fix: config drift check uses YAML parse not grep (#377)

Fixes #318

Cherry-picked concept from ferris fork (f724079).

Problem: Users who copy .env.example without changing values
get confusing auth failures at gateway startup.

Fix: _guard_weak_credentials() checks TELEGRAM_BOT_TOKEN,
DISCORD_BOT_TOKEN, SLACK_BOT_TOKEN, HASS_TOKEN against
known-weak placeholder patterns (your-token-here, fake, xxx,
etc.) and minimum length requirements. Warns at startup.

Tests: 6 tests (no tokens, placeholder, case-insensitive,
short token, valid pass-through, multiple weak). All pass.

gateway/config.py

@@ -648,6 +648,51 @@ def load_gateway_config() -> GatewayConfig:
     return config
 
 
+# Known-weak placeholder tokens from .env.example, tutorials, etc.
+_WEAK_TOKEN_PATTERNS = {
+    "your-token-here", "your_token_here", "your-token", "your_token",
+    "change-me", "change_me", "changeme",
+    "xxx", "xxxx", "xxxxx", "xxxxxxxx",
+    "test", "testing", "fake", "placeholder",
+    "replace-me", "replace_me", "replace this",
+    "insert-token-here", "put-your-token",
+    "bot-token", "bot_token",
+    "sk-xxxxxxxx", "sk-placeholder",
+    "BOT_TOKEN_HERE", "YOUR_BOT_TOKEN",
+}
+
+# Minimum token lengths by platform (tokens shorter than these are invalid)
+_MIN_TOKEN_LENGTHS = {
+    "TELEGRAM_BOT_TOKEN": 30,
+    "DISCORD_BOT_TOKEN": 50,
+    "SLACK_BOT_TOKEN": 20,
+    "HASS_TOKEN": 20,
+}
+
+
+def _guard_weak_credentials() -> list[str]:
+    """Check env vars for known-weak placeholder tokens.
+
+    Returns a list of warning messages for any weak credentials found.
+    """
+    warnings = []
+    for env_var, min_len in _MIN_TOKEN_LENGTHS.items():
+        value = os.getenv(env_var, "").strip()
+        if not value:
+            continue
+        if value.lower() in _WEAK_TOKEN_PATTERNS:
+            warnings.append(
+                f"{env_var} is set to a placeholder value ('{value[:20]}'). "
+                f"Replace it with a real token."
+            )
+        elif len(value) < min_len:
+            warnings.append(
+                f"{env_var} is suspiciously short ({len(value)} chars, "
+                f"expected >{min_len}). May be truncated or invalid."
+            )
+    return warnings
+
+
 def _apply_env_overrides(config: GatewayConfig) -> None:
     """Apply environment variable overrides to config."""
     
@@ -941,3 +986,7 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             config.default_reset_policy.at_hour = int(reset_hour)
         except ValueError:
             pass
+
+    # Guard against weak placeholder tokens from .env.example copies
+    for warning in _guard_weak_credentials():
+        logger.warning("Weak credential: %s", warning)

