fix(agent): gate microcompaction on context pressure

Extract model-facing context governance from AgentRunner.

Only compact in-flight tool results when the model request is over budget, keep compacted IDs stable within a turn, and allow the newest result to be compacted as a last resort when it is the remaining source of overflow.

nanobot/agent/context_governance.py

@@ -0,0 +1,391 @@
+"""Model-message governance for agent runner requests.
+
+This module owns model-facing message shaping and tool-result content normalization.
+It may return copied messages or persisted-result placeholders, but it must not
+mutate an existing session history list in place.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+
+from loguru import logger
+
+from nanobot.utils.helpers import (
+    estimate_message_tokens,
+    estimate_prompt_tokens_chain,
+    find_legal_message_start,
+    maybe_persist_tool_result,
+    truncate_text,
+)
+from nanobot.utils.runtime import ensure_nonempty_tool_result
+
+if TYPE_CHECKING:
+    from nanobot.providers.base import LLMProvider
+
+SNIP_SAFETY_BUFFER = 1024
+MICROCOMPACT_KEEP_RECENT = 10
+MICROCOMPACT_MIN_CHARS = 500
+INFLIGHT_COMPACT_TARGET_RATIO = 0.85
+COMPACTABLE_TOOLS = frozenset({
+    "read_file", "exec", "grep", "find_files",
+    "web_search", "web_fetch", "list_dir", "list_exec_sessions",
+})
+# read_file is the recovery path for persisted results; exempting it prevents persist->read->persist loops.
+TOOL_RESULT_OFFLOAD_EXEMPT_TOOLS = frozenset({"read_file"})
+BACKFILL_CONTENT = "[Tool result unavailable — call was interrupted or lost]"
+
+
+@dataclass(slots=True)
+class ContextGovernanceConfig:
+    provider: LLMProvider
+    model: str
+    tools: Any
+    workspace: Path | None
+    session_key: str | None
+    max_tool_result_chars: int
+    context_window_tokens: int | None = None
+    context_block_limit: int | None = None
+    max_tokens: int | None = None
+    inflight_start_index: int = 0
+
+
+class ContextGovernor:
+    """Prepare model-copy messages while preserving persisted history."""
+
+    def prepare_for_model(
+        self,
+        config: ContextGovernanceConfig,
+        messages: list[dict[str, Any]],
+        compacted_tool_call_ids: set[str],
+    ) -> list[dict[str, Any]]:
+        updated = self.drop_orphan_tool_results(messages)
+        updated = self.backfill_missing_tool_results(updated)
+        updated = self.apply_tool_result_budget(config, updated)
+        updated = self.compact_inflight_overflow(config, updated, compacted_tool_call_ids)
+        updated = self.snip_history(config, updated)
+        updated = self.drop_orphan_tool_results(updated)
+        return self.backfill_missing_tool_results(updated)
+
+    @staticmethod
+    def input_budget(config: ContextGovernanceConfig) -> int:
+        if not config.context_window_tokens:
+            return 0
+
+        provider_max_tokens = getattr(
+            getattr(config.provider, "generation", None),
+            "max_tokens",
+            4096,
+        )
+        max_output = config.max_tokens if isinstance(config.max_tokens, int) else (
+            provider_max_tokens if isinstance(provider_max_tokens, int) else 4096
+        )
+        budget = config.context_block_limit or (
+            config.context_window_tokens - max_output - SNIP_SAFETY_BUFFER
+        )
+        return budget if budget > 0 else 0
+
+    @staticmethod
+    def normalize_tool_result(
+        config: ContextGovernanceConfig,
+        tool_call_id: str,
+        tool_name: str,
+        result: Any,
+    ) -> Any:
+        result = ensure_nonempty_tool_result(tool_name, result)
+        if tool_name in TOOL_RESULT_OFFLOAD_EXEMPT_TOOLS:
+            return result
+        try:
+            content = maybe_persist_tool_result(
+                config.workspace,
+                config.session_key,
+                tool_call_id,
+                result,
+                max_chars=config.max_tool_result_chars,
+            )
+        except Exception:
+            logger.exception(
+                "Tool result persist failed for {} in {}; using raw result",
+                tool_call_id,
+                config.session_key or "default",
+            )
+            content = result
+        if isinstance(content, str) and len(content) > config.max_tool_result_chars:
+            return truncate_text(content, config.max_tool_result_chars)
+        return content
+
+    @staticmethod
+    def drop_orphan_tool_results(
+        messages: list[dict[str, Any]],
+    ) -> list[dict[str, Any]]:
+        """Drop tool results that have no matching assistant tool_call earlier in history."""
