Keep Leo candidate reasoning grounded and stop automatic chat reinforcement (#180)

* Preserve candidate reasoning and disable automatic chat reinforcement

* Retain exact-row guards for combined claim audits

deploy/auto-deploy.sh

@@ -19,8 +19,10 @@ AGENT_STATE_DIR="/opt/teleo-eval/ops/agent-state"
 LEOCLEAN_BIN_DIR="/home/teleo/.hermes/profiles/leoclean/bin"
 LEOCLEAN_SKILLS_DIR="/home/teleo/.hermes/profiles/leoclean/skills"
 LEOCLEAN_PLUGINS_DIR="/home/teleo/.hermes/profiles/leoclean/plugins"
+LEOCLEAN_CONFIG="/home/teleo/.hermes/profiles/leoclean/config.yaml"
 HERMES_AGENT_DIR="/home/teleo/.hermes/hermes-agent"
 HERMES_PATCH_DIR="/home/teleo/.hermes/teleo-runtime-patches"
+LEOCLEAN_POLICY_PATCH="$HERMES_PATCH_DIR/enforce_leoclean_learning_policy.py"
 SYSTEMD_DIR="/etc/systemd/system"
 STAMP_FILE="/opt/teleo-eval/.last-deploy-sha"
 LOG_TAG="auto-deploy"
@@ -55,6 +57,7 @@ NEW_SHA=$(git rev-parse "$DEPLOY_REMOTE/main")
 OLD_SHA=$(cat "$STAMP_FILE" 2>/dev/null || echo "none")
 
 if [ "$NEW_SHA" = "$OLD_SHA" ]; then
+  LEOCLEAN_DRIFT_REPAIRED=false
   # Hermes can be upgraded independently of this repository. Repair an
   # overwritten response-transform hook even when no infrastructure commit
   # changed, and restart only when the patch was actually reapplied.
@@ -65,14 +68,25 @@ if [ "$NEW_SHA" = "$OLD_SHA" ]; then
     fi
     if echo "$PATCH_RESULT" | grep -q '"status": "installed_now"'; then
       log "Hermes response-transform drift repaired: $PATCH_RESULT"
-      if systemctl is-active --quiet leoclean-gateway.service; then
-        sudo systemctl restart leoclean-gateway
-        sleep 10
-        if ! systemctl is-active --quiet leoclean-gateway.service; then
-          log "ERROR: leoclean-gateway failed after response-transform drift repair"
-          exit 1
-        fi
-      fi
+      LEOCLEAN_DRIFT_REPAIRED=true
+    fi
+  fi
+  if [ -f "$LEOCLEAN_POLICY_PATCH" ] && [ -f "$LEOCLEAN_CONFIG" ]; then
+    if ! POLICY_RESULT=$(python3 "$LEOCLEAN_POLICY_PATCH" "$LEOCLEAN_CONFIG"); then
+      log "ERROR: leoclean learning-policy drift could not be repaired"
+      exit 1
+    fi
+    if echo "$POLICY_RESULT" | grep -q '"status": "updated"'; then
+      log "Leoclean learning-policy drift repaired: $POLICY_RESULT"
+      LEOCLEAN_DRIFT_REPAIRED=true
+    fi
+  fi
+  if $LEOCLEAN_DRIFT_REPAIRED && systemctl is-active --quiet leoclean-gateway.service; then
+    sudo systemctl restart leoclean-gateway
+    sleep 10
+    if ! systemctl is-active --quiet leoclean-gateway.service; then
+      log "ERROR: leoclean-gateway failed after runtime-policy drift repair"
+      exit 1
     fi
   fi
   exit 0
@@ -152,6 +166,14 @@ if [ -f "$HERMES_PATCH_DIR/apply_response_transform_hook.py" ]; then
   PATCH_RESULT=$(python3 "$HERMES_PATCH_DIR/apply_response_transform_hook.py" "$HERMES_AGENT_DIR/run_agent.py")
   log "Hermes response transform: $PATCH_RESULT"
 fi
+LEOCLEAN_POLICY_CHANGED=false
+if [ -f "$LEOCLEAN_POLICY_PATCH" ] && [ -f "$LEOCLEAN_CONFIG" ]; then
+  POLICY_RESULT=$(python3 "$LEOCLEAN_POLICY_PATCH" "$LEOCLEAN_CONFIG")
+  log "Leoclean learning policy: $POLICY_RESULT"
+  if echo "$POLICY_RESULT" | grep -q '"status": "updated"'; then
+    LEOCLEAN_POLICY_CHANGED=true
+  fi
+fi
 
