loki-mode 7.81.0 → 7.82.0

package/docs/INSTALLATION.md

@@ -2,7 +2,7 @@
 
 The flagship product of [Autonomi](https://www.autonomi.dev/). Loki Mode is a spec-driven autonomous builder with a built-in trust layer that takes any spec to a deployed product and verifies completion with evidence (quality gates plus a completion council), not just a "done" claim. Complete installation instructions for all platforms and use cases.
 
-**Version:** v7.81.0
+**Version:** v7.82.0
 
 ---
 
@@ -395,7 +395,7 @@ provider works inside the container. Provide auth with your Anthropic API key:
 # Run Loki Mode in Docker (Claude provider, API-key auth)
 docker run --rm -e ANTHROPIC_API_KEY="$ANTHROPIC_API_KEY" \
   -v $(pwd):/workspace -w /workspace \
-  asklokesh/loki-mode:7.81.0 start ./my-spec.md
+  asklokesh/loki-mode:7.82.0 start ./my-spec.md
 ```
 
 ##### docker compose + .env (no host install)

package/events/bus.py

@@ -5,6 +5,7 @@ Events are written to .loki/events/pending/ and processed by subscribers.
 This enables CLI, API, VS Code, and MCP to communicate without shared memory.
 """
 
+import hashlib
 import json
 import logging
 import os
@@ -123,8 +124,23 @@ class LokiEvent:
         except ValueError:
             event_source = EventSource.CLI
 
+        # ID: prefer the canonical `id`. Flat events.jsonl lines (run.sh
+        # emit_event / emit_event_json) carry NO id, so __post_init__ would mint
+        # a fresh random uuid on every parse -- defeating import_from_jsonl()
+        # dedup (_processed_ids + pending-file glob) and re-importing every flat
+        # line on each call. Derive a DETERMINISTIC id from the record content
+        # so repeated imports of the same line are idempotent.
+        event_id = data.get('id', '')
+        if not event_id:
+            digest_src = '%s|%s|%s' % (
+                data.get('timestamp', ''),
+                raw_type,
+                json.dumps(payload, sort_keys=True),
+            )
+            event_id = hashlib.sha1(digest_src.encode('utf-8')).hexdigest()[:8]
+
         return cls(
-            id=data.get('id', ''),
+            id=event_id,
             type=event_type,
             source=event_source,
             timestamp=data.get('timestamp', ''),

package/events/emit.sh

@@ -118,8 +118,22 @@ else
 fi
 
 # JSON escape helper: handles \, ", and control characters including newlines
+#
+# The sed pass escapes the named short forms (\\ \" \t \r \b \f); the first awk
+# pass collapses embedded newlines to \n. Any OTHER C0 control byte
+# (0x01-0x07, 0x0B, 0x0E-0x1F) is invalid raw inside a JSON string and is
+# escaped as \uXXXX by the final awk pass -- otherwise json.loads / JSON.parse
+# reject the line and consumers (dashboard _read_events, learning aggregator)
+# silently drop it. The named short forms are already two-char ASCII by the
+# time the final pass runs, so they are never re-escaped.
