Merge remote-tracking branch 'origin/main' into nanobot-webui

nanobot/agent/loop.py

@@ -648,7 +648,10 @@ class AgentLoop:
 
             session, pending = self.auto_compact.prepare_session(session, key)
 
-            await self.consolidator.maybe_consolidate_by_tokens(session)
+            await self.consolidator.maybe_consolidate_by_tokens(
+                session,
+                session_summary=pending,
+            )
             # Persist subagent follow-ups into durable history BEFORE prompt
             # assembly. ContextBuilder merges adjacent same-role messages for
             # provider compatibility, which previously caused the follow-up to
@@ -709,7 +712,10 @@ class AgentLoop:
         if result := await self.commands.dispatch(ctx):
             return result
 
-        await self.consolidator.maybe_consolidate_by_tokens(session)
+        await self.consolidator.maybe_consolidate_by_tokens(
+            session,
+            session_summary=pending,
+        )
 
         self._set_tool_context(msg.channel, msg.chat_id, msg.metadata.get("message_id"))
         if message_tool := self.tools.get("message"):

nanobot/agent/memory.py

@@ -416,7 +416,12 @@ class Consolidator:
                 return idx
         return None
 
-    def estimate_session_prompt_tokens(self, session: Session) -> tuple[int, str]:
+    def estimate_session_prompt_tokens(
+        self,
+        session: Session,
+        *,
+        session_summary: str | None = None,
+    ) -> tuple[int, str]:
         """Estimate current prompt size for the normal session history view."""
         history = session.get_history(max_messages=0)
         channel, chat_id = (session.key.split(":", 1) if ":" in session.key else (None, None))
@@ -425,6 +430,7 @@ class Consolidator:
             current_message="[token-probe]",
             channel=channel,
             chat_id=chat_id,
+            session_summary=session_summary,
         )
         return estimate_prompt_tokens_chain(
             self.provider,
@@ -467,7 +473,12 @@ class Consolidator:
             self.store.raw_archive(messages)
             return None
 
-    async def maybe_consolidate_by_tokens(self, session: Session) -> None:
+    async def maybe_consolidate_by_tokens(
+        self,
+        session: Session,
+        *,
+        session_summary: str | None = None,
+    ) -> None:
         """Loop: archive old messages until prompt fits within safe budget.
 
         The budget reserves space for completion tokens and a safety buffer
@@ -481,7 +492,10 @@ class Consolidator:
             budget = self.context_window_tokens - self.max_completion_tokens - self._SAFETY_BUFFER
             target = budget // 2
             try:
-                estimated, source = self.estimate_session_prompt_tokens(session)
+                estimated, source = self.estimate_session_prompt_tokens(
+                    session,
+                    session_summary=session_summary,
+                )
             except Exception:
                 logger.exception("Token estimation failed for {}", session.key)
                 estimated, source = 0, "error"
@@ -499,9 +513,10 @@ class Consolidator:
                 )
                 return
 
+            last_summary = None
             for round_num in range(self._MAX_CONSOLIDATION_ROUNDS):
                 if estimated <= target:
-                    return
+                    break
 
                 boundary = self.pick_consolidation_boundary(session, max(1, estimated - target))
                 if boundary is None:
@@ -510,7 +525,7 @@ class Consolidator:
                         session.key,
                         round_num,
                     )
-                    return
+                    break
 
                 end_idx = boundary[0]
                 end_idx = self._cap_consolidation_boundary(session, end_idx)
@@ -520,11 +535,11 @@ class Consolidator:
                         session.key,
                         round_num,
                     )
-                    return
+                    break
 
                 chunk = session.messages[session.last_consolidated:end_idx]
                 if not chunk:
-                    return
+                    break
 
                 logger.info(
                     "Token consolidation round {} for {}: {}/{} via {}, chunk={} msgs",
@@ -535,18 +550,34 @@ class Consolidator:
                     source,
                     len(chunk),
                 )
-                if not await self.archive(chunk):
-                    return
+                summary = await self.archive(chunk)
+                if summary:
+                    last_summary = summary
+                else:
+                    break
                 session.last_consolidated = end_idx
                 self.sessions.save(session)
 
                 try:
-                    estimated, source = self.estimate_session_prompt_tokens(session)
+                    estimated, source = self.estimate_session_prompt_tokens(
+                        session,
+                        session_summary=session_summary,
+                    )
                 except Exception:
                     logger.exception("Token estimation failed for {}", session.key)
                     estimated, source = 0, "error"
                 if estimated <= 0:
-                    return
+                    break
+
+            # Persist the last summary to session metadata so it can be injected
+            # into the runtime context on the next prepare_session() call, aligning
+            # the summary injection strategy with AutoCompact._archive().
+            if last_summary and last_summary != "(nothing)":
+                session.metadata["_last_summary"] = {
+                    "text": last_summary,
+                    "last_active": session.updated_at.isoformat(),
+                }
+                self.sessions.save(session)
 
 
 # ---------------------------------------------------------------------------

nanobot/providers/openai_compat_provider.py

@@ -9,11 +9,13 @@ import importlib.util
 import os
 import secrets
 import string
+import time
 import uuid
 from collections.abc import Awaitable, Callable
 from typing import TYPE_CHECKING, Any
 
 import json_repair
+from loguru import logger
 
 if os.environ.get("LANGFUSE_SECRET_KEY") and importlib.util.find_spec("langfuse"):
     from langfuse.openai import AsyncOpenAI
@@ -143,6 +145,10 @@ def _uses_openrouter_attribution(spec: "ProviderSpec | None", api_base: str | No
     return bool(api_base and "openrouter" in api_base.lower())
 
 
+_RESPONSES_FAILURE_THRESHOLD = 3
+_RESPONSES_PROBE_INTERVAL_S = 300  # 5 minutes
+
+
 def _is_direct_openai_base(api_base: str | None) -> bool:
     """Return True for direct OpenAI endpoints, not generic OpenAI-compatible gateways."""
     if not api_base:
@@ -151,6 +157,16 @@ def _is_direct_openai_base(api_base: str | None) -> bool:
     return "api.openai.com" in normalized and "openrouter" not in normalized
 
 
+def _responses_circuit_key(
+    model: str | None,
+    default_model: str,
+    reasoning_effort: str | None,
+) -> str:
+    model_name = (model or default_model).lower()
+    effort = reasoning_effort.lower() if isinstance(reasoning_effort, str) else ""
+    return f"{model_name}:{effort}"
+
+
 class OpenAICompatProvider(LLMProvider):
     """Unified provider for all OpenAI-compatible APIs.
 
@@ -189,6 +205,11 @@ class OpenAICompatProvider(LLMProvider):
             max_retries=0,
         )
 
+        # Responses API circuit breaker: skip after repeated failures,
+        # probe again after _RESPONSES_PROBE_INTERVAL_S seconds.
+        self._responses_failures: dict[str, int] = {}
+        self._responses_tripped_at: dict[str, float] = {}
+
     def _setup_env(self, api_key: str, api_base: str | None) -> None:
         """Set environment variables based on provider spec."""
         spec = self._spec
@@ -414,9 +435,39 @@ class OpenAICompatProvider(LLMProvider):
             return False
 
         model_name = (model or self.default_model).lower()
+        wants = False
         if reasoning_effort and reasoning_effort.lower() != "none":
-            return True
-        return any(token in model_name for token in ("gpt-5", "o1", "o3", "o4"))
+            wants = True
+        elif any(token in model_name for token in ("gpt-5", "o1", "o3", "o4")):
+            wants = True
+        if not wants:
+            return False
+
+        # Circuit breaker: skip after repeated failures, probe periodically.
+        key = _responses_circuit_key(model, self.default_model, reasoning_effort)
+        failures = self._responses_failures.get(key, 0)
+        if failures >= _RESPONSES_FAILURE_THRESHOLD:
+            tripped = self._responses_tripped_at.get(key, 0.0)
+            if (time.monotonic() - tripped) < _RESPONSES_PROBE_INTERVAL_S:
+                return False
+            # Half-open: allow one probe attempt
+        return True
+
+    def _record_responses_failure(self, model: str | None, reasoning_effort: str | None) -> None:
+        key = _responses_circuit_key(model, self.default_model, reasoning_effort)
+        count = self._responses_failures.get(key, 0) + 1
+        self._responses_failures[key] = count
+        if count >= _RESPONSES_FAILURE_THRESHOLD:
+            self._responses_tripped_at[key] = time.monotonic()
+            logger.warning(
+                "Responses API circuit open for {} — falling back to Chat Completions",
+                key,
+            )
+
+    def _record_responses_success(self, model: str | None, reasoning_effort: str | None) -> None:
+        key = _responses_circuit_key(model, self.default_model, reasoning_effort)
+        self._responses_failures.pop(key, None)
+        self._responses_tripped_at.pop(key, None)
 
     @staticmethod
     def _should_fallback_from_responses_error(e: Exception) -> bool:
@@ -915,10 +966,13 @@ class OpenAICompatProvider(LLMProvider):
                         messages, tools, model, max_tokens, temperature,
                         reasoning_effort, tool_choice,
                     )
-                    return parse_response_output(await self._client.responses.create(**body))
+                    result = parse_response_output(await self._client.responses.create(**body))
+                    self._record_responses_success(model, reasoning_effort)
+                    return result
                 except Exception as responses_error:
                     if not self._should_fallback_from_responses_error(responses_error):
                         raise
+                    self._record_responses_failure(model, reasoning_effort)
 
             kwargs = self._build_kwargs(
                 messages, tools, model, max_tokens, temperature,
@@ -965,6 +1019,7 @@ class OpenAICompatProvider(LLMProvider):
                         _timed_stream(),
                         on_content_delta,
                     )
+                    self._record_responses_success(model, reasoning_effort)
                     return LLMResponse(
                         content=content or None,
                         tool_calls=tool_calls,
@@ -975,6 +1030,7 @@ class OpenAICompatProvider(LLMProvider):
                 except Exception as responses_error:
                     if not self._should_fallback_from_responses_error(responses_error):
                         raise
+                    self._record_responses_failure(model, reasoning_effort)
 
             kwargs = self._build_kwargs(
                 messages, tools, model, max_tokens, temperature,