@zilliz/memsearch-opencode 0.3.7 → 0.3.11

package/index.ts

@@ -13,13 +13,15 @@
 
 import type { Plugin } from "@opencode-ai/plugin";
 import { tool } from "@opencode-ai/plugin";
-import { execSync, exec, spawnSync } from "node:child_process";
+import { execSync, exec, spawn, spawnSync } from "node:child_process";
 import {
   readFileSync,
   existsSync,
   mkdirSync,
   readdirSync,
   realpathSync,
+  unlinkSync,
+  writeFileSync,
 } from "node:fs";
 import { join, dirname } from "node:path";
 import { fileURLToPath } from "node:url";
@@ -70,11 +72,48 @@ function deriveCollectionName(projectDir: string): string {
 
 /**
  * Summarize the N most recent daily .md files for cold-start context.
- * Extracts headings (## Session, ### turns) and bullet content from each
- * file so the agent sees the structure of past days (what sessions existed,
- * what topics came up), not just the tail of whichever file is newest.
+ * Extracts recent non-empty session sections so empty SessionStart headings
+ * do not crowd out useful context.
  */
-function getRecentMemories(
+function recentMemoryPreviewLines(content: string, maxLines: number): string[] {
+  const sections: string[][] = [];
+  let current: string[] = [];
+  let hasBody = false;
+
+  const flush = () => {
+    if (current.length > 0 && hasBody) {
+      sections.push(current);
+    }
+    current = [];
+    hasBody = false;
+  };
+
+  for (const rawLine of content.split("\n")) {
+    const line = rawLine.trimEnd();
+    if (/^##\s/.test(line)) {
+      flush();
+      current = [line];
+      continue;
+    }
+    if (/^#{3,4}\s/.test(line)) {
+      current.push(line);
+      continue;
+    }
+    if (line.startsWith("- ") || line.startsWith("[User]") || line.startsWith("[Assistant]")) {
+      current.push(line);
+      hasBody = true;
+    }
+  }
+
+  flush();
+  return sections.flat().slice(-maxLines);
+}
+
+export function isDailyJournalFile(file: string): boolean {
+  return /^\d{4}-\d{2}-\d{2}\.md$/.test(file);
+}
+
+export function getRecentMemories(
   memDir: string,
   count = 2,
   maxLinesPerFile = 30
@@ -82,7 +121,7 @@ function getRecentMemories(
   if (!existsSync(memDir)) return "";
 
   const files = readdirSync(memDir)
-    .filter((f) => f.endsWith(".md"))
+    .filter(isDailyJournalFile)
     .sort()
     .slice(-count);
 
@@ -92,9 +131,7 @@ function getRecentMemories(
   for (const file of files) {
     try {
       const content = readFileSync(join(memDir, file), "utf-8");
-      const lines = content.split("\n")
-        .filter((l) => /^#{2,4}\s/.test(l) || l.startsWith("- ") || l.startsWith("[User]") || l.startsWith("[Assistant]"))
-        .slice(0, maxLinesPerFile);
+      const lines = recentMemoryPreviewLines(content, maxLinesPerFile);
       if (lines.length > 0) {
         summary.push(`[${file}]`, ...lines);
       }
@@ -113,6 +150,34 @@ function shellEscape(s: string): string {
   return s.replace(/'/g, "'\\''");
 }
 