+        declared: set[str] = set()
+        updated: list[dict[str, Any]] | None = None
+        for idx, msg in enumerate(messages):
+            role = msg.get("role")
+            if role == "assistant":
+                for tc in msg.get("tool_calls") or []:
+                    if isinstance(tc, dict) and tc.get("id"):
+                        declared.add(str(tc["id"]))
+            if role == "tool":
+                tid = msg.get("tool_call_id")
+                if tid and str(tid) not in declared:
+                    if updated is None:
+                        updated = [dict(m) for m in messages[:idx]]
+                    continue
+            if updated is not None:
+                updated.append(dict(msg))
+
+        if updated is None:
+            return messages
+        return updated
+
+    @staticmethod
+    def backfill_missing_tool_results(
+        messages: list[dict[str, Any]],
+    ) -> list[dict[str, Any]]:
+        """Insert synthetic error results for assistant tool_calls with missing tool outputs."""
+        declared: list[tuple[int, str, str]] = []
+        fulfilled: set[str] = set()
+        for idx, msg in enumerate(messages):
+            role = msg.get("role")
+            if role == "assistant":
+                for tc in msg.get("tool_calls") or []:
+                    if isinstance(tc, dict) and tc.get("id"):
+                        name = ""
+                        func = tc.get("function")
+                        if isinstance(func, dict):
+                            name = func.get("name", "")
+                        declared.append((idx, str(tc["id"]), name))
+            elif role == "tool":
+                tid = msg.get("tool_call_id")
+                if tid:
+                    fulfilled.add(str(tid))
+
+        missing = [(ai, cid, name) for ai, cid, name in declared if cid not in fulfilled]
+        if not missing:
+            return messages
+
+        updated = list(messages)
+        offset = 0
+        for assistant_idx, call_id, name in missing:
+            insert_at = assistant_idx + 1 + offset
+            while insert_at < len(updated) and updated[insert_at].get("role") == "tool":
+                insert_at += 1
+            updated.insert(insert_at, {
+                "role": "tool",
+                "tool_call_id": call_id,
+                "name": name,
+                "content": BACKFILL_CONTENT,
+            })
+            offset += 1
+        return updated
+
+    def apply_tool_result_budget(
+        self,
+        config: ContextGovernanceConfig,
+        messages: list[dict[str, Any]],
+    ) -> list[dict[str, Any]]:
+        updated = messages
+        for idx, message in enumerate(messages):
+            if message.get("role") != "tool":
+                continue
+            normalized = self.normalize_tool_result(
+                config,
+                str(message.get("tool_call_id") or f"tool_{idx}"),
+                str(message.get("name") or "tool"),
+                message.get("content"),
+            )
+            if normalized != message.get("content"):
+                if updated is messages:
+                    updated = [dict(m) for m in messages]
+                updated[idx]["content"] = normalized
+        return updated
+
+    def compact_inflight_overflow(
+        self,
+        config: ContextGovernanceConfig,
+        messages: list[dict[str, Any]],
+        compacted_tool_call_ids: set[str],
+    ) -> list[dict[str, Any]]:
+        """Compact in-flight tool results only when the request would overflow."""