 json_escape() {
-    printf '%s' "$1" | sed 's/\\/\\\\/g; s/"/\\"/g; s/\t/\\t/g; s/\r/\\r/g; s//\\b/g; s//\\f/g' | awk '{if(NR>1) printf "\\n"; printf "%s", $0}'
+    printf '%s' "$1" \
+        | sed 's/\\/\\\\/g; s/"/\\"/g; s/\t/\\t/g; s/\r/\\r/g; s//\\b/g; s//\\f/g' \
+        | awk '{if(NR>1) printf "\\n"; printf "%s", $0}' \
+        | awk 'BEGIN{for(i=1;i<=31;i++)m[sprintf("%c",i)]=sprintf("\\u%04x",i)}
+               {s=""; n=length($0);
+                for(i=1;i<=n;i++){c=substr($0,i,1); s=s (c in m ? m[c] : c)}
+                printf "%s", s}'
 }
 
 # Build payload JSON

package/loki-ts/dist/loki.js

@@ -1,5 +1,5 @@
 // @bun
-var QQ=Object.defineProperty;var ZQ=($)=>$;function zQ($,Q){this[$]=ZQ.bind(null,Q)}var b=($,Q)=>{for(var Z in Q)QQ($,Z,{get:Q[Z],enumerable:!0,configurable:!0,set:zQ.bind(Q,Z)})};var L=($,Q)=>()=>($&&(Q=$($=0)),Q);var q$=import.meta.require;var h1={};b(h1,{lokiDir:()=>P,homeLokiDir:()=>i$,findRepoRootForVersion:()=>t$,REPO_ROOT:()=>g});import{resolve as a,dirname as r$}from"path";import{fileURLToPath as XQ}from"url";import{existsSync as R$}from"fs";import{homedir as KQ}from"os";function qQ(){let $=b1;for(let Q=0;Q<6;Q++){if(R$(a($,"VERSION"))&&R$(a($,"autonomy/run.sh")))return $;let Z=r$($);if(Z===$)break;$=Z}return a(b1,"..","..","..")}function t$($){let Q=$;for(let Z=0;Z<6;Z++){if(R$(a(Q,"VERSION"))&&R$(a(Q,"autonomy/run.sh")))return Q;let z=r$(Q);if(z===Q)break;Q=z}return a($,"..","..","..")}function P(){return process.env.LOKI_DIR??a(process.cwd(),".loki")}function i$(){return a(KQ(),".loki")}var b1,g;var C=L(()=>{b1=r$(XQ(import.meta.url));g=qQ()});import{readFileSync as VQ}from"fs";import{resolve as JQ,dirname as UQ}from"path";import{fileURLToPath as WQ}from"url";function E$(){if(Q$!==null)return Q$;let $="7.81.0";if(typeof $==="string"&&$.length>0)return Q$=$,Q$;try{let Q=UQ(WQ(import.meta.url)),Z=t$(Q);Q$=VQ(JQ(Z,"VERSION"),"utf-8").trim()}catch{Q$="unknown"}return Q$}var Q$=null;var e$=L(()=>{C()});var g1={};b(g1,{runOrThrow:()=>HQ,run:()=>F,readStreamCapped:()=>m1,commandVersion:()=>BQ,commandExists:()=>f,ShellError:()=>$1,MAX_STDOUT_BYTES:()=>v1});async function m1($,Q=v1){let Z=$.getReader(),z=new TextDecoder,X="",q=0;try{while(q<Q){let{done:K,value:U}=await Z.read();if(K)break;if(!U)continue;if(q+=U.byteLength,q>Q){let J=U.byteLength-(q-Q);X+=z.decode(U.subarray(0,J),{stream:!0});break}X+=z.decode(U,{stream:!0})}X+=z.decode()}finally{try{await Z.cancel()}catch{}Z.releaseLock()}return X}async function F($,Q={}){let Z=Bun.spawn({cmd:[...$],stdout:"pipe",stderr:"pipe",env:Q.env?{...process.env,...Q.env}:process.env,cwd:Q.cwd}),z,X;if(Q.timeoutMs&&Q.timeoutMs>0)z=setTimeout(()=>{try{Z.kill("SIGTERM")}catch{}X=setTimeout(()=>{try{Z.kill("SIGKILL")}catch{}},2000)},Q.timeoutMs);try{let[q,K,U]=await Promise.all([m1(Z.stdout),new Response(Z.stderr).text(),Z.exited]);return{stdout:q,stderr:K,exitCode:U}}finally{if(z)clearTimeout(z);if(X)clearTimeout(X)}}async function HQ($,Q={}){let Z=await F($,Q);if(Z.exitCode!==0)throw new $1(`command failed (${Z.exitCode}): ${$.join(" ")}`,Z.exitCode,Z.stdout,Z.stderr);return Z}async function f($){let Q=GQ($),Z=await F(["sh","-c",`command -v ${Q}`],{timeoutMs:5000});if(Z.exitCode===0)return Z.stdout.trim()||null;return null}function GQ($){if(!