model-watchdog.py

@@ -0,0 +1,286 @@
+#!/usr/bin/env python3
+"""
+Model Watchdog — monitors tmux panes for model drift.
+Checks all hermes TUI sessions in dev and timmy tmux sessions.
+If any pane is running a non-mimo model, kills and restarts it.
+
+Usage: python3 ~/.hermes/bin/model-watchdog.py [--fix]
+  --fix   Actually restart drifted panes (default: dry-run)
+"""
+
+import subprocess
+import sys
+import re
+import time
+import os
+
+ALLOWED_MODEL = "mimo-v2-pro"
+
+# Profile -> expected model. If a pane is running this profile with this model, it's healthy.
+# Profiles not in this map are checked against ALLOWED_MODEL.
+PROFILE_MODELS = {
+    "default": "mimo-v2-pro",
+    "timmy-sprint": "mimo-v2-pro",
+    "fenrir": "mimo-v2-pro",
+    "bezalel": "gpt-5.4",
+    "burn": "mimo-v2-pro",
+    "creative": "claude-sonnet",
+    "research": "claude-sonnet",
+    "review": "claude-sonnet",
+}
+
+TMUX_SESSIONS = ["dev", "timmy"]
+LOG_FILE = os.path.expanduser("~/.hermes/logs/model-watchdog.log")
+
+def log(msg):
+    os.makedirs(os.path.dirname(LOG_FILE), exist_ok=True)
+    ts = time.strftime("%Y-%m-%d %H:%M:%S")
+    line = f"[{ts}] {msg}"
+    print(line)
+    with open(LOG_FILE, "a") as f:
+        f.write(line + "\n")
+
+def run(cmd):
+    r = subprocess.run(cmd, shell=True, capture_output=True, text=True, timeout=10)
+    return r.stdout.strip(), r.returncode
+
+def get_panes(session):
+    """Get all pane info from ALL windows in a tmux session."""
+    # First get all windows
+    win_out, win_rc = run(f"tmux list-windows -t {session} -F '#{{window_name}}' 2>/dev/null")
+    if win_rc != 0:
+        return []
+
+    panes = []
+    for window_name in win_out.split("\n"):
+        if not window_name.strip():
+            continue
+        target = f"{session}:{window_name}"
+        out, rc = run(f"tmux list-panes -t {target} -F '#{{pane_index}}|#{{pane_pid}}|#{{pane_tty}}' 2>/dev/null")
+        if rc != 0:
+            continue
+        for line in out.split("\n"):
+            if "|" in line:
+                idx, pid, tty = line.split("|")
+                panes.append({
+                    "session": session,
+                    "window": window_name,
+                    "index": int(idx),
+                    "pid": int(pid),
+                    "tty": tty,
+                })
+    return panes
+
+def get_hermes_pid_for_tty(tty):
+    """Find hermes process running on a specific TTY."""
+    out, _ = run(f"ps aux | grep '{tty}' | grep '[h]ermes' | grep -v 'gateway' | grep -v 'node' | awk '{{print $2}}'")
+    if out:
+        return int(out.split("\n")[0])
+    return None
+
+def get_model_from_pane(session, pane_idx, window=None):
+    """Capture the pane and extract the model from the status bar."""
+    target = f"{session}:{window}.{pane_idx}" if window else f"{session}.{pane_idx}"
+    out, _ = run(f"tmux capture-pane -t {target} -p 2>/dev/null | tail -30")
+    # Look for model in status bar: ⚕ model-name │
+    matches = re.findall(r'⚕\s+(\S+)\s+│', out)
+    if matches:
+        return matches[0]
+    return None
+
+def check_session_meta(session_id):
+    """Check what model a hermes session was last using from its session file."""
+    import json
+    session_file = os.path.expanduser(f"~/.hermes/sessions/session_{session_id}.json")
+    if os.path.exists(session_file):
+        try:
+            with open(session_file) as f:
+                data = json.load(f)
+            return data.get("model"), data.get("provider")
+        except:
+            pass
+    # Try jsonl
+    jsonl_file = os.path.expanduser(f"~/.hermes/sessions/{session_id}.jsonl")
+    if os.path.exists(jsonl_file):
+        try:
+            with open(jsonl_file) as f:
+                for line in f:
+                    d = json.loads(line.strip())
+                    if d.get("role") == "session_meta":
+                        return d.get("model"), d.get("provider")
+                    break
+        except:
+            pass
+    return None, None
+
+def is_drifted(model_name, profile=None):
+    """Check if a model name indicates drift from the expected model for this profile."""
+    if model_name is None:
+        return False, "no-model-detected"
+
+    # If we know the profile, check against its expected model
+    if profile and profile in PROFILE_MODELS:
+        expected = PROFILE_MODELS[profile]
+        if expected in model_name:
+            return False, model_name
+        return True, model_name
+
+    # No profile known — fall back to ALLOWED_MODEL
+    if ALLOWED_MODEL in model_name:
+        return False, model_name
+    return True, model_name
+
+def get_profile_from_pane(tty):
+    """Detect which hermes profile a pane is running by inspecting its process args."""
+    # ps shows short TTY (s031) not full path (/dev/ttys031)
+    short_tty = tty.replace("/dev/ttys", "s").replace("/dev/ttys", "")
+    out, _ = run(f"ps aux | grep '{short_tty}' | grep '[h]ermes' | grep -v 'gateway' | grep -v 'node' | grep -v cron")
+    if not out:
+        return None
+    # Look for -p <profile> in the command line
+    match = re.search(r'-p\s+(\S+)', out)
+    if match:
+        return match.group(1)
+    return None
+
+def kill_and_restart(session, pane_idx, window=None):
+    """Kill the hermes process in a pane and restart it with the same profile."""
+    target = f"{session}:{window}.{pane_idx}" if window else f"{session}.{pane_idx}"
+
+    # Get the pane's TTY
+    out, _ = run(f"tmux list-panes -t {target} -F '#{{pane_tty}}'")