+        budget = self.input_budget(config)
+        if budget <= 0:
+            return messages
+
+        tools = config.tools.get_definitions()
+        updated = self._apply_recorded_compactions(messages, compacted_tool_call_ids)
+        estimate, source = estimate_prompt_tokens_chain(
+            config.provider,
+            config.model,
+            updated,
+            tools,
+        )
+        if estimate <= budget:
+            return updated
+
+        target = int(budget * INFLIGHT_COMPACT_TARGET_RATIO)
+        candidates = self._inflight_compaction_candidates(
+            config,
+            updated,
+            compacted_tool_call_ids,
+        )
+        if not candidates:
+            return updated
+
+        for candidate_idx, (idx, tool_call_id) in enumerate(candidates):
+            is_newest_candidate = candidate_idx == len(candidates) - 1
+            if is_newest_candidate and estimate <= budget:
+                break
+            if tool_call_id in compacted_tool_call_ids:
+                continue
+            if updated is messages:
+                updated = [dict(m) for m in messages]
+            compacted_tool_call_ids.add(tool_call_id)
+            self._compact_tool_result_at(updated, idx)
+            estimate, source = estimate_prompt_tokens_chain(
+                config.provider,
+                config.model,
+                updated,
+                tools,
+            )
+            if estimate <= target:
+                break
+
+        logger.debug(
+            "In-flight context compaction for {}: prompt={} budget={} target={} via {}, ids={}",
+            config.session_key or "default",
+            estimate,
+            budget,
+            target,
+            source,
+            len(compacted_tool_call_ids),
+        )
+        return updated
+
+    def snip_history(
+        self,
+        config: ContextGovernanceConfig,
+        messages: list[dict[str, Any]],
+    ) -> list[dict[str, Any]]:
+        if not messages or not config.context_window_tokens:
+            return messages
+
+        budget = self.input_budget(config)
+        if budget <= 0:
+            return messages
+
+        tools = config.tools.get_definitions()
+        estimate, _ = estimate_prompt_tokens_chain(
+            config.provider,
+            config.model,
+            messages,
+            tools,
+        )
+        if estimate <= budget:
+            return messages
+
+        system_messages = [dict(msg) for msg in messages if msg.get("role") == "system"]
+        non_system = [dict(msg) for msg in messages if msg.get("role") != "system"]
+        if not non_system:
+            return messages
+
+        system_tokens = sum(estimate_message_tokens(msg) for msg in system_messages)
+        fixed_tokens, _ = estimate_prompt_tokens_chain(
+            config.provider,
+            config.model,
+            system_messages,
+            tools,
+        )
+        remaining_budget = max(0, budget - max(system_tokens, fixed_tokens))
+        kept: list[dict[str, Any]] = []
+        kept_tokens = 0
+        for message in reversed(non_system):
+            msg_tokens = estimate_message_tokens(message)
+            if kept and kept_tokens + msg_tokens > remaining_budget:
+                break
+            kept.append(message)
+            kept_tokens += msg_tokens
+        kept.reverse()
+
+        return system_messages + self._legal_history_tail(kept, non_system)
+
+    @staticmethod
+    def _summary_for(message: dict[str, Any]) -> str:
+        name = message.get("name", "tool")
+        return f"[{name} result omitted from context]"
+
+    def _legal_history_tail(
+        self,
+        kept: list[dict[str, Any]],
+        non_system: list[dict[str, Any]],
+    ) -> list[dict[str, Any]]:
+        fallback = kept if kept else (non_system[-1:] if non_system else [])
+        kept = self._user_tail(kept) or self._user_tail(non_system, last=True) or fallback
+
+        start = find_legal_message_start(kept)
+        return kept[start:] if start else kept
+
+    @staticmethod