/^[A-Za-z0-9._/-]+$/.test($))throw Error(`refused to shell-escape suspect token: ${$}`);return $}async function BQ($,Q="--version"){if(!await f($))return null;let z=await F([$,Q],{timeoutMs:5000});if(z.exitCode!==0)return null;return((z.stdout||z.stderr).split(/\r?\n/)[0]?.trim()??"")||null}var v1=16777216,$1;var d=L(()=>{$1=class $1 extends Error{message;exitCode;stdout;stderr;constructor($,Q,Z,z){super($);this.message=$;this.exitCode=Q;this.stdout=Z;this.stderr=z;this.name="ShellError"}}});function s($){return YQ?"":$}var YQ,O,S,_,_Z,I,k,h,V;var c=L(()=>{YQ=(process.env.NO_COLOR??"").length>0;O=s("\x1B[0;31m"),S=s("\x1B[0;32m"),_=s("\x1B[1;33m"),_Z=s("\x1B[0;34m"),I=s("\x1B[0;36m"),k=s("\x1B[1m"),h=s("\x1B[2m"),V=s("\x1B[0m")});import{existsSync as jQ}from"fs";async function Z$(){if(Y$!==void 0)return Y$;let $="/opt/homebrew/bin/python3.12";if(jQ($))return Y$=$,$;let Q=await f("python3.12");if(Q)return Y$=Q,Q;let Z=await f("python3");return Y$=Z,Z}async function z$($,Q={}){let Z=await Z$();if(!Z)return{stdout:"",stderr:"python3 not found",exitCode:127};return F([Z,"-c",$],Q)}var Y$;var V$=L(()=>{d()});var X0={};b(X0,{runStatus:()=>oQ});import{existsSync as y,readFileSync as U$,readdirSync as r1,statSync as t1}from"fs";import{resolve as D,basename as vQ}from"path";import{homedir as mQ}from"os";function i1($){let Q=Math.trunc($);if(Q>=1e6)return`${(Math.trunc(Q/1e6*10)/10).toFixed(1)}M`;if(Q>=1000)return`${(Math.trunc(Q/1000*10)/10).toFixed(1)}K`;return String(Q)}function e1($,Q,Z){if(Q===0)return null;let z=Math.trunc($*100/Q),X=Math.trunc($*N$/Q);if(X>N$)X=N$;let q=N$-X,K=S;if(z>=80)K=O;else if(z>=50)K=_;let U="=".repeat(Math.max(0,X))+" ".repeat(Math.max(0,q)),J=i1($),W=i1(Q);return`  ${k}${Z}${V} ${K}[${U}]${V} ${z}% (${J} / ${W})`}async function fQ(){if(await f("jq"))return!0;return process.stdout.write(`${O}Error: jq is required but not installed.${V}
+var QQ=Object.defineProperty;var ZQ=($)=>$;function zQ($,Q){this[$]=ZQ.bind(null,Q)}var b=($,Q)=>{for(var Z in Q)QQ($,Z,{get:Q[Z],enumerable:!0,configurable:!0,set:zQ.bind(Q,Z)})};var L=($,Q)=>()=>($&&(Q=$($=0)),Q);var q$=import.meta.require;var h1={};b(h1,{lokiDir:()=>P,homeLokiDir:()=>i$,findRepoRootForVersion:()=>t$,REPO_ROOT:()=>g});import{resolve as a,dirname as r$}from"path";import{fileURLToPath as XQ}from"url";import{existsSync as R$}from"fs";import{homedir as KQ}from"os";function qQ(){let $=b1;for(let Q=0;Q<6;Q++){if(R$(a($,"VERSION"))&&R$(a($,"autonomy/run.sh")))return $;let Z=r$($);if(Z===$)break;$=Z}return a(b1,"..","..","..")}function t$($){let Q=$;for(let Z=0;Z<6;Z++){if(R$(a(Q,"VERSION"))&&R$(a(Q,"autonomy/run.sh")))return Q;let z=r$(Q);if(z===Q)break;Q=z}return a($,"..","..","..")}function P(){return process.env.LOKI_DIR??a(process.cwd(),".loki")}function i$(){return a(KQ(),".loki")}var b1,g;var C=L(()=>{b1=r$(XQ(import.meta.url));g=qQ()});import{readFileSync as VQ}from"fs";import{resolve as JQ,dirname as UQ}from"path";import{fileURLToPath as WQ}from"url";function E$(){if(Q$!