fix: preserve exploration tool history and cap context budget

artemis/agents/operator/operator.py

@@ -17,7 +17,7 @@ from collections.abc import Awaitable
 from typing import Any
 from uuid import uuid4
 
-from langchain_core.messages import SystemMessage, ToolMessage
+from langchain_core.messages import BaseMessage, SystemMessage, ToolMessage
 from langchain_core.tools import BaseTool
 
 from artemis.agents.validator.tool_declarations import normalize_coordinate_target
@@ -146,9 +146,9 @@ class OperatorNode:
                 from artemis.config.agent import MemoryTranscriptConfig
 
                 self._transcript_cfg = MemoryTranscriptConfig()
-        # Index into the rendered message list where this turn's tail begins;
-        # set by the transcript build and consumed after the tool loop.
-        self._transcript_turn_base: int | None = None
+        # Keep the actual observation, not an index: plugin/exploration context
+        # may be prepended after render, shifting every message's position.
+        self._transcript_turn_tail: BaseMessage | None = None
         self._turn_plan_baseline: str | None = None
         self._turn_ledger_streak: int = 0
         # Whether the previous turn's replies carried no visible reasoning text;
@@ -225,7 +225,7 @@ class OperatorNode:
         active_background_tasks: list[dict] = None,
         newly_finished_tasks: list[dict] = None,
     ) -> list:
-        self._transcript_turn_base = None
+        self._transcript_turn_tail = None
         if getattr(self._transcript_cfg, "enabled", False):
             try:
                 return await self._build_prompt_transcript(
@@ -243,7 +243,7 @@ class OperatorNode:
                     f"Transcript prompt path failed; falling back to the legacy"
                     f" 2-message build for this turn: {e}"
                 )
-                self._transcript_turn_base = None
+                self._transcript_turn_tail = None
 
         plan_and_history = self._build_plan_and_history(steps, task_plan)
 
@@ -468,18 +468,19 @@ class OperatorNode:
             mark_ephemeral(tail, builder.ephemeral_indices)
 
         messages = ledger.render([tail])
-        self._transcript_turn_base = len(messages) - 1
+        self._transcript_turn_tail = tail
         return messages
 
     def _stage_transcript_turn(self, messages: list) -> None:
         """Hand this turn's tail + tool-loop products to the ledger."""
-        base = self._transcript_turn_base
-        self._transcript_turn_base = None
-        if base is None:
+        tail = self._transcript_turn_tail
+        self._transcript_turn_tail = None
+        if tail is None:
             return
         try:
             ledger = getattr(self.ctx, "transcript_ledger", None)
             if ledger is not None:
+                base = next(i for i, message in enumerate(messages) if message is tail)
                 ledger.stage_turn(messages[base:])
         except Exception as e:
             logger.error(f"Failed to stage transcript turn: {e}")

artemis/config/agent.py

@@ -23,6 +23,7 @@ from pydantic import BaseModel, Field, model_validator
 from artemis.config.constants import (
     AGENT_CONFIG_FILENAME,
     DEFAULT_EXPLORER_VERSION,
+    MAX_CONTEXT_WINDOW_TOKENS,
     ExplorerVersion,
 )
 from artemis.config.paths import ROOT_DIR, get_config_path
@@ -483,6 +484,7 @@ class MemoryTranscriptConfig(BaseModel):
     context_budget_tokens: int = Field(
         default=80000,
         ge=1024,
+        le=MAX_CONTEXT_WINDOW_TOKENS,
         description=(
             "Measured-token budget for the transcript context (provisional"
             " default pending the on-device baseline; §3.4). The soft/hard"

artemis/config/constants.py

@@ -16,6 +16,9 @@
 
 from typing import Literal
 
+# Application context ceiling; independent from output and session token totals.
+MAX_CONTEXT_WINDOW_TOKENS = 300_000
+
 # ==============================================================================
 # Environment Variable Keys
 # ==============================================================================

artemis/llm/reliability.py

@@ -81,6 +81,11 @@ _AUTH_MARKERS = ("unauthorized", "forbidden", "api key", "credential")
 # "deadline exceeded"/"deadline_exceeded" cover google-genai SDK timeouts that
 # surface without an HTTP status code attached.
 _TIMEOUT_MARKERS = ("timed out", "timeout", "deadline exceeded", "deadline_exceeded")
+_TOOL_HISTORY_ERRORS = (
+    "no tool call found for function call output",
+    "no tool output found for function call",
+    "tool_call_id not found",
+)
 
 
 def classify_failure(error: BaseException) -> Failure:
@@ -91,6 +96,12 @@ def classify_failure(error: BaseException) -> Failure:
     code = extract_status_code(error)
     message = str(error).lower()
 
+    # Compatible Responses endpoints can report input validation inside an
+    # HTTP 200 stream as APIError without a status code. Retrying the same
+    # broken call/output history cannot repair it, even on a fallback model.
+    if any(marker in message for marker in _TOOL_HISTORY_ERRORS):
+        return Failure(FailureCategory.BAD_REQUEST, False, False)
+
     if code == 429 or any(marker in message for marker in _RATE_LIMIT_MARKERS):
         return Failure(FailureCategory.RATE_LIMIT, True, True)
     if code in {500, 502, 503, 504} or any(marker in message for marker in _UNAVAILABLE_MARKERS):

artemis/resources/config/artemis.jsonc

@@ -94,6 +94,7 @@
       },
       "transcript": {
         "enabled": true,
+        // 历史压缩预算，允许范围 1024–300000；上下文参照上限固定为 300k。
         "context_budget_tokens": 80000,
         "start_ratio": 0.35,
         "soft_ratio": 0.7,

artemis/services/token_meter.py

@@ -40,6 +40,7 @@ import threading
 from typing import Any
 
 from artemis.data_engine.context_vars import CURRENT_NODE_NAME
+from artemis.config.constants import MAX_CONTEXT_WINDOW_TOKENS
 from artemis.llm.usage import usage_from_message
 from artemis.utils.logger import get_logger
 
@@ -52,7 +53,7 @@ _METERS: dict[str, "SessionTokenMeter"] = {}
 # measured against when the UI reports "context used". The prompt size of the
 # last executor call (``context_base_tokens`` on its ``llm_usage`` trace) is
 # the numerator; this is the denominator.
-OPERATOR_CONTEXT_WINDOW_TOKENS = 1_000_000
+OPERATOR_CONTEXT_WINDOW_TOKENS = MAX_CONTEXT_WINDOW_TOKENS
 
 # Traced node names whose LLM calls carry the live executor context. Both the
 # Pro Operator and the Flash runner are "the operator" from the user's view.