+/** Marks the start of memsearch's injected block within a system message. */
+export const MEMSEARCH_SYSTEM_MARKER = "[memsearch] Memory available.";
+
+/**
+ * Merge memsearch's memory context into an `output.system` array without
+ * growing it. Some backends (litellm/vllm serving e.g. Qwen models) reject a
+ * multi-entry `output.system` array with "system message must be first", so
+ * the memory block is folded into the first entry instead of pushed as a new
+ * one. If a memsearch block from a previous transform call is already present
+ * (identified by MEMSEARCH_SYSTEM_MARKER), it is replaced in place rather than
+ * appended again, so repeated calls against the same output stay idempotent.
+ */
+export function mergeSystemMemoryContext(
+  system: string[] | undefined,
+  memoryText: string
+): string[] {
+  if (!Array.isArray(system) || system.length === 0) {
+    return [memoryText];
+  }
+  const result = [...system];
+  const existing = result[0];
+  const markerIndex = existing.indexOf(MEMSEARCH_SYSTEM_MARKER);
+  const base =
+    markerIndex === -1 ? existing : existing.slice(0, markerIndex).replace(/\n+$/, "");
+  result[0] = base ? `${base}\n\n${memoryText}` : memoryText;
+  return result;
+}
+
 /**
  * Start the capture daemon as a background process.
  * The daemon polls OpenCode's SQLite for completed turns and writes to daily .md files.
@@ -122,6 +187,7 @@ function startCaptureDaemon(
   collectionName: string,
   memsearchCmd: string
 ): void {
+  const stateDir = join(projectDir, ".memsearch");
   const pidFile = join(projectDir, ".memsearch", ".capture.pid");
   const daemonScript = join(PLUGIN_DIR, "scripts", "capture-daemon.py");
 
@@ -136,21 +202,40 @@ function startCaptureDaemon(
           return; // Already running
         } catch {
           // Process is dead, clean up stale PID file
+          try { unlinkSync(pidFile); } catch { /* ignore */ }
         }
       }
     } catch { /* ignore */ }
   }
 
-  // Start daemon in background
-  exec(
-    `python3 "${daemonScript}" "${projectDir}" "${collectionName}" ` +
-      `--memsearch-cmd "${shellEscape(memsearchCmd)}" --poll-interval 10 &`,
-    {
-      timeout: 5000,
-      env: { ...process.env, MEMSEARCH_NO_WATCH: "1" },
-    },
-    () => { /* ignore */ }
-  );
+  try {
+    mkdirSync(stateDir, { recursive: true });
+    const child = spawn(
+      "python3",
+      [
+        daemonScript,
+        projectDir,
+        collectionName,
+        "--memsearch-cmd",
+        memsearchCmd,
+        "--poll-interval",
+        "10",
+        "--parent-pid",
+        String(process.pid),
+      ],
+      {
+        detached: true,
+        stdio: "ignore",
+        env: { ...process.env, MEMSEARCH_NO_WATCH: "1" },
+      }
+    );
+    child.unref();
+    if (child.pid) {
+      try { writeFileSync(pidFile, String(child.pid), "utf-8"); } catch { /* ignore */ }
+    }
+  } catch {
+    // Capture is best-effort; tools still work without the background daemon.
+  }
 }
 
 /**
@@ -349,9 +434,10 @@ const MemsearchPlugin: Plugin = async ({ project, directory, worktree }) => {
             try {
               const context = getRecentMemories(memoryDir);
               if (context) {
-                output.system.push(
-                  `[memsearch] Memory available. You have access to memory_search, memory_get, and memory_transcript tools for recalling past sessions.\n\n${context}`
-                );
+                const memoryText =
+                  `${MEMSEARCH_SYSTEM_MARKER} You have access to memory_search, ` +
+                  `memory_get, and memory_transcript tools for recalling past sessions.\n\n${context}`;
+                output.system = mergeSystemMemoryContext(output.system, memoryText);
               }
             } catch { /* ignore */ }
           },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zilliz/memsearch-opencode",
-  "version": "0.3.7",
+  "version": "0.3.11",
   "description": "memsearch plugin for OpenCode — semantic memory search across sessions",
   "type": "module",
   "main": "index.ts",
@@ -23,6 +23,9 @@
     "semantic-search",
     "milvus"
   ],
+  "scripts": {
+    "test": "node --test"
+  },
   "author": "memsearch contributors",
   "license": "MIT",
   "peerDependencies": {

package/scripts/capture-daemon.py

@@ -481,6 +481,21 @@ def wake_maintenance(project_dir: str) -> None:
     )
 