==null)return Q$;let $="7.82.0";if(typeof $==="string"&&$.length>0)return Q$=$,Q$;try{let Q=UQ(WQ(import.meta.url)),Z=t$(Q);Q$=VQ(JQ(Z,"VERSION"),"utf-8").trim()}catch{Q$="unknown"}return Q$}var Q$=null;var e$=L(()=>{C()});var g1={};b(g1,{runOrThrow:()=>HQ,run:()=>F,readStreamCapped:()=>m1,commandVersion:()=>BQ,commandExists:()=>f,ShellError:()=>$1,MAX_STDOUT_BYTES:()=>v1});async function m1($,Q=v1){let Z=$.getReader(),z=new TextDecoder,X="",q=0;try{while(q<Q){let{done:K,value:U}=await Z.read();if(K)break;if(!U)continue;if(q+=U.byteLength,q>Q){let J=U.byteLength-(q-Q);X+=z.decode(U.subarray(0,J),{stream:!0});break}X+=z.decode(U,{stream:!0})}X+=z.decode()}finally{try{await Z.cancel()}catch{}Z.releaseLock()}return X}async function F($,Q={}){let Z=Bun.spawn({cmd:[...$],stdout:"pipe",stderr:"pipe",env:Q.env?{...process.env,...Q.env}:process.env,cwd:Q.cwd}),z,X;if(Q.timeoutMs&&Q.timeoutMs>0)z=setTimeout(()=>{try{Z.kill("SIGTERM")}catch{}X=setTimeout(()=>{try{Z.kill("SIGKILL")}catch{}},2000)},Q.timeoutMs);try{let[q,K,U]=await Promise.all([m1(Z.stdout),new Response(Z.stderr).text(),Z.exited]);return{stdout:q,stderr:K,exitCode:U}}finally{if(z)clearTimeout(z);if(X)clearTimeout(X)}}async function HQ($,Q={}){let Z=await F($,Q);if(Z.exitCode!==0)throw new $1(`command failed (${Z.exitCode}): ${$.join(" ")}`,Z.exitCode,Z.stdout,Z.stderr);return Z}async function f($){let Q=GQ($),Z=await F(["sh","-c",`command -v ${Q}`],{timeoutMs:5000});if(Z.exitCode===0)return Z.stdout.trim()||null;return null}function GQ($){if(!/^[A-Za-z0-9._/-]+$/.test($))throw Error(`refused to shell-escape suspect token: ${$}`);return $}async function BQ($,Q="--version"){if(!await f($))return null;let z=await F([$,Q],{timeoutMs:5000});if(z.exitCode!==0)return null;return((z.stdout||z.stderr).split(/\r?\n/)[0]?.trim()??"")||null}var v1=16777216,$1;var d=L(()=>{$1=class $1 extends Error{message;exitCode;stdout;stderr;constructor($,Q,Z,z){super($);this.message=$;this.exitCode=Q;this.stdout=Z;this.stderr=z;this.name="ShellError"}}});function s($){return YQ?"":$}var YQ,O,S,_,_Z,I,k,h,V;var c=L(()=>{YQ=(process.env.NO_COLOR??"").length>0;O=s("\x1B[0;31m"),S=s("\x1B[0;32m"),_=s("\x1B[1;33m"),_Z=s("\x1B[0;34m"),I=s("\x1B[0;36m"),k=s("\x1B[1m"),h=s("\x1B[2m"),V=s("\x1B[0m")});import{existsSync as jQ}from"fs";async function Z$(){if(Y$!==void 0)return Y$;let $="/opt/homebrew/bin/python3.12";if(jQ($))return Y$=$,$;let Q=await f("python3.12");if(Q)return Y$=Q,Q;let Z=await f("python3");return Y$=Z,Z}async function z$($,Q={}){let Z=await Z$();if(!Z)return{stdout:"",stderr:"python3 not found",exitCode:127};return F([Z,"-c",$],Q)}var Y$;var V$=L(()=>{d()});var X0={};b(X0,{runStatus:()=>oQ});import{existsSync as y,readFileSync as U$,readdirSync as r1,statSync as t1}from"fs";import{resolve as D,basename as vQ}from"path";import{homedir as mQ}from"os";function i1($){let Q=Math.trunc($);if(Q>=1e6)return`${(Math.trunc(Q/1e6*10)/10).toFixed(1)}M`;if(Q>=1000)return`${(Math.trunc(Q/1000*10)/10).toFixed(1)}K`;return String(Q)}function e1($,Q,Z){if(Q===0)return null;let z=Math.trunc($*100/Q),X=Math.trunc($*N$/Q);if(X>N$)X=N$;let q=N$-X,K=S;if(z>=80)K=O;else if(z>=50)K=_;let U="=".repeat(Math.max(0,X))+" ".repeat(Math.max(0,q)),J=i1($),W=i1(Q);return`  ${k}${Z}${V} ${K}[${U}]${V} ${z}% (${J} / ${W})`}async function fQ(){if(await f("jq"))return!0;return process.stdout.write(`${O}Error: jq is required but not installed.${V}
 `),process.stdout.write(`Install with:
 `),process.stdout.write(`  brew install jq    (macOS)
 `),process.stdout.write(`  apt install jq     (Debian/Ubuntu)
@@ -793,4 +793,4 @@ Set LOKI_LEGACY_BASH=1 to force the bash CLI for every command.
