loki-mode 7.68.0 → 7.70.0

package/mcp/__init__.py

@@ -57,4 +57,4 @@ try:
 except ImportError:
     __all__ = ['mcp']
 
-__version__ = '7.68.0'
+__version__ = '7.70.0'

package/mcp/server.py

@@ -1358,14 +1358,17 @@ async def loki_start_project(prd_content: str = "", prd_path: str = "") -> str:
             try:
                 resolved = validate_path(prd_path, allowed_dirs=['.'])
             except PathTraversalError as e:
+                _emit_tool_event_async('loki_start_project', 'complete', result_status='error', error=str(e))
                 return json.dumps({"error": str(e)})
             if os.path.exists(resolved) and os.path.isfile(resolved):
                 with open(resolved, 'r', encoding='utf-8') as f:
                     content = f.read()
             else:
+                _emit_tool_event_async('loki_start_project', 'complete', result_status='error', error='PRD file not found')
                 return json.dumps({"error": f"PRD file not found: {prd_path}"})
 
         if not content:
+            _emit_tool_event_async('loki_start_project', 'complete', result_status='error', error='No PRD content or path provided')
             return json.dumps({"error": "No PRD content or path provided"})
 
         # Initialize project state using safe path operations
@@ -1390,6 +1393,7 @@ async def loki_start_project(prd_content: str = "", prd_path: str = "") -> str:
         _emit_tool_event_async('loki_start_project', 'complete', result_status='success')
         return json.dumps({"success": True, **project})
     except PathTraversalError as e:
+        _emit_tool_event_async('loki_start_project', 'complete', result_status='error', error='Access denied')
         return json.dumps({"error": f"Access denied: {e}"})
     except Exception as e:
         logger.error(f"Start project failed: {e}")
@@ -1440,6 +1444,7 @@ async def loki_project_status() -> str:
         _emit_tool_event_async('loki_project_status', 'complete', result_status='success')
         return json.dumps(status, default=str)
     except PathTraversalError:
+        _emit_tool_event_async('loki_project_status', 'complete', result_status='error', error='Access denied')
         return json.dumps({"error": "Access denied"})
     except Exception as e:
         logger.error(f"Project status failed: {e}")
@@ -1478,6 +1483,7 @@ async def loki_agent_metrics() -> str:
         _emit_tool_event_async('loki_agent_metrics', 'complete', result_status='success')
         return json.dumps(metrics, default=str)
     except PathTraversalError:
+        _emit_tool_event_async('loki_agent_metrics', 'complete', result_status='error', error='Access denied')
         return json.dumps({"error": "Access denied"})
     except Exception as e:
         logger.error(f"Agent metrics failed: {e}")
@@ -1500,6 +1506,7 @@ async def loki_checkpoint_restore(checkpoint_id: str = "") -> str:
     try:
         cp_dir = safe_path_join('.loki', 'state', 'checkpoints')
         if not os.path.isdir(cp_dir):
+            _emit_tool_event_async('loki_checkpoint_restore', 'complete', result_status='success')
             return json.dumps({"checkpoints": [], "message": "No checkpoints directory"})
 
         checkpoints = []
@@ -1518,6 +1525,7 @@ async def loki_checkpoint_restore(checkpoint_id: str = "") -> str:
         # Find and restore specific checkpoint
         target = next((c for c in checkpoints if c["id"] == checkpoint_id), None)
         if not target:
+            _emit_tool_event_async('loki_checkpoint_restore', 'complete', result_status='error', error='Checkpoint not found')
             return json.dumps({"error": f"Checkpoint not found: {checkpoint_id}"})
 
