Compare commits
1 Commits
whip/251-1
...
fix/500-cl
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
568a735f65 |
@@ -41,6 +41,64 @@ from agent.model_metadata import is_local_endpoint
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# Minimum context tokens required for cron job execution
|
||||
CRON_MIN_CONTEXT_TOKENS = 500
|
||||
|
||||
|
||||
class ModelContextError(Exception):
|
||||
"""Raised when a model does not have enough context tokens for a cron job."""
|
||||
pass
|
||||
|
||||
|
||||
# =====================================================================
|
||||
# Cloud Context Warning — detect local service refs in cloud prompts
|
||||
# =====================================================================
|
||||
|
||||
import re as _re
|
||||
|
||||
_LOCAL_SERVICE_PATTERNS = [
|
||||
_re.compile(r'\blocalhost:\d+', _re.IGNORECASE),
|
||||
_re.compile(r'\b127\.\d+\.\d+\.\d+:\d+', _re.IGNORECASE),
|
||||
_re.compile(r'\b0\.0\.0\.0:\d+', _re.IGNORECASE),
|
||||
_re.compile(r'\bollama\b', _re.IGNORECASE),
|
||||
_re.compile(r'\bcurl\s+localhost\b', _re.IGNORECASE),
|
||||
_re.compile(r'\bwget\s+localhost\b', _re.IGNORECASE),
|
||||
_re.compile(r'\bhttp://localhost\b', _re.IGNORECASE),
|
||||
_re.compile(r'\bhttps?://127\.\d+\.\d+\.\d+\b', _re.IGNORECASE),
|
||||
_re.compile(r'\bcheck\s+ollama\b', _re.IGNORECASE),
|
||||
_re.compile(r'\bconnect\s+local\b', _re.IGNORECASE),
|
||||
_re.compile(r'\bhermes\s+gateway\s+local\b', _re.IGNORECASE),
|
||||
_re.compile(r'\blocal\s+model\b', _re.IGNORECASE),
|
||||
]
|
||||
|
||||
_CLOUD_CONTEXT_WARNING = (
|
||||
"\n\n[SYSTEM NOTE: This cron job is running on a CLOUD inference endpoint. "
|
||||
"Local services (Ollama, localhost, local gateway) are NOT accessible from "
|
||||
"this environment. Do not attempt to connect to localhost, run curl/wget "
|
||||
"against local ports, or check local model availability. Report the "
|
||||
"limitation and focus on tasks achievable remotely.]\n"
|
||||
)
|
||||
|
||||
|
||||
def _detect_local_service_refs(text: str) -> list[str]:
|
||||
"""Detect references to local services in prompt text."""
|
||||
refs = []
|
||||
for pat in _LOCAL_SERVICE_PATTERNS:
|
||||
if pat.search(text):
|
||||
refs.append(pat.pattern)
|
||||
return refs
|
||||
|
||||
|
||||
def _inject_cloud_context(prompt: str, base_url: str) -> str:
|
||||
"""If running on cloud but prompt references local services, inject warning."""
|
||||
if is_local_endpoint(base_url):
|
||||
return prompt
|
||||
refs = _detect_local_service_refs(prompt)
|
||||
if refs:
|
||||
logger.info("Cloud endpoint + local service refs detected (%d patterns), injecting warning", len(refs))
|
||||
return _CLOUD_CONTEXT_WARNING + prompt
|
||||
return prompt
|
||||
|
||||
|
||||
# =====================================================================
|
||||
# Deploy Sync Guard
|
||||
@@ -817,6 +875,9 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
|
||||
job_name,
|
||||
)
|
||||
|
||||
# Inject cloud-context warning if prompt references local services (#468)
|
||||
prompt = _inject_cloud_context(prompt, _runtime_base_url)
|
||||
|
||||
_agent_kwargs = _safe_agent_kwargs({
|
||||
"model": turn_route["model"],
|
||||
"api_key": turn_route["runtime"].get("api_key"),
|
||||
|
||||
@@ -234,7 +234,12 @@ class HolographicMemoryProvider(MemoryProvider):
|
||||
return self._handle_fact_feedback(args)
|
||||
return json.dumps({"error": f"Unknown tool: {tool_name}"})
|
||||
|
||||
|
||||
def on_session_end(self, messages: List[Dict[str, Any]]) -> None:
|
||||
if not self._config.get("auto_extract", False):
|
||||
return
|
||||
if not self._store or not messages:
|
||||
return
|
||||
self._auto_extract_facts(messages)
|
||||
|
||||
def on_memory_write(self, action: str, target: str, content: str) -> None:
|
||||
"""Mirror built-in memory writes as facts.
