langchain-agentx-python 2.3.3__py3-none-any.whl → 2.3.4__py3-none-any.whl

langchain_agentx/__init__.py

@@ -11,7 +11,7 @@ from langchain_agentx import create_loop_agent
 ```
 """
 
-__version__ = "2.3.3"
+__version__ = "2.3.4"
 
 from .loop import (  # noqa: F401
     create_loop_agent,

langchain_agentx/loop/config/compaction_config.py

@@ -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",
+]

langchain_agentx/loop/config/loop_config.py

@@ -253,6 +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

langchain_agentx/loop/context/pipeline.py

@@ -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

langchain_agentx/loop/context/settings.py

@@ -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",
 ]

langchain_agentx/loop/context/stages/autocompact.py

@@ -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,
         )
 
 

langchain_agentx/loop/context/types.py

@@ -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,
         )
 

langchain_agentx/loop/exit/exit_logic.py

@@ -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/loop/graph/factory.py

@@ -10,20 +10,15 @@ LangChain AgentX Agent Factory
 
 参考设计文档：.claude/design/create-opencode-style-agent-design.md
 
-LangChain AgentX Agent 状态图（文字版，含 loop_controller）：
+LangChain AgentX Agent 状态图（文字版，含 loop_controller · proactive compact gate）：
 
-    START
+    START → hook_loop_start → prune_and_compress → hook_before_model → model
       ↓
-    hook_loop_start
-      ↓
-    model 节点（同步/异步，带 OpenCode 退出判断）
-      ↓
-    hook_stop
-      ↓
-    loop_controller（集中处理 resolution / terminal / 工具调用）
-      ├─ resolution.kind=terminal → END（或 task reconcile / recovery 分支）
-      ├─ 有工具调用       → tools → (task_event_*/prune_and_compress) → model
-      └─ 无工具调用       → END
+    hook_stop → loop_controller
+      ├─ terminal → hook_loop_end → END
+      ├─ tools → (task_event_* →) prune_and_compress → hook_before_model → model
+      ├─ 再进 model（无 tools）→ prune_and_compress → hook_before_model → model
+      └─ reactive_compact → prune_and_compress（或 task 链）→ … → model
 """
 
 from __future__ import annotations
@@ -39,6 +34,16 @@ from langchain_core.messages import HumanMessage, SystemMessage
 from langchain_core.runnables import RunnableConfig
 from langgraph._internal._runnable import RunnableCallable
 from langgraph.constants import END, START
+
+_PRE_MODEL_COMPACT_NODE = "prune_and_compress"
+_BEFORE_MODEL_NODE = "hook_before_model"
+
+
+def _pre_model_gate_destination(loop_entry_node: str) -> str:
+    """CC 对齐：直连 hook_before_model 的 loop 续跑须先过 proactive compact gate。"""
+    if loop_entry_node == _BEFORE_MODEL_NODE:
+        return _PRE_MODEL_COMPACT_NODE
+    return loop_entry_node
 from langgraph.graph.state import StateGraph
 import asyncio
 import dataclasses
@@ -66,8 +71,9 @@ from ..model.schema_and_format import (
 )
 from ..model.model_node import ModelNode
 from ..model.model_nodes import make_model_nodes
-from ..context import default_loop_context_compaction
+from ..context import LoopContextCompaction, default_loop_context_compaction
 from ..context.settings import CompactionSettings
+from ..context.types import CompactProgress
 from ..injection.dedup import filter_attachment_dicts_by_task_command_id
 from ..message_contract import to_record
 from ..model.tool_call_degradation_corrector import (
@@ -102,6 +108,10 @@ from ...memory.memdir.paths import is_memory_enabled
 from ..prompt import SystemPromptSection, build_effective_system_prompt
 from ..prompt.session_context import get_default_session_context_sections
 from ..config import ModelContextResolver
+from ..config.compaction_config import (
+    compute_autocompact_trigger_token_threshold,
+    resolve_is_auto_compact_enabled,
+)
 from ..config.interaction_profile import (
     assemble_services,
     normalize_config,
