Merge branch 'main' of github.com:NousResearch/hermes-agent into feat/ink-refactor

run_agent.py

@@ -94,7 +94,7 @@ from agent.model_metadata import (
 from agent.context_compressor import ContextCompressor
 from agent.subdirectory_hints import SubdirectoryHintTracker
 from agent.prompt_caching import apply_anthropic_cache_control
-from agent.prompt_builder import build_skills_system_prompt, build_context_files_prompt, load_soul_md, TOOL_USE_ENFORCEMENT_GUIDANCE, TOOL_USE_ENFORCEMENT_MODELS, DEVELOPER_ROLE_MODELS, GOOGLE_MODEL_OPERATIONAL_GUIDANCE, OPENAI_MODEL_EXECUTION_GUIDANCE
+from agent.prompt_builder import build_skills_system_prompt, build_context_files_prompt, build_environment_hints, load_soul_md, TOOL_USE_ENFORCEMENT_GUIDANCE, TOOL_USE_ENFORCEMENT_MODELS, DEVELOPER_ROLE_MODELS, GOOGLE_MODEL_OPERATIONAL_GUIDANCE, OPENAI_MODEL_EXECUTION_GUIDANCE
 from agent.usage_pricing import estimate_usage_cost, normalize_usage
 from agent.display import (
     KawaiiSpinner, build_tool_preview as _build_tool_preview,
@@ -339,10 +339,7 @@ def _paths_overlap(left: Path, right: Path) -> bool:
 
 _SURROGATE_RE = re.compile(r'[\ud800-\udfff]')
 
-_BUDGET_WARNING_RE = re.compile(
-    r"\[BUDGET(?:\s+WARNING)?:\s+Iteration\s+\d+/\d+\..*?\]",
-    re.DOTALL,
-)
+
 
 
 def _sanitize_surrogates(text: str) -> str:
@@ -463,34 +460,7 @@ def _sanitize_messages_non_ascii(messages: list) -> bool:
     return found
 
 
-def _strip_budget_warnings_from_history(messages: list) -> None:
-    """Remove budget pressure warnings from tool-result messages in-place.
 
-    Budget warnings are turn-scoped signals that must not leak into replayed
-    history.  They live in tool-result ``content`` either as a JSON key
-    (``_budget_warning``) or appended plain text.
-    """
-    for msg in messages:
-        if not isinstance(msg, dict) or msg.get("role") != "tool":
-            continue
-        content = msg.get("content")
-        if not isinstance(content, str) or "_budget_warning" not in content and "[BUDGET" not in content:
-            continue
-
-        # Try JSON first (the common case: _budget_warning key in a dict)
-        try:
-            parsed = json.loads(content)
-            if isinstance(parsed, dict) and "_budget_warning" in parsed:
-                del parsed["_budget_warning"]
-                msg["content"] = json.dumps(parsed, ensure_ascii=False)
-                continue
-        except (json.JSONDecodeError, TypeError):
-            pass
-
-        # Fallback: strip the text pattern from plain-text tool results
-        cleaned = _BUDGET_WARNING_RE.sub("", content).strip()
-        if cleaned != content:
-            msg["content"] = cleaned
 
 
 # =========================================================================
@@ -579,6 +549,7 @@ class AIAgent:
         clarify_callback: callable = None,
         step_callback: callable = None,
         stream_delta_callback: callable = None,
+        interim_assistant_callback: callable = None,
         tool_gen_callback: callable = None,
         status_callback: callable = None,
         max_tokens: int = None,
@@ -728,6 +699,7 @@ class AIAgent:
         self.clarify_callback = clarify_callback
         self.step_callback = step_callback
         self.stream_delta_callback = stream_delta_callback
+        self.interim_assistant_callback = interim_assistant_callback
         self.status_callback = status_callback
         self.tool_gen_callback = tool_gen_callback
 
@@ -775,12 +747,14 @@ class AIAgent:
         self._use_prompt_caching = (is_openrouter and is_claude) or is_native_anthropic
         self._cache_ttl = "5m"  # Default 5-minute TTL (1.25x write cost)
         