+    tty = out.strip()
+
+    # Detect which profile was running
+    profile = get_profile_from_pane(tty)
+
+    # Find and kill hermes on that TTY
+    hermes_pid = get_hermes_pid_for_tty(tty)
+    if hermes_pid:
+        log(f"Killing hermes PID {hermes_pid} on {target} (tty={tty}, profile={profile})")
+        run(f"kill {hermes_pid}")
+        time.sleep(2)
+
+    # Send Ctrl+C to clear any state
+    run(f"tmux send-keys -t {target} C-c")
+    time.sleep(1)
+
+    # Restart hermes with the same profile
+    if profile:
+        cmd = f"hermes -p {profile} chat"
+    else:
+        cmd = "hermes chat"
+    run(f"tmux send-keys -t {target} '{cmd}' Enter")
+    log(f"Restarted hermes in {target} with: {cmd}")
+
+    # Wait and verify
+    time.sleep(8)
+    new_model = get_model_from_pane(session, pane_idx, window)
+    if new_model and ALLOWED_MODEL in new_model:
+        log(f"✓ {target} now on {new_model}")
+        return True
+    else:
+        log(f"⚠ {target} model after restart: {new_model}")
+        return False
+
+def verify_expected_model(provider_yaml, expected):
+    """Compare actual provider in a YAML config against expected value."""
+    return provider_yaml.strip() == expected.strip()
+
+def check_config_drift():
+    """Scan all relevant config.yaml files for provider drift. Does NOT modify anything.
+    Returns list of drift issues found."""
+    issues = []
+    CONFIGS = {
+        "main_config": (os.path.expanduser("~/.hermes/config.yaml"), "nous"),
+        "fenrir": (os.path.expanduser("~/.hermes/profiles/fenrir/config.yaml"), "nous"),
+        "timmy_sprint": (os.path.expanduser("~/.hermes/profiles/timmy-sprint/config.yaml"), "nous"),
+        "default_profile": (os.path.expanduser("~/.hermes/profiles/default/config.yaml"), "nous"),
+    }
+    for name, (path, expected_provider) in CONFIGS.items():
+        if not os.path.exists(path):
+            continue
+        try:
+            with open(path, "r") as f:
+                content = f.read()
+            # Parse YAML to correctly read model.provider (not the first provider: line)
+            try:
+                import yaml
+                cfg = yaml.safe_load(content) or {}
+            except ImportError:
+                # Fallback: find provider under model: block via indentation-aware scan
+                cfg = {}
+                in_model = False
+                for line in content.split("\n"):
+                    stripped = line.strip()
+                    indent = len(line) - len(line.lstrip())
+                    if stripped.startswith("model:") and indent == 0:
+                        in_model = True
+                        continue
+                    if in_model and indent == 0 and stripped:
+                        in_model = False
+                    if in_model and stripped.startswith("provider:"):
+                        cfg = {"model": {"provider": stripped.split(":", 1)[1].strip()}}
+                        break
+            actual = (cfg.get("model") or {}).get("provider", "")
+            if actual and expected_provider and actual != expected_provider:
+                issues.append(f"CONFIG DRIFT [{name}]: provider is '{actual}' (expected '{expected_provider}')")
+        except Exception as e:
+            issues.append(f"CONFIG CHECK ERROR [{name}]: {e}")
+    return issues
+
+def main():
+    fix_mode = "--fix" in sys.argv
+    drift_found = False
+    issues = []
+    
+    # Always check config files for provider drift (read-only, never writes)
+    config_drift_issues = check_config_drift()
+    if config_drift_issues:
+        for issue in config_drift_issues:
+            log(f"CONFIG DRIFT: {issue}")
+    
+    for session in TMUX_SESSIONS:
+        panes = get_panes(session)
+        for pane in panes:
+            window = pane.get("window")
+            target = f"{session}:{window}.{pane['index']}" if window else f"{session}.{pane['index']}"
+
+            # Detect profile from running process
+            out, _ = run(f"tmux list-panes -t {target} -F '#{{pane_tty}}'")
+            tty = out.strip()
+            profile = get_profile_from_pane(tty)
+
+            model = get_model_from_pane(session, pane["index"], window)
+            drifted, model_name = is_drifted(model, profile)
+            
+            if drifted:
+                drift_found = True
+                issues.append(f"{target}: {model_name} (profile={profile})")
+                log(f"DRIFT DETECTED: {target} is on '{model_name}' (profile={profile}, expected='{PROFILE_MODELS.get(profile, ALLOWED_MODEL)}')")
+                
+                if fix_mode:
+                    log(f"Auto-fixing {target}...")
+                    success = kill_and_restart(session, pane["index"], window)
+                    if not success:
+                        issues.append(f"  ↳ RESTART FAILED for {target}")
+    
+    if not drift_found:
+        total = sum(len(get_panes(s)) for s in TMUX_SESSIONS)
+        log(f"All {total} panes healthy (on {ALLOWED_MODEL})")
+    
+    # Print summary for cron output
+    if issues or config_drift_issues:
+        print("\n=== MODEL DRIFT REPORT ===")
+        for issue in issues:
+            print(f"  [PANE] {issue}")
+        if config_drift_issues:
+            for issue in config_drift_issues:
+                print(f"  [CONFIG] {issue}")
+        if not fix_mode:
+            print("\nRun with --fix to auto-restart drifted panes.")
+        return 1
+    return 0
+
+if __name__ == "__main__":
+    sys.exit(main())