@@ -556,7 +566,7 @@ class LoopGraphBuilder:
         # 运行时需要的额外值存储在 self._* 字段中
 
         self._compaction_settings = self._build_compaction_settings()
-        self._loop_context_compaction = default_loop_context_compaction().with_default_pipeline(
+        self._loop_context_compaction = LoopContextCompaction.with_default_pipeline(
             settings=self._compaction_settings
         )
         self._model = model
@@ -1354,15 +1364,29 @@ class LoopGraphBuilder:
 
     def _build_compaction_settings(self) -> CompactionSettings:
         base = CompactionSettings()
+        context_window = self._model_context_resolver.resolve_context_window(self._config)
+        trigger_threshold = compute_autocompact_trigger_token_threshold(
+            context_window_tokens=context_window,
+            max_output_tokens_reserved=self._config.max_output_tokens_reserved,
+            compact_buffer_tokens=self._config.compact_buffer_tokens,
+        )
+        auto_compact_enabled = resolve_is_auto_compact_enabled(
+            config_is_auto_compact=self._config.is_auto_compact
+        )
+        # G1-A (CC 0331): auto + reactive 路径下不在 call 前 synthetic blocking_limit。
+        enable_main_token_block = base.enable_main_model_context_token_block
+        if auto_compact_enabled:
+            enable_main_token_block = False
         return CompactionSettings(
             num_chars_per_token=base.num_chars_per_token,
             prune_protect_tokens=base.prune_protect_tokens,
             prune_minimum_tokens=base.prune_minimum_tokens,
             prune_protected_tools=base.prune_protected_tools,
-            default_max_context_tokens=self._model_context_resolver.resolve_context_window(
-                self._config
-            ),
+            default_max_context_tokens=context_window,
             default_compress_trigger_fraction=base.default_compress_trigger_fraction,
+            autocompact_trigger_token_threshold=trigger_threshold,
+            is_auto_compact_enabled=auto_compact_enabled,
+            max_consecutive_autocompact_failures=base.max_consecutive_autocompact_failures,
             default_compress_keep_recent_messages=base.default_compress_keep_recent_messages,
             autocompact_llm_prompt_kind=base.autocompact_llm_prompt_kind,
             compact_custom_instructions=base.compact_custom_instructions,
@@ -1382,7 +1406,7 @@ class LoopGraphBuilder:
             compaction_token_warn_fraction=base.compaction_token_warn_fraction,
             compaction_block_compact_llm_fraction=base.compaction_block_compact_llm_fraction,
             compaction_block_main_model_fraction=base.compaction_block_main_model_fraction,
-            enable_main_model_context_token_block=base.enable_main_model_context_token_block,
+            enable_main_model_context_token_block=enable_main_token_block,
             enable_tool_transcript_repair_before_llm=base.enable_tool_transcript_repair_before_llm,
             emit_compaction_pipeline_meta=base.emit_compaction_pipeline_meta,
         )
@@ -1426,19 +1450,21 @@ class LoopGraphBuilder:
             context_schema=self._context_schema,
         )
         graph_inner.add_node("model", model_node)
+        graph_inner.add_node(
+            _PRE_MODEL_COMPACT_NODE,
+            RunnableCallable(
+                self._prune_and_compress_node,
+                self._prune_and_compress_node_async,
+                trace=False,
+            ),
+        )
         if tool_node is not None:
             graph_inner.add_node("tools", tool_node)
-
-            graph_inner.add_node(
-                "prune_and_compress",
-                RunnableCallable(
-                    self._prune_and_compress_node,
-                    self._prune_and_compress_node_async,
-                    trace=False,
-                ),
-            )