-        # Iteration budget pressure: warn the LLM as it approaches max_iterations.
-        # Warnings are injected into the last tool result JSON (not as separate
-        # messages) so they don't break message structure or invalidate caching.
-        self._budget_caution_threshold = 0.7   # 70% — nudge to start wrapping up
-        self._budget_warning_threshold = 0.9   # 90% — urgent, respond now
-        self._budget_pressure_enabled = True
+        # Iteration budget: the LLM is only notified when it actually exhausts
+        # the iteration budget (api_call_count >= max_iterations).  At that
+        # point we inject ONE message, allow one final API call, and if the
+        # model doesn't produce a text response, force a user-message asking
+        # it to summarise.  No intermediate pressure warnings — they caused
+        # models to "give up" prematurely on complex tasks (#7915).
+        self._budget_exhausted_injected = False
+        self._budget_grace_call = False
 
         # Context pressure warnings: notify the USER (not the LLM) as context
         # fills up.  Purely informational — displayed in CLI output and sent via
@@ -831,6 +805,11 @@ class AIAgent:
         # Deferred paragraph break flag — set after tool iterations so a
         # single "\n\n" is prepended to the next real text delta.
         self._stream_needs_break = False
+        # Visible assistant text already delivered through live token callbacks
+        # during the current model response. Used to avoid re-sending the same
+        # commentary when the provider later returns it as a completed interim
+        # assistant message.
+        self._current_streamed_assistant_text = ""
 
         # Optional current-turn user-message override used when the API-facing
         # user message intentionally differs from the persisted transcript
@@ -1328,9 +1307,23 @@ class AIAgent:
                 api_key=getattr(self, "api_key", ""),
                 config_context_length=_config_context_length,
                 provider=self.provider,
+                api_mode=self.api_mode,
             )
         self.compression_enabled = compression_enabled
 
+        # Reject models whose context window is below the minimum required
+        # for reliable tool-calling workflows (64K tokens).
+        from agent.model_metadata import MINIMUM_CONTEXT_LENGTH
+        _ctx = getattr(self.context_compressor, "context_length", 0)
+        if _ctx and _ctx < MINIMUM_CONTEXT_LENGTH:
+            raise ValueError(
+                f"Model {self.model} has a context window of {_ctx:,} tokens, "
+                f"which is below the minimum {MINIMUM_CONTEXT_LENGTH:,} required "
+                f"by Hermes Agent.  Choose a model with at least "
+                f"{MINIMUM_CONTEXT_LENGTH // 1000}K context, or set "
+                f"model.context_length in config.yaml to override."
+            )
+
         # Inject context engine tool schemas (e.g. lcm_grep, lcm_describe, lcm_expand)
         self._context_engine_tool_names: set = set()
         if hasattr(self, "context_compressor") and self.context_compressor and self.tools is not None:
@@ -1571,6 +1564,7 @@ class AIAgent:
                 base_url=self.base_url,
                 api_key=getattr(self, "api_key", ""),
                 provider=self.provider,
+                api_mode=self.api_mode,
             )
 
         # ── Invalidate cached system prompt so it rebuilds next turn ──
@@ -1704,6 +1698,16 @@ class AIAgent:
             except Exception:
                 logger.debug("status_callback error in _emit_status", exc_info=True)
 
+    def _current_main_runtime(self) -> Dict[str, str]:
+        """Return the live main runtime for session-scoped auxiliary routing."""
+        return {
+            "model": getattr(self, "model", "") or "",
+            "provider": getattr(self, "provider", "") or "",
+            "base_url": getattr(self, "base_url", "") or "",
+            "api_key": getattr(self, "api_key", "") or "",
+            "api_mode": getattr(self, "api_mode", "") or "",
+        }
+
     def _check_compression_model_feasibility(self) -> None:
         """Warn at session start if the auxiliary compression model's context
         window is smaller than the main model's compression threshold.
@@ -1724,7 +1728,10 @@ class AIAgent:
             from agent.auxiliary_client import get_text_auxiliary_client
             from agent.model_metadata import get_model_context_length
 
-            client, aux_model = get_text_auxiliary_client("compression")
+            client, aux_model = get_text_auxiliary_client(
+                "compression",
+                main_runtime=self._current_main_runtime(),
+            )
             if client is None or not aux_model:
                 msg = (
                     "⚠ No auxiliary LLM provider configured — context "
@@ -3186,11 +3193,17 @@ class AIAgent:
                 f"not on any model name returned by the API."
             )
 
+        # Environment hints (WSL, Termux, etc.) — tell the agent about the
+        # execution environment so it can translate paths and adapt behavior.
+        _env_hints = build_environment_hints()
+        if _env_hints:
+            prompt_parts.append(_env_hints)
+
         platform_key = (self.platform or "").lower().strip()
         if platform_key in PLATFORM_HINTS:
             prompt_parts.append(PLATFORM_HINTS[platform_key])
 
-        return "\n\n".join(prompt_parts)
+        return "\n\n".join(p.strip() for p in prompt_parts if p.strip())
 
     # =========================================================================
     # Pre/post-call guardrails (inspired by PR #1321 — @alireza78a)
@@ -4730,6 +4743,49 @@ class AIAgent:
 
     # ── Unified streaming API call ─────────────────────────────────────────
 
+    def _reset_stream_delivery_tracking(self) -> None:
+        """Reset tracking for text delivered during the current model response."""
+        self._current_streamed_assistant_text = ""
+
+    def _record_streamed_assistant_text(self, text: str) -> None:
+        """Accumulate visible assistant text emitted through stream callbacks."""
+        if isinstance(text, str) and text:
+            self._current_streamed_assistant_text = (
+                getattr(self, "_current_streamed_assistant_text", "") + text
+            )
+
+    @staticmethod
+    def _normalize_interim_visible_text(text: str) -> str:
+        if not isinstance(text, str):
+            return ""
+        return re.sub(r"\s+", " ", text).strip()
+
+    def _interim_content_was_streamed(self, content: str) -> bool:
+        visible_content = self._normalize_interim_visible_text(
+            self._strip_think_blocks(content or "")
+        )
+        if not visible_content:
+            return False
+        streamed = self._normalize_interim_visible_text(
+            self._strip_think_blocks(getattr(self, "_current_streamed_assistant_text", "") or "")
+        )
+        return bool(streamed) and streamed == visible_content
+
+    def _emit_interim_assistant_message(self, assistant_msg: Dict[str, Any]) -> None:
+        """Surface a real mid-turn assistant commentary message to the UI layer."""
+        cb = getattr(self, "interim_assistant_callback", None)
+        if cb is None or not isinstance(assistant_msg, dict):
+            return
+        content = assistant_msg.get("content")
+        visible = self._strip_think_blocks(content or "").strip()
+        if not visible or visible == "(empty)":
+            return
+        already_streamed = self._interim_content_was_streamed(visible)
+        try:
+            cb(visible, already_streamed=already_streamed)
+        except Exception:
+            logger.debug("interim_assistant_callback error", exc_info=True)
+
     def _fire_stream_delta(self, text: str) -> None:
         """Fire all registered stream delta callbacks (display + TTS)."""
         # If a tool iteration set the break flag, prepend a single paragraph
@@ -4739,12 +4795,16 @@ class AIAgent:
         if getattr(self, "_stream_needs_break", False) and text and text.strip():
             self._stream_needs_break = False
             text = "\n\n" + text
-        for cb in (self.stream_delta_callback, self._stream_callback):
-            if cb is not None:
-                try:
-                    cb(text)
-                except Exception:
-                    pass
+        callbacks = [cb for cb in (self.stream_delta_callback, self._stream_callback) if cb is not None]
+        delivered = False
+        for cb in callbacks:
+            try:
+                cb(text)
+                delivered = True
+            except Exception:
+                pass
+        if delivered:
+            self._record_streamed_assistant_text(text)
 
     def _fire_reasoning_delta(self, text: str) -> None:
         """Fire reasoning callback if registered."""
@@ -4928,6 +4988,7 @@ class AIAgent:
                         if self.stream_delta_callback:
                             try:
                                 self.stream_delta_callback(delta.content)
+                                self._record_streamed_assistant_text(delta.content)
                             except Exception:
                                 pass
 
@@ -6509,17 +6570,23 @@ class AIAgent:
             if messages and messages[-1].get("_flush_sentinel") == _sentinel:
                 messages.pop()
 
-    def _compress_context(self, messages: list, system_message: str, *, approx_tokens: int = None, task_id: str = "default") -> tuple:
+    def _compress_context(self, messages: list, system_message: str, *, approx_tokens: int = None, task_id: str = "default", focus_topic: str = None) -> tuple:
         """Compress conversation context and split the session in SQLite.
 
+        Args:
+            focus_topic: Optional focus string for guided compression — the
+                summariser will prioritise preserving information related to
+                this topic.  Inspired by Claude Code's ``/compact <focus>``.
+
         Returns:
             (compressed_messages, new_system_prompt) tuple
         """
         _pre_msg_count = len(messages)
         logger.info(
-            "context compression started: session=%s messages=%d tokens=~%s model=%s",
+            "context compression started: session=%s messages=%d tokens=~%s model=%s focus=%r",
             self.session_id or "none", _pre_msg_count,
             f"{approx_tokens:,}" if approx_tokens else "unknown", self.model,
+            focus_topic,
         )
         # Pre-compression memory flush: let the model save memories before they're lost
         self.flush_memories(messages, min_turns=0)
@@ -6531,7 +6598,7 @@ class AIAgent:
             except Exception:
                 pass
 
-        compressed = self.context_compressor.compress(messages, current_tokens=approx_tokens)
+        compressed = self.context_compressor.compress(messages, current_tokens=approx_tokens, focus_topic=focus_topic)
 
         todo_snapshot = self._todo_store.format_for_injection()
         if todo_snapshot:
@@ -6920,24 +6987,6 @@ class AIAgent:
             turn_tool_msgs = messages[-num_tools:]
             enforce_turn_budget(turn_tool_msgs, env=get_active_env(effective_task_id))
 
-        # ── Budget pressure injection ────────────────────────────────────
-        budget_warning = self._get_budget_warning(api_call_count)
-        if budget_warning and messages and messages[-1].get("role") == "tool":
-            last_content = messages[-1]["content"]
-            try:
-                parsed = json.loads(last_content)
-                if isinstance(parsed, dict):
-                    parsed["_budget_warning"] = budget_warning
-                    messages[-1]["content"] = json.dumps(parsed, ensure_ascii=False)
-                else:
-                    messages[-1]["content"] = last_content + f"\n\n{budget_warning}"
-            except (json.JSONDecodeError, TypeError):
-                messages[-1]["content"] = last_content + f"\n\n{budget_warning}"
-            if not self.quiet_mode:
-                remaining = self.max_iterations - api_call_count
-                tier = "⚠️  WARNING" if remaining <= self.max_iterations * 0.1 else "💡 CAUTION"
-                print(f"{self.log_prefix}{tier}: {remaining} iterations remaining")
-
     def _execute_tool_calls_sequential(self, assistant_message, messages: list, effective_task_id: str, api_call_count: int = 0) -> None:
         """Execute tool calls sequentially (original behavior). Used for single calls or interactive tools."""
         for i, tool_call in enumerate(assistant_message.tool_calls, 1):
@@ -6986,6 +7035,15 @@ class AIAgent:
             self._current_tool = function_name
             self._touch_activity(f"executing tool: {function_name}")
 
+            # Set activity callback for long-running tool execution (terminal
+            # commands, etc.) so the gateway's inactivity monitor doesn't kill
+            # the agent while a command is running.
+            try:
+                from tools.environments.base import set_activity_callback
+                set_activity_callback(self._touch_activity)
+            except Exception:
+                pass
+
             if self.tool_progress_callback:
                 try:
                     preview = _build_tool_preview(function_name, function_args)
@@ -7275,50 +7333,7 @@ class AIAgent:
         if num_tools_seq > 0:
             enforce_turn_budget(messages[-num_tools_seq:], env=get_active_env(effective_task_id))
 
-        # ── Budget pressure injection ─────────────────────────────────
-        # After all tool calls in this turn are processed, check if we're
-        # approaching max_iterations. If so, inject a warning into the LAST
-        # tool result's JSON so the LLM sees it naturally when reading results.
-        budget_warning = self._get_budget_warning(api_call_count)
-        if budget_warning and messages and messages[-1].get("role") == "tool":
-            last_content = messages[-1]["content"]
-            try:
-                parsed = json.loads(last_content)
-                if isinstance(parsed, dict):
-                    parsed["_budget_warning"] = budget_warning
-                    messages[-1]["content"] = json.dumps(parsed, ensure_ascii=False)
-                else:
-                    messages[-1]["content"] = last_content + f"\n\n{budget_warning}"
-            except (json.JSONDecodeError, TypeError):
-                messages[-1]["content"] = last_content + f"\n\n{budget_warning}"
-            if not self.quiet_mode:
-                remaining = self.max_iterations - api_call_count
-                tier = "⚠️  WARNING" if remaining <= self.max_iterations * 0.1 else "💡 CAUTION"
-                print(f"{self.log_prefix}{tier}: {remaining} iterations remaining")
 
-    def _get_budget_warning(self, api_call_count: int) -> Optional[str]:
-        """Return a budget pressure string, or None if not yet needed.
-
-        Two-tier system:
-          - Caution (70%): nudge to consolidate work
-          - Warning (90%): urgent, must respond now
-        """
-        if not self._budget_pressure_enabled or self.max_iterations <= 0:
-            return None
-        progress = api_call_count / self.max_iterations
-        remaining = self.max_iterations - api_call_count
-        if progress >= self._budget_warning_threshold:
-            return (
-                f"[BUDGET WARNING: Iteration {api_call_count}/{self.max_iterations}. "
-                f"Only {remaining} iteration(s) left. "
-                "Provide your final response NOW. No more tool calls unless absolutely critical.]"
-            )
-        if progress >= self._budget_caution_threshold:
-            return (
-                f"[BUDGET: Iteration {api_call_count}/{self.max_iterations}. "
-                f"{remaining} iterations left. Start consolidating your work.]"
-            )
-        return None
 
     def _emit_context_pressure(self, compaction_progress: float, compressor) -> None:
         """Notify the user that context is approaching the compaction threshold.
@@ -7542,6 +7557,11 @@ class AIAgent:
         # Installed once, transparent when streams are healthy, prevents crash on write.
         _install_safe_stdio()
 
+        # Tag all log records on this thread with the session ID so
+        # ``hermes logs --session <id>`` can filter a single conversation.
+        from hermes_logging import set_session_context
+        set_session_context(self.session_id)
+
         # If the previous turn activated fallback, restore the primary
         # runtime so this turn gets a fresh attempt with the preferred model.
         # No-op when _fallback_activated is False (gateway, first turn, etc.).
@@ -7611,14 +7631,6 @@ class AIAgent:
         # Initialize conversation (copy to avoid mutating the caller's list)
         messages = list(conversation_history) if conversation_history else []
 
-        # Strip budget pressure warnings from previous turns.  These are
-        # turn-scoped signals injected by _get_budget_warning() into tool
-        # result content.  If left in the replayed history, models (especially
-        # GPT-family) interpret them as still-active instructions and avoid
-        # making tool calls in ALL subsequent turns.
-        if messages:
-            _strip_budget_warnings_from_history(messages)
-        
         # Hydrate todo store from conversation history (gateway creates a fresh
         # AIAgent per message, so the in-memory store is empty -- we need to
         # recover the todo state from the most recent todo tool response in history)
@@ -7835,7 +7847,7 @@ class AIAgent:
             except Exception:
                 pass
 
-        while api_call_count < self.max_iterations and self.iteration_budget.remaining > 0:
+        while (api_call_count < self.max_iterations and self.iteration_budget.remaining > 0) or self._budget_grace_call:
             # Reset per-turn checkpoint dedup so each iteration can take one snapshot
             self._checkpoint_mgr.new_turn()
 
@@ -7850,7 +7862,13 @@ class AIAgent:
             api_call_count += 1
             self._api_call_count = api_call_count
             self._touch_activity(f"starting API call #{api_call_count}")
-            if not self.iteration_budget.consume():
+
+            # Grace call: the budget is exhausted but we gave the model one
+            # more chance.  Consume the grace flag so the loop exits after
+            # this iteration regardless of outcome.
+            if self._budget_grace_call:
+                self._budget_grace_call = False
+            elif not self.iteration_budget.consume():
                 _turn_exit_reason = "budget_exhausted"
                 if not self.quiet_mode:
                     self._safe_print(f"\n⚠️  Iteration budget exhausted ({self.iteration_budget.used}/{self.iteration_budget.max_total} iterations used)")
@@ -7977,9 +7995,39 @@ class AIAgent:
             # manual message manipulation are always caught.
             api_messages = self._sanitize_api_messages(api_messages)
 
+            # Normalize message whitespace and tool-call JSON for consistent
+            # prefix matching.  Ensures bit-perfect prefixes across turns,
+            # which enables KV cache reuse on local inference servers
+            # (llama.cpp, vLLM, Ollama) and improves cache hit rates for
+            # cloud providers.  Operates on api_messages (the API copy) so
+            # the original conversation history in `messages` is untouched.
+            for am in api_messages:
+                if isinstance(am.get("content"), str):
+                    am["content"] = am["content"].strip()
+            for am in api_messages:
+                tcs = am.get("tool_calls")
+                if not tcs:
+                    continue
+                new_tcs = []
+                for tc in tcs:
+                    if isinstance(tc, dict) and "function" in tc:
+                        try:
+                            args_obj = json.loads(tc["function"]["arguments"])
+                            tc = {**tc, "function": {
+                                **tc["function"],
+                                "arguments": json.dumps(
+                                    args_obj, separators=(",", ":"),
+                                    sort_keys=True,
+                                ),
+                            }}
+                        except Exception:
+                            pass
+                    new_tcs.append(tc)
+                am["tool_calls"] = new_tcs
+
             # Calculate approximate request size for logging
             total_chars = sum(len(str(msg)) for msg in api_messages)
-            approx_tokens = total_chars // 4  # Rough estimate: 4 chars per token
+            approx_tokens = estimate_messages_tokens_rough(api_messages)
             
             # Thinking spinner for quiet mode (animated during API call)
             thinking_spinner = None
@@ -8028,6 +8076,7 @@ class AIAgent:
 
             while retry_count < max_retries:
                 try:
+                    self._reset_stream_delivery_tracking()
                     api_kwargs = self._build_api_kwargs(api_messages)
                     if self.api_mode == "codex_responses":
                         api_kwargs = self._preflight_codex_api_kwargs(api_kwargs, allow_stream=False)
@@ -8186,6 +8235,8 @@ class AIAgent:
                             self._emit_status("⚠️ Empty/malformed response — switching to fallback...")
                         if self._try_activate_fallback():
                             retry_count = 0
+                            compression_attempts = 0
+                            primary_recovery_attempted = False
                             continue
 
                         # Check for error field in response (some providers include this)
@@ -8221,6 +8272,8 @@ class AIAgent:
                             self._emit_status(f"⚠️ Max retries ({max_retries}) for invalid responses — trying fallback...")
                             if self._try_activate_fallback():
                                 retry_count = 0
+                                compression_attempts = 0
+                                primary_recovery_attempted = False
                                 continue
                             self._emit_status(f"❌ Max retries ({max_retries}) exceeded for invalid responses. Giving up.")
                             logging.error(f"{self.log_prefix}Invalid API response after {max_retries} retries.")
@@ -8875,6 +8928,8 @@ class AIAgent:
                             self._emit_status("⚠️ Rate limited — switching to fallback provider...")
                             if self._try_activate_fallback():
                                 retry_count = 0
+                                compression_attempts = 0
+                                primary_recovery_attempted = False
                                 continue
 
                     is_payload_too_large = (
@@ -9088,6 +9143,8 @@ class AIAgent:
                         self._emit_status(f"⚠️ Non-retryable error (HTTP {status_code}) — trying fallback...")
                         if self._try_activate_fallback():
                             retry_count = 0
+                            compression_attempts = 0
+                            primary_recovery_attempted = False
                             continue
                         if api_kwargs is not None:
                             self._dump_api_request_debug(
@@ -9153,6 +9210,8 @@ class AIAgent:
                         self._emit_status(f"⚠️ Max retries ({max_retries}) exhausted — trying fallback...")
                         if self._try_activate_fallback():
                             retry_count = 0
+                            compression_attempts = 0
+                            primary_recovery_attempted = False
                             continue
                         _final_summary = self._summarize_api_error(api_error)
                         if is_rate_limited:
@@ -9376,8 +9435,6 @@ class AIAgent:
                 # Check for incomplete <REASONING_SCRATCHPAD> (opened but never closed)
                 # This means the model ran out of output tokens mid-reasoning — retry up to 2 times
                 if has_incomplete_scratchpad(assistant_message.content or ""):
-                    if not hasattr(self, '_incomplete_scratchpad_retries'):
-                        self._incomplete_scratchpad_retries = 0
                     self._incomplete_scratchpad_retries += 1
                     
                     self._vprint(f"{self.log_prefix}⚠️  Incomplete <REASONING_SCRATCHPAD> detected (opened but never closed)")
@@ -9405,12 +9462,9 @@ class AIAgent:
                         }
                 
                 # Reset incomplete scratchpad counter on clean response
-                if hasattr(self, '_incomplete_scratchpad_retries'):
-                    self._incomplete_scratchpad_retries = 0
+                self._incomplete_scratchpad_retries = 0
 
                 if self.api_mode == "codex_responses" and finish_reason == "incomplete":
-                    if not hasattr(self, "_codex_incomplete_retries"):
-                        self._codex_incomplete_retries = 0
                     self._codex_incomplete_retries += 1
 
                     interim_msg = self._build_assistant_message(assistant_message, finish_reason)
@@ -9437,6 +9491,7 @@ class AIAgent:
                         )
                         if not duplicate_interim:
                             messages.append(interim_msg)
+                            self._emit_interim_assistant_message(interim_msg)
 
                     if self._codex_incomplete_retries < 3:
                         if not self.quiet_mode:
@@ -9481,8 +9536,6 @@ class AIAgent:
                     ]
                     if invalid_tool_calls:
                         # Track retries for invalid tool calls
-                        if not hasattr(self, '_invalid_tool_retries'):
-                            self._invalid_tool_retries = 0
                         self._invalid_tool_retries += 1
 
                         # Return helpful error to model — model can self-correct next turn
@@ -9518,8 +9571,7 @@ class AIAgent:
                             })
                         continue
                     # Reset retry counter on successful tool call validation
-                    if hasattr(self, '_invalid_tool_retries'):
-                        self._invalid_tool_retries = 0
+                    self._invalid_tool_retries = 0
                     
                     # Validate tool call arguments are valid JSON
                     # Handle empty strings as empty objects (common model quirk)
@@ -9659,6 +9711,7 @@ class AIAgent:
                         messages.pop()
 
                     messages.append(assistant_msg)
+                    self._emit_interim_assistant_message(assistant_msg)
 
                     # Close any open streaming display (response box, reasoning
                     # box) before tool execution begins.  Intermediate turns may
@@ -9921,8 +9974,7 @@ class AIAgent:
                         break
                     
                     # Reset retry counter/signature on successful content
-                    if hasattr(self, '_empty_content_retries'):
-                        self._empty_content_retries = 0
+                    self._empty_content_retries = 0
                     self._thinking_prefill_retries = 0
 
                     if (
@@ -9938,6 +9990,7 @@ class AIAgent:
                         codex_ack_continuations += 1
                         interim_msg = self._build_assistant_message(assistant_message, "incomplete")
                         messages.append(interim_msg)
+                        self._emit_interim_assistant_message(interim_msg)
 
                         continue_msg = {
                             "role": "user",
@@ -9988,8 +10041,7 @@ class AIAgent:
                 except (OSError, ValueError):
                     logger.error(error_msg)
                 
-                if self.verbose_logging:
-                    logging.exception("Detailed error information:")
+                logger.debug("Outer loop error in API call #%d", api_call_count, exc_info=True)
                 
                 # If an assistant message with tool_calls was already appended,
                 # the API expects a role="tool" result for every tool_call_id.
@@ -10035,7 +10087,31 @@ class AIAgent:
         if final_response is None and (
             api_call_count >= self.max_iterations
             or self.iteration_budget.remaining <= 0
-        ):
+        ) and not self._budget_exhausted_injected:
+            # Budget exhausted but we haven't tried asking the model to
+            # summarise yet.  Inject a user message and give it one grace
+            # API call to produce a text response.
+            self._budget_exhausted_injected = True
+            self._budget_grace_call = True
+            _grace_msg = (
+                "Your tool budget ran out. Please give me the information "
+                "or actions you've completed so far."
+            )
+            messages.append({"role": "user", "content": _grace_msg})
+            self._emit_status(
+                f"⚠️ Iteration budget exhausted ({api_call_count}/{self.max_iterations}) "
+                "— asking model to summarise"
+            )
+            if not self.quiet_mode:
+                self._safe_print(
+                    f"\n⚠️  Iteration budget exhausted ({api_call_count}/{self.max_iterations}) "
+                    "— requesting summary..."
+                )
+
+        if final_response is None and (
+            api_call_count >= self.max_iterations
+            or self.iteration_budget.remaining <= 0
+        ) and not self._budget_grace_call:
             _turn_exit_reason = f"max_iterations_reached({api_call_count}/{self.max_iterations})"
             if self.iteration_budget.remaining <= 0 and not self.quiet_mode:
                 print(f"\n⚠️  Iteration budget exhausted ({self.iteration_budget.used}/{self.iteration_budget.max_total} iterations used)")