 `),2}default:return process.stderr.write(`Unknown command: ${Q}
 `),process.stderr.write($Q),2}}s1();process.on("SIGINT",()=>process.exit(130));process.on("SIGTERM",()=>process.exit(143));var qZ=await KZ(Bun.argv.slice(2));process.exit(qZ);
 
-//# debugId=C0A7A027BE2CCC6B64756E2164756E21
+//# debugId=3E8A7AD89D7DCCD164756E2164756E21

package/lokistore/cloud.py

@@ -63,6 +63,7 @@ class S3Store(LokiStore):
         bucket: str,
         prefix: Optional[str] = None,
         region: Optional[str] = None,
+        endpoint: Optional[str] = None,
     ):
         if not bucket:
             raise ValueError("S3Store requires a bucket name")
@@ -78,7 +79,22 @@ class S3Store(LokiStore):
         self._prefix = _normalize_prefix(prefix)
         # boto3 picks up credentials from its default chain (env, shared
         # config, instance/role metadata). region_name is optional.
-        kwargs = {"region_name": region} if region else {}
+        # endpoint_url targets an S3-COMPATIBLE store (MinIO, Ceph, R2,
+        # Wasabi, ...) instead of real AWS S3; without it the docstring's
+        # "(or S3-compatible)" claim was false. Read from the arg or the
+        # AWS_ENDPOINT_URL env (boto3's own convention) as a fallback.
+        kwargs = {}
+        endpoint = endpoint or os.environ.get("AWS_ENDPOINT_URL")
+        if endpoint:
+            kwargs["endpoint_url"] = endpoint
+        if region:
+            kwargs["region_name"] = region
+        elif endpoint:
+            # A custom S3-compatible endpoint (MinIO/Ceph/...) still needs a
+            # region for SigV4 request signing even though the store ignores it.
+            # Default to us-east-1 (the conventional MinIO default) so a MinIO
+            # config without an explicit region does not fail signing.
+            kwargs["region_name"] = "us-east-1"
         self._client = boto3.client("s3", **kwargs)
 
     def _object_key(self, key: str) -> str:

package/lokistore/factory.py

@@ -60,6 +60,9 @@ def _config_from_env() -> Dict[str, Any]:
     region = os.environ.get("LOKI_STORAGE_REGION")
     if region:
         cfg["region"] = region
+    endpoint = os.environ.get("LOKI_STORAGE_ENDPOINT")
+    if endpoint:
+        cfg["endpoint"] = endpoint
     return cfg
 
 
@@ -75,6 +78,7 @@ def build_store(config: Optional[Dict[str, Any]] = None) -> LokiStore:
       bucket  : bucket/container name (cloud backends)
       prefix  : key prefix within the bucket (optional)
       region  : region (s3, optional)
+      endpoint: custom S3-compatible endpoint URL (s3 only; e.g. MinIO/Ceph/R2)
       base_dir: local base directory (local backend only; overrides resolution)
     """
     config = dict(config or {})
@@ -86,11 +90,12 @@ def build_store(config: Optional[Dict[str, Any]] = None) -> LokiStore:
     bucket = config.get("bucket")
     prefix = config.get("prefix")
     region = config.get("region")
+    endpoint = config.get("endpoint")
 
     if backend in ("s3", "aws", "aws-s3"):
         from .cloud import S3Store
 
-        return S3Store(bucket=bucket, prefix=prefix, region=region)
+        return S3Store(bucket=bucket, prefix=prefix, region=region, endpoint=endpoint)
 
     if backend in ("gcs", "gcp", "google", "google-cloud-storage"):
         from .cloud import GCSStore

package/lokistore/local.py

@@ -178,16 +178,24 @@ class LocalStore(LokiStore):
     def list(self, prefix: str = "") -> List[str]:
         # Normalize the prefix to a path under the base. An empty prefix lists
         # everything under the base.
+        #
+        # Walk the realpath of the base consistently with how relpath is taken
+        # below. If the base is reached through a symlink (macOS /tmp ->
+        # /private/tmp, k8s bind mounts, a symlinked LOKI_DIR), walking the
+        # symlink path while computing relpath against the realpath would emit
+        # "../../.." garbage keys that downstream sync then rejects as path
+        # traversal -- silently breaking the whole object-store sync.
+        real_base = os.path.realpath(self._base)
         if prefix:
             clean_prefix = normalize_key(prefix)
-            search_root = self._base / clean_prefix
+            search_root = Path(real_base) / clean_prefix
         else:
-            search_root = self._base
+            search_root = Path(real_base)
 
         if not search_root.exists():
             return []
 
-        base_str = os.path.realpath(self._base)
+        base_str = real_base
         results: List[str] = []
 
         if search_root.is_file():

package/mcp/__init__.py

@@ -57,4 +57,4 @@ try:
 except ImportError:
     __all__ = ['mcp']
 
-__version__ = '7.81.0'
+__version__ = '7.82.0'

package/mcp/server.py

@@ -1919,20 +1919,32 @@ async def mem_search(
         context = {"goal": query, "task_type": "exploration"}
         results = retriever.retrieve_task_aware(context, top_k=limit)
 
-        # BUG-MCP-006: Filter results by collection parameter when not "all"
-        # The retrieve_task_aware method returns all collections, but the user
-        # may have requested only a specific collection type
+        # Filter results by collection parameter when not "all".
+        # retrieve_task_aware tags every item with _source in
+        # {episodic, semantic, skills, anti_patterns}; it does NOT emit
+        # _type/type. Map _source to the public collection vocabulary so the
+        # filter actually matches (previously every result was dropped because
+        # result_type defaulted to "unknown").
         collection_type_map = {
             "episodes": "episode",
             "patterns": "pattern",
             "skills": "skill",
         }
         filter_type = collection_type_map.get(collection)
+        source_to_type = {
+            "episodic": "episode",
+            "semantic": "pattern",
+            "skills": "skill",
+            "anti_patterns": "pattern",
+        }
 
         # Compact results for token efficiency
         compact = []
         for r in results:
-            result_type = r.get("_type", r.get("type", "unknown"))
+            result_type = source_to_type.get(
+                r.get("_source"),
+                r.get("_type", r.get("type", "unknown")),
+            )
             # Apply collection filter
             if filter_type and result_type != filter_type:
                 continue

package/memory/cross_project.py

@@ -6,6 +6,7 @@ builds a unified index with memory statistics per project.
 
 import json
 import os
+import tempfile
 from pathlib import Path
 from datetime import datetime, timezone
 
@@ -49,6 +50,55 @@ class CrossProjectIndex:
                     })
         return projects
 
+    @staticmethod
+    def _count_episodes(episodic_dir):
+        """Count episode files in an episodic store.
+
+        Episodes are persisted by the storage layer under date subdirectories
+        (episodic/{YYYY-MM-DD}/task-*.json), not directly under episodic/. A
+        non-recursive episodic/*.json glob therefore reported 0 for every real
+        store. Walk recursively and skip the per-directory index.json sidecars.
+        """
+        if not episodic_dir.exists():
+            return 0
+        return sum(
+            1
+            for f in episodic_dir.rglob('*.json')
+            if f.name != 'index.json'
+        )
+
+    @staticmethod
+    def _count_patterns(semantic_dir):
+        """Count semantic patterns in a semantic store.
+
+        Production stores all patterns as a list inside a single
+        semantic/patterns.json, so a semantic/*.json file glob counted 1 (the
+        container) regardless of how many patterns it held. Count the entries
+        in patterns.json when present, and add any other *.json pattern files
+        (legacy one-file-per-pattern layout) so both shapes are counted.
