Skip to content

"Vibe-Trading: Your Personal Trading Agent"

fix(agent): recover from native claude-cli tool attempts

Yufeng He9c3bf7bbb47f
diff --git a/README.md b/README.md
index 6cc68b354..af43cfd93 100644
--- a/README.md
+++ b/README.md
@@ -1087,7 +1087,7 @@ Authentication priority is `COPILOT_GITHUB_TOKEN`, `GH_TOKEN`, `GITHUB_TOKEN`, s
### Claude Code CLI provider
-The `claude-cli` provider uses the [Claude Code](https://claude.com/claude-code) CLI you already have installed and logged in. Each agent step runs one headless `claude -p` process that only chooses the next step — call these tools, or give the final answer. Vibe-Trading runs the tools itself, so the evidence ledger, grounding checks and swarm deliverable checks behave exactly as with an API provider. It does not read `ANTHROPIC_API_KEY`.
+The `claude-cli` provider uses the [Claude Code](https://claude.com/claude-code) CLI you already have installed and logged in. Each agent step runs one headless `claude -p` process that only chooses the next step — call these tools, or give the final answer. Vibe-Trading runs the tools itself, so the evidence ledger, grounding checks and swarm deliverable checks behave exactly as with an API provider. It does not read `ANTHROPIC_API_KEY`. Native CLI tools stay disabled. If the model tries one, Vibe-Trading feeds the refusal back and asks for a host tool call instead of accepting a claim that the tools are unavailable.
Log in once with Claude Code, then check that Vibe-Trading can see the login:
diff --git a/agent/src/agent/context.py b/agent/src/agent/context.py
index f1f00abd6..69e32e30f 100644
--- a/agent/src/agent/context.py
+++ b/agent/src/agent/context.py
@@ -350,8 +350,10 @@ class ContextBuilder:
tool_discovery = (
"The complete available-tool catalog follows. Only loaded tools have full schemas "
"in this request. Use load_tools(names=[...]) to load the exact names you need, "
- "preferably together, then call them directly. Loading does not execute a tool. "
- "No catalog entry is hidden by search or ranking.\n\n" + self.registry.catalog()
+ "preferably together, then request them through the host's tool-call protocol. "
+ "Loading does not execute a tool. "
+ "No catalog entry is hidden by search or ranking.\n\n"
+ + self.registry.catalog()
)
return _SYSTEM_PROMPT.format(
diff --git a/agent/src/agent/tool_loading.py b/agent/src/agent/tool_loading.py
index 5a1d4fe72..4eb09c9dd 100644
--- a/agent/src/agent/tool_loading.py
+++ b/agent/src/agent/tool_loading.py
@@ -78,8 +78,9 @@ class DeferredToolRegistry(ToolRegistry):
# prefix/order for providers that cache common prompt prefixes.
self._loaded.update(dict.fromkeys(names))
result: dict[str, Any] = {
- "status": "ok", "loaded": list(dict.fromkeys(names)),
- "message": "Full schemas are available in the next request. Call the tools directly; loading has not executed them.",
+ "status": "ok",
+ "loaded": list(dict.fromkeys(names)),
+ "message": "Full schemas are available in the next request. Request them through the host's tool-call protocol in your next reply; loading has not executed them.",
}
return result
@@ -90,13 +91,16 @@ class LoadToolsTool(BaseTool):
name = "load_tools"
description = (
"Load full argument schemas for tools named in the system catalog. "
- "Select all tools needed for the next steps in one call, then call them "
- "directly in the next turn. This only exposes schemas; it executes no tool."
+ "Select all tools needed for the next steps in one call, then request them "
+ "through the host's tool-call protocol in the next turn. This only exposes "
+ "schemas; it executes no tool."