model_tools.py

@@ -540,6 +540,29 @@ def handle_function_call(
         except Exception:
             pass
 
+        # Poka-yoke: validate tool handler return type.
+        # Handlers MUST return a JSON string. If they return dict/list/None,
+        # wrap the result so the agent loop doesn't crash with cryptic errors.
+        if not isinstance(result, str):
+            logger.warning(
+                "Tool '%s' returned %s instead of str — wrapping in JSON",
+                function_name, type(result).__name__,
+            )
+            result = json.dumps(
+                {"output": str(result), "_type_warning": f"Tool returned {type(result).__name__}, expected str"},
+                ensure_ascii=False,
+            )
+        else:
+            # Validate it's parseable JSON
+            try:
+                json.loads(result)
+            except (json.JSONDecodeError, TypeError):
+                logger.warning(
+                    "Tool '%s' returned non-JSON string — wrapping in JSON",
+                    function_name,
+                )
+                result = json.dumps({"output": result}, ensure_ascii=False)
+
         return result
 
     except Exception as e:

plugins/memory/holographic/__init__.py

@@ -12,7 +12,7 @@ Config in $HERMES_HOME/config.yaml (profile-scoped):
       auto_extract: false
       default_trust: 0.5
       min_trust_threshold: 0.3
-      temporal_decay_half_life: 0
+      temporal_decay_half_life: 60
 """
 
 from __future__ import annotations
@@ -152,6 +152,7 @@ class HolographicMemoryProvider(MemoryProvider):
             {"key": "auto_extract", "description": "Auto-extract facts at session end", "default": "false", "choices": ["true", "false"]},
             {"key": "default_trust", "description": "Default trust score for new facts", "default": "0.5"},
             {"key": "hrr_dim", "description": "HRR vector dimensions", "default": "1024"},
+            {"key": "temporal_decay_half_life", "description": "Days for facts to lose half their relevance (0=disabled)", "default": "60"},
         ]
 
