fix(compactor): shrink oversized tool_calls arguments so trim_for_context can fit a tool-only turn (#2949)

src/context_compactor.py

@@ -5,6 +5,7 @@ Auto-compacts conversation history when approaching context window limits.
 Summarizes older messages via the same LLM, preserving key context.
 """
 
+import json
 import logging
 from typing import Any, Dict, List, Optional
 
@@ -146,15 +147,53 @@ def _truncate_text_to_token_budget(text: str, token_budget: int) -> str:
     return text[:head_len].rstrip() + notice + "\n\n" + text[-tail_len:].lstrip()
 
 
+def _truncate_tool_call_args(msg: Dict[str, Any], token_budget: int) -> Dict[str, Any]:
+    """Shrink oversized assistant ``tool_calls`` arguments to fit ``token_budget``.
+
+    A tool-only turn persists ``content=None`` with its whole payload in
+    ``tool_calls[].function.arguments`` (e.g. a large create_document body), which
+    the text-content truncation can't reach — so the message could stay over
+    budget and the upstream call would 400. Replace each argument string that
+    overflows its share of the budget with a small valid-JSON placeholder,
+    preserving ``id``/``type``/``function.name`` so tool/result pairing and
+    provider validation are unaffected. Returns msg unchanged when there is
+    nothing oversized.
+    """
+    tool_calls = msg.get("tool_calls")
+    if not isinstance(tool_calls, list) or not tool_calls:
+        return msg
+    # Budget left after whatever content survived (estimate_tokens counts tool
+    # arguments too, so measure content alone here).
+    content_tokens = estimate_tokens([{"role": msg.get("role", "assistant"), "content": msg.get("content")}])
+    per_call = max(16, (max(0, token_budget - content_tokens)) // len(tool_calls))
+    new_calls = []
+    changed = False
+    for tc in tool_calls:
+        fn = tc.get("function") if isinstance(tc, dict) else None
+        args = fn.get("arguments") if isinstance(fn, dict) else None
+        if isinstance(args, str) and int(len(args) * 0.3) > per_call:
+            new_fn = dict(fn)
+            new_fn["arguments"] = json.dumps({"_truncated_for_context": len(args)})
+            new_tc = dict(tc)
+            new_tc["function"] = new_fn
+            new_calls.append(new_tc)
+            changed = True
+        else:
+            new_calls.append(tc)
+    if not changed:
+        return msg
+    out = dict(msg)
+    out["tool_calls"] = new_calls
+    return out
+
+
 def _truncate_message_to_token_budget(msg: Dict[str, Any], token_budget: int) -> Dict[str, Any]:
-    """Return a copy of msg whose text content fits inside token_budget."""
+    """Return a copy of msg whose text content (and tool-call args) fit token_budget."""
     out = dict(msg)
     content = out.get("content", "")
     if isinstance(content, str):
         out["content"] = _truncate_text_to_token_budget(content, token_budget)
-        return out
-
-    if isinstance(content, list):
+    elif isinstance(content, list):
         remaining = token_budget
         new_content = []
         for item in content:
@@ -168,7 +207,9 @@ def _truncate_message_to_token_budget(msg: Dict[str, Any], token_budget: int) ->
             new_content.append(cloned)
             remaining -= _message_text_token_estimate(truncated)
         out["content"] = new_content
-    return out
+    # A tool-only turn (content=None) carries its payload in tool_calls args,
+    # which the branches above can't shrink — handle it so the message can fit.
+    return _truncate_tool_call_args(out, token_budget)
 
 
 def trim_for_context(messages: List[Dict], context_length: int, reserve_tokens: int = 512) -> List[Dict]:

tests/test_compact_truncate_tool_call_args.py

@@ -0,0 +1,62 @@
+"""Issue #2947 — _truncate_message_to_token_budget must shrink oversized tool_calls
+arguments, not just text content.
+
+A tool-only assistant turn persists content=None with its whole payload in
+tool_calls[].function.arguments. The text-content truncation can't reach it, so
+trim_for_context's last-resort message shrink left the message over budget and the
+upstream call 400'd. This pins that oversized args are bounded (so the message
+fits) while id/type/function.name are preserved, and that small args / plain text
+are untouched.
+"""
+import json
+import sys
+from unittest.mock import MagicMock
+
+import pytest
+
+for mod in [
+    'sqlalchemy', 'sqlalchemy.orm', 'sqlalchemy.ext', 'sqlalchemy.ext.declarative',
+    'sqlalchemy.ext.hybrid', 'sqlalchemy.sql', 'sqlalchemy.sql.expression',
+    'src.database',
+    'core.models', 'core.database',
+]:
+    if mod not in sys.modules:
+        sys.modules[mod] = MagicMock()
+
+from src.context_compactor import _truncate_message_to_token_budget  # noqa: E402
+from src.model_context import estimate_tokens  # noqa: E402
+
+
+def _tool_msg(arg_len):
+    return {
+        "role": "assistant",
+        "content": None,
+        "tool_calls": [{
+            "id": "c1", "type": "function",
+            "function": {"name": "create_document", "arguments": "x" * arg_len},
+        }],
+    }
+
+
+def test_oversized_tool_call_args_are_truncated_to_fit_budget():
+    budget = 200
+    out = _truncate_message_to_token_budget(_tool_msg(40000), budget)
+    # The message now fits the budget (before the fix it stayed ~12k tokens).
+    assert estimate_tokens([out]) <= budget, estimate_tokens([out])
+    tc = out["tool_calls"][0]
+    # Structure preserved so tool/result pairing + provider validation still hold.
+    assert tc["id"] == "c1" and tc["type"] == "function"
+    assert tc["function"]["name"] == "create_document"
+    # Arguments remain valid JSON, just bounded.
+    parsed = json.loads(tc["function"]["arguments"])
+    assert parsed.get("_truncated_for_context") == 40000
+
+
+def test_small_tool_call_args_are_left_untouched():
+    out = _truncate_message_to_token_budget(_tool_msg(20), 500)
+    assert out["tool_calls"][0]["function"]["arguments"] == "x" * 20
+
+
+def test_plain_text_content_still_truncates():
+    out = _truncate_message_to_token_budget({"role": "user", "content": "y" * 40000}, 200)
+    assert len(out["content"]) < 2000  # truncated, not left at 40k