Grayscale/nanobot
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(session): enforce replay/file-cap invariants for history lifecycle

Browse Source
This commit is contained in:
hanyuanling
2026-04-27 00:53:32 +08:00
parent c64ec3e73c
commit 59dfd74842
8 changed files with 374 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
nanobot/agent/loop.py
+71 -2
View File
@@ -192,6 +192,9 @@ class AgentLoop:
timezone: str | None = None,
session_ttl_minutes: int = 0,
consolidation_ratio: float = 0.5,
session_history_max_messages: int | None = None,
session_history_max_tokens: int | None = None,
session_file_max_messages: int | None = None,
hooks: list[AgentHook] | None = None,
unified_session: bool = False,
disabled_skills: list[str] | None = None,
@@ -224,6 +227,21 @@ class AgentLoop:
if max_tool_result_chars is not None
else defaults.max_tool_result_chars
)
self.session_history_max_messages = (
session_history_max_messages
if session_history_max_messages is not None
else defaults.session_history_max_messages
)
self.session_history_max_tokens = (
session_history_max_tokens
if session_history_max_tokens is not None
else defaults.session_history_max_tokens
)
self.session_file_max_messages = (
session_file_max_messages
if session_file_max_messages is not None
else defaults.session_file_max_messages
)
self.provider_retry_mode = provider_retry_mode
self.web_config = web_config or WebToolsConfig()
self.exec_config = exec_config or ExecToolConfig()
@@ -452,6 +470,49 @@ class AgentLoop:
return UNIFIED_SESSION_KEY
return msg.session_key
def _history_token_budget(self) -> int:
"""Resolve token budget for session history replay."""
if self.session_history_max_tokens > 0:
return self.session_history_max_tokens
if self.context_window_tokens <= 0:
return 0
max_output = getattr(getattr(self.provider, "generation", None), "max_tokens", 4096)
try:
reserved_output = int(max_output)
except (TypeError, ValueError):
reserved_output = 4096
budget = self.context_window_tokens - max(1, reserved_output) - 1024
if budget > 0:
return budget
return max(128, self.context_window_tokens // 2)
def _enforce_session_file_cap(self, session: Session) -> None:
"""Bound session.jsonl growth by archiving and trimming old prefixes."""
limit = self.session_file_max_messages
if limit <= 0 or len(session.messages) <= limit:
return
before = list(session.messages)
before_last_consolidated = session.last_consolidated
before_count = len(before)
session.retain_recent_legal_suffix(limit)
dropped_count = before_count - len(session.messages)
if dropped_count <= 0:
return
dropped = before[:dropped_count]
already_consolidated = min(before_last_consolidated, dropped_count)
archive_chunk = dropped[already_consolidated:]
if archive_chunk:
self.context.memory.raw_archive(archive_chunk)
logger.info(
"Session file cap hit for {}: dropped {}, raw-archived {}, kept {}",
session.key,
dropped_count,
len(archive_chunk),
len(session.messages),
)
async def _run_agent_loop(
self,
initial_messages: list[dict],
@@ -832,7 +893,10 @@ class AgentLoop:
if is_subagent and self._persist_subagent_followup(session, msg):
self.sessions.save(session)
self._set_tool_context(channel, chat_id, msg.metadata.get("message_id"))
history = session.get_history(max_messages=0)
history = session.get_history(
max_messages=self.session_history_max_messages,
max_tokens=self._history_token_budget(),
)
current_role = "assistant" if is_subagent else "user"
# Subagent content is already in `history` above; passing it again
@@ -851,6 +915,7 @@ class AgentLoop:
pending_queue=pending_queue,
)
self._save_turn(session, all_msgs, 1 + len(history))
self._enforce_session_file_cap(session)
self._clear_runtime_checkpoint(session)
self.sessions.save(session)
self._schedule_background(self.consolidator.maybe_consolidate_by_tokens(session))
@@ -901,7 +966,10 @@ class AgentLoop:
if isinstance(message_tool, MessageTool):
message_tool.start_turn()
history = session.get_history(max_messages=0)
history = session.get_history(
max_messages=self.session_history_max_messages,
max_tokens=self._history_token_budget(),
)
pending_ask_id = pending_ask_user_id(history)
if pending_ask_id:
@@ -987,6 +1055,7 @@ class AgentLoop:
# Skip the already-persisted user message when saving the turn
save_skip = 1 + len(history) + (1 if user_persisted_early else 0)
self._save_turn(session, all_msgs, save_skip)
self._enforce_session_file_cap(session)
self._clear_pending_user_turn(session)
self._clear_runtime_checkpoint(session)
self.sessions.save(session)
nanobot/cli/commands.py
+9
View File
@@ -538,6 +538,9 @@ def serve(
disabled_skills=runtime_config.agents.defaults.disabled_skills,
session_ttl_minutes=runtime_config.agents.defaults.session_ttl_minutes,
consolidation_ratio=runtime_config.agents.defaults.consolidation_ratio,
session_history_max_messages=runtime_config.agents.defaults.session_history_max_messages,
session_history_max_tokens=runtime_config.agents.defaults.session_history_max_tokens,
session_file_max_messages=runtime_config.agents.defaults.session_file_max_messages,
tools_config=runtime_config.tools,
)
@@ -651,6 +654,9 @@ def _run_gateway(
disabled_skills=config.agents.defaults.disabled_skills,
session_ttl_minutes=config.agents.defaults.session_ttl_minutes,
consolidation_ratio=config.agents.defaults.consolidation_ratio,
session_history_max_messages=config.agents.defaults.session_history_max_messages,
session_history_max_tokens=config.agents.defaults.session_history_max_tokens,
session_file_max_messages=config.agents.defaults.session_file_max_messages,