|
||||
@@ -261,44 +266,6 @@ class HolographicMemoryProvider(MemoryProvider):
|
||||
except Exception as e:
|
||||
logger.debug("Holographic memory_write mirror failed: %s", e)
|
||||
|
||||
def on_session_end(self, messages: List[Dict[str, Any]]) -> None:
|
||||
"""Run auto-extraction and periodic contradiction detection."""
|
||||
if self._config.get("auto_extract", False):
|
||||
self._auto_extract_facts(messages)
|
||||
|
||||
# Periodic contradiction detection (run every ~50 sessions or on first session)
|
||||
self._maybe_run_contradiction_scan()
|
||||
|
||||
def _maybe_run_contradiction_scan(self) -> None:
|
||||
"""Run contradiction detection if enough sessions have passed since last run."""
|
||||
if not self._store or not self._retriever:
|
||||
return
|
||||
try:
|
||||
# Use a counter file to track sessions since last scan
|
||||
from hermes_constants import get_hermes_home
|
||||
counter_path = get_hermes_home() / ".contradiction_scan_counter"
|
||||
|
||||
count = 0
|
||||
if counter_path.exists():
|
||||
try:
|
||||
count = int(counter_path.read_text().strip())
|
||||
except (ValueError, OSError):
|
||||
count = 0
|
||||
|
||||
count += 1
|
||||
counter_path.write_text(str(count))
|
||||
|
||||
# Run every 50 sessions
|
||||
if count >= 50:
|
||||
counter_path.write_text("0")
|
||||
from .resolver import ContradictionResolver
|
||||
resolver = ContradictionResolver(self._store, self._retriever)
|
||||
report = resolver.run_detection_and_resolution(limit=50, auto_resolve=True)
|
||||
if report.contradictions_found > 0:
|
||||
logger.info("Periodic contradiction scan: %s", report.summary())
|
||||
except Exception as e:
|
||||
logger.debug("Periodic contradiction scan failed: %s", e)
|
||||
|
||||
def shutdown(self) -> None:
|
||||
self._store = None
|
||||
self._retriever = None
|
||||
|
||||
@@ -1,179 +0,0 @@
|
||||
"""Contradiction detection and resolution for holographic memory.
|
||||
|
||||
Periodically scans the fact store for contradictions using the retriever's
|
||||
contradict() method, then resolves obvious cases and flags ambiguous ones.
|
||||
|
||||
Resolution strategy:
|
||||
- Obvious: same entity, newer fact supersedes older → lower trust on older
|
||||
- Ambiguous: different claims about same entity → flag for review, don't auto-resolve
|
||||
- High-confidence contradiction (>0.7 score): lower trust on both, log warning
|
||||
|
||||
Usage:
|
||||
from plugins.memory.holographic.resolver import ContradictionResolver
|
||||
resolver = ContradictionResolver(store, retriever)
|
||||
report = resolver.run_detection_and_resolution()
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import logging
|
||||
from datetime import datetime
|
||||
from typing import Any, Dict, List, Optional
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# Thresholds
|
||||
_OBVIOUS_THRESHOLD = 0.5 # contradiction_score >= this → obvious
|
||||
_AMBIGUOUS_THRESHOLD = 0.3 # contradiction_score >= this → flag
|
||||
_HIGH_CONFIDENCE = 0.7 # contradiction_score >= this → high confidence
|
||||
_TRUST_PENALTY_OLD = 0.3 # trust reduction for superseded fact
|
||||
_TRUST_PENALTY_CONFLICT = 0.15 # trust reduction for ambiguous conflicts
|
||||
|
||||
|
||||
class ContradictionReport:
|
||||
"""Results of a contradiction detection and resolution run."""
|
||||
|
||||
def __init__(self):
|
||||
self.total_scanned: int = 0
|
||||
self.contradictions_found: int = 0
|
||||
self.auto_resolved: int = 0
|
||||
self.flagged_for_review: int = 0
|
||||
self.high_confidence: int = 0
|
||||
self.resolved_pairs: List[Dict[str, Any]] = []
|
||||
self.flagged_pairs: List[Dict[str, Any]] = []
|
||||
|
||||
def summary(self) -> str:
|
||||
lines = [
|
||||
f"Contradiction scan: {self.total_scanned} facts, "
|
||||
f"{self.contradictions_found} contradictions found",
|
||||
f" Auto-resolved: {self.auto_resolved} (newer supersedes older)",
|
||||
f" High-confidence: {self.high_confidence} (trust lowered on both)",
|
||||
f" Flagged for review: {self.flagged_for_review}",
|
||||
]
|
||||
for pair in self.flagged_pairs[:5]:
|
||||
lines.append(
|
||||
f" [REVIEW] score={pair['contradiction_score']:.2f} "
|
||||
f"entities={pair['shared_entities']} "
|
||||
f"A: {pair['fact_a']['content'][:50]}… "
|
||||
f"B: {pair['fact_b']['content'][:50]}…"
|
||||
)
|
||||
return "\n".join(lines)
|
||||
|
||||
def to_dict(self) -> dict:
|
||||
return {
|
||||
"total_scanned": self.total_scanned,
|
||||
"contradictions_found": self.contradictions_found,
|
||||
"auto_resolved": self.auto_resolved,
|
||||
"high_confidence": self.high_confidence,
|
||||
"flagged_for_review": self.flagged_for_review,
|
||||
"resolved_pairs": self.resolved_pairs,
|
||||
"flagged_pairs": self.flagged_pairs,
|
||||
}
|
||||
|
||||
|
||||
class ContradictionResolver:
|
||||
"""Detects and resolves contradictions in the holographic fact store."""