)
parameters = {
- "type": "object", "properties": {
+ "type": "object",
+ "properties": {
"names": {"type": "array", "items": {"type": "string"}, "minItems": 1},
- }, "required": ["names"],
+ },
+ "required": ["names"],
}
repeatable = True
diff --git a/agent/src/providers/claude_cli.py b/agent/src/providers/claude_cli.py
index 012b28132..46f72d11e 100644
--- a/agent/src/providers/claude_cli.py
+++ b/agent/src/providers/claude_cli.py
@@ -72,7 +72,9 @@ def child_env() -> dict[str, str]:
_PROTOCOL = (
"\n\n## Tool-calling protocol\n"
"You cannot execute tools yourself; the host executes them and returns the "
- "results in the next turn. Respond with action=tool_calls and the calls you "
+ "results in the next turn. The CLI's native tools are disabled: never call "
+ "them, and never interpret their refusal as host tools being unavailable. "
+ "Respond with action=tool_calls and the calls you "
"want (name + arguments that match the tool's JSON schema), or action=final "
"with your complete answer in content once you have what you need.\n"
"Available tools (OpenAI function format):\n"
@@ -351,6 +353,18 @@ def _usage(payload: dict[str, Any]) -> dict[str, int]:
}
+def _native_attempt_reason(payload: dict[str, Any]) -> str:
+ """Identify a native tool attempt without parsing the model's prose."""
+ if payload.get("permission_denials"):
+ return "permission_denials is non-empty"
+ turns = payload.get("num_turns")
+ if isinstance(turns, int) and turns > 1:
+ # Native tools are disabled for a one-decision request. Extra turns are
+ # the remaining signal when the CLI does not itemize the denied calls.
+ return f"num_turns={turns}"
+ return ""
+
+
def _decide(
*,
model: str,
@@ -406,6 +420,17 @@ def _decide(
else []
)
if not calls:
+ native_reason = _native_attempt_reason(payload) if allow_tools else ""
+ if native_reason:
+ user_input += (
+ "\n\n[host] Native tool calls were attempted and refused; none was executed "
+ f"({native_reason}). This does not mean the host tools are unavailable or broken. "
+ "Do not call the CLI's native tools. Reply again with action=tool_calls "
+ "using the host-provided schemas, or action=final with a complete answer "
+ "in content if no tool is needed."
+ )
+ failure = f"native tool calls kept bypassing the host protocol ({native_reason})"
+ continue
# Under --json-schema, payload["result"] is the structured output as text, so
# an empty content must not fall back to it: the user would see raw JSON.
content = str(output.get("content") or "")
diff --git a/agent/tests/test_claude_cli_provider.py b/agent/tests/test_claude_cli_provider.py
index 6bedc1493..7090418cb 100644
--- a/agent/tests/test_claude_cli_provider.py
+++ b/agent/tests/test_claude_cli_provider.py
@@ -136,6 +136,103 @@ def test_final_without_content_is_fed_back_and_retried(fake_cli) -> None:
assert message.usage_metadata["output_tokens"] == 10
+def test_native_attempt_is_fed_back_with_the_original_prompt(fake_cli) -> None:
+ first = payload({"action": "final", "content": "No such tool available"})
+ first["num_turns"] = 2
+ set_replies, calls = fake_cli
+ set_replies(
+ {"payload": first},
+ {
+ "payload": payload(
+ {
+ "action": "tool_calls",
+ "tool_calls": [
+ {
+ "name": "get_market_data",
+ "arguments": {"codes": ["HYPE-USDT"]},
+ }
+ ],
+ }
+ )
+ },
+ )
+
+ message = claude_cli.ClaudeCliLLM(tools=[tool("get_market_data")]).invoke(
+ [{"role": "user", "content": "run the backtest"}]
+ )
+
+ assert message.tool_calls[0]["name"] == "get_market_data"
+ stdin = calls()[1]["stdin"]