         return graph_inner
 
+    def _wire_prune_and_compress_to_before_model(self, graph_inner: Any) -> None:
+        graph_inner.add_edge(_PRE_MODEL_COMPACT_NODE, _BEFORE_MODEL_NODE)
+
     def _prune_and_compress_node(self, state: dict[str, Any], runtime: Any) -> dict[str, Any]:
         return self._run_prune_and_compress_with_trace(state)
 
@@ -1461,6 +1487,31 @@ class LoopGraphBuilder:
             "skipping wait_for_extraction (use ainvoke for CC-aligned wait)"
         )
 
+    @staticmethod
+    def _compact_progress_phase(
+        progress: CompactProgress,
+        *,
+        seen: bool,
+    ) -> tuple[str, bool]:
+        if not seen:
+            return "start", True
+        if progress.stage == "done" and progress.percent >= 100:
+            return "end", True
+        return "tick", True
+
+    def _merge_compact_progress_into_patch(
+        self,
+        patch: dict[str, Any],
+        progress_events: list[dict[str, Any]],
+    ) -> None:
+        if not progress_events:
+            return
+        meta = patch.get("compaction_pipeline_meta")
+        if not isinstance(meta, dict):
+            meta = {}
+            patch["compaction_pipeline_meta"] = meta
+        meta["compact_progress"] = progress_events
+
     def _apply_prune_and_compress(self, state: dict[str, Any]) -> dict[str, Any]:
         if self._session_memory_compact_bridge is not None:
             compacted_messages = self._session_memory_compact_bridge.try_compact(
@@ -1481,6 +1532,55 @@ class LoopGraphBuilder:
             patch["collapse_retry_available"] = bool(compaction_meta.get("tokens_freed_sum", 0))
         return patch
 
+    async def _apply_prune_and_compress_async(self, state: dict[str, Any]) -> dict[str, Any]:
+        if self._session_memory_compact_bridge is not None:
+            compacted_messages = self._session_memory_compact_bridge.try_compact(
+                list(state.get("messages") or [])
+            )
+            if compacted_messages is not None:
+                return {
+                    "messages": compacted_messages,
+                    "collapse_retry_available": True,
+                    "compaction_pipeline_meta": {
+                        "autocompact_mode": "session_memory",
+                        "session_memory_compaction": "applied",
+                    },
+                }
+        progress_events: list[dict[str, Any]] = []
+        seen_progress = False
+
+        def _on_progress(progress: CompactProgress) -> None:
+            nonlocal seen_progress
+            phase, seen_progress = self._compact_progress_phase(
+                progress, seen=seen_progress
+            )
+            progress_events.append(
+                {
+                    "phase": phase,
+                    "stage": progress.stage,
+                    "stage_index": progress.stage_index,
+                    "total_stages": progress.total_stages,
+                    "percent": progress.percent,
+                    "detail": progress.detail,
+                }
+            )
+
+        emit_progress = bool(
+            self._compaction_settings.is_auto_compact_enabled
+            and self._compaction_settings.emit_compaction_pipeline_meta
+        )
+        patch = await self._loop_context_compaction.arun(
+            state,
+            model=self._model,
+            progress_callback=_on_progress if emit_progress else None,
+        )
+        if emit_progress:
+            self._merge_compact_progress_into_patch(patch, progress_events)
+        compaction_meta = patch.get("compaction_pipeline_meta")
+        if isinstance(compaction_meta, dict):
+            patch["collapse_retry_available"] = bool(compaction_meta.get("tokens_freed_sum", 0))
+        return patch
+
     def _run_prune_and_compress_with_trace(self, state: dict[str, Any]) -> dict[str, Any]:
         """执行 compaction，并在当前 session 内记录 compaction span（sync 路径）。"""
         collector = state.get("_trace_collector")
@@ -1546,7 +1646,7 @@ class LoopGraphBuilder:
         try:
             if self._session_memory_compact_bridge is not None:
                 await self._session_memory_compact_bridge.wait_for_extraction()
-            patch = self._apply_prune_and_compress(state)
+            patch = await self._apply_prune_and_compress_async(state)
         except Exception as exc:
             if span_id is not None:
                 collector.span_end(
@@ -1593,6 +1693,7 @@ class LoopGraphBuilder:
             skill_prefetch_provider=skill_prefetch_provider,
             entry_node=entry_node,
         )
+        self._wire_prune_and_compress_to_before_model(graph_inner)
         graph_inner.add_edge("hook_before_model", "model")
         return {
             "entry_node": entry_node,
@@ -1896,12 +1997,8 @@ class LoopGraphBuilder:
             RunnableCallable(_task_event_reserve_node, None, trace=False),
         )
         graph_inner.add_edge("task_event_commit_ack", "task_event_reserve")
-        # Phase 5A：压缩收束到「每轮进入 model 之前」唯一入口；有 task 链时在注入附件之后再压缩。
-        if tool_node is not None:
-            graph_inner.add_edge("task_event_reserve", "prune_and_compress")
-            graph_inner.add_edge("prune_and_compress", raw_loop_entry_node)
-        else:
-            graph_inner.add_edge("task_event_reserve", raw_loop_entry_node)
+        # Phase 5A / proactive autocompact：task 注入后再压缩，再进 hook_before_model。
+        graph_inner.add_edge("task_event_reserve", _PRE_MODEL_COMPACT_NODE)
         return "task_event_commit_ack"
 
     def _find_skill_prefetch_provider(
@@ -1984,10 +2081,10 @@ class LoopGraphBuilder:
             )
             graph_inner.add_edge(START, "skill_init")
             graph_inner.add_edge("skill_init", "hook_loop_start")
-            graph_inner.add_edge("hook_loop_start", entry_node)
+            graph_inner.add_edge("hook_loop_start", _PRE_MODEL_COMPACT_NODE)
             return
         graph_inner.add_edge(START, "hook_loop_start")
-        graph_inner.add_edge("hook_loop_start", entry_node)
+        graph_inner.add_edge("hook_loop_start", _PRE_MODEL_COMPACT_NODE)
 
     def _wire_main_edges(self, *, graph_inner: Any, tool_node: Any, routing: dict[str, Any]) -> None:
         loop_entry_node = routing["loop_entry_node"]
@@ -2040,8 +2137,8 @@ class LoopGraphBuilder:
             tools_model_destination = loop_entry_node
             tools_to_model_destinations = [loop_entry_node, exit_node]
         else:
-            tools_model_destination = "prune_and_compress"
-            tools_to_model_destinations = ["prune_and_compress", exit_node]
+            tools_model_destination = _PRE_MODEL_COMPACT_NODE
+            tools_to_model_destinations = [_PRE_MODEL_COMPACT_NODE, exit_node]
 
         graph_inner.add_conditional_edges(
             "tools",
@@ -2056,8 +2153,6 @@ class LoopGraphBuilder:
             ),
             tools_to_model_destinations,
         )
-        if not wire_task_event_nodes:
-            graph_inner.add_edge("prune_and_compress", loop_entry_node)
         graph_inner.add_edge(loop_exit_node, "hook_after_model")
         graph_inner.add_edge("hook_after_model", "hook_stop")
         self._wire_loop_controller(
@@ -2079,11 +2174,12 @@ class LoopGraphBuilder:
         wire_task_event_nodes: bool,
         tools_destination: str | None = None,
     ) -> None:
+        model_gate = _pre_model_gate_destination(loop_entry_node)
         graph_inner.add_node(
             "loop_controller",
             RunnableCallable(
                 _make_loop_controller_node(
-                    model_destination=loop_entry_node,
+                    model_destination=model_gate,