|
||||
|
||||
def __init__(self, store, retriever):
|
||||
self._store = store
|
||||
self._retriever = retriever
|
||||
|
||||
def run_detection_and_resolution(
|
||||
self,
|
||||
limit: int = 50,
|
||||
auto_resolve: bool = True,
|
||||
) -> ContradictionReport:
|
||||
"""Run a full contradiction detection and resolution pass.
|
||||
|
||||
Args:
|
||||
limit: Max contradiction pairs to process.
|
||||
auto_resolve: If True, auto-resolve obvious cases.
|
||||
|
||||
Returns:
|
||||
ContradictionReport with all findings and actions taken.
|
||||
"""
|
||||
report = ContradictionReport()
|
||||
|
||||
try:
|
||||
contradictions = self._retriever.contradict(limit=limit)
|
||||
except Exception as e:
|
||||
logger.warning("Contradiction detection failed: %s", e)
|
||||
return report
|
||||
|
||||
report.total_scanned = len(contradictions)
|
||||
report.contradictions_found = len(contradictions)
|
||||
|
||||
for pair in contradictions:
|
||||
score = pair.get("contradiction_score", 0.0)
|
||||
|
||||
if score >= _HIGH_CONFIDENCE:
|
||||
report.high_confidence += 1
|
||||
if auto_resolve:
|
||||
self._resolve_high_confidence(pair, report)
|
||||
elif score >= _OBVIOUS_THRESHOLD:
|
||||
if auto_resolve:
|
||||
self._resolve_obvious(pair, report)
|
||||
elif score >= _AMBIGUOUS_THRESHOLD:
|
||||
report.flagged_for_review += 1
|
||||
report.flagged_pairs.append(pair)
|
||||
# Lower trust slightly on both to reduce retrieval weight
|
||||
self._penalize_both(pair, _TRUST_PENALTY_CONFLICT)
|
||||
|
||||
return report
|
||||
|
||||
def _resolve_obvious(self, pair: dict, report: ContradictionReport) -> None:
|
||||
"""Resolve obvious contradiction: newer fact supersedes older.
|
||||
|
||||
Same entity, clearly contradictory claims. Newer wins.
|
||||
"""
|
||||
fact_a = pair["fact_a"]
|
||||
fact_b = pair["fact_b"]
|
||||
|
||||
# Determine which is newer
|
||||
a_time = fact_a.get("updated_at") or fact_a.get("created_at", "")
|
||||
b_time = fact_b.get("updated_at") or fact_b.get("created_at", "")
|
||||
|
||||
if a_time >= b_time:
|
||||
newer, older = fact_a, fact_b
|
||||
else:
|
||||
newer, older = fact_b, fact_a
|
||||
|
||||
# Lower trust on the older (superseded) fact
|
||||
try:
|
||||
self._store.update_fact(
|
||||
older["fact_id"],
|
||||
trust_delta=-_TRUST_PENALTY_OLD,
|
||||
)
|
||||
report.auto_resolved += 1
|
||||
report.resolved_pairs.append({
|
||||
"action": "superseded",
|
||||
"kept": newer["fact_id"],
|
||||
"demoted": older["fact_id"],
|
||||
"reason": f"Newer fact supersedes older (score={pair['contradiction_score']:.2f})",
|
||||
})
|
||||
logger.info(
|
||||
"Contradiction resolved: fact#%d supersedes fact#%d (entities: %s)",
|
||||
newer["fact_id"], older["fact_id"],
|
||||
pair.get("shared_entities", []),
|
||||
)
|
||||
except Exception as e:
|
||||
logger.debug("Failed to resolve contradiction: %s", e)
|
||||
|
||||
def _resolve_high_confidence(self, pair: dict, report: ContradictionReport) -> None:
|
||||
"""Resolve high-confidence contradiction: lower trust on both facts.
|
||||
|
||||
We can't determine which is correct, so both get penalized.