 
+def process_is_alive(pid: int) -> bool:
+    """Return whether a process exists without sending it a signal."""
+    if pid <= 0:
+        return True
+    try:
+        os.kill(pid, 0)
+    except ProcessLookupError:
+        return False
+    except PermissionError:
+        return True
+    except OSError:
+        return True
+    return True
+
+
 def get_session_ids(conn: sqlite3.Connection, project_dir: str) -> list[str]:
     """Find OpenCode sessions that belong to the given project directory."""
     sessions = conn.execute(
@@ -764,6 +779,7 @@ def main() -> None:
     parser.add_argument("collection_name", help="Milvus collection name")
     parser.add_argument("--memsearch-cmd", default="memsearch", help="memsearch command")
     parser.add_argument("--poll-interval", type=int, default=10, help="Poll interval in seconds")
+    parser.add_argument("--parent-pid", type=int, default=0, help="Exit when this parent process is gone")
     args = parser.parse_args()
 
     db_path = get_db_path()
@@ -791,6 +807,9 @@ def main() -> None:
     tail_turn_cache: dict[str, TailTurnObservation] = {}
 
     while True:
+        if args.parent_pid and not process_is_alive(args.parent_pid):
+            cleanup()
+
         any_new = False
         conn = None
         try:
@@ -820,6 +839,8 @@ def main() -> None:
             if conn is not None:
                 conn.close()
 
+        if args.parent_pid and not process_is_alive(args.parent_pid):
+            cleanup()
         time.sleep(args.poll_interval)
 
 

package/skills/memory-config/SKILL.md

@@ -15,7 +15,7 @@ When this skill is triggered, inspect the user's request text. If there is no co
 
 - Empty request or "check": diagnose current MemSearch setup.
 - "Show/get setting": read the requested resolved/global/project value.
-- "Set/enable/disable/change": make a safe project-scoped change after confirming ambiguous choices.
+- "Set/enable/disable/change": choose global vs project scope explicitly; use global config for trusted plugin automation/provider/prompt/endpoint settings and project config only for allowlisted local indexing knobs.
 - "Not capturing/search empty/no memory": troubleshoot files, config, and index health.
 - "Use OpenAI/Gemini/Anthropic/native/model": configure provider routing.
 - "PROJECT.md/USER.md/profile/review": configure advanced maintenance.
@@ -114,14 +114,33 @@ Config is resolved from built-in defaults, global config, project config, env re
 
 Use `memsearch config list --resolved` for effective behavior, `--global` for global overrides, and `--project` for repository-specific overrides.
 
+Since v0.4.11, project-local `.memsearch.toml` is restricted before it is merged.
+Only these low-risk local indexing keys are honored from project config:
+
+- `milvus.collection`
+- `embedding.batch_size`
+- `chunking.max_chunk_size`
+- `chunking.overlap_lines`
+- `watch.debounce_ms`
+
+Trusted settings are ignored or rejected in project config. Put these in global
+config (`~/.memsearch/config.toml`) or pass explicit CLI flags instead:
+
+- provider/model/API endpoint/API key settings
+- `[llm]` and `[llm.providers.*]`
+- `[prompts]`
+- plugin automation such as `plugins.opencode.project_review.enabled`,
+  `plugins.opencode.user_profile.enabled`, and
+  `plugins.opencode.memory_to_skill.enabled`
+
 Default recommendation:
 
 - Put reusable defaults in global config so users do not repeat setup in every project.
-- Put only project-specific overrides in project config.
-- Use global config for named LLM providers, common model choices, normal input/output defaults, and shared prompt defaults.
-- Use project config for one repo's enable/disable switches, unusual output files, special intervals, or repo-specific providers.
+- Put only allowlisted local indexing overrides in project config.
+- Use global config for named LLM providers, common model choices, plugin enable/disable switches, task intervals, install paths, and shared prompt defaults.
+- If the user wants advanced maintenance enabled for all projects, set the plugin keys globally. The default relative `input_dir` / `output_file` values still resolve inside each current project.
 
-Maintenance `input_dir` and `output_file` may be relative even when configured globally. They are resolved from the current project directory at runtime. For custom prompt paths, prefer absolute paths in global config and relative paths in project config.
+Maintenance `input_dir` and `output_file` may be relative even when configured globally. They are resolved from the current project directory at runtime, so a global `output_file = ".memsearch/PROJECT.md"` writes to each project's own `.memsearch/PROJECT.md`. For custom prompt paths, prefer absolute paths in global config; project prompt paths are not trusted.
 
