "checkpoint_id": self.checkpoint_id,
+            "trigger": self.trigger,
+            "compacted_until_seq_id": self.compacted_until_seq_id,
+            "summary": self.summary,
+            "retained_critical_facts": list(self.retained_critical_facts),
+            "dropped_critical_facts": list(self.dropped_critical_facts),
+            "reinject": bool(self.reinjection),
+        }
+
+
+@dataclass
+class ContextOverflow:
+    """模型明确返回 context overflow 时的恢复请求。"""
+
+    error_message: str
+    retry_count: int = 0
+
+
+class ContextEngineError(RuntimeError):
+    """上下文引擎错误（含紧急压缩超限：只允许一次，§8.4）。"""
+
+
+class HarnessContextEngine:
+    """默认 ContextEngine：Stable/Dynamic 分层 + 确定性规划 + 压缩后校验/重注入。"""
+
+    def __init__(self, *, policy: ContextBudgetPolicy | None = None) -> None:
+        self._planner = ContextPlanner(policy=policy)
+        self._assembler = ContextAssembler()
+        self._counter = get_default_token_counter()
+        self._budget_policy = policy or ContextBudgetPolicy()
+
+    # ------------------------------------------------------------- stable
+
+    def stable_prompt(self, spec: HarnessSpec) -> str:
+        """Stable Prompt（§8.1）：Agent 身份/角色 + 稳定指令，与动态上下文分离。"""
+        if spec.prompt.instructions:
+            return spec.prompt.instructions
+        return ""  # instructions_ref 的正文由宿主解析后经 stable_prompt_override 注入
+
+    def stable_prompt_hash(self, spec: HarnessSpec) -> str:
+        material = self.stable_prompt(spec) + (spec.prompt.instructions_ref or "")
+        return "sha256:" + hashlib.sha256(material.encode("utf-8")).hexdigest()
+
+    # --------------------------------------------------------------- plan
+
+    def plan(self, request: ContextRequest) -> ContextPlan:
+        """构建一次模型调用的上下文计划（§8.2 plan）。
+
+        Stable Prompt 与动态 Context 分开构建、分别计 Token（§8.1）；
+        Tool Call/Result 经 group_id 原子保留（§8.4 压缩约束提前到规划层）。
+        """
+        stable = self.stable_prompt(request.spec)
+        items: list[ContextItem] = [self._stable_item(stable)]
+        items.extend(self._dynamic_items(request))
+        budget = self.build_budget(request.spec, request.context_window_tokens)
+        return self._planner.plan(
+            items,
+            budget=budget,
+            integration_mode="ksadk_harness",
+            stable_prefix_hash=self.stable_prompt_hash(request.spec),
+        )
+
+    def assemble_chat(self, plan: ContextPlan) -> AssembledInput:
+        """把计划投影为 Chat 输入（Stable Prompt → system，其余按序）。"""
+        return self._assembler.assemble_chat(plan)
+
+    # ------------------------------------------------------------- budget
+
+    def build_budget(self, spec: HarnessSpec, context_window_tokens: int) -> ContextBudget:
+        """动态 Token Budget（§8.3）：窗口来自调用方（Model Profile 优先），无固定常数。"""
+        if context_window_tokens <= 0:
+            raise ContextEngineError(
+                f"上下文窗口必须为正（Model Profile/目录/静态元数据/安全默认值链失效）: "
+                f"{context_window_tokens}"
+            )
+        return build_budget(
+            policy=self._budget_policy,
+            context_window_tokens=context_window_tokens,
+            reserved_output_tokens=int(
+                context_window_tokens * spec.context_policy.reserved_output_ratio
+            ),
+        )
+
+    # ------------------------------------------------------------ compact
+
+    def compact(self, request: CompactionRequest) -> ContextCheckpoint:
+        """压缩（§8.2 compact）：边界切分不拆 Tool Pair + 后校验 + 重注入。
+
+        本方法不内嵌模型调用：摘要文本由调用方（引擎 reason 节点侧）生成后传入，
+        引擎负责边界、校验与重注入——与 PCM 的"纯计算 + 受控调用"分层一致。
+        """
+        if request.trigger == "emergency":
+            if request.summary is None:
+                raise ContextEngineError("紧急压缩必须提供摘要（更强压缩由调用方执行）")
+        summary = request.summary or ""
+        messages = list(request.messages)
+        if len(messages) <= request.keep_recent:
+            return ContextCheckpoint(
+                checkpoint_id=self._checkpoint_id(summary),
+                trigger=request.trigger,