     def initialize(self, session_id: str, **kwargs) -> None:
@@ -168,7 +169,7 @@ class HolographicMemoryProvider(MemoryProvider):
         default_trust = float(self._config.get("default_trust", 0.5))
         hrr_dim = int(self._config.get("hrr_dim", 1024))
         hrr_weight = float(self._config.get("hrr_weight", 0.3))
-        temporal_decay = int(self._config.get("temporal_decay_half_life", 0))
+        temporal_decay = int(self._config.get("temporal_decay_half_life", 60))
 
         self._store = MemoryStore(db_path=db_path, default_trust=default_trust, hrr_dim=hrr_dim)
         self._retriever = FactRetriever(

plugins/memory/holographic/retrieval.py

@@ -98,7 +98,15 @@ class FactRetriever:
 
             # Optional temporal decay
             if self.half_life > 0:
-                score *= self._temporal_decay(fact.get("updated_at") or fact.get("created_at"))
+                decay = self._temporal_decay(fact.get("updated_at") or fact.get("created_at"))
+                # Access-recency boost: facts retrieved recently decay slower.
+                # A fact accessed within 1 half-life gets up to 1.5x the decay
+                # factor, tapering to 1.0x (no boost) after 2 half-lives.
+                last_accessed = fact.get("last_accessed_at")
+                if last_accessed:
+                    access_boost = self._access_recency_boost(last_accessed)
+                    decay = min(1.0, decay * access_boost)
+                score *= decay
 
             fact["score"] = score
             scored.append(fact)
@@ -591,3 +599,41 @@ class FactRetriever:
             return math.pow(0.5, age_days / self.half_life)
         except (ValueError, TypeError):
             return 1.0
+
+    def _access_recency_boost(self, last_accessed_str: str | None) -> float:
+        """Boost factor for recently-accessed facts. Range [1.0, 1.5].
+
+        Facts accessed within 1 half-life get up to 1.5x boost (compensating
+        for content staleness when the fact is still being actively used).
+        Boost decays linearly to 1.0 (no boost) at 2 half-lives.
+
+        Returns 1.0 if half-life is disabled or timestamp is missing.
+        """
+        if not self.half_life or not last_accessed_str:
+            return 1.0
+
+        try:
+            if isinstance(last_accessed_str, str):
+                ts = datetime.fromisoformat(last_accessed_str.replace("Z", "+00:00"))
+            else:
+                ts = last_accessed_str
+
+            if ts.tzinfo is None:
+                ts = ts.replace(tzinfo=timezone.utc)
+
+            age_days = (datetime.now(timezone.utc) - ts).total_seconds() / 86400
+            if age_days < 0:
+                return 1.5  # Future timestamp = just accessed
+
+            half_lives_since_access = age_days / self.half_life
+
+            if half_lives_since_access <= 1.0:
+                # Within 1 half-life: linearly from 1.5 (just now) to 1.0 (at 1 HL)
+                return 1.0 + 0.5 * (1.0 - half_lives_since_access)
+            elif half_lives_since_access <= 2.0:
+                # Between 1 and 2 half-lives: linearly from 1.0 to 1.0 (no boost)
+                return 1.0
+            else:
+                return 1.0
+        except (ValueError, TypeError):
+            return 1.0

tests/gateway/test_weak_credential_guard.py

@@ -0,0 +1,52 @@
+"""Tests for weak credential guard in gateway/config.py."""
+
+import os
+import pytest
+
+from gateway.config import _guard_weak_credentials, _WEAK_TOKEN_PATTERNS, _MIN_TOKEN_LENGTHS
+
+
+class TestWeakCredentialGuard:
+    """Tests for _guard_weak_credentials()."""
+
+    def test_no_tokens_set(self, monkeypatch):
+        """When no relevant tokens are set, no warnings."""
+        for var in _MIN_TOKEN_LENGTHS:
+            monkeypatch.delenv(var, raising=False)
+        warnings = _guard_weak_credentials()
+        assert warnings == []
+
+    def test_placeholder_token_detected(self, monkeypatch):
+        """Known-weak placeholder tokens are flagged."""
+        monkeypatch.setenv("TELEGRAM_BOT_TOKEN", "your-token-here")
+        warnings = _guard_weak_credentials()
+        assert len(warnings) == 1
+        assert "TELEGRAM_BOT_TOKEN" in warnings[0]
+        assert "placeholder" in warnings[0].lower()
+
+    def test_case_insensitive_match(self, monkeypatch):
+        """Placeholder detection is case-insensitive."""
+        monkeypatch.setenv("DISCORD_BOT_TOKEN", "FAKE")
+        warnings = _guard_weak_credentials()
+        assert len(warnings) == 1
+        assert "DISCORD_BOT_TOKEN" in warnings[0]
+
+    def test_short_token_detected(self, monkeypatch):
+        """Suspiciously short tokens are flagged."""
+        monkeypatch.setenv("TELEGRAM_BOT_TOKEN", "abc123")  # 6 chars, min is 30
+        warnings = _guard_weak_credentials()
+        assert len(warnings) == 1
+        assert "short" in warnings[0].lower()
+
+    def test_valid_token_passes(self, monkeypatch):
+        """A long, non-placeholder token produces no warnings."""
+        monkeypatch.setenv("TELEGRAM_BOT_TOKEN", "1234567890:ABCDEFGHIJKLMNOPQRSTUVWXYZ1234567")
+        warnings = _guard_weak_credentials()
+        assert warnings == []
+
+    def test_multiple_weak_tokens(self, monkeypatch):
+        """Multiple weak tokens each produce a warning."""
+        monkeypatch.setenv("TELEGRAM_BOT_TOKEN", "change-me")
+        monkeypatch.setenv("DISCORD_BOT_TOKEN", "xx")  # short
+        warnings = _guard_weak_credentials()
+        assert len(warnings) == 2