                     tools_destination=tools_destination or loop_entry_node,
                     structured_output_tools=self._structured_output_tools,
                     end_destination=exit_node,
@@ -2107,7 +2203,7 @@ class LoopGraphBuilder:
         def _reactive_compact_node(state, runtime):
             node = ReactiveCompactNode(
                 compaction_service=default_loop_context_compaction(),
-                loop_entry_destination=loop_entry_node,
+                loop_entry_destination=model_gate,
                 end_destination=exit_node,
             )
             return node(state)
@@ -2116,13 +2212,14 @@ class LoopGraphBuilder:
             "reactive_compact",
             RunnableCallable(_reactive_compact_node, None, trace=False),
         )
+        reactive_destinations = [model_gate, exit_node]
         graph_inner.add_conditional_edges(
             "reactive_compact",
-            lambda state: state.get("destination", loop_entry_node),
-            [loop_entry_node, exit_node],
+            lambda state: state.get("destination", model_gate),
+            reactive_destinations,
         )
 
-        loop_controller_destinations = [loop_entry_node, exit_node, "reactive_compact"]
+        loop_controller_destinations = [model_gate, exit_node, "reactive_compact"]
         if tools_destination is not None:
             loop_controller_destinations.insert(0, tools_destination)
         if wire_task_event_nodes and "task_event_commit_ack" not in loop_controller_destinations:
@@ -2144,16 +2241,17 @@ class LoopGraphBuilder:
         wire_task_event_nodes: bool,
     ) -> None:
         graph_inner.add_edge(loop_exit_node, "hook_after_model")
+        model_gate = _pre_model_gate_destination(loop_entry_node)
         graph_inner.add_conditional_edges(
             "hook_after_model",
             RunnableCallable(
                 _make_model_to_model_edge(
-                    model_destination=loop_entry_node,
+                    model_destination=model_gate,
                     end_destination="hook_stop",
                 ),
                 trace=False,
             ),
-            [loop_entry_node, "hook_stop"],
+            [model_gate, "hook_stop"],
         )
         self._wire_loop_controller(
             graph_inner=graph_inner,

langchain_agentx/observability/events/agent_event_projector.py

@@ -128,6 +128,7 @@ class LangchainAgentEventType(Enum):
     HOOK_PROGRESS = "hook-progress"
     HOOK_FINISHED = "hook-finished"
     COMPACT_BOUNDARY = "compact-boundary"
+    COMPACT_PROGRESS = "compact-progress"
     MEMORY_SAVED = "memory-saved"
     SUBAGENT_START = "subagent-start"
     SUBAGENT_END = "subagent-end"
@@ -850,6 +851,16 @@ class AgentEventProjector:
 
         if self._is_compact_node(name):
             meta = extract_compaction_meta_from_chain_end_data(data)
+            if isinstance(meta, dict) and self.enable_compact_events:
+                raw_progress = meta.get("compact_progress")
+                if isinstance(raw_progress, list):
+                    for item in raw_progress:
+                        if isinstance(item, dict):
+                            yield self._event(
+                                event_type=LangchainAgentEventType.COMPACT_PROGRESS,
+                                data=dict(item),
+                                step_index=self.state.step_index or None,
+                            )
             if is_user_visible_compact(meta):
                 if self.enable_compact_events:
                     compact_data = self._extract_compact_summary(data)

{langchain_agentx_python-2.3.3.dist-info → langchain_agentx_python-2.3.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: langchain-agentx-python
-Version: 2.3.3
+Version: 2.3.4
 Summary: LangChain/LangGraph-based agent utilities for CodeBaseX.
