fix
This commit is contained in:
Mikan
@@ -312,7 +312,16 @@ async def _run_tool_loop(
|
||||
max_substeps: int,
|
||||
settings: dict[str, Any],
|
||||
) -> dict[str, Any] | None:
|
||||
"""Generic tool-calling loop. Returns the result of the terminal tool call."""
|
||||
"""Generic tool-calling loop. Returns the result of the terminal tool call.
|
||||
|
||||
Rules:
|
||||
- The terminal tool (e.g. submit_plan) cannot be called on the first substep.
|
||||
The model must do at least some work first.
|
||||
- If the terminal tool is called in the same response as other tools that
|
||||
FAILED, the terminal call is cancelled (the model should fix errors first).
|
||||
- If the model returns no tool calls, retry with a system nudge (up to
|
||||
`llm.tool_retry_attempts` times, default 3).
|
||||
"""
|
||||
registry = get_registry()
|
||||
ctx = ToolContext(
|
||||
db=db, world=world, stage=stage,
|
||||
@@ -324,6 +333,8 @@ async def _run_tool_loop(
|
||||
]
|
||||
tools = registry.to_openai_format(stage)
|
||||
last_terminal_result: dict[str, Any] | None = None
|
||||
tool_retry_attempts = int(settings.get("llm.tool_retry_attempts", 3)) if settings else 3
|
||||
no_tool_retries = 0
|
||||
|
||||
for substep in range(max_substeps):
|
||||
await sse.emit("llm_call_start", {"stage": stage, "model": getattr(llm, "_model", "mock")})
|
||||
@@ -350,8 +361,16 @@ async def _run_tool_loop(
|
||||
msg["content"] = content
|
||||
tool_calls = msg.get("tool_calls") or []
|
||||
if not tool_calls:
|
||||
# No tool calls — append assistant message and retry with a nudge.
|
||||
# Up to 3 retries.
|
||||
# No tool calls — retry with nudge (up to tool_retry_attempts)
|
||||
no_tool_retries += 1
|
||||
if no_tool_retries > tool_retry_attempts:
|
||||
# Exhausted retries — force terminal
|
||||
messages.append({"role": "assistant", "content": msg.get("content", "")})
|
||||
messages.append({
|
||||
"role": "user",
|
||||
"content": f"No more retries. You MUST call {terminal_tool} now to end this stage.",
|
||||
})
|
||||
continue
|
||||
messages.append({"role": "assistant", "content": msg.get("content", "")})
|
||||
messages.append({
|
||||
"role": "user",
|
||||
@@ -363,13 +382,43 @@ async def _run_tool_loop(
|
||||
),
|
||||
})
|
||||
continue
|
||||
# Reset retry counter on success
|
||||
no_tool_retries = 0
|
||||
|
||||
# Separate terminal from non-terminal calls
|
||||
non_terminal_calls = [tc for tc in tool_calls if isinstance(tc, dict)
|
||||
and (tc.get("function", {}) if isinstance(tc.get("function"), dict) else {})
|
||||
.get("name") != terminal_tool]
|
||||
terminal_calls = [tc for tc in tool_calls if isinstance(tc, dict)
|
||||
and (tc.get("function", {}) if isinstance(tc.get("function"), dict) else {})
|
||||
.get("name") == terminal_tool]
|
||||
|
||||
# Rule 1: terminal tool not allowed on first substep
|
||||
if substep == 0 and terminal_calls and not non_terminal_calls:
|
||||
messages.append(msg)
|
||||
messages.append({
|
||||
"role": "user",
|
||||
"content": (
|
||||
f"You called {terminal_tool} without doing any work first. "
|
||||
f"You MUST call other tools (entity_create, env_update, etc.) to accomplish the task. "
|
||||
f"Only call {terminal_tool} after you have completed the work."
|
||||
),
|
||||
})
|
||||
continue
|
||||
|
||||
# Execute non-terminal calls first
|
||||
messages.append(msg)
|
||||
for tc in tool_calls:
|
||||
any_failed = False
|
||||
for tc in non_terminal_calls:
|
||||
fn = tc.get("function", {}) if isinstance(tc, dict) else {}
|
||||
if not isinstance(fn, dict):
|
||||
fn = {"name": tc.get("name", ""), "arguments": tc.get("arguments", "{}")}
|
||||
tname = fn.get("name", "")
|
||||
args_str = fn.get("arguments", "{}")
|
||||
if not isinstance(args_str, str):
|
||||
args_str = json.dumps(args_str)
|
||||
try:
|
||||
targs = json.loads(fn.get("arguments") or "{}")
|
||||
targs = json.loads(args_str) if args_str else {}
|
||||
except json.JSONDecodeError:
|
||||
targs = {}
|
||||
result = await registry.execute(tname, targs, ctx)
|
||||
@@ -379,9 +428,49 @@ async def _run_tool_loop(
|
||||
"name": tname,
|
||||
"content": json.dumps(result.to_dict(), ensure_ascii=False),
|
||||
})
|
||||
if tname == terminal_tool:
|
||||
last_terminal_result = result.to_dict()
|
||||
return last_terminal_result
|
||||
if not result.ok:
|
||||
any_failed = True
|
||||
|
||||
# Rule 2: if any non-terminal tool failed, cancel terminal calls
|
||||
if terminal_calls:
|
||||
if any_failed:
|
||||
# Cancel terminal — tell the model to fix errors first
|
||||
for tc in terminal_calls:
|
||||
messages.append({
|
||||
"role": "tool",
|
||||
"tool_call_id": tc.get("id", ""),
|
||||
"name": terminal_tool,
|
||||
"content": json.dumps({
|
||||
"ok": False,
|
||||
"error": {
|
||||
"code": "cancelled",
|
||||
"message": f"{terminal_tool} cancelled because other tools in this response failed. Fix the errors first, then call {terminal_tool}.",
|
||||
},
|
||||
}),
|
||||
})
|
||||
continue
|
||||
else:
|
||||
# Execute terminal call
|
||||
tc = terminal_calls[0]
|
||||
fn = tc.get("function", {}) if isinstance(tc, dict) else {}
|
||||
if not isinstance(fn, dict):
|
||||
fn = {"name": tc.get("name", ""), "arguments": tc.get("arguments", "{}")}
|
||||
args_str = fn.get("arguments", "{}")
|
||||
if not isinstance(args_str, str):
|
||||
args_str = json.dumps(args_str)
|
||||
try:
|
||||
targs = json.loads(args_str) if args_str else {}
|
||||
except json.JSONDecodeError:
|
||||
targs = {}
|
||||
result = await registry.execute(terminal_tool, targs, ctx)
|
||||
messages.append({
|
||||
"role": "tool",
|
||||
"tool_call_id": tc.get("id", ""),
|
||||
"name": terminal_tool,
|
||||
"content": json.dumps(result.to_dict(), ensure_ascii=False),
|
||||
})
|
||||
if result.ok:
|
||||
last_terminal_result = result.to_dict()
|
||||
return last_terminal_result
|
||||
|
||||
# If we exhausted substeps without terminal, return None
|
||||
return last_terminal_result
|
||||
|
||||
Reference in New Issue
Block a user