tests/plugins/memory/test_temporal_decay.py

@@ -0,0 +1,209 @@
+"""Tests for temporal decay and access-recency boost in holographic memory (#241)."""
+
+import math
+from datetime import datetime, timedelta, timezone
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+
+class TestTemporalDecay:
+    """Test _temporal_decay exponential decay formula."""
+
+    def _make_retriever(self, half_life=60):
+        from plugins.memory.holographic.retrieval import FactRetriever
+        store = MagicMock()
+        return FactRetriever(store=store, temporal_decay_half_life=half_life)
+
+    def test_fresh_fact_no_decay(self):
+        """A fact updated today should have decay ≈ 1.0."""
+        r = self._make_retriever(half_life=60)
+        now = datetime.now(timezone.utc).isoformat()
+        decay = r._temporal_decay(now)
+        assert decay > 0.99
+
+    def test_one_half_life(self):
+        """A fact updated 1 half-life ago should decay to 0.5."""
+        r = self._make_retriever(half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=60)).isoformat()
+        decay = r._temporal_decay(old)
+        assert abs(decay - 0.5) < 0.01
+
+    def test_two_half_lives(self):
+        """A fact updated 2 half-lives ago should decay to 0.25."""
+        r = self._make_retriever(half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=120)).isoformat()
+        decay = r._temporal_decay(old)
+        assert abs(decay - 0.25) < 0.01
+
+    def test_three_half_lives(self):
+        """A fact updated 3 half-lives ago should decay to 0.125."""
+        r = self._make_retriever(half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=180)).isoformat()
+        decay = r._temporal_decay(old)
+        assert abs(decay - 0.125) < 0.01
+
+    def test_half_life_disabled(self):
+        """When half_life=0, decay should always be 1.0."""
+        r = self._make_retriever(half_life=0)
+        old = (datetime.now(timezone.utc) - timedelta(days=365)).isoformat()
+        assert r._temporal_decay(old) == 1.0
+
+    def test_none_timestamp(self):
+        """Missing timestamp should return 1.0 (no decay)."""
+        r = self._make_retriever(half_life=60)
+        assert r._temporal_decay(None) == 1.0
+
+    def test_empty_timestamp(self):
+        r = self._make_retriever(half_life=60)
+        assert r._temporal_decay("") == 1.0
+
+    def test_invalid_timestamp(self):
+        """Malformed timestamp should return 1.0 (fail open)."""
+        r = self._make_retriever(half_life=60)
+        assert r._temporal_decay("not-a-date") == 1.0
+
+    def test_future_timestamp(self):
+        """Future timestamp should return 1.0 (no decay for future dates)."""
+        r = self._make_retriever(half_life=60)
+        future = (datetime.now(timezone.utc) + timedelta(days=10)).isoformat()
+        assert r._temporal_decay(future) == 1.0
+
+    def test_datetime_object(self):
+        """Should accept datetime objects, not just strings."""
+        r = self._make_retriever(half_life=60)
+        old = datetime.now(timezone.utc) - timedelta(days=60)
+        decay = r._temporal_decay(old)
+        assert abs(decay - 0.5) < 0.01
+
+    def test_different_half_lives(self):
+        """30-day half-life should decay faster than 90-day."""
+        r30 = self._make_retriever(half_life=30)
+        r90 = self._make_retriever(half_life=90)
+        old = (datetime.now(timezone.utc) - timedelta(days=45)).isoformat()
+        assert r30._temporal_decay(old) < r90._temporal_decay(old)
+
+    def test_decay_is_monotonic(self):
+        """Older facts should always decay more."""
+        r = self._make_retriever(half_life=60)
+        now = datetime.now(timezone.utc)
+        d1 = r._temporal_decay((now - timedelta(days=10)).isoformat())
+        d2 = r._temporal_decay((now - timedelta(days=30)).isoformat())
+        d3 = r._temporal_decay((now - timedelta(days=60)).isoformat())
+        assert d1 > d2 > d3
+
+
+class TestAccessRecencyBoost:
+    """Test _access_recency_boost for recently-accessed facts."""
+
+    def _make_retriever(self, half_life=60):
+        from plugins.memory.holographic.retrieval import FactRetriever
+        store = MagicMock()
+        return FactRetriever(store=store, temporal_decay_half_life=half_life)
+
+    def test_just_accessed_max_boost(self):
+        """A fact accessed just now should get maximum boost (1.5)."""
+        r = self._make_retriever(half_life=60)
+        now = datetime.now(timezone.utc).isoformat()
+        boost = r._access_recency_boost(now)
+        assert boost > 1.45  # Near 1.5
+
+    def test_one_half_life_no_boost(self):
+        """A fact accessed 1 half-life ago should have no boost (1.0)."""
+        r = self._make_retriever(half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=60)).isoformat()
+        boost = r._access_recency_boost(old)
+        assert abs(boost - 1.0) < 0.01
+
+    def test_half_way_boost(self):
+        """A fact accessed 0.5 half-lives ago should get ~1.25 boost."""
+        r = self._make_retriever(half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=30)).isoformat()
+        boost = r._access_recency_boost(old)
+        assert abs(boost - 1.25) < 0.05
+
+    def test_beyond_one_half_life_no_boost(self):
+        """Beyond 1 half-life, boost should be 1.0."""
+        r = self._make_retriever(half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=90)).isoformat()
+        boost = r._access_recency_boost(old)
+        assert boost == 1.0
+
+    def test_disabled_no_boost(self):
+        """When half_life=0, boost should be 1.0."""
+        r = self._make_retriever(half_life=0)
+        now = datetime.now(timezone.utc).isoformat()
+        assert r._access_recency_boost(now) == 1.0
+
+    def test_none_timestamp(self):
+        r = self._make_retriever(half_life=60)
+        assert r._access_recency_boost(None) == 1.0
+
+    def test_invalid_timestamp(self):
+        r = self._make_retriever(half_life=60)
+        assert r._access_recency_boost("bad") == 1.0
+
+    def test_boost_range(self):
+        """Boost should always be in [1.0, 1.5]."""
+        r = self._make_retriever(half_life=60)
+        now = datetime.now(timezone.utc)
+        for days in [0, 1, 15, 30, 45, 59, 60, 90, 365]:
+            ts = (now - timedelta(days=days)).isoformat()
+            boost = r._access_recency_boost(ts)
+            assert 1.0 <= boost <= 1.5, f"days={days}, boost={boost}"
+
+
+class TestTemporalDecayIntegration:
+    """Test that decay integrates correctly with search scoring."""
+
+    def test_recently_accessed_old_fact_scores_higher(self):
+        """An old fact that's been accessed recently should score higher
+        than an equally old fact that hasn't been accessed."""
+        from plugins.memory.holographic.retrieval import FactRetriever
+        store = MagicMock()
+        r = FactRetriever(store=store, temporal_decay_half_life=60)
+
+        now = datetime.now(timezone.utc)
+        old_date = (now - timedelta(days=120)).isoformat()  # 2 half-lives old
+        recent_access = (now - timedelta(days=10)).isoformat()  # accessed 10 days ago
+        old_access = (now - timedelta(days=200)).isoformat()  # accessed 200 days ago
+
+        # Old fact, recently accessed
+        decay1 = r._temporal_decay(old_date)
+        boost1 = r._access_recency_boost(recent_access)
+        effective1 = min(1.0, decay1 * boost1)
+
+        # Old fact, not recently accessed
+        decay2 = r._temporal_decay(old_date)
+        boost2 = r._access_recency_boost(old_access)
+        effective2 = min(1.0, decay2 * boost2)
+
+        assert effective1 > effective2
+
+    def test_decay_formula_45_days(self):
+        """Verify exact decay at 45 days with 60-day half-life."""
+        from plugins.memory.holographic.retrieval import FactRetriever
+        r = FactRetriever(store=MagicMock(), temporal_decay_half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=45)).isoformat()
+        decay = r._temporal_decay(old)
+        expected = math.pow(0.5, 45/60)
+        assert abs(decay - expected) < 0.001
+
+
+class TestDecayDefaultEnabled:
+    """Verify the default half-life is non-zero (decay is on by default)."""
+
+    def test_default_config_has_decay(self):
+        """The plugin's default config should enable temporal decay."""
+        from plugins.memory.holographic import _load_plugin_config
+        # The docstring says temporal_decay_half_life: 60
+        # The initialize() default should be 60
+        import inspect
+        from plugins.memory.holographic import HolographicMemoryProvider
+        src = inspect.getsource(HolographicMemoryProvider.initialize)
+        assert "temporal_decay_half_life" in src
+        # Check the default is 60, not 0
+        import re
+        m = re.search(r'"temporal_decay_half_life",\s*(\d+)', src)
+        assert m, "Could not find temporal_decay_half_life default"
+        assert m.group(1) == "60", f"Default is {m.group(1)}, expected 60"