 Author-email: GoodMood2008 <GoodMood2008@users.noreply.github.com>
 License: Apache License

{langchain_agentx_python-2.3.3.dist-info → langchain_agentx_python-2.3.4.dist-info}/RECORD

@@ -1,4 +1,4 @@
-langchain_agentx/__init__.py,sha256=6ftpvlLKaDt-1ujsfwn1CpOzcZoVDPlRwOsS7KZDCR0,1614
+langchain_agentx/__init__.py,sha256=t5u-2g4slzGHVtYyJTJcFQgv_d5BfqHyJWeSDe9FMHY,1614
 langchain_agentx/command/__init__.py,sha256=Ej260S5uFQcmEtGZQG_NH7BYjHoStrpBLtGUsBTxyZk,631
 langchain_agentx/command/allowed_tools.py,sha256=TVA0VM8rm98H-QbLK_GRiX5RhEAZFxKawliMi4kk96A,3359
 langchain_agentx/command/context.py,sha256=DIGOGPBw5UGDBm0WNNJlKx1h_9rCRmcdROv4DT61Qug,731
@@ -31,8 +31,9 @@ langchain_agentx/loop/message_contract.py,sha256=Twl83TneU2PRdxgGRBxs6S9_ovRT6cb
 langchain_agentx/loop/run_context.py,sha256=dhNiLhvqD_QYacGMA0B5P3wWpSe3O6wU_XcVEzb86Z4,8303
 langchain_agentx/loop/config/__init__.py,sha256=0pf3Mml7LZpQ1C3_AtROFQClrBQegaGCG4YCTfSr9v8,1137
 langchain_agentx/loop/config/agent_config.py,sha256=tOnAHFor6dkle3FvUOVx0YsmiYVTH11UrWSyTmOLV0c,3100
+langchain_agentx/loop/config/compaction_config.py,sha256=MmXqY47-9nPDhdXvnU92Xp0XeaZ605ErQwX9d2emEkM,1176
 langchain_agentx/loop/config/interaction_profile.py,sha256=UWrlI3aTWjYRq1mDeXL-xL-iCrQU-bXDXhETP6q9icY,10020
-langchain_agentx/loop/config/loop_config.py,sha256=b87oJNcy9rYi-bYnw32OFCndXOPj54-ysrpe0F1fR2w,16147
+langchain_agentx/loop/config/loop_config.py,sha256=3kYDXg91PSDNVUdHGm6In0mHJRpfrkITs4OLKdOngYY,16180
 langchain_agentx/loop/config/loop_runtime_overlay.py,sha256=PuB4vFdqCHXRbAFhAJIrtwq9pUIrHafyn0ApD4s06KU,2440
 langchain_agentx/loop/config/model_context_resolver.py,sha256=o9RR0Vnu07TD4Y3V8jk5V1NfkMcPcLjE8uaaLOe22Qo,4067
 langchain_agentx/loop/config/model_retry_config.py,sha256=bakCbGeNRO-_SXMT_LLE3nu3m9scZe1MSxPdyEGO99w,8178
@@ -45,13 +46,13 @@ langchain_agentx/loop/context/__init__.py,sha256=RXVTFjHcH6_8q1x13Qk9ZP8RVbqZoyh
 langchain_agentx/loop/context/blocking_guard.py,sha256=Rmt5qZdE8Zf0bt-Q6WwMSfG869772PE8w1mNPvBQRCQ,3789
 langchain_agentx/loop/context/compaction_service.py,sha256=MwAPPaT31XbzgwqpuknGBZlYBRKxwyhWXDifEnBgh2E,3725
 langchain_agentx/loop/context/message_utils.py,sha256=6qIwtHeh63OlMqOnPUpEEWRknU9KJTGkcY1v3tz6feY,2010
-langchain_agentx/loop/context/pipeline.py,sha256=Ry90p-lrno2jgoKBLEUJle9evJCFJ_IBSe9sru0XjPM,7486
+langchain_agentx/loop/context/pipeline.py,sha256=PrhCWWKUGGwMGd-F8UX6vt3U7tlZo2ZZDLV0CTTDSJ4,8279
 langchain_agentx/loop/context/query_source.py,sha256=-wqVpST64jXF1ZelcGdvIsJE4ngv54eco4gmeS3pyyg,5118
 langchain_agentx/loop/context/recovery_mode.py,sha256=ytHGnE-g91pDQIkyVcfxEw6XgJGZ5eVJ12jb3zaeXWU,1100
