"Vibe-Trading: Your Personal Trading Agent"
Python ai-agentalgorithmic-tradingbacktestingfintechllmmcpmulti-agentpythonquantitative-financetrading
fix(loop): size compaction by tool-call arguments, fold cleared-pair payloads
cgycorey224ebdddc6a7
diff --git a/agent/src/agent/loop.py b/agent/src/agent/loop.pyindex 923c108cb..d1823563a 100644--- a/agent/src/agent/loop.py+++ b/agent/src/agent/loop.py@@ -471,6 +471,72 @@ def _context_collapse(messages: list) -> None:trimmed = len(content) - COLLAPSE_HEAD - COLLAPSE_TAILmsg["content"] = f"{head}\n\n...[{trimmed} chars collapsed]...\n\n{tail}"+ # Zero-cost relief for oversized tool-call payloads whose paired result+ # was already compacted away (``[cleared]``): the arguments blob is now+ # useless to the model (the data it requested is gone), so fold it to a+ # valid JSON stub. The call id/name survive, so tool pairing and the+ # model's "I called tool X" memory are intact, and the provider still+ # receives well-formed ``arguments``. Nothing re-reads historical+ # arguments for re-dispatch, so this is safe.+ cleared_ids = {+ m.get("tool_call_id")+ for m in messages+ if m.get("role") == "tool" and m.get("content") == "[cleared]"+ }+ for msg in messages[1:-COLLAPSE_PRESERVE_RECENT]:+ for tc in msg.get("tool_calls") or []:+ fn = tc.get("function")+ if not isinstance(fn, dict):+ continue+ args = fn.get("arguments")+ if (+ isinstance(args, str)+ and len(args) > COLLAPSE_TEXT_MIN+ and tc.get("id") in cleared_ids+ ):+ fn["arguments"] = "{}"+++def _msg_estimate_chars(msg: dict) -> int:+ """Rough character size of a message for token budgeting.++ Sizes ``content`` plus every tool-call ``arguments`` payload. Assistant+ tool-call messages carry their payload in ``tool_calls[].function.+ arguments`` with empty ``content``; sizing them by content alone made the+ layer-3 tail budget count a 100 KB arguments blob as ~10 tokens.+ """+ size = len(str(msg.get("content", "")))+ for tc in msg.get("tool_calls") or []:+ fn = tc.get("function")+ if isinstance(fn, dict):+ args = fn.get("arguments")+ if isinstance(args, str):+ size += len(args)+ return size+++def _tail_cut_index(body: list, budget: int = TAIL_TOKEN_BUDGET) -> int:+ """First index of the preserved ``body`` tail that fits ``budget`` tokens.++ Walks back from the end accumulating each message's estimated tokens and+ returns the earliest index that fits, never splitting a tool_call /+ tool_result pair. Sized with ``_msg_estimate_chars`` so oversized tool-call+ arguments push their message into the folded head instead of being counted+ as a handful of tokens in the preserved tail.+ """+ accumulated = 0+ cut_idx = len(body)+ for i in range(len(body) - 1, -1, -1):+ msg_tokens = (_msg_estimate_chars(body[i]) // 4) + 10+ if accumulated + msg_tokens > budget:+ cut_idx = i + 1+ break+ accumulated += msg_tokens+ cut_idx = i+ while 0 < cut_idx < len(body) and body[cut_idx].get("role") == "tool":+ cut_idx += 1+ return cut_idx+def _fix_tool_pairs(messages: list) -> None:"""Repair orphaned tool_call / tool_result pairs after compression.@@ -2646,21 +2712,9 @@ class AgentLoop:system_msg = messages[0]body = messages[1:]- # Token-budget tail: walk backward to find how many recent messages to preserve- accumulated = 0- cut_idx = len(body)- for i in range(len(body) - 1, -1, -1):- content = body[i].get("content", "")- msg_tokens = (len(str(content)) // 4) + 10- if accumulated + msg_tokens > TAIL_TOKEN_BUDGET:- cut_idx = i + 1- break- accumulated += msg_tokens- cut_idx = i-- # Don't split in the middle of a tool_call/tool_result pair- while 0 < cut_idx < len(body) and body[cut_idx].get("role") == "tool":- cut_idx += 1+ # Token-budget tail: size messages with their tool-call arguments so+ # oversized tool calls are folded instead of hiding in the tail.