 OpenCode plugin keys:
 
@@ -193,7 +212,7 @@ If advanced maintenance or `memory_to_skill` seems silent, check
 `.memsearch/.maintenance-state.json` for `<plugin>.<task>.last_error` and
 `last_failed_at`; background hook errors may not surface in the chat.
 
-Before enabling advanced maintenance, ask which provider to use, whether the default 24-hour interval is acceptable, and whether `.memsearch/PROJECT.md` / `.memsearch/USER.md` are acceptable output files.
+Before enabling advanced maintenance, ask which provider to use, whether the default 24-hour interval is acceptable, whether `.memsearch/PROJECT.md` / `.memsearch/USER.md` are acceptable output files, and whether the user wants the enablement global. Do not write plugin automation keys with `--project`; v0.4.11+ project config ignores or rejects them.
 
 Prompt overrides:
 
@@ -207,8 +226,8 @@ memory_to_skill = ""
 
 Empty prompt paths mean use the built-in MemSearch prompts. Custom prompt files may use `{{AGENT_NAME}}`, `{{TASK_NAME}}`, `{{PROJECT_DIR}}`, `{{INPUT_DIR}}`, and `{{OUTPUT_FILE}}`; the runner appends existing output, recent journals, and digest automatically.
 
-Use `memsearch config set` for changes. After changing anything, show the command, the resolved value, and whether a new session is needed.
+Use `memsearch config set` for changes. For trusted keys such as `plugins.*`, `[llm.providers.*]`, `[prompts]`, `embedding.provider`, or `milvus.uri`, set global config by omitting `--project`. Use `--project` only for allowlisted local indexing keys. After changing anything, show the command, the resolved value, and whether a new session is needed.
 
 MemSearch TOML changes are read lazily by the CLI, capture daemon, and maintenance runner, so values such as `plugins.opencode.summarize.*`, `plugins.opencode.project_review.*`, `plugins.opencode.user_profile.*`, `[llm.providers.*]`, `[prompts]`, `milvus.*`, and `embedding.*` usually apply on the next capture, recall, index, or maintenance invocation. Restart OpenCode after `opencode.json` or plugin package changes; if capture behavior still looks stale after TOML edits, restart OpenCode or the capture daemon. In final diagnostic/change summaries, make clear that this is MemSearch memory configuration, not OpenCode's own memory/config system.
 
-When useful, remind the user that they can either continue using this `memory-config` skill for guided configuration, or manually run `memsearch config init` for global interactive setup, `memsearch config init --project` for project interactive setup, and `memsearch config set/get/list` for direct CLI changes.
+When useful, remind the user that they can either continue using this `memory-config` skill for guided configuration, or manually run `memsearch config init` for global interactive setup, `memsearch config init --project` for allowlisted project indexing setup, and `memsearch config set/get/list` for direct CLI changes.

package/skills/memory-to-skill/SKILL.md

@@ -90,14 +90,19 @@ The background pass mines automatically when enabled, starting from the summarie
 
 ```bash
 memsearch config get plugins.opencode.memory_to_skill.enabled 2>/dev/null || echo "false"
-# enable the background pass (do not enable silently)
-memsearch config set plugins.opencode.memory_to_skill.enabled true --project
+# enable the background pass globally (do not enable silently)
+memsearch config set plugins.opencode.memory_to_skill.enabled true
 # how eagerly history-mining distils (default 3; lower = more eager)
-memsearch config set plugins.opencode.memory_to_skill.min_occurrences 3 --project
+memsearch config set plugins.opencode.memory_to_skill.min_occurrences 3
 # pre-set install targets (otherwise you are asked at install time)
-memsearch config set plugins.opencode.memory_to_skill.paths '[".agents/skills"]' --project
+memsearch config set plugins.opencode.memory_to_skill.paths '[".agents/skills"]'
 ```
 
+Since v0.4.11, project-local `.memsearch.toml` accepts only allowlisted local
+indexing keys. Do not use `--project` for `plugins.*` settings such as
+`memory_to_skill.enabled`, `min_occurrences`, or `paths`; put them in global
+config instead.
+
 Note: `enabled` only gates the **background** (session-end) pass. The explicit
 commands above (`skills add`, `skills install`) always work, and you can mine history (C) directly.