-langchain_agentx/loop/context/settings.py,sha256=K9WRpvyUEo57gDxt4qpo0ouE8ihN5L0vGy3NmdIbssw,5987
-langchain_agentx/loop/context/types.py,sha256=UpnurkJfYsmbZXkrwbaDuVpZYJ4ydPHRNiUijL6fh3s,3382
+langchain_agentx/loop/context/settings.py,sha256=EO35T2P5tIUIqkO42ffLTRD-U-w2aRxeAYuHFHYc0nE,6589
+langchain_agentx/loop/context/types.py,sha256=jWXibGiJY6sQnd5-OQu9z7ZZM8Efi1X-6M8zJHNutqw,3637
 langchain_agentx/loop/context/stages/__init__.py,sha256=wk9CFh8_Akw8qs3Dldnx14VE1MhyF5i9Cg3phTVChhg,892
-langchain_agentx/loop/context/stages/autocompact.py,sha256=C2CW-H5NQbNSN7fCcxJ8o-liAc6uuPTBvoy4PJjdkGk,12014
+langchain_agentx/loop/context/stages/autocompact.py,sha256=7segcCEl0kk9hK6YRHAbM67e8As1sE_CRaUUfUnHFh0,13581
 langchain_agentx/loop/context/stages/base.py,sha256=LVfCp-DLySrTExy0J-M40tDJvzbOXxfZxXvi7XQayA8,1380
 langchain_agentx/loop/context/stages/collapse.py,sha256=yOf3nfWZr83rq2qnhUpuTCLFHcGvF7my8klMLmfZ7jc,2971
 langchain_agentx/loop/context/stages/microcompact.py,sha256=WnT58w85xcaSzsGEww5TEnU1noAzkao3ZgAARo65a7Y,3142
@@ -60,7 +61,7 @@ langchain_agentx/loop/context/stages/snip.py,sha256=AZy3D2XPrL1GQU_gE9isKApcphXw
 langchain_agentx/loop/context/stages/tool_result_budget.py,sha256=zQcJynRnC8QF30cEDduHkso6jChYjckVcnOjYSrCrtg,3833
 langchain_agentx/loop/exit/__init__.py,sha256=3-CMz5ApkJeBc549vRfgNABenyYKVbib4IdkErVtGcM,528
 langchain_agentx/loop/exit/api_error_detection.py,sha256=QzyMTov4VrkpHmn288Tw7DN00hiyJNXMKlOrCbvsiAA,1479
-langchain_agentx/loop/exit/exit_logic.py,sha256=nJzerVtTmS1iIY3PyBRet7jZ9v9pi-jUO3mJwzqxcnw,13922
+langchain_agentx/loop/exit/exit_logic.py,sha256=3I63n9pyjHD9CfpboWSLNm2O2J8c5PlVjcxesir40k8,14068
 langchain_agentx/loop/exit/query_source.py,sha256=b_VB13zxNbLAqUAuDt1E3SII3jNLagY-FdN6SbczTsI,2082
 langchain_agentx/loop/exit/reason_codes.py,sha256=Nk1x2ipftRIRa4U1Z2goyXXWB9TsPFL9TW6WfvtMCEo,4810
 langchain_agentx/loop/exit/resolution.py,sha256=taIMHzAUtgrQoU9446-CUXBN0PyCujhl6wBQfbxR1qI,1427
@@ -68,7 +69,7 @@ langchain_agentx/loop/exit/terminal.py,sha256=_DAZk-kS4Xd0nxMlayLo60-uj1Y5ozgBjL
 langchain_agentx/loop/exit/withheld_error.py,sha256=LfX3j7Rgd9VoMzlz7GOs9TPuRe7RXr_Gxweyotx3tqQ,1846
 langchain_agentx/loop/graph/__init__.py,sha256=9pUUqn7G87n3lV45iYIU24j0wcTHyCD--SSdtv3urh4,136
 langchain_agentx/loop/graph/builtin_loop_control.py,sha256=31o4H6SDGC_vECSliXecuVEnzlD1ahlVZ5JzEpxYY04,9248
-langchain_agentx/loop/graph/factory.py,sha256=-RDgiprUAydk_tSA86CyI7zgj9wC6gPen2p39OFscIU,105591
+langchain_agentx/loop/graph/factory.py,sha256=n0lszl2jYBq1G_rzgmt0-0mCmTa3zuTqrVCow4lUMso,109909