+    def _user_tail(messages: list[dict[str, Any]], *, last: bool = False) -> list[dict[str, Any]]:
+        indexes = range(len(messages) - 1, -1, -1) if last else range(len(messages))
+        for idx in indexes:
+            if messages[idx].get("role") == "user":
+                return messages[idx:]
+        return []
+
+    def _apply_recorded_compactions(
+        self,
+        messages: list[dict[str, Any]],
+        compacted_tool_call_ids: set[str],
+    ) -> list[dict[str, Any]]:
+        if not compacted_tool_call_ids:
+            return messages
+        updated = messages
+        for idx, msg in enumerate(messages):
+            if msg.get("role") != "tool":
+                continue
+            tool_call_id = msg.get("tool_call_id")
+            if not tool_call_id or str(tool_call_id) not in compacted_tool_call_ids:
+                continue
+            summary = self._summary_for(msg)
+            if msg.get("content") == summary:
+                continue
+            if updated is messages:
+                updated = [dict(m) for m in messages]
+            updated[idx]["content"] = summary
+        return updated
+
+    def _inflight_compaction_candidates(
+        self,
+        config: ContextGovernanceConfig,
+        messages: list[dict[str, Any]],
+        compacted_tool_call_ids: set[str],
+    ) -> list[tuple[int, str]]:
+        compactable: list[tuple[int, str]] = []
+        for idx, msg in enumerate(messages):
+            if idx < config.inflight_start_index:
+                continue
+            if msg.get("role") != "tool" or msg.get("name") not in COMPACTABLE_TOOLS:
+                continue
+            tool_call_id = msg.get("tool_call_id")
+            if not tool_call_id or str(tool_call_id) in compacted_tool_call_ids:
+                continue
+            content = msg.get("content")
+            if not isinstance(content, str) or len(content) < MICROCOMPACT_MIN_CHARS:
+                continue
+            compactable.append((idx, str(tool_call_id)))
+
+        if not compactable:
+            return []
+        primary_count = max(0, len(compactable) - MICROCOMPACT_KEEP_RECENT)
+        primary = compactable[:primary_count]
+        # Hard overflow beats the keep-recent preference. Return recent results
+        # after stale ones so the newest result is naturally last.
+        fallback = compactable[primary_count:]
+        return primary + fallback
+
+    def _compact_tool_result_at(self, messages: list[dict[str, Any]], idx: int) -> None:
+        messages[idx]["content"] = self._summary_for(messages[idx])

nanobot/agent/runner.py

@@ -13,6 +13,10 @@ from typing import Any, Callable
 
 from loguru import logger
 
+from nanobot.agent.context_governance import (
+    ContextGovernanceConfig,
+    ContextGovernor,
+)
 from nanobot.agent.hook import AgentHook, AgentHookContext, AgentRunHookContext
 from nanobot.agent.tools.registry import ToolRegistry
 from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
@@ -32,11 +36,8 @@ from nanobot.utils.helpers import (
     estimate_message_tokens,
     estimate_prompt_tokens_chain,
     extract_reasoning,
-    find_legal_message_start,
-    maybe_persist_tool_result,
     strip_reasoning_tags,
     strip_think,
-    truncate_text,
 )
 from nanobot.utils.progress_events import (
     invoke_file_edit_progress,
@@ -49,7 +50,6 @@ from nanobot.utils.runtime import (
     build_finalization_retry_message,
     build_goal_continue_message,
     build_length_recovery_message,
-    ensure_nonempty_tool_result,
     is_blank_text,
     repeated_external_lookup_error,
     repeated_workspace_violation_error,
@@ -67,17 +67,6 @@ _MAX_EMPTY_RETRIES = 2
 _MAX_LENGTH_RECOVERIES = 3
 _MAX_INJECTIONS_PER_TURN = 3
 _MAX_INJECTION_CYCLES = 5
-_SNIP_SAFETY_BUFFER = 1024
-_MICROCOMPACT_KEEP_RECENT = 10
-_MICROCOMPACT_MIN_CHARS = 500
-_COMPACTABLE_TOOLS = frozenset({
-    "read_file", "exec", "grep", "find_files",
-    "web_search", "web_fetch", "list_dir", "list_exec_sessions",
-})
-# read_file is the recovery path for persisted results; exempting it prevents persist->read->persist loops.