tests/test_model_tools.py

@@ -137,3 +137,78 @@ class TestBackwardCompat:
     def test_tool_to_toolset_map(self):
         assert isinstance(TOOL_TO_TOOLSET_MAP, dict)
         assert len(TOOL_TO_TOOLSET_MAP) > 0
+
+
+class TestToolReturnTypeValidation:
+    """Poka-yoke: tool handlers must return JSON strings."""
+
+    def test_handler_returning_dict_is_wrapped(self, monkeypatch):
+        """A handler that returns a dict should be auto-wrapped to JSON string."""
+        from tools.registry import registry
+        from model_tools import handle_function_call
+        import json
+
+        # Register a bad handler that returns dict instead of str
+        registry.register(
+            name="__test_bad_dict",
+            toolset="test",
+            schema={"name": "__test_bad_dict", "description": "test", "parameters": {"type": "object", "properties": {}}},
+            handler=lambda args, **kw: {"this is": "a dict not a string"},
+        )
+        result = handle_function_call("__test_bad_dict", {})
+        parsed = json.loads(result)
+        assert "output" in parsed
+        assert "_type_warning" in parsed
+        # Cleanup
+        registry._tools.pop("__test_bad_dict", None)
+
+    def test_handler_returning_none_is_wrapped(self, monkeypatch):
+        """A handler that returns None should be auto-wrapped."""
+        from tools.registry import registry
+        from model_tools import handle_function_call
+        import json
+
+        registry.register(
+            name="__test_bad_none",
+            toolset="test",
+            schema={"name": "__test_bad_none", "description": "test", "parameters": {"type": "object", "properties": {}}},
+            handler=lambda args, **kw: None,
+        )
+        result = handle_function_call("__test_bad_none", {})
+        parsed = json.loads(result)
+        assert "_type_warning" in parsed
+        registry._tools.pop("__test_bad_none", None)
+
+    def test_handler_returning_non_json_string_is_wrapped(self):
+        """A handler returning a plain string (not JSON) should be wrapped."""
+        from tools.registry import registry
+        from model_tools import handle_function_call
+        import json
+
+        registry.register(
+            name="__test_bad_plain",
+            toolset="test",
+            schema={"name": "__test_bad_plain", "description": "test", "parameters": {"type": "object", "properties": {}}},
+            handler=lambda args, **kw: "just a plain string, not json",
+        )
+        result = handle_function_call("__test_bad_plain", {})
+        parsed = json.loads(result)
+        assert "output" in parsed
+        registry._tools.pop("__test_bad_plain", None)
+
+    def test_handler_returning_valid_json_passes_through(self):
+        """A handler returning valid JSON string passes through unchanged."""
+        from tools.registry import registry
+        from model_tools import handle_function_call
+        import json
+
+        registry.register(
+            name="__test_good",
+            toolset="test",
+            schema={"name": "__test_good", "description": "test", "parameters": {"type": "object", "properties": {}}},
+            handler=lambda args, **kw: json.dumps({"status": "ok", "data": [1, 2, 3]}),
+        )
+        result = handle_function_call("__test_good", {})
+        parsed = json.loads(result)
+        assert parsed == {"status": "ok", "data": [1, 2, 3]}
+        registry._tools.pop("__test_good", None)

tests/tools/test_memory_tool.py

@@ -144,7 +144,8 @@ class TestMemoryStoreReplace:
     def test_replace_no_match(self, store):
         store.add("memory", "fact A")
         result = store.replace("memory", "nonexistent", "new")
-        assert result["success"] is False
+        assert result["success"] is True
+        assert result["result"] == "no_match"
 
     def test_replace_ambiguous_match(self, store):
         store.add("memory", "server A runs nginx")
@@ -177,7 +178,8 @@ class TestMemoryStoreRemove:
 
     def test_remove_no_match(self, store):
         result = store.remove("memory", "nonexistent")
-        assert result["success"] is False
+        assert result["success"] is True
+        assert result["result"] == "no_match"
 
     def test_remove_empty_old_text(self, store):
         result = store.remove("memory", "  ")

tests/tools/test_syntax_preflight.py

@@ -0,0 +1,107 @@
+"""Tests for syntax preflight check in execute_code (issue #312)."""
+
+import ast
+import json
+import pytest
+
+
+class TestSyntaxPreflight:
+    """Verify that execute_code catches syntax errors before sandbox execution."""
+
+    def test_valid_syntax_passes_parse(self):
+        """Valid Python should pass ast.parse."""
+        code = "print('hello')\nx = 1 + 2\n"
+        ast.parse(code)  # should not raise
+
+    def test_syntax_error_indentation(self):
+        """IndentationError is a subclass of SyntaxError."""
+        code = "def foo():\nbar()\n"
+        with pytest.raises(SyntaxError):
+            ast.parse(code)
+
+    def test_syntax_error_missing_colon(self):
+        code = "if True\n  pass\n"
+        with pytest.raises(SyntaxError):
+            ast.parse(code)
+
+    def test_syntax_error_unmatched_paren(self):
+        code = "x = (1 + 2\n"
+        with pytest.raises(SyntaxError):
+            ast.parse(code)
+
+    def test_syntax_error_invalid_token(self):
+        code = "x = 1 +*\n"
+        with pytest.raises(SyntaxError):
+            ast.parse(code)
+
+    def test_syntax_error_details(self):
+        """SyntaxError should provide line, offset, msg."""
+        code = "if True\n  pass\n"
+        with pytest.raises(SyntaxError) as exc_info:
+            ast.parse(code)
+        e = exc_info.value
+        assert e.lineno is not None
+        assert e.msg is not None
+
+    def test_empty_string_passes(self):
+        """Empty string is valid Python (empty module)."""
+        ast.parse("")
+
+    def test_comments_only_passes(self):
+        ast.parse("# just a comment\n# another\n")
+
+    def test_complex_valid_code(self):
+        code = '''
+import os
+def foo(x):
+    if x > 0:
+        return x * 2
+    return 0
+
+result = [foo(i) for i in range(10)]
+print(result)
+'''
+        ast.parse(code)
+
+
+class TestSyntaxPreflightResponse:
+    """Test the error response format from the preflight check."""
+
+    def _check_syntax(self, code):
+        """Mimic the preflight check logic from execute_code."""
+        try:
+            ast.parse(code)
+            return None
+        except SyntaxError as e:
+            return json.dumps({
+                "error": f"Python syntax error: {e.msg}",
+                "line": e.lineno,
+                "offset": e.offset,
+                "text": (e.text or "").strip()[:200],
+            })
+
+    def test_returns_json_error(self):
+        result = self._check_syntax("if True\n  pass\n")
+        assert result is not None
+        data = json.loads(result)
+        assert "error" in data
+        assert "syntax error" in data["error"].lower()
+
+    def test_includes_line_number(self):
+        result = self._check_syntax("x = 1\nif True\n  pass\n")
+        data = json.loads(result)
+        assert data["line"] == 2  # error on line 2
+
+    def test_includes_offset(self):
+        result = self._check_syntax("x = (1 + 2\n")
+        data = json.loads(result)
+        assert data["offset"] is not None
+
+    def test_includes_snippet(self):
+        result = self._check_syntax("if True\n")
+        data = json.loads(result)
+        assert "if True" in data["text"]
+
+    def test_none_for_valid_code(self):
+        result = self._check_syntax("print('ok')")
+        assert result is None

tools/code_execution_tool.py

@@ -28,6 +28,7 @@ Platform: Linux / macOS only (Unix domain sockets for local). Disabled on Window
 Remote execution additionally requires Python 3 in the terminal backend.
 """
 
+import ast
 import base64
 import json
 import logging
@@ -893,6 +894,20 @@ def execute_code(
     if not code or not code.strip():
         return json.dumps({"error": "No code provided."})
 
+    # Poka-yoke (#312): Syntax check before execution.
+    # 83.2% of execute_code errors are Python exceptions; most are syntax
+    # errors the LLM generated.  ast.parse() is sub-millisecond and catches
+    # them before we spin up a sandbox child process.
+    try:
+        ast.parse(code)
+    except SyntaxError as e:
+        return json.dumps({
+            "error": f"Python syntax error: {e.msg}",
+            "line": e.lineno,
+            "offset": e.offset,
+            "text": (e.text or "").strip()[:200],
+        })
+
     # Dispatch: remote backends use file-based RPC, local uses UDS
     from tools.terminal_tool import _get_env_config
     env_type = _get_env_config()["env_type"]

tools/memory_tool.py

@@ -260,8 +260,12 @@ class MemoryStore:
             entries = self._entries_for(target)
             matches = [(i, e) for i, e in enumerate(entries) if old_text in e]
 
-            if len(matches) == 0:
-                return {"success": False, "error": f"No entry matched '{old_text}'."}
+            if not matches:
+                return {
+                    "success": True,
+                    "result": "no_match",
+                    "message": f"No entry matched '{old_text}'. The search substring was not found in any existing entry.",
+                }
 
             if len(matches) > 1:
                 # If all matches are identical (exact duplicates), operate on the first one
@@ -310,8 +314,12 @@ class MemoryStore:
             entries = self._entries_for(target)
             matches = [(i, e) for i, e in enumerate(entries) if old_text in e]
 
-            if len(matches) == 0:
-                return {"success": False, "error": f"No entry matched '{old_text}'."}
+            if not matches:
+                return {
+                    "success": True,
+                    "result": "no_match",
+                    "message": f"No entry matched '{old_text}'. The search substring was not found in any existing entry.",
+                }
 
             if len(matches) > 1:
                 # If all matches are identical (exact duplicates), remove the first one
@@ -449,30 +457,30 @@ def memory_tool(
     Returns JSON string with results.
     """
     if store is None:
-        return json.dumps({"success": False, "error": "Memory is not available. It may be disabled in config or this environment."}, ensure_ascii=False)
+        return tool_error("Memory is not available. It may be disabled in config or this environment.", success=False)
 
     if target not in ("memory", "user"):
-        return json.dumps({"success": False, "error": f"Invalid target '{target}'. Use 'memory' or 'user'."}, ensure_ascii=False)
+        return tool_error(f"Invalid target '{target}'. Use 'memory' or 'user'.", success=False)
 
     if action == "add":
         if not content:
-            return json.dumps({"success": False, "error": "Content is required for 'add' action."}, ensure_ascii=False)
+            return tool_error("Content is required for 'add' action.", success=False)
         result = store.add(target, content)
 
     elif action == "replace":
         if not old_text:
-            return json.dumps({"success": False, "error": "old_text is required for 'replace' action."}, ensure_ascii=False)
+            return tool_error("old_text is required for 'replace' action.", success=False)
         if not content:
-            return json.dumps({"success": False, "error": "content is required for 'replace' action."}, ensure_ascii=False)
+            return tool_error("content is required for 'replace' action.", success=False)
         result = store.replace(target, old_text, content)
 
     elif action == "remove":
         if not old_text:
-            return json.dumps({"success": False, "error": "old_text is required for 'remove' action."}, ensure_ascii=False)
+            return tool_error("old_text is required for 'remove' action.", success=False)
         result = store.remove(target, old_text)
 
     else:
-        return json.dumps({"success": False, "error": f"Unknown action '{action}'. Use: add, replace, remove"}, ensure_ascii=False)
+        return tool_error(f"Unknown action '{action}'. Use: add, replace, remove", success=False)
 
     return json.dumps(result, ensure_ascii=False)
 
@@ -539,7 +547,7 @@ MEMORY_SCHEMA = {
 
 
 # --- Registry ---
-from tools.registry import registry
+from tools.registry import registry, tool_error
 
 registry.register(
     name="memory",