feat: harden Qwen agentic tool orchestration
This commit is contained in:
Mikei386
@@ -1,7 +1,7 @@
|
||||
"""
|
||||
title: MikeAI Auto Tool Selector
|
||||
author: MikeAI
|
||||
version: 3.0.0
|
||||
version: 3.2.0
|
||||
description: Selects a small, relevant set of MCP servers for each user request.
|
||||
"""
|
||||
|
||||
@@ -16,8 +16,11 @@ class Filter:
|
||||
class Valves(BaseModel):
|
||||
priority: int = 25
|
||||
enabled: bool = True
|
||||
max_automatic_tools: int = 2
|
||||
max_automatic_tools: int = 3
|
||||
show_selection_status: bool = True
|
||||
enable_multidomain_reasoning: bool = True
|
||||
multidomain_reasoning_effort: str = "medium"
|
||||
multidomain_reasoning_budget: int = 3072
|
||||
|
||||
TOOL_IDS = {
|
||||
"github": "server:mcp:github-local",
|
||||
@@ -121,8 +124,9 @@ class Filter:
|
||||
homeassistant = self._matches(
|
||||
text,
|
||||
(
|
||||
r"\bhome\s*assist(?:ant|ent)\b", r"\bhomeassistant\b", r"\bhass\b",
|
||||
r"\bhome\s*assist(?:ant|ent)s?\b", r"\bhomeassistants?\b", r"\bhass\b",
|
||||
r"\bautomatisierung(?:en)?\b", r"\bentit[aä]t(?:en)?\b",
|
||||
r"\bautomations?\.ya?ml\b", r"\bconfiguration\.ya?ml\b",
|
||||
r"\b(?:sensor|light|switch|climate|automation)\.[\w.-]+",
|
||||
r"\b(?:temperatur|luftfeuchtigkeit|wie warm)\b.*\b"
|
||||
r"(?:k[uü]che|wohnzimmer|schlafzimmer|bad|toilette|keller|haus)\b",
|
||||
@@ -166,35 +170,30 @@ class Filter:
|
||||
),
|
||||
)
|
||||
|
||||
# Existing infrastructure must be inspected before a new integration
|
||||
# is designed. For a GitHub-backed MCP that targets a service already
|
||||
# running on Unraid, source plus Unraid inventory are the two most
|
||||
# useful bounded capabilities; the Operator follows at implementation.
|
||||
if github and unraid:
|
||||
selected.extend(("github", "unraid"))
|
||||
elif operator:
|
||||
# A real integration or incident task can legitimately span three
|
||||
# trust domains (for example GitHub source + Unraid runtime + Home
|
||||
# Assistant configuration). Attach each relevant specialist once;
|
||||
# the evidence-plan rule keeps the model breadth-first and bounded.
|
||||
if operator:
|
||||
selected.append("operator")
|
||||
elif platform:
|
||||
selected.append("platform")
|
||||
elif homeassistant:
|
||||
selected.append("homeassistant")
|
||||
elif arr:
|
||||
selected.append("arr")
|
||||
elif navidrome:
|
||||
selected.append("navidrome")
|
||||
elif unraid:
|
||||
selected.append("unraid")
|
||||
elif github:
|
||||
if github:
|
||||
selected.append("github")
|
||||
if unraid:
|
||||
selected.append("unraid")
|
||||
if homeassistant:
|
||||
selected.append("homeassistant")
|
||||
if arr:
|
||||
selected.append("arr")
|
||||
if navidrome:
|
||||
selected.append("navidrome")
|
||||
if platform and not operator:
|
||||
selected.append("platform")