+ assert "run the backtest" in stdin
+ assert "Native tool calls" in stdin
+ assert "action=tool_calls" in stdin
+
+
+def test_permission_denials_are_fed_back_instead_of_accepting_final(fake_cli) -> None:
+ first = payload({"action": "final", "content": "tools are unavailable"})
+ first["permission_denials"] = [{"tool_name": "bash"}]
+ set_replies, calls = fake_cli
+ set_replies(
+ {"payload": first},
+ {"payload": payload({"action": "final", "content": "recovered"})},
+ )
+
+ message = claude_cli.ClaudeCliLLM(tools=[tool("get_market_data")]).invoke(
+ [{"role": "user", "content": "x"}]
+ )
+
+ assert message.content == "recovered"
+ assert "Native tool calls" in calls()[1]["stdin"]
+
+
+def test_repeated_native_attempts_exhaust_the_existing_retry_budget(fake_cli) -> None:
+ reply = payload({"action": "final", "content": "tools are broken"})
+ reply["num_turns"] = 2
+ set_replies, calls = fake_cli
+ set_replies({"payload": reply})
+
+ with pytest.raises(RuntimeError, match="native tool calls kept bypassing"):
+ claude_cli.ClaudeCliLLM(tools=[tool("get_market_data")]).invoke(
+ [{"role": "user", "content": "x"}]
+ )
+
+ assert len(calls()) == claude_cli._MAX_RETRIES + 1
+
+
+def test_valid_structured_calls_are_not_retried_after_native_metadata(fake_cli) -> None:
+ reply = payload(
+ {
+ "action": "tool_calls",
+ "tool_calls": [{"name": "get_market_data", "arguments": {}}],
+ }
+ )
+ reply["permission_denials"] = [{"tool_name": "bash"}]
+ set_replies, calls = fake_cli
+ set_replies({"payload": reply})
+
+ message = claude_cli.ClaudeCliLLM(tools=[tool("get_market_data")]).invoke(
+ [{"role": "user", "content": "x"}]
+ )
+
+ assert message.tool_calls[0]["name"] == "get_market_data"
+ assert len(calls()) == 1
+
+
+def test_single_turn_final_answer_does_not_trigger_native_retry(fake_cli) -> None:
+ reply = payload({"action": "final", "content": "done"})
+ reply["num_turns"] = 1
+ reply["permission_denials"] = []
+ set_replies, calls = fake_cli
+ set_replies({"payload": reply})
+
+ message = claude_cli.ClaudeCliLLM(tools=[tool("get_market_data")]).invoke(
+ [{"role": "user", "content": "x"}]
+ )
+
+ assert message.content == "done"
+ assert len(calls()) == 1
+
+
def test_final_never_echoes_the_structured_output_as_content(fake_cli) -> None:
set_replies, _ = fake_cli
set_replies({"payload": payload({"action": "final", "content": " "})})
diff --git a/agent/tests/test_deferred_tool_loading.py b/agent/tests/test_deferred_tool_loading.py
index bdb6cae75..01e0aff49 100644
--- a/agent/tests/test_deferred_tool_loading.py
+++ b/agent/tests/test_deferred_tool_loading.py
@@ -47,6 +47,21 @@ def test_every_tool_is_visible_and_loading_preserves_exact_schema():
assert len(deferred.get_definitions()) == len(source) + 1 # Full manifest.
+def test_loading_message_uses_the_host_tool_protocol():
+ deferred = DeferredToolRegistry(registry())
+
+ result = deferred.expose(["read_observation_2"])
+
+ assert result["message"] == (
+ "Full schemas are available in the next request. Request them through "
+ "the host's tool-call protocol in your next reply; loading has not executed them."
+ )
+ assert "host's tool-call protocol" in deferred.get("load_tools").description
+ prompt = ContextBuilder(deferred, WorkspaceMemory()).build_system_prompt()
+ assert "then request them through the host's tool-call protocol" in prompt
+ assert "then call them directly" not in prompt
+
+
def test_loading_cannot_add_a_tool_outside_the_current_registry():
deferred = DeferredToolRegistry(registry())
before = request_definitions(deferred)