+        """
+        if not semantic_dir.exists():
+            return 0
+        count = 0
+        patterns_file = semantic_dir / 'patterns.json'
+        if patterns_file.exists():
+            try:
+                data = json.loads(patterns_file.read_text())
+                if isinstance(data, dict):
+                    patterns = data.get('patterns', [])
+                    if isinstance(patterns, list):
+                        count += len(patterns)
+                elif isinstance(data, list):
+                    count += len(data)
+            except (json.JSONDecodeError, OSError, ValueError):
+                pass
+        # Count legacy per-pattern files, excluding known container files.
+        for f in semantic_dir.glob('*.json'):
+            if f.name in ('patterns.json', 'anti-patterns.json'):
+                continue
+            count += 1
+        return count
+
     def build_index(self):
         """Build a cross-project index with memory statistics.
 
@@ -70,8 +120,8 @@ class CrossProjectIndex:
             semantic_dir = memory_dir / 'semantic'
             skills_dir = memory_dir / 'skills'
 
-            episodic_count = len(list(episodic_dir.glob('*.json'))) if episodic_dir.exists() else 0
-            semantic_count = len(list(semantic_dir.glob('*.json'))) if semantic_dir.exists() else 0
+            episodic_count = self._count_episodes(episodic_dir)
+            semantic_count = self._count_patterns(semantic_dir)
             skills_count = len(list(skills_dir.glob('*.json'))) if skills_dir.exists() else 0
 
             project['episodic_count'] = episodic_count
@@ -86,12 +136,31 @@ class CrossProjectIndex:
         return index
 
     def save_index(self):
-        """Save index to disk."""
+        """Save index to disk atomically.
+
+        Writes to a temp file in the destination directory then os.replace()s
+        it over the target, so a crash or a concurrent load_index() never sees
+        a truncated/torn file (the previous direct open('w') truncated in place
+        before the new bytes landed).
+        """
         if self._index is None:
             return
         self.index_file.parent.mkdir(parents=True, exist_ok=True)
-        with open(self.index_file, 'w') as f:
-            json.dump(self._index, f, indent=2)
+        tmp_fd, tmp_path = tempfile.mkstemp(
+            dir=str(self.index_file.parent), suffix='.tmp'
+        )
+        try:
+            with os.fdopen(tmp_fd, 'w') as f:
+                json.dump(self._index, f, indent=2)
+                f.flush()
+                os.fsync(f.fileno())
+            os.replace(tmp_path, str(self.index_file))
+        except BaseException:
+            try:
+                os.unlink(tmp_path)
+            except OSError:
+                pass
+            raise
 
     def load_index(self):
         """Load index from disk."""

package/memory/embeddings.py

@@ -1008,8 +1008,26 @@ class EmbeddingEngine:
             logger.warning("Primary provider failed: %s, trying fallback", e)
             old_dimension = self.dimension
             self._use_fallback()
-            embedding = self._primary_provider.embed(text)
+            # Run the fallback through the SAME chunk + weighted-average path as
+            # the success branch so the vector is computed consistently (the old
+            # code embedded the raw, un-chunked text, producing a different vector
+            # for multi-chunk inputs).
+            if len(chunks) == 1:
+                embedding = self._primary_provider.embed(chunks[0])
+            else:
+                chunk_embeddings = self._primary_provider.embed_batch(chunks)
+                weights = np.array([len(c) for c in chunks], dtype=np.float32)
+                weights = weights / weights.sum()
+                embedding = np.average(chunk_embeddings, axis=0, weights=weights)
             embedding = self._normalize(embedding)
+            # The cache key computed above used the pre-fallback provider name.
+            # _use_fallback() switched the provider, so recompute the key to
+            # reflect the provider that actually produced this vector. Without
+            # this, the entry is stored under the old key while the next call
+            # looks it up under the new provider key, causing a permanent cache
+            # miss (and re-embedding) on every fallback request.
+            if self.config.cache_enabled:
+                cache_key = self._get_cache_key(text)
             # If dimension changed after fallback, log a warning so callers
             # know existing vector indices may be incompatible (BUG-MEM-006).
             if self.dimension != old_dimension: