PyPI - langchain-agentx-python - Versions diffs - 2.3.2__tar.gz → 2.3.4__tar.gz - Mend

Files changed (780) hide show

{langchain-agentx-python-2.3.2 → langchain-agentx-python-2.3.4}/langchain_agentx/loop/context/stages/autocompact.py RENAMED Viewed

@@ -25,7 +25,11 @@ from langchain_core.messages import HumanMessage, SystemMessage
 from ..blocking_guard import CompactionBlockingGuard
 from ..message_utils import stringify_content, total_estimated_tokens_for_messages
 from ...prompt.compact import build_autocompact_llm_instruction_prompt, get_compact_user_summary_message
-from ..settings import CompactionSettings, get_active_compaction_settings
+from ..settings import (
+    CompactionSettings,
+    autocompact_trigger_threshold_tokens,
+    get_active_compaction_settings,
+)
 from ..types import CompactProgress, ContextCompactionContext, StageResult
 from .base import CompactionStage
@@ -45,10 +49,25 @@ class AutocompactStage(CompactionStage):
     def run(self, messages: list[Any], ctx: ContextCompactionContext) -> StageResult:
         s = self._effective_settings()
+        if not s.is_auto_compact_enabled:
+            return StageResult(
+                messages=messages,
+                tokens_freed=0,
+                extra={"autocompact_skipped": "auto_compact_disabled"},
+            )
+        if ctx.consecutive_autocompact_failures >= s.max_consecutive_autocompact_failures:
+            return StageResult(
+                messages=messages,
+                tokens_freed=0,
+                extra={
+                    "autocompact_skipped": "consecutive_failures",
+                    "consecutive_autocompact_failures": ctx.consecutive_autocompact_failures,
+                },
+            )
         if ctx.query_source in s.autocompact_skip_query_sources:
             return StageResult(messages=messages, tokens_freed=0, extra={"autocompact_skipped": "query_source"})
         total = total_estimated_tokens_for_messages(messages, s.num_chars_per_token)
-        threshold = int(s.default_max_context_tokens * s.default_compress_trigger_fraction)
+        threshold = autocompact_trigger_threshold_tokens(s)
         slack = ctx.snip_tokens_freed_carry if s.autocompact_snip_carry_threshold_slack else 0
         if total <= threshold + slack:
             return StageResult(messages=messages, tokens_freed=0)
@@ -66,11 +85,12 @@ class AutocompactStage(CompactionStage):
         tail = rest[-k:]
         block_llm = CompactionBlockingGuard(settings=s).should_block_compact_llm(messages)
-        if (
+        attempted_llm = (
             s.autocompact_use_llm_when_available
             and ctx.model is not None
             and not block_llm
-        ):
+        )
+        if attempted_llm:
             summary = _invoke_summary(ctx.model, s, head)
             if summary:
                 # 修复后：不包含 leading，state 不含 SystemMessage
@@ -104,20 +124,38 @@ class AutocompactStage(CompactionStage):
         ] + tail
         new_total = total_estimated_tokens_for_messages(folded, s.num_chars_per_token)
         mode = "prune_blocked_llm" if block_llm else "prune"
+        extra: dict[str, Any] = {"autocompact_mode": mode, "folded_messages": len(head)}
+        if attempted_llm:
+            extra["autocompact_llm_failed"] = True
         return StageResult(
             messages=folded,
             tokens_freed=max(0, total - new_total),
-            extra={"autocompact_mode": mode, "folded_messages": len(head)},
+            extra=extra,
         )
     async def arun(self, messages: list[Any], ctx: ContextCompactionContext) -> StageResult:
         """Async 版本：LLM 摘要使用 astream 流式生成，支持进度回调。"""
         s = self._effective_settings()
+        if not s.is_auto_compact_enabled:
+            return StageResult(
+                messages=messages,
+                tokens_freed=0,
+                extra={"autocompact_skipped": "auto_compact_disabled"},
+            )
+        if ctx.consecutive_autocompact_failures >= s.max_consecutive_autocompact_failures:
+            return StageResult(
+                messages=messages,
+                tokens_freed=0,
+                extra={
+                    "autocompact_skipped": "consecutive_failures",
+                    "consecutive_autocompact_failures": ctx.consecutive_autocompact_failures,
+                },
+            )
         if ctx.query_source in s.autocompact_skip_query_sources:
             return StageResult(messages=messages, tokens_freed=0, extra={"autocompact_skipped": "query_source"})
         total = total_estimated_tokens_for_messages(messages, s.num_chars_per_token)
-        threshold = int(s.default_max_context_tokens * s.default_compress_trigger_fraction)
+        threshold = autocompact_trigger_threshold_tokens(s)
         slack = ctx.snip_tokens_freed_carry if s.autocompact_snip_carry_threshold_slack else 0
         if total <= threshold + slack:
             return StageResult(messages=messages, tokens_freed=0)
@@ -135,11 +173,12 @@ class AutocompactStage(CompactionStage):
         tail = rest[-k:]
         block_llm = CompactionBlockingGuard(settings=s).should_block_compact_llm(messages)
-        if (
+        attempted_llm = (
             s.autocompact_use_llm_when_available
             and ctx.model is not None
             and not block_llm
-        ):
+        )
+        if attempted_llm:
             summary = await _astream_summary(ctx.model, s, head, ctx.progress_callback)
             if summary:
                 folded = [
@@ -172,10 +211,13 @@ class AutocompactStage(CompactionStage):
         ] + tail
         new_total = total_estimated_tokens_for_messages(folded, s.num_chars_per_token)
         mode = "prune_blocked_llm" if block_llm else "prune"
+        extra: dict[str, Any] = {"autocompact_mode": mode, "folded_messages": len(head)}
+        if attempted_llm:
+            extra["autocompact_llm_failed"] = True
         return StageResult(
             messages=folded,
             tokens_freed=max(0, total - new_total),
-            extra={"autocompact_mode": mode, "folded_messages": len(head)},
+            extra=extra,
         )

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: langchain-agentx-python
-Version: 2.3.2
+Version: 2.3.4
 Summary: LangChain/LangGraph-based agent utilities for CodeBaseX.
 Author-email: GoodMood2008 <GoodMood2008@users.noreply.github.com>
 License:                                  Apache License

@@ -11,7 +11,7 @@ from langchain_agentx import create_loop_agent
 ```
 """
-__version__ = "2.3.2"
+__version__ = "2.3.4"
 from .loop import (  # noqa: F401
     create_loop_agent,

@@ -4,6 +4,7 @@ from .loop_config import (
     AgentStateConfig,
     LoopServices,
     PromptConfig,
+    SkillDiscoveryConfig,
     TraceConfig,
     SubagentLinkage,
 )
@@ -25,6 +26,7 @@ from .runtime_settings import (
 __all__ = [
     "AgentLoopConfig",
     "AgentStateConfig",
+    "SkillDiscoveryConfig",
     "LoopServices",
     "PromptConfig",
     "TraceConfig",

@@ -0,0 +1,38 @@
+"""Compaction 控制面 env 解析（对齐 CC DISABLE_COMPACT / DISABLE_AUTO_COMPACT）。"""
+from __future__ import annotations
+import os
+def _parse_bool_env(name: str) -> bool | None:
+    raw = os.environ.get(name, "").strip().lower()
+    if not raw:
+        return None
+    return raw in ("1", "true", "yes", "on")
+def resolve_is_auto_compact_enabled(*, config_is_auto_compact: bool = True) -> bool:
+    """CC：DISABLE_COMPACT 关全部；DISABLE_AUTO_COMPACT 仅关 proactive autocompact。"""
+    if _parse_bool_env("DISABLE_COMPACT") is True:
+        return False
+    if _parse_bool_env("DISABLE_AUTO_COMPACT") is True:
+        return False
+    return config_is_auto_compact
+def compute_autocompact_trigger_token_threshold(
+    *,
+    context_window_tokens: int,
+    max_output_tokens_reserved: int,
+    compact_buffer_tokens: int,
+) -> int:
+    """对齐 CC：effectiveWindow - AUTOCOMPACT_BUFFER_TOKENS。"""
+    effective = max(0, int(context_window_tokens) - int(max_output_tokens_reserved))
+    return max(0, effective - int(compact_buffer_tokens))
+__all__ = [
+    "compute_autocompact_trigger_token_threshold",
+    "resolve_is_auto_compact_enabled",
+]

@@ -52,6 +52,16 @@ class PromptConfig:
     thinking_generation_enabled: bool = True  # 是否启用 request-side thinking 生成（对齐 CC thinkingConfig）
+@dataclass(frozen=True)
+class SkillDiscoveryConfig:
+    """Workflow / 双 workspace 节点的 Skill 发现层级（对齐 bug.md ComponentScan）。"""
+    project_skills_dir: str | Path | None = None
+    global_skills_dir: str | Path | None = None
+    include_managed: bool = False
+    include_user_home: bool = False
 @dataclass(frozen=True)
 class AgentStateConfig:
     """Agent 状态配置（Agent 状态目录落点 / 派生目录视图）。
@@ -64,6 +74,7 @@ class AgentStateConfig:
         runtime_cwd: 工具运行时 cwd / 相对路径展开基准（None 时回退到 workspace_root）
         read_roots: 只读根目录
         write_roots: 只写根目录
+        skill_discovery: 可选项目级 / 全局 skill 目录（覆盖 legacy 三源）
     CC 对照：无直接对应，SDK 特化的状态层配置。
     """
@@ -74,6 +85,7 @@ class AgentStateConfig:
     write_roots: tuple[str, ...] = ()
     workspace_trust_accepted: bool | None = None
     enable_dynamic_root_grant: bool = False
+    skill_discovery: SkillDiscoveryConfig | None = None
 @dataclass(frozen=True)
@@ -241,6 +253,7 @@ class AgentLoopConfig:
     token_budget_max_continuations: int = 3
     compact_buffer_tokens: int = 13_000
     max_output_tokens_reserved: int = 20_000
+    is_auto_compact: bool = True
     # -----------------------------------------------------------------------
     # 运行模式

@@ -33,17 +33,33 @@ from .stages.base import CompactionStage
 from .types import CompactProgress, ContextCompactionContext
+def _apply_autocompact_failure_patch(
+    state: dict[str, Any],
+    stages_meta: list[dict[str, Any]],
+    patch: dict[str, Any],
+) -> None:
+    ac = next((m for m in stages_meta if m.get("name") == "autocompact"), None)
+    if ac is None:
+        return
+    extra = ac.get("extra") or {}
+    prev = int(state.get("consecutive_autocompact_failures") or 0)
+    if extra.get("autocompact_llm_failed"):
+        patch["consecutive_autocompact_failures"] = prev + 1
+    elif extra.get("autocompact_mode") == "llm":
+        patch["consecutive_autocompact_failures"] = 0
 class DefaultCompactionStages:
     """CC 五段默认装配（顺序即契约）。"""
     @staticmethod
-    def build() -> tuple[CompactionStage, ...]:
+    def build(settings: CompactionSettings | None = None) -> tuple[CompactionStage, ...]:
         return (
-            ToolResultBudgetStage(),
-            SnipCompactStage(),
-            MicrocompactStage(),
-            ContextCollapseStage(),
-            AutocompactStage(),
+            ToolResultBudgetStage(settings),
+            SnipCompactStage(settings),
+            MicrocompactStage(settings),
+            ContextCollapseStage(settings),
+            AutocompactStage(settings),
         )
@@ -65,7 +81,7 @@ class ContextPipeline:
         settings: CompactionSettings | None = None,
     ) -> ContextPipeline:
         """使用 `DefaultCompactionStages` 构建管道。"""
-        return cls(DefaultCompactionStages.build(), settings=settings)
+        return cls(DefaultCompactionStages.build(settings), settings=settings)
     @property
     def stages(self) -> tuple[CompactionStage, ...]:
@@ -117,6 +133,7 @@ class ContextPipeline:
                 "mode": ctx.mode,
             }
+        _apply_autocompact_failure_patch(state, stages_meta, patch)
         if not patch:
             return {}
         return patch
@@ -198,6 +215,7 @@ class ContextPipeline:
                 "mode": ctx.mode,
             }
+        _apply_autocompact_failure_patch(state, stages_meta, patch)
         if not patch:
             return {}
         return patch

@@ -32,6 +32,8 @@ class CompactionSettings:
     prune_protected_tools: FrozenSet[str] = field(default_factory=lambda: frozenset({"Skill"}))
     default_max_context_tokens: int = 120_000
     default_compress_trigger_fraction: float = 0.8
+    # 非 None 时 AutocompactStage 用绝对触发线（由 AgentLoopConfig compact_buffer 派生）。
+    autocompact_trigger_token_threshold: int | None = None
     default_compress_keep_recent_messages: int = 8
     # --- autocompact LLM 提示词（对齐 `compact.ts` + `prompt.ts`）---
     # 摘要插在保留 tail 之前 → 默认 `partial_up_to`；全量折叠无尾窗时用 `full`；与 CC partial pivot「from」一致时用 `partial_from`。
@@ -62,6 +64,8 @@ class CompactionSettings:
         default_factory=QuerySourcePolicy.autocompact_skip_sources
     )
     autocompact_use_llm_when_available: bool = True
+    is_auto_compact_enabled: bool = True
+    max_consecutive_autocompact_failures: int = 3
     autocompact_llm_excerpt_chars_per_message: int = 20_000
     # --- snip carry → autocompact 阈值放宽（对齐 CC snipTokensFreed 计入阈值）---
     # True：将本流水线已累计的 `snip_tokens_freed_carry` 加到触发阈值上，避免 snip 已释放后仍立刻 autocompact。
@@ -89,6 +93,12 @@ class CompactionSettings:
 active_compaction_settings: CompactionSettings = CompactionSettings()
+def autocompact_trigger_threshold_tokens(settings: CompactionSettings) -> int:
+    if settings.autocompact_trigger_token_threshold is not None:
+        return settings.autocompact_trigger_token_threshold
+    return int(settings.default_max_context_tokens * settings.default_compress_trigger_fraction)
 def get_active_compaction_settings() -> CompactionSettings:
     """返回当前生效的策略（默认即模块级 `active_compaction_settings`）。"""
     return active_compaction_settings
@@ -104,6 +114,7 @@ __all__ = [
     "AutocompactLlmPromptKind",
     "CompactionSettings",
     "active_compaction_settings",
+    "autocompact_trigger_threshold_tokens",
     "get_active_compaction_settings",
     "set_active_compaction_settings",
 ]

@@ -38,6 +38,7 @@ class ContextCompactionContext:
     model: Any | None = None
     snip_tokens_freed_carry: int = 0
     query_source: str | None = None
+    consecutive_autocompact_failures: int = 0
     config: CompactionConfig = field(default_factory=CompactionConfig)
     progress_callback: CompactProgressCallback = None
     mode: str = "proactive"
@@ -47,11 +48,14 @@ class ContextCompactionContext:
     def from_state(
         cls, state: Mapping[str, Any], model: Any | None, mode: str = "proactive"
     ) -> ContextCompactionContext:
+        raw_failures = state.get("consecutive_autocompact_failures")
+        failures = int(raw_failures) if isinstance(raw_failures, int) else 0
         return cls(
             model=model,
             query_source=state.get("query_source")
             if isinstance(state.get("query_source"), str)
             else None,
+            consecutive_autocompact_failures=max(0, failures),
             mode=mode,
         )

@@ -204,6 +204,7 @@ def create_initial_state(
         # 恢复路径守卫状态
         "collapse_retry_available": False,
         "has_attempted_reactive_compact": False,
+        "consecutive_autocompact_failures": 0,
         "compaction_pipeline_meta": None,
         "max_output_tokens_override": None,
         "max_output_tokens_recovery_count": 0,
@@ -248,6 +249,7 @@ class AgentLoopState(TypedDict):
     withhold_retry_count: int  # withhold 已重试次数
     collapse_retry_available: bool  # collapse_drain_retry 相关守卫
     has_attempted_reactive_compact: bool  # reactive_compact 重试守卫
+    consecutive_autocompact_failures: int  # proactive autocompact LLM 连续失败（CC 熔断）
     compaction_pipeline_meta: dict[str, Any] | None  # prune_and_compress 观测
     max_output_tokens_override: int | None  # max_output_tokens 升档参数
     max_output_tokens_recovery_count: int  # max_output_tokens 恢复计数

langchain-agentx-python 2.3.2__tar.gz → 2.3.4__tar.gz

langchain-agentx-python 2.3.2tar.gz → 2.3.4tar.gz