         # Write checkpoint state as current state, stripping the injected "id" field
@@ -1529,6 +1537,7 @@ async def loki_checkpoint_restore(checkpoint_id: str = "") -> str:
         _emit_tool_event_async('loki_checkpoint_restore', 'complete', result_status='success')
         return json.dumps({"restored": True, "checkpoint_id": checkpoint_id})
     except PathTraversalError:
+        _emit_tool_event_async('loki_checkpoint_restore', 'complete', result_status='error', error='Access denied')
         return json.dumps({"error": "Access denied"})
     except Exception as e:
         logger.error(f"Checkpoint restore failed: {e}")
@@ -1569,6 +1578,7 @@ async def loki_quality_report() -> str:
         _emit_tool_event_async('loki_quality_report', 'complete', result_status='success')
         return json.dumps(report, default=str)
     except PathTraversalError:
+        _emit_tool_event_async('loki_quality_report', 'complete', result_status='error', error='Access denied')
         return json.dumps({"error": "Access denied"})
     except Exception as e:
         logger.error(f"Quality report failed: {e}")
@@ -2057,6 +2067,7 @@ async def mem_get(
 
         id_list = [i.strip() for i in ids.split(",") if i.strip()]
         if not id_list:
+            _emit_tool_event_async('mem_get', 'complete', result_status='error', error='No IDs provided')
             return json.dumps({"entries": {}, "error": "No IDs provided"})
 
         # Cap at 20 to prevent abuse
@@ -2401,6 +2412,7 @@ async def loki_learnings(limit: int = 50) -> str:
     try:
         path = safe_path_join('.loki', 'state', 'relevant-learnings.json')
         if not os.path.exists(path):
+            _emit_tool_event_async('loki_learnings', 'complete', result_status='success')
             return json.dumps({"version": 1, "learnings": [], "total": 0})
         try:
             with safe_open(path, 'r') as f:

package/memory/engine.py

@@ -244,7 +244,11 @@ class MemoryEngine:
         for pattern in patterns_data.get("patterns", []):
             if not isinstance(pattern, dict):
                 continue
-            referenced_ids.update(pattern.get("source_episodes", []))
+            # `or []` guards an explicit null source_episodes (corrupt or
+            # hand-edited record): .get(..., []) returns None on a stored
+            # null, and set.update(None) raises TypeError, crashing the whole
+            # cleanup pass. A null and an empty list are equivalent here.
+            referenced_ids.update(pattern.get("source_episodes") or [])
 
         # Scan episodic directories
         episodic_path = Path(self.base_path) / "episodic"
@@ -376,17 +380,28 @@ class MemoryEngine:
                         "last_accessed": now,
                         "relevance_score": 0.5,
                     })
-                    index["total_memories"] = index.get("total_memories", 0) + 1
+                    # total_memories counts memories (episodes), not topics, to
+                    # match the canonical rebuild_index semantics
+                    # (total_memories += 1 per episode at line ~860). Previously
+                    # this only incremented when a NEW topic was created, so N
+                    # episodes sharing one phase reported total_memories=1 until
+                    # the user ran rebuild_index, which then jumped it to N
+                    # (two functions in this file disagreeing on the same field,
+                    # surfaced by get_stats). Increment per distinct episode.
+                    if episode_id:
+                        index["total_memories"] = index.get("total_memories", 0) + 1
                 else:
                     # Only count a given episode once. On resume/checkpoint the same
                     # trace id can be re-saved; without this guard episode_count,
-                    # total_cost_usd, and total_tokens would inflate on every re-save
-                    # even though episode_ids is already de-duplicated.
+                    # total_cost_usd, total_tokens, and total_memories would inflate
+                    # on every re-save even though episode_ids is already
+                    # de-duplicated.
                     if episode_id and episode_id not in found.get("episode_ids", []):
                         found.setdefault("episode_ids", []).append(episode_id)
                         found["episode_count"] = found.get("episode_count", 0) + 1
                         found["total_cost_usd"] = float(found.get("total_cost_usd", 0) or 0) + cost
                         found["total_tokens"] = int(found.get("total_tokens", 0) or 0) + tokens
+                        index["total_memories"] = index.get("total_memories", 0) + 1
                     merged = set(found.get("files_touched", []) or []) | set(files[:20])
                     found["files_touched"] = sorted(merged)[:50]
                     found["last_accessed"] = now
@@ -878,7 +893,12 @@ class MemoryEngine:
                             }
 
                         topics[phase]["token_count"] += tokens
-                        if data.get("timestamp", "") > topics[phase].get("last_accessed", ""):
+                        # `or ""` on BOTH sides: an episode with a missing or
+                        # explicit-null timestamp leaves last_accessed=None (set
+                        # from data.get("timestamp") above), and `str > None`
+                        # raises TypeError, crashing rebuild_index. A single
+                        # such episode is enough to break the whole rebuild.
+                        if (data.get("timestamp") or "") > (topics[phase].get("last_accessed") or ""):
                             topics[phase]["last_accessed"] = data.get("timestamp")
 
         # Index semantic patterns

package/memory/layers/loader.py

@@ -156,7 +156,6 @@ class ProgressiveLoader:
             remaining_tokens -= layer2_tokens
 
             # Collect timeline context for each relevant topic
-            topic_ids = {t.id for t in relevant_topics}
             timeline_context: Dict[str, List[Dict[str, Any]]] = {}
 
             for topic in relevant_topics:
@@ -177,12 +176,17 @@ class ProgressiveLoader:
                 self._metrics.calculate_savings(index.get("total_tokens_available", 0))
                 return memories, self._metrics
 
-        # Layer 3: Load full memories for high-relevance topics
+        # Layer 3: Load full memories for high-relevance topics.
+        # Gate on effective_score (boosted match score when set, else stored
+        # relevance), not the stored relevance_score. The Layer-1 keyword
+        # boost lives on the transient match_score precisely so a strongly
+        # matching topic can clear this Layer-3 gate; reading the un-boosted
+        # relevance_score here would silently drop exactly those topics the
+        # boost was meant to surface.
         if remaining_tokens > 0:
-            # Sort topics by relevance
             high_relevance = [
                 t for t in relevant_topics
-                if t.relevance_score >= self.HIGH_RELEVANCE_THRESHOLD
+                if t.effective_score >= self.HIGH_RELEVANCE_THRESHOLD
             ]
 
             storage = self._get_storage()

package/memory/token_economics.py

@@ -256,6 +256,15 @@ def optimize_context(
             relevance = memory.get("_weighted_score", 0.5)
         else:
             relevance = memory.get("_score", 0.5)
+        # Guard against an explicit null score in a corrupt/hand-edited
+        # record: .get(key, default) returns None when the key is present
+        # but null, and the `relevance > 1.0` comparison below would then
+        # raise "'>' not supported between instances of 'NoneType' and
+        # 'float'". Use an is-None check (not `or`) so a legitimate stored
+        # 0.0 relevance is preserved, mirroring the confidence guard above
+        # and retrieval.py's own base_score guard.
+        if relevance is None:
+            relevance = 0.5
         if relevance > 1.0:
             # Normalize high scores
             relevance = min(1.0, relevance / 10.0)

package/memory/vector_index.py

@@ -42,7 +42,20 @@ class VectorIndex:
         Args:
             dimension: The dimensionality of vectors. Default is 384
                       which matches MiniLM embedding size.
+
+        Raises:
+            ImportError: If numpy is not installed. Every operation in this
+                index (add, search, save, load) relies on numpy, so fail
+                fast with a clear message here rather than letting a later
+                call crash with an opaque ``AttributeError: 'NoneType'
+                object has no attribute ...`` once ``np`` is None. The
+                module-level NUMPY_AVAILABLE flag is what gates this check.
         """
+        if not NUMPY_AVAILABLE:
+            raise ImportError(
+                "numpy is required for the vector index. "
+                "Install it with: pip install numpy"
+            )
         self.dimension = dimension
         self.embeddings: List[np.ndarray] = []
         self.ids: List[str] = []

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "loki-mode",
   "mcpName": "io.github.asklokesh/loki-mode",
-  "version": "7.68.0",
+  "version": "7.70.0",
   "description": "Loki Mode by Autonomi. Autonomous spec-to-product system: takes a PRD, GitHub issue, OpenAPI/JSON/YAML, or one-line brief to a deployed app via the RARV-C closure loop with 8 quality gates. Provider-agnostic (Claude Code, OpenAI Codex, Cline, Aider).",
   "keywords": [
     "agent",

package/plugins/loki-mode/.claude-plugin/plugin.json

@@ -2,7 +2,7 @@
   "$schema": "https://json.schemastore.org/claude-code-plugin-manifest.json",
   "name": "loki-mode",
   "displayName": "Loki Mode",
-  "version": "7.68.0",
+  "version": "7.70.0",
   "description": "Autonomous spec-to-product build system with a built-in trust layer (RARV-C closure loop, 8 quality gates, completion council). Ships Loki's spec-hardening, drift-detection, and deterministic PR verification commands plus the Loki MCP server.",
   "author": {
     "name": "Autonomi",

package/providers/aider.sh

@@ -89,7 +89,7 @@ PROVIDER_COST_OUTPUT_FAST=0.015
 PROVIDER_DEGRADED=true
 PROVIDER_DEGRADED_REASONS=(
     "No subagent support"
-    "Sequential execution only"
+    "Sequential execution only - no parallel agents (PROVIDER_MAX_PARALLEL=1)"
     "No Task tool or MCP"
 )
 

package/providers/claude.sh

@@ -377,7 +377,15 @@ loki_apply_max_tier_clamp() {
 
 # Dynamic model resolution (v6.0.0)
 # Resolves a capability tier to a concrete model name at runtime.
-# Respects LOKI_MAX_TIER to cap cost (e.g., maxTier=sonnet prevents opus usage).
+# Respects LOKI_MAX_TIER to cap cost via loki_apply_max_tier_clamp. NOTE the
+# ceiling clamps DOWN to the provider's configured tier model, not to the alias
+# named by the cap: with the stock config (CLAUDE_DEFAULT_DEVELOPMENT=opus, see
+# line 56), LOKI_MAX_TIER=sonnet resolves planning/fable DOWN to
+# PROVIDER_MODEL_DEVELOPMENT, which is still opus. To actually pin sonnet as the
+# ceiling, also set LOKI_ALLOW_HAIKU=true (which makes PROVIDER_MODEL_DEVELOPMENT
+# sonnet) or override LOKI_CLAUDE_MODEL_DEVELOPMENT=sonnet. This is intentional
+# and is mirrored byte-for-byte by the dashboard/estimator ports (parity-locked
+# in tests/test-model-override.sh).
 # Capability aliases: "best" -> planning tier, "fast" -> fast tier, "balanced" -> development tier
 resolve_model_for_tier() {
     local tier="$1"

package/providers/codex.sh

@@ -161,6 +161,13 @@ resolve_model_for_tier() {
     esac
 
     local max_tier="${LOKI_MAX_TIER:-}"
+    # Normalize EXACTLY like claude.sh:356 (loki_apply_max_tier_clamp): trim +
+    # lowercase BEFORE the case match. Without this, a user-typed cap like "Haiku"
+    # or " haiku " (settings.json maxTier exports verbatim) fell through to the
+    # default arm and the cost ceiling was silently bypassed for codex while
+    # claude honored it. Both routes (this + applyCodexMaxTier in providers.ts)
+    # normalize identically. Parity fix.
+    max_tier="$(printf '%s' "$max_tier" | tr '[:upper:]' '[:lower:]' | sed -e 's/^[[:space:]]*//' -e 's/[[:space:]]*$//')"
     local effort=""
 
     case "$tier" in