+ cut_idx = _tail_cut_index(body)head = body[:cut_idx]tail = body[cut_idx:]diff --git a/agent/tests/test_loop_helpers.py b/agent/tests/test_loop_helpers.pyindex b3e1db506..df631f27d 100644--- a/agent/tests/test_loop_helpers.py+++ b/agent/tests/test_loop_helpers.py@@ -598,3 +598,87 @@ def test_stall_timeout_seconds_default_and_override(monkeypatch) -> None:assert _stall_timeout_seconds() == 42.0monkeypatch.delattr(loop_module, "STALL_TIMEOUT_SECONDS", raising=False)assert _stall_timeout_seconds() > 0+++# ---------------------------------------------------------------------------+# tool_calls[].function.arguments must count toward compaction sizing/relief+# ---------------------------------------------------------------------------++def test_tail_cut_index_counts_tool_call_arguments() -> None:+ """#tail-budget: a fat tool_call (empty content, huge arguments) must be+ pushed into the folded head, not counted as ~10 tokens in the tail."""+ from src.agent.loop import _tail_cut_index # new helper: absent on pre-fix code++ fat = "X" * 90_000 # ~22.5K tokens once arguments are counted+ body = [+ {"role": "user", "content": "old header"},+ {+ "role": "assistant",+ "content": "",+ "tool_calls": [+ {"id": "c1", "type": "function", "function": {"name": "t", "arguments": fat}}+ ],+ },+ ]+ # Old content-only sizing: both messages fit the 20K budget -> cut at 0.+ # New sizing: the call alone exceeds it -> cut at len(body), tail empty,+ # the oversized call goes into the folded head.+ assert _tail_cut_index(body) == len(body)+ assert _tail_cut_index(body) != 0+++def test_context_collapse_stubs_args_of_cleared_paired_call() -> None:+ """#layer-2: a tool_call whose paired result was [cleared] gets its huge+ arguments folded to a valid JSON '{}' stub (call id/name preserved)."""+ fat = "Y" * (COLLAPSE_TEXT_MIN * 3)+ call_msg = {+ "role": "assistant",+ "content": "",+ "tool_calls": [{"id": "c9", "type": "function", "function": {"name": "f", "arguments": fat}}],+ }+ msgs = [+ {"role": "system", "content": "sys"},+ {"role": "user", "content": "u0"},+ call_msg, # index 2 — inside the collapse window+ {"role": "tool", "tool_call_id": "c9", "content": "[cleared]"},+ {"role": "assistant", "content": "a0"},+ {"role": "user", "content": "u1"},+ {"role": "assistant", "content": "a1"},+ {"role": "user", "content": "u2"},+ {"role": "assistant", "content": "a2"},+ {"role": "user", "content": "u3"},+ ]+ _context_collapse(msgs)+ assert call_msg["tool_calls"][0]["function"]["arguments"] == "{}"+ assert call_msg["tool_calls"][0]["function"]["name"] == "f"+++def test_context_collapse_keeps_args_when_result_intact_or_pending() -> None:+ """#layer-2 negative arm: arguments are NOT stubbed when the result still+ carries data, nor when the call is pending (no result in the transcript)."""+ fat = "Y" * (COLLAPSE_TEXT_MIN * 3)+ c1 = {+ "role": "assistant",+ "content": "",+ "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "f", "arguments": fat}}],+ }+ c2 = {+ "role": "assistant",+ "content": "",+ "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "g", "arguments": fat}}],+ }+ msgs = [+ {"role": "system", "content": "sys"},+ {"role": "user", "content": "u0"},+ c1, # index 2 — inside the collapse window+ {"role": "tool", "tool_call_id": "c1", "content": "still have the data"},+ c2, # pending — result never arrived+ {"role": "assistant", "content": "continue"},+ {"role": "user", "content": "u1"},+ {"role": "assistant", "content": "a1"},+ {"role": "user", "content": "u2"},+ {"role": "assistant", "content": "a2"},+ ]+ _context_collapse(msgs)+ assert c1["tool_calls"][0]["function"]["arguments"] == fat+ assert c2["tool_calls"][0]["function"]["arguments"] == fat