fix(agent): prevent infinite compression loop when aux model context is too small (#53008)

agent/conversation_compression.py

@@ -270,12 +270,15 @@ def check_compression_model_feasibility(agent: Any) -> None:
             # above guarantees aux_context >= MINIMUM_CONTEXT_LENGTH,
             # so the new threshold is always >= 64K.
             #
-            # The compression summariser sends a single user-role
-            # prompt (no system prompt, no tools) to the aux model, so
-            # new_threshold == aux_context is safe: the request is
-            # the raw messages plus a small summarisation instruction.
+            # Apply an 80 % safety margin on the aux model's context
+            # window: the summarisation prompt template, system
+            # instructions, and tool schemas consume part of the
+            # context, so reserving 20 % headroom avoids passing a
+            # request that exceeds the aux model's actual capacity
+            # (see issue #53008 — a too-high threshold combined with
+            # no headroom can produce an infinite compression loop).
             old_threshold = threshold
-            new_threshold = aux_context
+            new_threshold = max(int(aux_context * 0.8), MINIMUM_CONTEXT_LENGTH)
             agent.context_compressor.threshold_tokens = new_threshold
             # Keep threshold_percent in sync so future main-model
             # context_length changes (update_model) re-derive from a
@@ -285,7 +288,7 @@ def check_compression_model_feasibility(agent: Any) -> None:
                 agent.context_compressor.threshold_percent = (
                     new_threshold / main_ctx
                 )
-            safe_pct = int((aux_context / main_ctx) * 100) if main_ctx else 50
+            safe_pct = int((new_threshold / main_ctx) * 100) if main_ctx else 50
             # Build human-readable "model (provider)" labels for both
             # the main model and the compression model so users can
             # tell at a glance which provider each side is actually

agent/turn_context.py

@@ -397,6 +397,7 @@ def build_turn_context(
                 f">= {_compressor.threshold_tokens:,} threshold. "
                 "This may take a moment."
             )
+            _preflight_original_tokens = _preflight_tokens
             for _pass in range(3):
                 _orig_len = len(messages)
                 _orig_tokens = _preflight_tokens
@@ -428,7 +429,28 @@ def build_turn_context(
                 if not _compressor.should_compress(_preflight_tokens):
                     break
 
-    # Plugin hook: pre_llm_call (context injected into user message, not system prompt).
+            # Guard against an infinite compression loop when the aux
+            # model is too small to compress the current session
+            # effectively (issue #53008).  Token reduction below 10 %
+            # means every turn will re-trigger compression without
+            # making progress — raise the threshold to break the cycle.
+            _reduction = 1.0 - (_preflight_tokens / _preflight_original_tokens)
+            if _reduction < 0.10 and _compressor.should_compress(_preflight_tokens):
+                _old_threshold = _compressor.threshold_tokens
+                _compressor.threshold_tokens = max(
+                    _preflight_tokens + 1,
+                    int(_old_threshold * 1.5),
+                )
+                logger.warning(
+                    "Preflight compression ineffective "
+                    "(%.1f%% token reduction — aux model may be too "
+                    "small for current session). "
+                    "Raising threshold %s → %s to prevent "
+                    "compression loop.",
+                    _reduction * 100,
+                    f"{_old_threshold:,}",
+                    f"{_compressor.threshold_tokens:,}",
+                )
     plugin_user_context = ""
     try:
         from hermes_cli.plugins import invoke_hook as _invoke_hook