-_TOOL_RESULT_OFFLOAD_EXEMPT_TOOLS = frozenset({"read_file"})
-_BACKFILL_CONTENT = "[Tool result unavailable — call was interrupted or lost]"
-
 # Backward-compatible module attribute for tests/extensions that monkeypatch
 # the former single-file tracker hook. Runtime uses prepare_file_edit_trackers.
 prepare_file_edit_tracker = _prepare_file_edit_tracker
@@ -135,6 +124,7 @@ class AgentRunner:
 
     def __init__(self, provider: LLMProvider):
         self.provider = provider
+        self.context_governor = ContextGovernor()
 
     @staticmethod
     def _merge_message_content(left: Any, right: Any) -> str | list[dict[str, Any]]:
@@ -367,6 +357,19 @@ class AgentRunner:
         length_recovery_count = 0
         had_injections = False
         injection_cycles = 0
+        compacted_tool_call_ids: set[str] = set()
+        governance_config = ContextGovernanceConfig(
+            provider=self.provider,
+            model=spec.model,
+            tools=spec.tools,
+            workspace=spec.workspace,
+            session_key=spec.session_key,
+            max_tool_result_chars=spec.max_tool_result_chars,
+            context_window_tokens=spec.context_window_tokens,
+            context_block_limit=spec.context_block_limit,
+            max_tokens=spec.max_tokens,
+            inflight_start_index=len(spec.initial_messages),
+        )
 
         for iteration in range(spec.max_iterations):
             try:
@@ -374,14 +377,11 @@ class AgentRunner:
                 # may repair or compact historical messages for the model, but
                 # those synthetic edits must not shift the append boundary used
                 # later when the caller saves only the new turn.
-                messages_for_model = self._drop_orphan_tool_results(messages)
-                messages_for_model = self._backfill_missing_tool_results(messages_for_model)
-                messages_for_model = self._microcompact(messages_for_model)
-                messages_for_model = self._apply_tool_result_budget(spec, messages_for_model)
-                messages_for_model = self._snip_history(spec, messages_for_model)
-                # Snipping may have created new orphans; clean them up.
-                messages_for_model = self._drop_orphan_tool_results(messages_for_model)
-                messages_for_model = self._backfill_missing_tool_results(messages_for_model)
+                messages_for_model = self.context_governor.prepare_for_model(
+                    governance_config,
+                    messages,
+                    compacted_tool_call_ids,
+                )
             except Exception:
                 logger.exception(
                     "Context governance failed on turn {} for {}; applying minimal repair",
@@ -389,8 +389,10 @@ class AgentRunner:
                     spec.session_key or "default",
                 )
                 try:
-                    messages_for_model = self._drop_orphan_tool_results(messages)
-                    messages_for_model = self._backfill_missing_tool_results(messages_for_model)
+                    messages_for_model = ContextGovernor.drop_orphan_tool_results(messages)
+                    messages_for_model = ContextGovernor.backfill_missing_tool_results(
+                        messages_for_model
+                    )
                 except Exception:
                     messages_for_model = messages
             context = AgentHookContext(
@@ -463,8 +465,8 @@ class AgentRunner:
                         "role": "tool",
                         "tool_call_id": tool_call.id,
                         "name": tool_call.name,
-                        "content": self._normalize_tool_result(
-                            spec,
+                        "content": self.context_governor.normalize_tool_result(
+                            governance_config,
                             tool_call.id,
                             tool_call.name,
                             result,
@@ -1334,225 +1336,6 @@ class AgentRunner:
             return
         messages.append(build_assistant_message(_PERSISTED_MODEL_ERROR_PLACEHOLDER))
 
-    def _normalize_tool_result(
-        self,
-        spec: AgentRunSpec,
-        tool_call_id: str,
-        tool_name: str,
-        result: Any,
-    ) -> Any:
-        result = ensure_nonempty_tool_result(tool_name, result)
-        if tool_name in _TOOL_RESULT_OFFLOAD_EXEMPT_TOOLS:
-            # Exempt tools bound their own output; skip generic offload and truncation.