|
||||
"""
|
||||
self._penalize_both(pair, _TRUST_PENALTY_CONFLICT * 2)
|
||||
report.flagged_pairs.append(pair)
|
||||
|
||||
def _penalize_both(self, pair: dict, penalty: float) -> None:
|
||||
"""Lower trust on both contradictory facts."""
|
||||
for key in ("fact_a", "fact_b"):
|
||||
fact = pair.get(key, {})
|
||||
fid = fact.get("fact_id")
|
||||
if fid:
|
||||
try:
|
||||
self._store.update_fact(fid, trust_delta=-penalty)
|
||||
except Exception as e:
|
||||
logger.debug("Failed to penalize fact#%d: %s", fid, e)
|
||||
83
tests/cron/test_cron_cloud_context.py
Normal file
83
tests/cron/test_cron_cloud_context.py
Normal file
@@ -0,0 +1,83 @@
|
||||
"""Tests for cron cloud-context warning injection (#468)."""
|
||||
|
||||
import pytest
|
||||
|
||||
from cron.scheduler import (
|
||||
_LOCAL_SERVICE_PATTERNS,
|
||||
_detect_local_service_refs,
|
||||
_inject_cloud_context,
|
||||
_CLOUD_CONTEXT_WARNING,
|
||||
)
|
||||
|
||||
|
||||
class TestDetectLocalServiceRefs:
|
||||
"""Test local service reference detection."""
|
||||
|
||||
def test_detects_localhost_with_port(self):
|
||||
refs = _detect_local_service_refs("Connect to localhost:11434")
|
||||
assert len(refs) > 0
|
||||
|
||||
def test_detects_127_address(self):
|
||||
refs = _detect_local_service_refs("Check http://127.0.0.1:8080/health")
|
||||
assert len(refs) > 0
|
||||
|
||||
def test_detects_ollama(self):
|
||||
refs = _detect_local_service_refs("Run ollama pull gemma4")
|
||||
assert len(refs) > 0
|
||||
|
||||
def test_detects_curl_localhost(self):
|
||||
refs = _detect_local_service_refs("curl localhost:11434/api/tags")
|
||||
assert len(refs) > 0
|
||||
|
||||
def test_detects_wget_localhost(self):
|
||||
refs = _detect_local_service_refs("wget localhost:8080/data")
|
||||
assert len(refs) > 0
|
||||
|
||||
def test_detects_http_localhost(self):
|
||||
refs = _detect_local_service_refs("http://localhost:3000")
|
||||
assert len(refs) > 0
|
||||
|
||||
def test_detects_local_model(self):
|
||||
refs = _detect_local_service_refs("Use the local model for inference")
|
||||
assert len(refs) > 0
|
||||
|
||||
def test_no_refs_returns_empty(self):
|
||||
refs = _detect_local_service_refs("Search the web for Python tutorials")
|
||||
assert len(refs) == 0
|
||||
|
||||
def test_case_insensitive(self):
|
||||
refs = _detect_local_service_refs("OLLAMA is running on LocalHost:11434")
|
||||
assert len(refs) > 0
|
||||
|
||||
|
||||
class TestInjectCloudContext:
|
||||
"""Test cloud context warning injection."""
|
||||
|
||||
def test_no_warning_on_local_endpoint(self):
|
||||
prompt = "Check ollama on localhost:11434"
|
||||
result = _inject_cloud_context(prompt, "http://localhost:11434/v1")
|
||||
assert result == prompt # No injection for local endpoints
|
||||
|
||||
def test_no_warning_when_no_local_refs(self):
|
||||
prompt = "Search the web for news"
|
||||
result = _inject_cloud_context(prompt, "https://api.openai.com/v1")
|
||||
assert result == prompt
|
||||
|
||||
def test_injects_warning_on_cloud_with_local_refs(self):
|
||||
prompt = "Check ollama status on localhost:11434"
|
||||
result = _inject_cloud_context(prompt, "https://api.openai.com/v1")
|
||||
assert _CLOUD_CONTEXT_WARNING in result
|
||||
assert prompt in result
|
||||
assert result.startswith(_CLOUD_CONTEXT_WARNING)
|
||||
|
||||
def test_nous_cloud_injects_warning(self):
|
||||
prompt = "curl localhost:11434/api/tags"
|
||||
result = _inject_cloud_context(prompt, "https://inference-api.nousresearch.com/v1")
|
||||
assert _CLOUD_CONTEXT_WARNING in result
|
||||
|
||||
def test_warning_content(self):
|
||||
prompt = "local model check"
|
||||
result = _inject_cloud_context(prompt, "https://api.example.com/v1")
|
||||
assert "CLOUD" in result
|
||||
assert "NOT accessible" in result
|
||||
assert "localhost" in result
|
||||
Reference in New Issue
Block a user