Add Cursor Agent session indexing (Closes #9)

CHANGELOG.md

@@ -1,5 +1,25 @@
 # Changelog
 
+## 0.5.0
+
+Closes #9.
+
+- Add **Cursor Agent** session support (Cursor IDE + `agent` CLI)
+- Indexes `~/.cursor/projects/*/agent-transcripts/<uuid>/<uuid>.jsonl`
+- New `--source cursor` filter; results tagged `[cursor]`
+- Decodes workspace slug to filesystem path for `--project` matching
+- Session timestamps use file mtime (Cursor JSONL has no message timestamps)
+- `read_session.py` detects Cursor transcripts by path under `agent-transcripts/`
+- Resume: `agent --resume <session-id>` (see SKILL.md)
+
+### Upgrading to 0.5.0
+
+Run `--reindex` once to pull Cursor sessions into the index:
+
+```bash
+python3 ~/.claude/skills/recall/scripts/recall.py --reindex "test"
+```
+
 ## 0.4.1
 
 - Make the positional `query` argument optional. When omitted, list every

README.md

@@ -1,6 +1,6 @@
 # recall
 
-Ever lost a conversation session with Claude Code, Codex, or pi and wish you could resume it? This skill lets your agents search across all your past conversations with full-text search. Builds a SQLite FTS5 index over `~/.claude/projects/`, `~/.codex/sessions/`, and `~/.pi/agent/sessions/` with BM25 ranking, Porter stemming, CJK support, and incremental updates.
+Ever lost a conversation session with Claude Code, Codex, pi, or Cursor Agent and wish you could resume it? This skill lets your agents search across all your past conversations with full-text search. Builds a SQLite FTS5 index over `~/.claude/projects/`, `~/.codex/sessions/`, `~/.pi/agent/sessions/`, and `~/.cursor/projects/*/agent-transcripts/` with BM25 ranking, Porter stemming, CJK support, and incremental updates.
 
 ## Install
 
@@ -14,11 +14,13 @@ Then use `/recall` in Claude Code (or Codex, or pi) or ask "find a past session
 ### Index
 
 ```
-  ~/.claude/projects/**/*.jsonl ──┐
-                                  │
-  ~/.codex/sessions/**/*.jsonl ───┼─▶ Index ──▶ ~/.recall.db (SQLite FTS5)
-                                  │   [incremental - mtime-based]
-  ~/.pi/agent/sessions/**/*.jsonl ┘
+  ~/.claude/projects/**/*.jsonl ──────────────┐
+                                              │
+  ~/.codex/sessions/**/*.jsonl ───────────────┼─▶ Index ──▶ ~/.recall.db (SQLite FTS5)
+                                              │   [incremental - mtime-based]
+  ~/.pi/agent/sessions/**/*.jsonl ────────────┤
+                                              │
+  ~/.cursor/projects/**/agent-transcripts/ ───┘
 ```
 ### Query
 ```
@@ -43,7 +45,7 @@ Then use `/recall` in Claude Code (or Codex, or pi) or ask "find a past session
 - CJK messages are selectively indexed into the trigram table; query routing is automatic
 - Skips tool_use, tool_result, thinking, and image blocks
 - Results ranked by BM25 with a slight recency bias (recent sessions get up to a 20% boost, decaying with a 30-day half-life)
-- Results tagged `[claude]`, `[codex]`, or `[pi]` with highlighted excerpts
+- Results tagged `[claude]`, `[codex]`, `[pi]`, or `[cursor]` with highlighted excerpts
 - No dependencies — Python 3.9+ stdlib only (sqlite3, json, argparse)
 
 ## Tests
@@ -53,9 +55,9 @@ python3 -m unittest discover tests -v
 ```
 
 Stdlib `unittest` only — no test deps. Synthetic JSONL fixtures generated
-in `tmpdir` from the suite itself (no fixture files committed). An
-integration test runs against any real pi sessions in `~/.pi/agent/sessions/`
-on the host and is skipped if none are present.
+in `tmpdir` from the suite itself (no fixture files committed). Integration
+tests run against real sessions on the host when present (`~/.pi/agent/sessions/`
+for pi, `~/.cursor/projects/` for Cursor) and are skipped otherwise.
 