-            return result
-        try:
-            content = maybe_persist_tool_result(
-                spec.workspace,
-                spec.session_key,
-                tool_call_id,
-                result,
-                max_chars=spec.max_tool_result_chars,
-            )
-        except Exception:
-            logger.exception(
-                "Tool result persist failed for {} in {}; using raw result",
-                tool_call_id,
-                spec.session_key or "default",
-            )
-            content = result
-        if isinstance(content, str) and len(content) > spec.max_tool_result_chars:
-            return truncate_text(content, spec.max_tool_result_chars)
-        return content
-
-    @staticmethod
-    def _drop_orphan_tool_results(
-        messages: list[dict[str, Any]],
-    ) -> list[dict[str, Any]]:
-        """Drop tool results that have no matching assistant tool_call earlier in the history."""
-        declared: set[str] = set()
-        updated: list[dict[str, Any]] | None = None
-        for idx, msg in enumerate(messages):
-            role = msg.get("role")
-            if role == "assistant":
-                for tc in msg.get("tool_calls") or []:
-                    if isinstance(tc, dict) and tc.get("id"):
-                        declared.add(str(tc["id"]))
-            if role == "tool":
-                tid = msg.get("tool_call_id")
-                if tid and str(tid) not in declared:
-                    if updated is None:
-                        updated = [dict(m) for m in messages[:idx]]
-                    continue
-            if updated is not None:
-                updated.append(dict(msg))
-
-        if updated is None:
-            return messages
-        return updated
-
-    @staticmethod
-    def _backfill_missing_tool_results(
-        messages: list[dict[str, Any]],
-    ) -> list[dict[str, Any]]:
-        """Insert synthetic error results for orphaned tool_use blocks."""
-        declared: list[tuple[int, str, str]] = []  # (assistant_idx, call_id, name)
-        fulfilled: set[str] = set()
-        for idx, msg in enumerate(messages):
-            role = msg.get("role")
-            if role == "assistant":
-                for tc in msg.get("tool_calls") or []:
-                    if isinstance(tc, dict) and tc.get("id"):
-                        name = ""
-                        func = tc.get("function")
-                        if isinstance(func, dict):
-                            name = func.get("name", "")
-                        declared.append((idx, str(tc["id"]), name))
-            elif role == "tool":
-                tid = msg.get("tool_call_id")
-                if tid:
-                    fulfilled.add(str(tid))
-
-        missing = [(ai, cid, name) for ai, cid, name in declared if cid not in fulfilled]
-        if not missing:
-            return messages
-
-        updated = list(messages)
-        offset = 0
-        for assistant_idx, call_id, name in missing:
-            insert_at = assistant_idx + 1 + offset
-            while insert_at < len(updated) and updated[insert_at].get("role") == "tool":
-                insert_at += 1
-            updated.insert(insert_at, {
-                "role": "tool",
-                "tool_call_id": call_id,
-                "name": name,
-                "content": _BACKFILL_CONTENT,
-            })
-            offset += 1
-        return updated
-
-    @staticmethod
-    def _microcompact(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
-        """Replace old compactable tool results with one-line summaries."""