 # Safety net: ensure synced .sh files are executable after rsync.
 # Keep this bounded to deploy-owned paths: /opt/teleo-eval also contains
@@ -216,6 +238,9 @@ if [ "$OLD_SHA" != "none" ]; then
     | bash deploy/leoclean-gateway-restart-required.sh; then
     add_restart_if_unit_active leoclean-gateway
   fi
+  if $LEOCLEAN_POLICY_CHANGED; then
+    add_restart_if_unit_active leoclean-gateway
+  fi
 else
   RESTART="teleo-pipeline teleo-diagnostics"
   add_restart_if_unit_active teleo-agent@leo
@@ -247,6 +272,14 @@ if [ -n "$RESTART" ]; then
         FAIL=1
       fi
     fi
+    if [ "$svc" = "leoclean-gateway" ] && [ -f "$LEOCLEAN_POLICY_PATCH" ]; then
+      if POLICY_CHECK=$(python3 "$LEOCLEAN_POLICY_PATCH" --check "$LEOCLEAN_CONFIG"); then
+        log "leoclean-gateway learning policy: $POLICY_CHECK"
+      else
+        log "ERROR: leoclean-gateway learning policy check failed: $POLICY_CHECK"
+        FAIL=1
+      fi
+    fi
   done
 
   if echo "$RESTART" | grep -q "teleo-pipeline"; then

deploy/deploy.sh

@@ -13,8 +13,10 @@ VPS_AGENT_STATE="/opt/teleo-eval/ops/agent-state"
 VPS_LEOCLEAN_BIN="/home/teleo/.hermes/profiles/leoclean/bin"
 VPS_LEOCLEAN_SKILLS="/home/teleo/.hermes/profiles/leoclean/skills"
 VPS_LEOCLEAN_PLUGINS="/home/teleo/.hermes/profiles/leoclean/plugins"
+VPS_LEOCLEAN_CONFIG="/home/teleo/.hermes/profiles/leoclean/config.yaml"
 VPS_HERMES_AGENT="/home/teleo/.hermes/hermes-agent"
 VPS_HERMES_PATCHES="/home/teleo/.hermes/teleo-runtime-patches"
+VPS_LEOCLEAN_POLICY_PATCH="$VPS_HERMES_PATCHES/enforce_leoclean_learning_policy.py"
 VPS_SYSTEMD="/etc/systemd/system"
 REPO_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
 
@@ -109,12 +111,18 @@ rsync "${RSYNC_OPTS[@]}" "$REPO_ROOT/hermes-agent/leoclean-plugins/vps/" "$VPS_H
 echo ""
 
 echo "=== Hermes runtime patches ==="
+LEOCLEAN_POLICY_CHANGED=false
 if $DRY_RUN; then
-  echo "Would sync hermes-agent/patches/ and verify the installed response-transform hook."
+  echo "Would sync hermes-agent/patches/, verify the response-transform hook, and enforce the leoclean learning policy."
 else
   ssh "$VPS_HOST" "mkdir -p '$VPS_HERMES_PATCHES'"
   rsync "${RSYNC_OPTS[@]}" "$REPO_ROOT/hermes-agent/patches/" "$VPS_HOST:$VPS_HERMES_PATCHES/"
   ssh "$VPS_HOST" "python3 '$VPS_HERMES_PATCHES/apply_response_transform_hook.py' '$VPS_HERMES_AGENT/run_agent.py'"
+  POLICY_RESULT=$(ssh "$VPS_HOST" "python3 '$VPS_LEOCLEAN_POLICY_PATCH' '$VPS_LEOCLEAN_CONFIG'")
+  echo "$POLICY_RESULT"
+  if echo "$POLICY_RESULT" | grep -q '"status": "updated"'; then
+    LEOCLEAN_POLICY_CHANGED=true
+  fi
 fi
 echo ""
 
@@ -142,4 +150,9 @@ if $RESTART; then
   echo "=== Restarting services ==="
   ssh "$VPS_HOST" "sudo systemctl restart teleo-pipeline teleo-diagnostics; if systemctl is-active --quiet teleo-agent@leo.service; then sudo systemctl restart teleo-agent@leo; fi; if systemctl list-units --all --full teleo-agent@leo-wallet-test.service --no-legend | grep -q .; then sudo systemctl restart teleo-agent@leo-wallet-test; fi; if systemctl is-active --quiet leoclean-gateway.service; then sudo systemctl restart leoclean-gateway; fi"
   echo "Services restarted."
+elif $LEOCLEAN_POLICY_CHANGED; then
+  echo ""
+  echo "=== Restarting leoclean for learning-policy change ==="
+  ssh "$VPS_HOST" "if systemctl is-active --quiet leoclean-gateway.service; then sudo systemctl restart leoclean-gateway; fi"
+  echo "Leoclean gateway restarted."
 fi

hermes-agent/install-hermes.sh

@@ -59,6 +59,10 @@ terminal:
 memory:
   enabled: true
   search: sqlite_fts5
+  nudge_interval: 0
+
+skills:
+  creation_nudge_interval: 0
 
 tools:
   web_search: true

hermes-agent/leoclean-bin/kb_tool.py

@@ -587,9 +587,7 @@ def operational_contracts(
 
     apply_receipt_question = any(
         term in lowered for term in ("canonical receipt", "postflight", "closure proof", "before/after")
-    ) or bool(
-        re.search(r"\bapplied_at\b\s+(?:readback|proof|receipt|postflight)\b", lowered)
-    )
+    ) or bool(re.search(r"\bapplied_at\b\s+(?:readback|proof|receipt|postflight)\b", lowered))
     if apply_receipt_question:
         contracts.append(
             {
@@ -685,8 +683,26 @@ def operational_contracts(
             )
         )
     )
+    existing_claim_audit_question = any(
+        term in lowered
+        for term in (
+            "inspect one exact claim",
+            "exact claim body and its evidence",
+            "one retrieved canonical claim",
+            "linked source or evidence id",
+        )
+    )
+    candidate_claim_composition_question = bool(
+        any(
+            term in lowered
+            for term in ("candidate claim", "reviewable candidate", "candidate only", "revised exact claim body")
+        )
+        and any(term in lowered for term in ("draft", "propose", "revise", "rewrite", "narrow", "return"))
+        and not existing_claim_audit_question
+    )
     claim_evidence_challenge_question = bool(
-        any(term in lowered for term in ("claim", "exact body", "claim body"))
+        not candidate_claim_composition_question
+        and any(term in lowered for term in ("claim", "exact body", "claim body"))
         and any(term in lowered for term in ("evidence", "source"))
         and any(
             term in lowered

hermes-agent/patches/enforce_leoclean_learning_policy.py

@@ -0,0 +1,114 @@
+#!/usr/bin/env python3
+"""Disable automatic chat-to-memory and chat-to-skill reinforcement for leoclean."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import shutil
+import tempfile
+from pathlib import Path
+from typing import Any
+
+import yaml
+
+POLICY_VERSION = "teleo-leoclean-learning-policy-v1"
+BACKUP_SUFFIX = ".before-teleo-learning-policy"
+MEMORY_NUDGE_INTERVAL = 0
+SKILL_CREATION_NUDGE_INTERVAL = 0
+
+
+def _load_config(path: Path) -> dict[str, Any]:
+    if path.is_symlink():
+        raise ValueError("refusing symlinked profile config")
+    raw = yaml.safe_load(path.read_text(encoding="utf-8")) or {}
+    if not isinstance(raw, dict):
+        raise ValueError("profile config must be a YAML mapping")
+    return raw
+
+
+def policy_projection(config: dict[str, Any]) -> dict[str, int | None]:
+    memory = config.get("memory") if isinstance(config.get("memory"), dict) else {}
+    skills = config.get("skills") if isinstance(config.get("skills"), dict) else {}
+    return {
+        "memory_nudge_interval": memory.get("nudge_interval"),
+        "skill_creation_nudge_interval": skills.get("creation_nudge_interval"),
+    }
+
+
+def enforce_policy(path: Path, *, check: bool = False) -> tuple[dict[str, Any], int]:
+    config = _load_config(path)
+    before = policy_projection(config)
+    expected = {
+        "memory_nudge_interval": MEMORY_NUDGE_INTERVAL,
+        "skill_creation_nudge_interval": SKILL_CREATION_NUDGE_INTERVAL,
+    }
+    changed = before != expected
+
+    if check:
+        return {
+            "policy_version": POLICY_VERSION,
+            "status": "change_required" if changed else "compliant",
+            "target": str(path),
+            "policy": before,
+        }, 1 if changed else 0
+
+    backup_path = path.with_name(path.name + BACKUP_SUFFIX)
+    if changed:
+        memory = config.setdefault("memory", {})
+        skills = config.setdefault("skills", {})
+        if not isinstance(memory, dict) or not isinstance(skills, dict):
+            raise ValueError("memory and skills config sections must be YAML mappings")
+        memory["nudge_interval"] = MEMORY_NUDGE_INTERVAL
+        skills["creation_nudge_interval"] = SKILL_CREATION_NUDGE_INTERVAL
+
+        if not backup_path.exists():
+            shutil.copy2(path, backup_path)
+
+        mode = path.stat().st_mode & 0o777
+        rendered = yaml.safe_dump(config, sort_keys=False, allow_unicode=False)
+        with tempfile.NamedTemporaryFile(
+            "w", encoding="utf-8", dir=path.parent, prefix=f".{path.name}.", delete=False
+        ) as handle:
+            handle.write(rendered)
+            temp_path = Path(handle.name)
+        try:
+            temp_path.chmod(mode)
+            os.replace(temp_path, path)
+        finally:
+            temp_path.unlink(missing_ok=True)
+
+    after = policy_projection(_load_config(path))
+    if after != expected:
+        raise ValueError("profile config did not retain the required learning policy")
+    return {
+        "policy_version": POLICY_VERSION,
+        "status": "updated" if changed else "already_compliant",
+        "target": str(path),
+        "backup": str(backup_path) if changed else None,
+        "policy": after,
+    }, 0
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("target", type=Path)
+    parser.add_argument("--check", action="store_true")
+    args = parser.parse_args()
+    try:
+        result, code = enforce_policy(args.target, check=args.check)
+    except (OSError, ValueError, yaml.YAMLError) as exc:
+        result = {
+            "policy_version": POLICY_VERSION,
+            "status": "error",
+            "target": str(args.target),
+            "error": type(exc).__name__,
+        }
+        code = 2
+    print(json.dumps(result, sort_keys=True))
+    return code
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/leo_behavior_manifest.py

@@ -235,7 +235,15 @@ def identity_config_projection(raw: dict[str, Any]) -> dict[str, Any]:
     section_fields = {
         "model": ("provider", "default", "smart_routing", "smart_routing_model", "max_tokens"),
         "agent": ("reasoning_effort",),
-        "memory": ("enabled", "search"),
+        "memory": (
+            "enabled",
+            "memory_enabled",
+            "user_profile_enabled",
+            "search",
+            "nudge_interval",
+            "flush_min_turns",
+        ),
+        "skills": ("creation_nudge_interval",),
         "tools": ("allowed", "write_enabled"),
         "identity_runtime": ("network_send_enabled", "database_write_enabled", "allowed_tools"),
     }
@@ -328,6 +336,8 @@ def safe_model_config(config_path: Path) -> dict[str, Any]:
         "reasoning_effort": _nested(identity_config, "agent", "reasoning_effort"),
         "memory_enabled": _nested(identity_config, "memory", "enabled"),
         "memory_search": _nested(identity_config, "memory", "search"),
+        "automatic_memory_review_interval": _nested(identity_config, "memory", "nudge_interval"),
+        "automatic_skill_review_interval": _nested(identity_config, "skills", "creation_nudge_interval"),
     }
 
 
@@ -495,6 +505,8 @@ def identity_runtime_payload(manifest: dict[str, Any]) -> dict[str, Any]:
             "reasoning_effort",
             "memory_enabled",
             "memory_search",
+            "automatic_memory_review_interval",
+            "automatic_skill_review_interval",
             "base_model_weights",
             "actual_per_turn_model_must_be_recorded_by_the_call_receipt",
         )

tests/test_enforce_leoclean_learning_policy.py

@@ -0,0 +1,76 @@
+from __future__ import annotations
+
+import importlib.util
+import json
+from pathlib import Path
+
+import yaml
+
+ROOT = Path(__file__).resolve().parents[1]
+PATCH_PATH = ROOT / "hermes-agent" / "patches" / "enforce_leoclean_learning_policy.py"
+
+
+def _load_module():
+    spec = importlib.util.spec_from_file_location("enforce_leoclean_learning_policy", PATCH_PATH)
+    assert spec and spec.loader
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def test_learning_policy_disables_automatic_reviews_without_removing_explicit_tools(tmp_path: Path) -> None:
+    module = _load_module()
+    config_path = tmp_path / "config.yaml"
+    original = """model:
+  provider: openrouter
+gateway:
+  telegram:
+    bot_token: private-test-token
+memory:
+  memory_enabled: true
+  user_profile_enabled: true
+  nudge_interval: 10
+skills:
+  external_dirs: []
+  creation_nudge_interval: 10
+"""
+    config_path.write_text(original, encoding="utf-8")
+    config_path.chmod(0o600)
+
+    result, code = module.enforce_policy(config_path)
+
+    assert code == 0
+    assert result["status"] == "updated"
+    assert result["policy"] == {
+        "memory_nudge_interval": 0,
+        "skill_creation_nudge_interval": 0,
+    }
+    assert "private-test-token" not in json.dumps(result)
+    updated = yaml.safe_load(config_path.read_text(encoding="utf-8"))
+    assert updated["memory"] == {
+        "memory_enabled": True,
+        "user_profile_enabled": True,
+        "nudge_interval": 0,
+    }
+    assert updated["skills"] == {"external_dirs": [], "creation_nudge_interval": 0}
+    assert updated["gateway"]["telegram"]["bot_token"] == "private-test-token"
+    assert config_path.stat().st_mode & 0o777 == 0o600
+    assert config_path.with_name(config_path.name + module.BACKUP_SUFFIX).read_text(encoding="utf-8") == original
+
+    check_result, check_code = module.enforce_policy(config_path, check=True)
+    assert check_code == 0
+    assert check_result["status"] == "compliant"
+
+
+def test_learning_policy_check_reports_drift_without_writing(tmp_path: Path) -> None:
+    module = _load_module()
+    config_path = tmp_path / "config.yaml"
+    original = "memory:\n  memory_enabled: true\n"
+    config_path.write_text(original, encoding="utf-8")
+
+    result, code = module.enforce_policy(config_path, check=True)
+
+    assert code == 1
+    assert result["status"] == "change_required"
+    assert config_path.read_text(encoding="utf-8") == original
+    assert not config_path.with_name(config_path.name + module.BACKUP_SUFFIX).exists()

tests/test_hermes_leoclean_kb_bridge_source.py

@@ -1669,6 +1669,39 @@ def test_vps_bridge_binds_claim_challenge_to_exact_retrieved_rows() -> None:
     assert "no database write was made" in unbound_response
 
 
+def test_vps_bridge_does_not_replace_candidate_composition_with_retrieved_claim() -> None:
+    module = _load_module(BRIDGE_DIR / "kb_tool.py")
+    candidate_queries = [
+        (
+            "Draft the strongest reviewable candidate claim this discussion supports. Give the exact claim body, "
+            "initial confidence, falsifier, minimum source and evidence requirements, and relations to existing "
+            "canonical rows. Do not write or stage it yet."
+        ),
+        (
+            "That candidate still smuggles in unsupported assumptions. Narrow the claim. Return the revised exact "
+            "claim body, revised initial confidence, a concrete falsifier, minimum source and evidence requirements, "
+            "and relations to existing rows. Do not write or stage anything."
+        ),
+    ]
+
+    for query in candidate_queries:
+        contracts = module.operational_contracts(query)
+        contract_ids = {contract["id"] for contract in contracts}
+
+        assert "claim_evidence_challenge" not in contract_ids
+        assert module.compile_operational_response(contracts) is None
+
+
+def test_vps_bridge_keeps_exact_row_guard_for_combined_audit_and_candidate_request() -> None:
+    module = _load_module(BRIDGE_DIR / "kb_tool.py")
+    contracts = module.operational_contracts(
+        "Inspect one exact claim body and its evidence, challenge its unsupported assumption, and then draft a "
+        "narrower candidate claim."
+    )
+
+    assert "claim_evidence_challenge" in {contract["id"] for contract in contracts}
+
+
 def test_vps_prepare_source_retrieves_canonical_candidates_before_model_extraction(
     tmp_path: Path, monkeypatch, capsys
 ) -> None:

tests/test_leo_behavior_manifest.py

@@ -30,6 +30,9 @@ def test_behavior_manifest_is_deterministic_and_never_emits_config_secrets(tmp_p
 memory:
   enabled: true
   search: sqlite_fts5
+  nudge_interval: 0
+skills:
+  creation_nudge_interval: 0
 gateway:
   telegram:
     bot_token: super-secret-token
@@ -52,6 +55,8 @@ gateway:
 
     assert first["behavior_sha256"] == second["behavior_sha256"]
     assert first["model_runtime"]["default_model"] == "anthropic/claude-sonnet-4-6"
+    assert first["model_runtime"]["automatic_memory_review_interval"] == 0
+    assert first["model_runtime"]["automatic_skill_review_interval"] == 0
     assert first["teleo_infrastructure_runtime"]["source_tree"]["file_count"] == 2
     assert first["components"]["persistent_memory"]["content"]["file_count"] == 1
     encoded = json.dumps(first)

tests/test_teleo_agent_systemd.py

@@ -128,6 +128,10 @@ def test_deploy_scripts_sync_leoclean_skills_and_restart_gateway_for_runtime_cha
     assert "hermes-agent/leoclean-skills/vps/" in auto_deploy
     assert "hermes-agent/leoclean-plugins/vps/" in auto_deploy
     assert 'HERMES_PATCH_DIR="/home/teleo/.hermes/teleo-runtime-patches"' in auto_deploy
+    assert 'LEOCLEAN_CONFIG="/home/teleo/.hermes/profiles/leoclean/config.yaml"' in auto_deploy
+    assert 'LEOCLEAN_POLICY_PATCH="$HERMES_PATCH_DIR/enforce_leoclean_learning_policy.py"' in auto_deploy
+    assert 'python3 "$LEOCLEAN_POLICY_PATCH" "$LEOCLEAN_CONFIG"' in auto_deploy
+    assert "Leoclean learning-policy drift repaired" in auto_deploy
     assert '"status": "installed_now"' in auto_deploy
     assert "Hermes response-transform drift repaired" in auto_deploy
     assert "deploy/leoclean-gateway-restart-required.sh" in auto_deploy
@@ -145,6 +149,8 @@ def test_deploy_scripts_sync_leoclean_skills_and_restart_gateway_for_runtime_cha
     assert "hermes-agent/leoclean-skills/vps/" in manual_deploy
     assert "hermes-agent/leoclean-plugins/vps/" in manual_deploy
     assert 'VPS_HERMES_PATCHES="/home/teleo/.hermes/teleo-runtime-patches"' in manual_deploy
+    assert 'VPS_LEOCLEAN_POLICY_PATCH="$VPS_HERMES_PATCHES/enforce_leoclean_learning_policy.py"' in manual_deploy
+    assert "Leoclean gateway restarted." in manual_deploy
     assert "systemctl is-active --quiet leoclean-gateway.service" in manual_deploy