+                compacted_until_seq_id=0,
+                summary="",
+            )
+        head = messages[: len(messages) - request.keep_recent]
+        tail = messages[len(messages) - request.keep_recent :]
+
+        # 后校验（§8.4.1 差距表：正则抽关键 ID 比对压缩前后存在性）。
+        head_text = "\n".join(m.content for m in head)
+        retained_text = summary + "\n" + "\n".join(m.content for m in tail)
+        retained = extract_critical_facts(retained_text)
+        expected = extract_critical_facts(head_text)
+        dropped = tuple(sorted(expected - retained))
+
+        # 重注入（§8.4.1：关键约束压缩后重新注入）。
+        reinjection = ""
+        if dropped:
+            reinjection = "【必须保留的关键事实】\n" + "\n".join(f"- {fact}" for fact in dropped)
+
+        return ContextCheckpoint(
+            checkpoint_id=self._checkpoint_id(summary + request.trigger),
+            trigger=request.trigger,
+            compacted_until_seq_id=request.compacted_count,
+            summary=summary,
+            retained_critical_facts=tuple(sorted(retained & expected)) if expected else (),
+            dropped_critical_facts=dropped,
+            reinjection=reinjection,
+        )
+
+    # ------------------------------------------------------------- recover
+
+    def recover(self, failure: ContextOverflow) -> ContextPlan:
+        """紧急恢复（§8.2 recover）：更强压缩 + 只重试一次，防无限循环。"""
+        if failure.retry_count >= 1:
+            raise ContextEngineError(
+                f"紧急压缩只允许重试一次（§8.4），已重试 {failure.retry_count} 次"
+            )
+        # 更强压缩 = 收紧到 60% 窗口的预算，交回 planner 重新规划。
+        tighter = build_budget(
+            policy=self._budget_policy,
+            context_window_tokens=max(1024, int(32768 * 0.6)),
+        )
+        return self._planner.plan([], budget=tighter, integration_mode="ksadk_harness")
+
+    # ------------------------------------------------------------ helpers
+
+    @staticmethod
+    def _checkpoint_id(material: str) -> str:
+        return f"ctxcp_{hashlib.sha256(material.encode('utf-8')).hexdigest()[:12]}"
+
+    def _stable_item(self, stable: str) -> ContextItem:
+        return ContextItem(
+            item_id="stable_prompt",
+            kind="compiled_prompt",
+            content=stable,
+            source="harness:stable_prompt",
+            trust_level="developer",
+            priority=1,
+            estimated_tokens=self._counter.count_text(stable),
+            required=True,
+            stable=True,
+        )
+
+    def _dynamic_items(self, request: ContextRequest) -> list[ContextItem]:
+        items: list[ContextItem] = []
+        state = request.state
+        working = state.working_context
+        working_text = "\n".join(
+            part
+            for part in (
+                f"当前目标：{working.goal}" if working.goal else "",
+                f"当前计划：{working.plan}" if working.plan else "",
+                "\n".join(f"已确认约束：{c}" for c in working.confirmed_constraints),
+                "\n".join(f"未解决问题：{q}" for q in working.open_questions),
+                "\n".join(f"已验证事实：{f}" for f in working.verified_facts),
+                "\n".join(f"最近工具失败：{t}" for t in working.recent_tool_failures),
+            )
+            if part
+        )
+        if working_text:
+            items.append(
+                ContextItem(
+                    item_id="working_context",
+                    kind="working_state",
+                    content=working_text,
+                    source="harness:working_context",
+                    trust_level="developer",
+                    priority=7,
+                    estimated_tokens=self._counter.count_text(working_text),
+                )
+            )
+        if request.section_items:
+            items.extend(request.section_items)
+        for ref in state.memory_refs:
+            items.append(
+                ContextItem(
+                    item_id=f"core_memory:{ref.memory_ref}",
+                    kind="core_memory",
+                    content=ref.memory_ref,
+                    source=f"harness:core_memory:{ref.scope}",