tools_config=config.tools,
provider_snapshot_loader=load_provider_snapshot,
provider_signature=provider_snapshot.signature,
@@ -1028,6 +1034,9 @@ def agent(
disabled_skills=config.agents.defaults.disabled_skills,
session_ttl_minutes=config.agents.defaults.session_ttl_minutes,
consolidation_ratio=config.agents.defaults.consolidation_ratio,
session_history_max_messages=config.agents.defaults.session_history_max_messages,
session_history_max_tokens=config.agents.defaults.session_history_max_tokens,
session_file_max_messages=config.agents.defaults.session_file_max_messages,
tools_config=config.tools,
)
restart_notice = consume_restart_notice_from_env()
nanobot/config/schema.py
+12
View File
@@ -97,6 +97,18 @@ class AgentDefaults(Base):
validation_alias=AliasChoices("consolidationRatio"),
serialization_alias="consolidationRatio",
) # Consolidation target ratio (0.5 = 50% of budget retained after compression)
session_history_max_messages: int = Field(
default=120,
ge=0,
) # Per-turn session history window for prompt replay (0 = unlimited)
session_history_max_tokens: int = Field(
default=0,
ge=0,
) # Per-turn token budget for replay history (0 = auto based on context window)
session_file_max_messages: int = Field(
default=2000,
ge=0,
) # Hard cap for on-disk session.jsonl messages (0 = disabled)
dream: DreamConfig = Field(default_factory=DreamConfig)
nanobot/nanobot.py
+3
View File
@@ -85,6 +85,9 @@ class Nanobot:
disabled_skills=defaults.disabled_skills,
session_ttl_minutes=defaults.session_ttl_minutes,
consolidation_ratio=defaults.consolidation_ratio,
session_history_max_messages=defaults.session_history_max_messages,
session_history_max_tokens=defaults.session_history_max_tokens,
session_file_max_messages=defaults.session_file_max_messages,
tools_config=config.tools,
)
return cls(loop)
nanobot/session/manager.py
+67 -9
View File
@@ -12,6 +12,7 @@ from loguru import logger
from nanobot.config.paths import get_legacy_sessions_dir
from nanobot.utils.helpers import (
estimate_message_tokens,
ensure_dir,
find_legal_message_start,
image_placeholder_text,
@@ -41,8 +42,17 @@ class Session:
self.messages.append(msg)
self.updated_at = datetime.now()
def get_history(self, max_messages: int = 500) -> list[dict[str, Any]]:
"""Return unconsolidated messages for LLM input, aligned to a legal tool-call boundary."""
def get_history(
self,
max_messages: int = 500,
*,
max_tokens: int = 0,
) -> list[dict[str, Any]]:
"""Return unconsolidated messages for LLM input.
History is sliced by message count first (``max_messages``), then by
token budget from the tail (``max_tokens``) when provided.
"""
unconsolidated = self.messages[self.last_consolidated:]
sliced = unconsolidated[-max_messages:]
@@ -80,6 +90,38 @@ class Session:
if key in message:
entry[key] = message[key]
out.append(entry)
if max_tokens > 0 and out:
kept: list[dict[str, Any]] = []
used = 0
for message in reversed(out):
tokens = estimate_message_tokens(message)
if kept and used + tokens > max_tokens:
break
kept.append(message)
used += tokens
kept.reverse()
# Keep history aligned to the first visible user turn.
first_user = next((i for i, m in enumerate(kept) if m.get("role") == "user"), None)
if first_user is not None:
kept = kept[first_user:]
else:
# Tight token budgets can otherwise leave assistant-only tails.
# If a user turn exists in the unsliced output, recover the
# nearest one even if it slightly exceeds the token budget.
recovered_user = next(
(i for i in range(len(out) - 1, -1, -1) if out[i].get("role") == "user"),
None,
)
if recovered_user is not None:
kept = out[recovered_user:]
# And keep a legal tool-call boundary at the front.
start = find_legal_message_start(kept)
if start:
kept = kept[start:]
out = kept
return out
def clear(self) -> None:
@@ -89,26 +131,42 @@ class Session:
self.updated_at = datetime.now()
def retain_recent_legal_suffix(self, max_messages: int) -> None:
"""Keep a legal recent suffix, mirroring get_history boundary rules."""
"""Keep a legal recent suffix constrained by a hard message cap."""
if max_messages <= 0:
self.clear()
return
if len(self.messages) <= max_messages:
return
start_idx = max(0, len(self.messages) - max_messages)
retained = list(self.messages[-max_messages:])
# If the cutoff lands mid-turn, extend backward to the nearest user turn.
while start_idx > 0 and self.messages[start_idx].get("role") != "user":
start_idx -= 1
retained = self.messages[start_idx:]
# Prefer starting at a user turn when one exists within the tail.
first_user = next((i for i, m in enumerate(retained) if m.get("role") == "user"), None)
if first_user is not None:
retained = retained[first_user:]
else:
# If the tail is assistant/tool-only, anchor to the latest user in
# the full session and take a capped forward window from there.
latest_user = next(
(i for i in range(len(self.messages) - 1, -1, -1)
if self.messages[i].get("role") == "user"),
None,
)
if latest_user is not None:
retained = list(self.messages[latest_user: latest_user + max_messages])
# Mirror get_history(): avoid persisting orphan tool results at the front.
start = find_legal_message_start(retained)
if start:
retained = retained[start:]
# Hard-cap guarantee: never keep more than max_messages.
if len(retained) > max_messages:
retained = retained[-max_messages:]
start = find_legal_message_start(retained)
if start:
retained = retained[start:]
dropped = len(self.messages) - len(retained)
self.messages = retained
self.last_consolidated = max(0, self.last_consolidated - dropped)