 ## Contributing
 

SKILL.md

@@ -1,23 +1,23 @@
 ---
 name: recall
 description: >
-  Search past Claude Code, Codex, and pi sessions. Triggers: /recall, "search old conversations",
-  "find a past session", "recall a previous conversation", "search session history",
-  "what did we discuss", "remember when we"
+  Search past Claude Code, Codex, pi, and Cursor Agent sessions. Triggers: /recall,
+  "search old conversations", "find a past session", "recall a previous conversation",
+  "search session history", "what did we discuss", "remember when we"
 metadata:
   author: arjunkmrm
-  version: "0.4.1"
+  version: "0.5.0"
   license: MIT
 ---
 
-# /recall — Search Past Claude, Codex & pi Sessions
+# /recall — Search Past Claude, Codex, pi & Cursor Sessions
 
-Search all past Claude Code, Codex, and pi sessions using full-text search with BM25 ranking.
+Search all past Claude Code, Codex, pi, and **Cursor Agent** (IDE + CLI) sessions using full-text search with BM25 ranking.
 
 ## Usage
 
 ```bash
-python3 ~/.claude/skills/recall/scripts/recall.py [QUERY] [--project PATH] [--days N] [--source claude|codex|pi] [--limit N] [--reindex]
+python3 ~/.claude/skills/recall/scripts/recall.py [QUERY] [--project PATH] [--days N] [--source claude|codex|pi|cursor] [--limit N] [--reindex]
 ```
 
 ## Examples
@@ -26,8 +26,8 @@ python3 ~/.claude/skills/recall/scripts/recall.py [QUERY] [--project PATH] [--da
 # List every session in the last day (no text search)
 python3 ~/.claude/skills/recall/scripts/recall.py --days 1
 
-# List every pi session in the last week
-python3 ~/.claude/skills/recall/scripts/recall.py --days 7 --source pi
+# List every Cursor session in the last week
+python3 ~/.claude/skills/recall/scripts/recall.py --days 7 --source cursor
 
 # Simple keyword search
 python3 ~/.claude/skills/recall/scripts/recall.py "bufferStore"
@@ -53,7 +53,10 @@ python3 ~/.claude/skills/recall/scripts/recall.py "buffer" --source codex
 # Search only pi sessions
 python3 ~/.claude/skills/recall/scripts/recall.py "buffer" --source pi
 
-# Force reindex
+# Search only Cursor Agent sessions (IDE + CLI)
+python3 ~/.claude/skills/recall/scripts/recall.py "sorting" --source cursor --project quick-gtasks
+
+# Force reindex (required once after upgrading to 0.5.0 for Cursor)
 python3 ~/.claude/skills/recall/scripts/recall.py --reindex "test"
 ```
 
@@ -82,6 +85,12 @@ codex resume SESSION_ID
 # Pi sessions [pi]
 cd /path/to/project
 pi --session SESSION_ID         # full or partial id; pi resolves prefix matches
+
+# Cursor Agent sessions [cursor] — IDE or CLI (agent)
+cd /path/to/project
+agent --resume SESSION_ID       # UUID from recall output
+# or: agent resume              # latest session
+# or: agent ls                  # pick interactively
 ```
 
 Each result includes a `File:` path. Use it to read the raw transcript (auto-detects format):
@@ -95,12 +104,23 @@ If results are missing `File:` paths, run `--reindex` to backfill.
 ## Notes
 
 - Index is stored at `~/.recall.db` (SQLite FTS5, auto-migrated from `~/.claude/recall.db`)
-- Indexes three sources: `~/.claude/projects/` (Claude Code), `~/.codex/sessions/` (Codex), and `~/.pi/agent/sessions/` (pi)
+- Indexes four sources:
+  - `~/.claude/projects/` — Claude Code
+  - `~/.codex/sessions/` — Codex
+  - `~/.pi/agent/sessions/` — pi
+  - `~/.cursor/projects/*/agent-transcripts/` — **Cursor Agent** (IDE + `agent` CLI)
 - First run indexes all sessions (a few seconds); subsequent runs are incremental
 - Only user and assistant messages are indexed (tool calls, thinking blocks, state snapshots skipped)
-- Results show `[claude]`, `[codex]`, or `[pi]` tags to indicate the source
+- Results show `[claude]`, `[codex]`, `[pi]`, or `[cursor]` tags to indicate the source
 - Dual-table FTS: English queries use Porter stemming, CJK queries use trigram matching
 - Omit the query argument for **list mode** — every session in the window, sorted by recency, no FTS
 - Provide a query for full-text search; both modes accept `--project`, `--days`, `--source`, `--limit`
+- **Upgrading to 0.5.0**: run `--reindex` once to pull in Cursor Agent sessions
 - **Upgrading from 0.3.x**: run `--reindex` once to pull in pi sessions
 - **Upgrading from 0.2.x**: run `--reindex` once to build the CJK index
+
+## Cursor-specific behavior
+
+- **Project path**: prefer `working_directory` from Shell tool calls in the transcript; fall back to decoding the encoded folder name under `~/.cursor/projects/` (hyphens in folder names like `quick-gtasks` are not always recoverable from the slug alone).
+- Session timestamps use file modification time (Cursor JSONL has no per-message timestamps).
+- `--project` matches path prefix **or** substring (e.g. `quick-gtasks` matches `C:\Users\you\src\quick-gtasks`).

scripts/read_session.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Pretty-print a Claude Code, Codex, or pi session transcript."""
+"""Pretty-print a Claude Code, Codex, pi, or Cursor Agent session transcript."""
 
 import json
 import sys
@@ -44,7 +44,16 @@ def iter_messages(path):
             if entry.get("record_type") == "state":
                 continue
 
-            if fmt == "pi":
+            if fmt == "cursor":
+                role = entry.get("role", "")
+                if role not in ("user", "assistant"):
+                    continue
+                msg = entry.get("message", {})
+                if not isinstance(msg, dict):
+                    continue
+                content = msg.get("content", "")
+
+            elif fmt == "pi":
                 # Pi: {type, id, parentId, timestamp, message: {role, content, ...}}
                 # Header is {type: "session", id, cwd, version, ...} — skip.
                 etype = entry.get("type", "")
@@ -105,14 +114,19 @@ def iter_messages(path):
 
 
 def detect_format(path):
-    """Detect whether a session file is Claude Code, Codex, or pi format.
+    """Detect whether a session file is Claude Code, Codex, pi, or Cursor format.
 
     Detection runs on the first non-empty parseable line and returns one of
-    "pi", "claude", or "codex". Order matters: pi headers carry both `type:
-    "session"` and `cwd`, which is the most distinctive signature; Claude
-    files have `parentUuid` or a top-level `message`; Codex files have
-    `record_type`, `instructions`, or `type: "session_meta"`.
+    "cursor", "pi", "claude", or "codex". Cursor transcripts live under
+    ~/.cursor/projects/.../agent-transcripts/; pi headers carry `type:
+    "session"` and `cwd`; Claude files have `parentUuid` or a top-level
+    `message`; Codex files have `record_type`, `instructions`, or
+    `type: "session_meta"`.
     """
+    norm = str(path).replace("\\", "/")
+    if "/.cursor/projects/" in norm and "/agent-transcripts/" in norm:
+        return "cursor"
+
     with open(path, "r", encoding="utf-8", errors="replace") as f:
         for line in f:
             line = line.strip()
@@ -145,7 +159,9 @@ def detect_format(path):
 
 def main():
     import argparse
-    parser = argparse.ArgumentParser(description="Pretty-print a Claude Code, Codex, or pi session transcript")
+    parser = argparse.ArgumentParser(
+        description="Pretty-print a Claude Code, Codex, pi, or Cursor Agent session transcript"
+    )
     parser.add_argument("path", help="Path to a session .jsonl file")
     parser.add_argument("--pretty", action="store_true", help="Human-readable output instead of JSON")
     args = parser.parse_args()