+                    trust_level="untrusted",
+                    priority=9,
+                    estimated_tokens=self._counter.count_text(ref.memory_ref),
+                )
+            )
+        for index, message in enumerate(state.messages):
+            if message.role is MessageRole.SYSTEM:
+                # 宿主注入的 system 段（如 Memory recall）——高优先级保留。
+                if not message.content.strip():
+                    continue
+                items.append(
+                    ContextItem(
+                        item_id=f"history_system:{index}",
+                        kind="history_round",
+                        content=message.content,
+                        source="harness:history",
+                        trust_level="developer",
+                        priority=3,
+                        estimated_tokens=self._counter.count_text(message.content),
+                        # seq_start 承载物理顺序：同 kind 内 planner 按 seq 排序，
+                        # 缺省时回退 item_id 字典序，10 条以上历史会被乱排。
+                        seq_start=index,
+                        metadata={"role": "system"},
+                    )
+                )
+                continue
+            if message.role not in (MessageRole.USER, MessageRole.ASSISTANT):
+                continue
+            group = None
+            if message.role is MessageRole.ASSISTANT and message.tool_call_id:
+                # Tool Call/Result 原子组（§8.4：不拆分）。
+                group = f"turn:{index}"
+            items.append(
+                ContextItem(
+                    item_id=f"history:{index}",
+                    kind="history_round",
+                    content=message.content,
+                    source="harness:history",
+                    trust_level="user",
+                    priority=8,
+                    estimated_tokens=self._counter.count_text(message.content),
+                    # seq_start 承载物理顺序（见上方 history_system 同注）。
+                    seq_start=index,
+                    group_id=group,
+                    # 组装层按原角色投影（否则 user 历史会被当成 assistant）。
+                    metadata={"role": message.role.value},
+                )
+            )
+        if request.user_input:
+            items.append(
+                ContextItem(
+                    item_id="current_input",
+                    kind="current_input",
+                    content=request.user_input,
+                    source="harness:current_input",
+                    trust_level="user",
+                    priority=2,
+                    estimated_tokens=self._counter.count_text(request.user_input),
+                    required=True,
+                )
+            )
+        return items
+
+
+#: 预算来源优先级链（§8.3）——宿主按序解析，落到哪层记录哪层。
+def resolve_context_window(
+    *,
+    model_profile_window: int | None = None,
+    provider_catalog_window: int | None = None,
+    static_metadata_window: int | None = None,
+) -> tuple[int, str]:
+    """返回 (window, source)。安全默认值 32k 并标注 fallback。"""
+    if model_profile_window and model_profile_window > 0:
+        return int(model_profile_window), "model_profile"
+    if provider_catalog_window and provider_catalog_window > 0:
+        return int(provider_catalog_window), "provider_catalog"
+    if static_metadata_window and static_metadata_window > 0:
+        return int(static_metadata_window), "static_metadata"
+    return 32768, "fallback_default"
+
+
+__all__ = [
+    "CompactionRequest",
+    "ContextCheckpoint",
+    "ContextEngineError",
+    "ContextOverflow",
+    "ContextRequest",
+    "HarnessContextEngine",
+    "extract_critical_facts",
+    "resolve_context_window",
+]

ksadk/harness/context_maintenance.py

@@ -0,0 +1,259 @@
+"""Deterministic background maintenance for session-scoped Working Context.
+
+Maintenance produces a versioned :class:`WorkingContextPatch`; callers decide when
+to apply it.  It never edits the Stable Prompt, never writes long-term Memory and
+never retries a version conflict by silently overwriting newer state.