 langchain_agentx/loop/graph/graph_edges.py,sha256=kW6Npx-nMIPzcaisCxVPkTwf5Y9GOKy1NWCNkJWeJBQ,47370
 langchain_agentx/loop/graph/reactive_compact_node.py,sha256=633b5hTnR3BK8GBEKWhzOwGB0BQrVRKozNxDomWFjLo,8762
 langchain_agentx/loop/graph/runtime_tools_node.py,sha256=wQkbc0ophyefpPZEc0q1vKoJOIdS4S6d6hSsrkdRLDk,7387
@@ -217,7 +218,7 @@ langchain_agentx/observability/evaluation/checkers/tool_behavior.py,sha256=o1nH8
 langchain_agentx/observability/events/__init__.py,sha256=Zr9Iguf9XsJQLsCJy13YFdBtXHPEnI9aSWXWmA2MisA,1497
 langchain_agentx/observability/events/accumulated_text_scope.py,sha256=M4OvuQTgReF8SFrOmr-M5vyOJpzfvxj_DKD5bVpGb70,3169
 langchain_agentx/observability/events/agent_defaults.py,sha256=gmX0P__lux5xbCb8dtAlVCb3wZ7avaqFleoW4Q5anw4,1187
-langchain_agentx/observability/events/agent_event_projector.py,sha256=Rktmy2D3qizCejAChbuOiAa06Tnup2Z6cVNATfrrxkU,63995
+langchain_agentx/observability/events/agent_event_projector.py,sha256=SthsPhWuM__Bfm3GoxHw-JCoMQ5xgn_eDgi5n4ib6-4,64600
 langchain_agentx/observability/events/api_retry_projection.py,sha256=XGb09sNOi6k1ZGW4YqMLBE8GDxNwNMprk62jV9TS9dQ,2433
 langchain_agentx/observability/events/compact_event_gate.py,sha256=50RJemwB1lnQn_NFWOFSgU418ZUR8dbd_v0538PT7zM,4751
 langchain_agentx/observability/events/factory.py,sha256=gqMATIK9tBQzaPWSPhCCSPuCq-WF562ZDTnFUaMm9Qc,1570
@@ -766,8 +767,8 @@ langchain_agentx/workspace/root_grant_manager.py,sha256=W3gy8HTevbERbXqIgRcjzOXO
 langchain_agentx/workspace/tool_boundary.py,sha256=UDwX6swpSLsx9HNkYuuRRiV5o7ZZG_Bru2Yn0c5kNX8,5724
 langchain_agentx/workspace/validators.py,sha256=tQt-6TOcL8Fw7Ig5ebA9S7vGWh1rby920eFW6x8Tk9E,1439
 langchain_agentx/workspace/view.py,sha256=PGasqTaqhlD03SXXazHuw4RHfV681AIdlsYqL70pEjc,4774
-langchain_agentx_python-2.3.3.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-langchain_agentx_python-2.3.3.dist-info/METADATA,sha256=e_CGB6uG3nhxDnh2VADLt975KFC-uVAARQlRtha-3es,23942
-langchain_agentx_python-2.3.3.dist-info/WHEEL,sha256=51RkbunBAw4BWsgaQWTpPhg4Diwp3c9P5iaLk67Hdtg,92
-langchain_agentx_python-2.3.3.dist-info/top_level.txt,sha256=Ge284pniNt8xea0OLk2o9o32GqVpDhOYk20fwE-0xxA,17
-langchain_agentx_python-2.3.3.dist-info/RECORD,,
+langchain_agentx_python-2.3.4.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+langchain_agentx_python-2.3.4.dist-info/METADATA,sha256=zMKaBp2iVa62_ASW-y0xHpoqDFuMu64NDF414ESG2XA,23942
+langchain_agentx_python-2.3.4.dist-info/WHEEL,sha256=51RkbunBAw4BWsgaQWTpPhg4Diwp3c9P5iaLk67Hdtg,92
+langchain_agentx_python-2.3.4.dist-info/top_level.txt,sha256=Ge284pniNt8xea0OLk2o9o32GqVpDhOYk20fwE-0xxA,17
+langchain_agentx_python-2.3.4.dist-info/RECORD,,