-        compactable_indices: list[int] = []
-        for idx, msg in enumerate(messages):
-            if msg.get("role") == "tool" and msg.get("name") in _COMPACTABLE_TOOLS:
-                compactable_indices.append(idx)
-
-        if len(compactable_indices) <= _MICROCOMPACT_KEEP_RECENT:
-            return messages
-
-        stale = compactable_indices[: len(compactable_indices) - _MICROCOMPACT_KEEP_RECENT]
-        updated: list[dict[str, Any]] | None = None
-        for idx in stale:
-            msg = messages[idx]
-            content = msg.get("content")
-            if not isinstance(content, str) or len(content) < _MICROCOMPACT_MIN_CHARS:
-                continue
-            name = msg.get("name", "tool")
-            summary = f"[{name} result omitted from context]"
-            if updated is None:
-                updated = [dict(m) for m in messages]
-            updated[idx]["content"] = summary
-
-        return updated if updated is not None else messages
-
-    def _apply_tool_result_budget(
-        self,
-        spec: AgentRunSpec,
-        messages: list[dict[str, Any]],
-    ) -> list[dict[str, Any]]:
-        updated = messages
-        for idx, message in enumerate(messages):
-            if message.get("role") != "tool":
-                continue
-            normalized = self._normalize_tool_result(
-                spec,
-                str(message.get("tool_call_id") or f"tool_{idx}"),
-                str(message.get("name") or "tool"),
-                message.get("content"),
-            )
-            if normalized != message.get("content"):
-                if updated is messages:
-                    updated = [dict(m) for m in messages]
-                updated[idx]["content"] = normalized
-        return updated
-
-    def _snip_history(
-        self,
-        spec: AgentRunSpec,
-        messages: list[dict[str, Any]],
-    ) -> list[dict[str, Any]]:
-        if not messages or not spec.context_window_tokens:
-            return messages
-
-        provider_max_tokens = getattr(getattr(self.provider, "generation", None), "max_tokens", 4096)
-        max_output = spec.max_tokens if isinstance(spec.max_tokens, int) else (
-            provider_max_tokens if isinstance(provider_max_tokens, int) else 4096
-        )
-        budget = spec.context_block_limit or (
-            spec.context_window_tokens - max_output - _SNIP_SAFETY_BUFFER
-        )
-        if budget <= 0:
-            return messages
-
-        estimate, _ = estimate_prompt_tokens_chain(
-            self.provider,
-            spec.model,
-            messages,
-            spec.tools.get_definitions(),
-        )
-        if estimate <= budget:
-            return messages
-
-        system_messages = [dict(msg) for msg in messages if msg.get("role") == "system"]
-        non_system = [dict(msg) for msg in messages if msg.get("role") != "system"]
-        if not non_system:
-            return messages
-
-        system_tokens = sum(estimate_message_tokens(msg) for msg in system_messages)
-        fixed_tokens, _ = estimate_prompt_tokens_chain(
-            self.provider,
-            spec.model,
-            system_messages,
-            spec.tools.get_definitions(),
-        )
-        remaining_budget = max(0, budget - max(system_tokens, fixed_tokens))
-        kept: list[dict[str, Any]] = []
-        kept_tokens = 0
-        for message in reversed(non_system):
-            msg_tokens = estimate_message_tokens(message)
-            if kept and kept_tokens + msg_tokens > remaining_budget:
-                break
-            kept.append(message)
-            kept_tokens += msg_tokens
-        kept.reverse()
-
-        if kept:
-            for i, message in enumerate(kept):
-                if message.get("role") == "user":
-                    kept = kept[i:]
-                    break
-            else:
-                # Recover nearest user message from outside the kept window;
-                # GLM rejects system→assistant (error 1214).  Budget is
-                # intentionally exceeded — oversized beats invalid.
-                for idx in range(len(non_system) - 1, -1, -1):
-                    if non_system[idx].get("role") == "user":
-                        kept = non_system[idx:]
-                        break
-                # If no user exists at all, _enforce_role_alternation
-                # will insert a synthetic one as a safety net.
-            start = find_legal_message_start(kept)
-            if start:
-                kept = kept[start:]
-        if not kept:
-            kept = non_system[-min(len(non_system), 4) :]
-            start = find_legal_message_start(kept)
-            if start:
-                kept = kept[start:]
-        return system_messages + kept
-
     def _partition_tool_batches(
         self,
         spec: AgentRunSpec,