feat(mem0-plugin): smart retrieval, citations, pre-commit capture (Tier 7)

Smart retrieval tier: - on_user_prompt.sh: stack-trace detection (auto-searches anti_patterns on errors) - on_user_prompt.sh: file-aware retrieval (extracts file paths, suggests scoped search) - on_user_prompt.sh: recency boost guidance for state-related queries - mem0-mcp/SKILL.md: inline citation format [mem0:<8char_id>] - on_pre_commit.py: captures staged diff as commit_context memory via REST API
2026-05-21 18:13:59 +05:30
parent 81211be473
commit cd66050748
4 changed files with 219 additions and 0 deletions
@@ -0,0 +1,116 @@
+#!/usr/bin/env python3
+"""Pre-commit memory capture.
+
+Captures a summary of staged changes as a mem0 memory before each commit.
+Runs as a background fire-and-forget call — never blocks the commit.
+
+Usage:
+  git diff --cached --stat | python3 on_pre_commit.py
+  # or with full diff:
+  git diff --cached | python3 on_pre_commit.py --full
+
+Env vars required: MEM0_API_KEY (or CLAUDE_PLUGIN_OPTION_MEM0_API_KEY)
+Env vars optional: MEM0_RESOLVED_USER_ID, MEM0_PROJECT_ID, MEM0_BRANCH
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import subprocess
+import sys
+import urllib.error
+import urllib.request
+
+sys.path.insert(0, os.path.dirname(__file__))
+from _identity import resolve_api_key, resolve_user_id
+
+try:
+    from _project import resolve_branch, resolve_project_id
+except ImportError:
+    def resolve_project_id() -> str:
+        return os.path.basename(os.getcwd())
+
+    def resolve_branch() -> str:
+        return "unknown"
+
+
+def get_commit_message() -> str:
+    try:
+        result = subprocess.run(
+            ["git", "log", "-1", "--format=%s", "HEAD"],
+            capture_output=True, text=True, timeout=5,
+        )
+        return result.stdout.strip() if result.returncode == 0 else ""
+    except Exception:
+        return ""
+
+
+def get_staged_summary() -> str:
+    try:
+        result = subprocess.run(
+            ["git", "diff", "--cached", "--stat"],
+            capture_output=True, text=True, timeout=10,
+        )
+        return result.stdout.strip() if result.returncode == 0 else ""
+    except Exception:
+        return ""
+
+
+def main() -> int:
+    api_key = resolve_api_key()
+    if not api_key:
+        return 0
+
+    diff_input = sys.stdin.read().strip() if not sys.stdin.isatty() else ""
+    staged = diff_input or get_staged_summary()
+    if not staged or len(staged) < 10:
+        return 0
+
+    user_id = os.environ.get("MEM0_RESOLVED_USER_ID") or resolve_user_id()
+    project_id = os.environ.get("MEM0_PROJECT_ID") or resolve_project_id()
+    branch = os.environ.get("MEM0_BRANCH") or resolve_branch()
+    commit_msg = get_commit_message()
+
+    lines = staged.splitlines()
+    if len(lines) > 30:
+        staged = "\n".join(lines[:30]) + f"\n... ({len(lines) - 30} more lines)"
+
+    content = f"## Commit Context\n\nBranch: {branch}\n"
+    if commit_msg:
+        content += f"Message: {commit_msg}\n"
+    content += f"\n### Staged Changes\n```\n{staged}\n```"
+
+    body = json.dumps({
+        "messages": [{"role": "user", "content": content}],
+        "user_id": user_id,
+        "app_id": project_id,
+        "metadata": {
+            "type": "commit_context",
+            "branch": branch,
+            "source": "pre-commit",
+        },
+        "infer": False,
+    }).encode()
+
+    req = urllib.request.Request(
+        "https://api.mem0.ai/v3/memories/add/",
+        data=body,
+        headers={
+            "Authorization": f"Token {api_key}",
+            "Content-Type": "application/json",
+        },
+        method="POST",
+    )
+
+    try:
+        with urllib.request.urlopen(req, timeout=10):
+            pass
+    except Exception:
+        pass
+
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
@@ -35,6 +35,15 @@ if [ -z "${MEM0_API_KEY:-}" ]; then
 fi
 USER_ID="$MEM0_RESOLVED_USER_ID"

+# Detect stack traces and error patterns in the prompt
+HAS_ERROR=""
+if echo "$PROMPT" | grep -qiE '(Traceback|Error:|Exception:|panic:|FAILED|fatal:| at .+\.[a-z]+:[0-9]+)'; then
+  HAS_ERROR="true"
+fi
+
+# Detect file paths in the prompt
+FILE_PATHS=$(echo "$PROMPT" | grep -oE '([a-zA-Z0-9_./-]+\.(py|ts|tsx|js|jsx|rs|go|rb|java|sh|yaml|yml|json|toml|md|sql|css|html))\b' 2>/dev/null | head -5 || echo "")
+
 cat <<EOF
 ## Memory check

@@ -53,6 +62,29 @@ improve your answer. The agent -- not this hook -- owns this decision.
 - the user is *stating* new info -- that's a write trigger (\`add_memory\`), not a search
 - it's a pure syntax / factual question answerable from general knowledge
 - you already searched this scope earlier in the turn
+EOF
+
+if [ -n "$HAS_ERROR" ]; then
+  cat <<EOF
+
+**ERROR DETECTED in prompt.** You SHOULD search mem0 for prior occurrences:
+- \`search_memories(query="<error class or message>", filters={"AND": [{"user_id": "$USER_ID"}, {"app_id": "$MEM0_PROJECT_ID"}, {"metadata": {"type": "anti_pattern"}}]})\`
+- \`search_memories(query="<module or file from stack trace>", filters={"AND": [{"user_id": "$USER_ID"}, {"app_id": "$MEM0_PROJECT_ID"}, {"metadata": {"type": "task_learning"}}]})\`
+This surfaces past debugging context and known failure modes.
+EOF
+fi
+
+if [ -n "$FILE_PATHS" ]; then
+  cat <<EOF
+
+**FILE PATHS detected:** \`$FILE_PATHS\`
+Search mem0 for context about these files:
+- \`search_memories(query="<filename without extension>", filters={"AND": [{"user_id": "$USER_ID"}, {"app_id": "$MEM0_PROJECT_ID"}]})\`
+Memories tagged with \`metadata.files\` containing these paths will surface via text match.
+EOF
+fi
+
+cat <<EOF

 **If searching, do it well:**
 - Run **2-4 parallel** \`search_memories\` calls with different angles, not one
@@ -66,6 +98,7 @@ improve your answer. The agent -- not this hook -- owns this decision.
  - \`{"AND": [{"user_id": "$USER_ID"}, {"app_id": "$MEM0_PROJECT_ID"}, {"metadata": {"type": "user_preference"}}]}\` -- tooling, stack, style
  - \`{"AND": [{"user_id": "$USER_ID"}, {"app_id": "$MEM0_PROJECT_ID"}, {"metadata": {"type": "convention"}}]}\` -- established patterns
  - Or scope with just \`{"AND": [{"user_id": "$USER_ID"}, {"app_id": "$MEM0_PROJECT_ID"}]}\` when no metadata filter fits.
+- **Recency boost:** For state-related queries ("where were we", "current task", "latest"), add a \`created_at\` filter: \`{"created_at": {"gte": "<90 days ago YYYY-MM-DD>"}}\`. Skip recency for durable facts (conventions, decisions).
 - Empty results are normal -- proceed without context.
 EOF