+"""
+
+from __future__ import annotations
+
+import time
+import uuid
+from collections import deque
+from dataclasses import dataclass, replace
+from typing import Any, Literal, Protocol, Sequence
+
+from ksadk.harness.state import WorkingContext
+from ksadk.harness.working_context_patch import (
+    PatchOperation,
+    WorkingContextPatch,
+    apply_patch,
+)
+
+_FIELD_LIMITS = {
+    "confirmed_constraints": 64,
+    "open_questions": 64,
+    "verified_facts": 32,
+    "recent_tool_failures": 8,
+}
+
+
+@dataclass(frozen=True)
+class ContextMaintenancePlan:
+    patch: WorkingContextPatch | None
+    before_items: int
+    after_items: int
+    removed_empty: int
+    removed_duplicates: int
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "schemaVersion": 1,
+            "baseVersion": self.patch.base_version if self.patch else None,
+            "changed": self.patch is not None,
+            "beforeItems": self.before_items,
+            "afterItems": self.after_items,
+            "removedEmpty": self.removed_empty,
+            "removedDuplicates": self.removed_duplicates,
+            "patch": self.patch.to_payload() if self.patch else None,
+        }
+
+
+class WorkingContextSnapshotStore(Protocol):
+    """Checkpoint/control-plane boundary used by the background worker."""
+
+    def load(self, *, tenant_id: str, session_id: str) -> WorkingContext | None: ...
+
+    def compare_and_set(
+        self,
+        *,
+        tenant_id: str,
+        session_id: str,
+        expected_version: int,
+        value: WorkingContext,
+    ) -> bool: ...
+
+
+MaintenanceStatus = Literal[
+    "pending", "running", "succeeded", "unchanged", "conflicted", "failed"
+]
+
+
+@dataclass(frozen=True)
+class ContextMaintenanceTask:
+    task_id: str
+    tenant_id: str
+    session_id: str
+    status: MaintenanceStatus = "pending"
+    attempts: int = 0
+    before_items: int = 0
+    after_items: int = 0
+    error_code: str = ""
+    created_at: float = 0.0
+    updated_at: float = 0.0
+
+
+class ContextMaintenanceQueue:
+    """Bounded background organizer with compare-and-set persistence.
+
+    Submitting never performs storage I/O.  A stale task is marked ``conflicted``
+    and never retries by overwriting the newer session snapshot.
+    """
+
+    def __init__(self, store: WorkingContextSnapshotStore, *, max_pending: int = 1024) -> None:
+        self._store = store
+        self._max_pending = max(1, int(max_pending))
+        self._pending: deque[str] = deque()
+        self._tasks: dict[str, ContextMaintenanceTask] = {}
+        self._session_keys: dict[tuple[str, str], str] = {}
+
+    def submit(self, *, tenant_id: str, session_id: str) -> ContextMaintenanceTask:
+        key = (tenant_id, session_id)
+        existing_id = self._session_keys.get(key)
+        if existing_id:
+            existing = self._tasks[existing_id]
+            if existing.status in {"pending", "running"}:
+                return existing
+        if len(self._pending) >= self._max_pending:
+            raise RuntimeError("context_maintenance_queue_full")
+        now = time.time()
+        task = ContextMaintenanceTask(
+            task_id=f"ctxorg_{uuid.uuid4().hex[:20]}",
+            tenant_id=tenant_id,
+            session_id=session_id,
+            created_at=now,
+            updated_at=now,
+        )
+        self._tasks[task.task_id] = task
+        self._session_keys[key] = task.task_id
+        self._pending.append(task.task_id)
+        return task
+
+    def get(self, task_id: str) -> ContextMaintenanceTask | None:
+        return self._tasks.get(task_id)
+
+    @property
+    def pending_count(self) -> int:
+        return len(self._pending)
+
+    def run_pending(self, *, limit: int = 1) -> list[ContextMaintenanceTask]:
+        completed: list[ContextMaintenanceTask] = []
+        for _ in range(max(0, int(limit))):
+            if not self._pending:
+                break
+            task_id = self._pending.popleft()
+            task = replace(
+                self._tasks[task_id],
+                status="running",
+                attempts=self._tasks[task_id].attempts + 1,