|
||||
|
||||
# A direct GitHub reference should still use GitHub even if broader web
|
||||
# research is also requested. General public web research is provided
|
||||
# by Open WebUI's native search_web/fetch_url tools and therefore must
|
||||
# not auto-attach the older specialist web MCP.
|
||||
if github and "github" not in selected:
|
||||
selected.insert(0, "github")
|
||||
|
||||
return selected[: max(0, self.valves.max_automatic_tools)]
|
||||
return list(dict.fromkeys(selected))[: max(0, self.valves.max_automatic_tools)]
|
||||
|
||||
async def _notify(self, emitter, labels: list[str]) -> None:
|
||||
if emitter is None or not self.valves.show_selection_status:
|
||||
@@ -230,6 +229,20 @@ class Filter:
|
||||
merged.append(tool_id)
|
||||
body["tool_ids"] = merged
|
||||
|
||||
# Qwen3.8 is markedly more reliable on long, cross-domain tool tasks
|
||||
# when its reasoning channel is preserved. Keep ordinary one-tool
|
||||
# requests fast, but give requests that genuinely span two specialist
|
||||
# MCPs a bounded reasoning budget. An explicit user-selected effort is
|
||||
# never overwritten.
|
||||
if (
|
||||
self.valves.enable_multidomain_reasoning
|
||||
and len(selected) >= 2
|
||||
and body.get("reasoning_effort") in (None, "", "none")
|
||||
):
|
||||
body["reasoning_effort"] = self.valves.multidomain_reasoning_effort
|
||||
if not body.get("reasoning_budget"):
|
||||
body["reasoning_budget"] = self.valves.multidomain_reasoning_budget
|
||||
|
||||
labels = [self.LABELS[category] for category in selected]
|
||||
self._add_system_rule(body, labels)
|
||||
await self._notify(__event_emitter__, labels)
|
||||
|
||||
@@ -1,7 +1,7 @@
|
||||
"""
|
||||
title: MikeAI Stability Guard
|
||||
author: MikeAI
|
||||
version: 2.2.0
|
||||
version: 2.3.0
|
||||
description: Bounds tool output and context use and breaks repeated tool-call loops.
|
||||
"""
|
||||
|
||||
@@ -32,7 +32,7 @@ class Filter:
|
||||
# Secondary protection for histories that re-enter the filter. The
|
||||
# live internal tool loop is bounded and finalized by the derived
|
||||
# OpenWebUI image because inlet filters do not run between its rounds.
|
||||
max_tool_calls_per_turn: int = 10
|
||||
max_tool_calls_per_turn: int = 12
|
||||
max_private_table_tool_calls: int = 6
|
||||
|
||||
def __init__(self):
|
||||
|
||||
@@ -95,7 +95,7 @@ functions = [
|
||||
("thinking", "Thinking", "filter", 20, filter_dir, ""),
|
||||
(
|
||||
"auto_tool_selector", "MikeAI Auto Tool Selector", "filter", 25, filter_dir,
|
||||
"Stellt pro Anfrage höchstens zwei passende MCP-Werkzeuge bereit. "
|
||||
"Stellt pro Anfrage höchstens drei passende MCP-Werkzeuge bereit und aktiviert bei echten Mehrdomänen-Aufgaben begrenztes Reasoning. "
|
||||
"Die Auswahl ist keine Freigabe für schreibende Aktionen.",
|
||||
),
|
||||
("stability_guard", "MikeAI Stability Guard", "filter", 30, filter_dir, ""),
|
||||
|
||||
@@ -225,6 +225,18 @@ params = {
|
||||
"public information you may make at most one focused fallback attempt with "
|
||||
"the general web tool, then synthesize the available evidence or stop clearly; "
|
||||
"never enter a fallback or synonym-search loop. "
|
||||
"For a multi-step or multi-domain request, make a compact evidence plan before "
|
||||
"the first tool call. Reserve at least one call for every requested domain instead "
|
||||
"of exhausting the budget in the first system. Work breadth-first: obtain one "
|
||||
"bounded inventory or overview from each relevant domain, then make only targeted "
|
||||
"follow-up calls for facts still missing. Every call must answer a specific unresolved "
|
||||
"question. Check cheap pass/fail constraints that can invalidate a candidate before "
|
||||
"spending calls on deep research. If a candidate fails a mandatory constraint, switch "
|
||||
"immediately; do not produce the explicitly forbidden candidate as the main result. "
|
||||
"Never invoke the same operation with identical arguments twice, and normally "
|
||||
"use one operation no more than four times. If the user asks for a simulation or plan, "
|
||||
"perform read-only discovery only and do not execute the proposed state changes. Stop "
|
||||
"research as soon as the evidence is sufficient and synthesize the complete answer. "
|
||||
"Never invent tool results, system state, files, measurements, or actions. "
|
||||
"For claims about current external or system state, you must successfully "
|
||||
"use the relevant domain tool during the current request before saying "
|
||||
|
||||
@@ -26,10 +26,15 @@ needle = """ tool_call_iterations = 0
|
||||
"""
|
||||
replacement = """ tool_call_iterations = 0
|
||||
# Open WebUI counts batches, while one model turn may request many
|
||||
# functions in parallel. Bound actual executions as well so a small
|
||||
# local model cannot expand eight rounds into dozens of API calls.
|
||||
# functions in parallel. Keep an execution budget as a second bound,
|
||||
# but make it large enough for genuinely agentic multi-domain work.
|
||||
# Per-tool and exact-repeat limits below prevent one low-level MCP
|
||||
# operation from consuming the whole turn.
|
||||
tool_call_executions = 0
|
||||
max_tool_call_executions = 6
|
||||
max_tool_call_executions = 12
|
||||
max_executions_per_tool = 4
|
||||
tool_execution_counts = {}
|
||||
seen_tool_signatures = set()
|
||||
max_tool_call_iterations = getattr(
|
||||
"""
|
||||
if source.count(needle) != 1:
|
||||
@@ -44,11 +49,30 @@ needle = """ response_tool_calls = tool_calls.pop(0)
|
||||
"""
|
||||
replacement = """ response_tool_calls = tool_calls.pop(0)
|
||||
|
||||
remaining_tool_calls = max(
|
||||
0, max_tool_call_executions - tool_call_executions
|
||||
)
|
||||
skipped_tool_calls = response_tool_calls[remaining_tool_calls:]
|
||||
response_tool_calls = response_tool_calls[:remaining_tool_calls]
|
||||
accepted_tool_calls = []
|
||||
skipped_tool_calls = []
|
||||
skip_reasons = []
|
||||
for candidate in response_tool_calls:
|
||||
function = candidate.get('function') or candidate
|
||||
tool_name = function.get('name') or candidate.get('name') or 'unknown'
|
||||
arguments = function.get('arguments') or candidate.get('arguments') or ''
|
||||
signature = f'{tool_name}:{arguments}'
|
||||
if tool_call_executions + len(accepted_tool_calls) >= max_tool_call_executions:
|
||||
skipped_tool_calls.append(candidate)
|
||||
skip_reasons.append('total execution budget reached')
|
||||
continue
|
||||
if tool_execution_counts.get(tool_name, 0) >= max_executions_per_tool:
|
||||
skipped_tool_calls.append(candidate)
|
||||
skip_reasons.append(f'per-tool budget reached for {tool_name}')
|
||||
continue
|
||||
if signature in seen_tool_signatures:
|
||||
skipped_tool_calls.append(candidate)
|
||||
skip_reasons.append(f'exact duplicate suppressed for {tool_name}')
|
||||
continue
|
||||
accepted_tool_calls.append(candidate)
|
||||
seen_tool_signatures.add(signature)
|
||||
tool_execution_counts[tool_name] = tool_execution_counts.get(tool_name, 0) + 1
|
||||
response_tool_calls = accepted_tool_calls
|
||||
if skipped_tool_calls:
|
||||
skipped_ids = {call.get('id', '') for call in skipped_tool_calls}
|
||||
# Responses API streaming may already have exposed all calls
|
||||
@@ -91,6 +115,7 @@ replacement = """ # The upstream loop otherwise stops wit
|
||||
and tool_call_iterations >= max_tool_call_iterations
|
||||
)
|
||||
or tool_call_executions >= max_tool_call_executions
|
||||
or bool(skipped_tool_calls)
|
||||
)
|
||||
if force_final_response:
|
||||
new_form_data.pop('tools', None)
|
||||
@@ -101,16 +126,32 @@ replacement = """ # The upstream loop otherwise stops wit
|
||||
final_metadata['tool_ids'] = []
|
||||
final_metadata['tool_servers'] = []
|
||||
new_form_data['metadata'] = final_metadata
|
||||
new_form_data['messages'] = add_or_update_system_message(
|
||||
'The tool budget for this turn is exhausted. Do not call '
|
||||
'or imitate any more tools. Give the user a concise final '
|
||||
'answer now using only the tool results already present. '
|
||||
'Explicitly distinguish verified findings from inference, '
|
||||
'and state what could not be verified. Never end without a '
|
||||
'visible answer.',
|
||||
new_form_data['messages'],
|
||||
append=True,
|
||||
final_reason = '; '.join(dict.fromkeys(skip_reasons)) or 'execution budget reached'
|
||||
final_instruction = (
|
||||
'The tool research phase is finished (' + final_reason + '). '
|
||||
'Do not call or imitate any more tools. Give the user a concise '
|
||||
'final answer now using only the tool results already present. '
|
||||
'Explicitly distinguish verified findings from inference, state '
|
||||
'what could not be verified, and never end without a visible answer.'
|
||||
)
|
||||
# Qwen3.8 requires system messages to precede the conversation.
|
||||
# Merge into the first system message (or create it at index 0)
|
||||
# instead of appending a mid-conversation system message.
|
||||
system_message = next(
|
||||
(
|
||||
message
|
||||
for message in new_form_data['messages']
|
||||
if message.get('role') == 'system'
|
||||
and isinstance(message.get('content'), str)
|
||||
),
|
||||
None,
|
||||
)
|
||||
if system_message is None:
|
||||
new_form_data['messages'].insert(
|
||||
0, {'role': 'system', 'content': final_instruction}
|
||||
)
|
||||
else:
|
||||
system_message['content'] += '\\n\\n' + final_instruction
|
||||
# A final user-role instruction is deliberately stronger
|
||||
# than another system suffix after a long tool-call pattern.
|
||||
# It is request-local and is not added to the saved chat.
|
||||
@@ -145,10 +186,46 @@ needle = """ await stream_body_handler(res, new_form_
|
||||
output[:0] = prior_output
|
||||
"""
|
||||
replacement = """ await stream_body_handler(res, new_form_data)
|
||||
if force_final_response:
|
||||
# A model may still print tool-shaped output after the
|
||||
# schemas were removed. It must not trigger the upstream
|
||||
# hard-error path or another execution attempt.
|
||||
if force_final_response and tool_calls:
|
||||
# Some local models still emit one more function-shaped
|
||||
# request after schemas were removed. Discard that
|
||||
# invisible attempt and grant exactly one clean,
|
||||
# tool-less synthesis retry. This is bounded and cannot
|
||||
# execute another function.
|
||||
tool_calls.clear()
|
||||
output = []
|
||||
retry_form_data = {
|
||||
**new_form_data,
|
||||
'messages': [
|
||||
*new_form_data['messages'],
|
||||
{
|
||||
'role': 'user',
|
||||
'content': (
|
||||
'Your previous synthesis attempt was not '
|
||||
'visible because it looked like another '
|
||||
'function call. Tools are no longer available. '
|
||||
'Write the final human-readable answer now, '
|
||||
'starting immediately with the conclusion.'
|
||||
),
|
||||
},
|
||||
],
|
||||
}
|
||||
retry_res = await generate_chat_completion(
|
||||
request,
|
||||
retry_form_data,
|
||||
user,
|
||||
bypass_system_prompt=True,
|
||||
)
|
||||
if isinstance(retry_res, StreamingResponse):
|
||||
await stream_body_handler(retry_res, retry_form_data)
|
||||
# Never let tool-shaped retry output re-enter execution.
|
||||
tool_calls.clear()
|
||||
output[:] = [
|
||||
item
|
||||
for item in output
|
||||
if item.get('type') != 'function_call'
|
||||
]
|
||||
elif force_final_response:
|
||||
tool_calls.clear()
|
||||
output[:0] = prior_output
|
||||
"""
|
||||
|
||||
Reference in New Issue
Block a user