+                updated_at=time.time(),
+            )
+            self._tasks[task_id] = task
+            try:
+                final = self._run(task)
+            except Exception:  # noqa: BLE001 - background failure is observable only
+                final = replace(
+                    task,
+                    status="failed",
+                    error_code="snapshot_store_error",
+                    updated_at=time.time(),
+                )
+            self._tasks[task_id] = final
+            completed.append(final)
+        return completed
+
+    def _run(self, task: ContextMaintenanceTask) -> ContextMaintenanceTask:
+        working = self._store.load(
+            tenant_id=task.tenant_id, session_id=task.session_id
+        )
+        if working is None:
+            return replace(
+                task,
+                status="failed",
+                error_code="session_not_found",
+                updated_at=time.time(),
+            )
+        plan = plan_context_maintenance(working)
+        common = {
+            "before_items": plan.before_items,
+            "after_items": plan.after_items,
+            "updated_at": time.time(),
+        }
+        if plan.patch is None:
+            return replace(task, status="unchanged", **common)
+        maintained = apply_patch(working, plan.patch)
+        stored = self._store.compare_and_set(
+            tenant_id=task.tenant_id,
+            session_id=task.session_id,
+            expected_version=working.version,
+            value=maintained,
+        )
+        if not stored:
+            return replace(
+                task,
+                status="conflicted",
+                error_code="version_conflict",
+                **common,
+            )
+        return replace(task, status="succeeded", **common)
+
+
+def plan_context_maintenance(
+    working: WorkingContext,
+    *,
+    source_event_ids: Sequence[str] = (),
+) -> ContextMaintenancePlan:
+    """Create a bounded, deterministic cleanup patch for tuple fields."""
+
+    operations: list[PatchOperation] = []
+    before_items = 0
+    after_items = 0
+    removed_empty = 0
+    removed_duplicates = 0
+    for field_name, limit in _FIELD_LIMITS.items():
+        original = tuple(getattr(working, field_name))
+        before_items += len(original)
+        cleaned, empty_count, duplicate_count = _clean(original, limit=limit)
+        after_items += len(cleaned)
+        removed_empty += empty_count
+        removed_duplicates += duplicate_count
+        if cleaned != original:
+            operations.append(
+                PatchOperation(op="set", field_name=field_name, value=cleaned)
+            )
+
+    patch = None
+    if operations:
+        patch = WorkingContextPatch(
+            base_version=working.version,
+            operations=tuple(operations),
+            source_event_ids=tuple(dict.fromkeys(item for item in source_event_ids if item)),
+            reason="background_context_maintenance",
+        )
+    return ContextMaintenancePlan(
+        patch=patch,
+        before_items=before_items,
+        after_items=after_items,
+        removed_empty=removed_empty,
+        removed_duplicates=removed_duplicates,
+    )
+
+
+def _clean(values: tuple[str, ...], *, limit: int) -> tuple[tuple[str, ...], int, int]:
+    normalized: list[str] = []
+    keys: set[str] = set()
+    removed_empty = 0
+    removed_duplicates = 0
+    # Keep the newest spelling/value when the same normalized item occurs repeatedly.
+    for raw in reversed(values):
+        value = " ".join(str(raw).split())
+        if not value:
+            removed_empty += 1
+            continue
+        key = value.casefold()
+        if key in keys:
+            removed_duplicates += 1
+            continue
+        keys.add(key)
+        normalized.append(value)
+    cleaned = tuple(reversed(normalized))[-limit:]
+    return cleaned, removed_empty, removed_duplicates
+
+
+__all__ = [
+    "ContextMaintenancePlan",
+    "ContextMaintenanceQueue",
+    "ContextMaintenanceTask",
+    "WorkingContextSnapshotStore",
+    "plan_context_maintenance",
+]