crispy-recall 0.1.3 → 0.1.5

package/README.md

@@ -19,7 +19,7 @@ recall install
 llama embedding binary, the model, and the SQLite DB), wires a Stop hook into
 Claude Code so your transcripts are indexed automatically as sessions end, and
 installs the `recall` skill. If Codex is detected it also gets the `recall`
-skill (so the agent can search), but not an automatic per-turn hook in v0.1.0 —
+skill (so the agent can search), but not an automatic per-turn hook (yet) —
 index Codex history with `recall backfill --vendor codex`. After that, recall
 runs passively — you only invoke `recall` directly for `status`, `doctor`,
 `repair`, or `uninstall`.
@@ -37,8 +37,8 @@ Windows-native are separate installs.** If you use both, run the install in each
 Prerequisites: Node ≥ 20 and Claude Code installed. If Codex is detected
 (`~/.codex/` exists), recall installs the `recall` skill into Codex (so the
 agent can search) and you can index your Codex history with
-`recall backfill --vendor codex`. Real-time per-turn Codex indexing is **not**
-in v0.1.0.
+`recall backfill --vendor codex`. Real-time per-turn Codex indexing is **not
+yet** supported.
 
 ## What it does
 
@@ -52,23 +52,38 @@ in v0.1.0.
 Two steps. First, search for the relevant session:
 
 ```bash
-$RECALL_BIN "the thing you're trying to remember"
+recall "the thing you're trying to remember"
 ```
 
 That returns a table of matching sessions. Then read one, centered on the match:
 
 ```bash
-$RECALL_BIN <session-id> <message-id>
+recall <session-id> <message-id>
+```
+
+Search and list default to the **current directory's project** — from inside one
+repo you only see that repo's sessions. Add `--all` to search across every
+indexed project (use it for cross-repo questions, or when a scoped search comes
+back thin), or `--project <path>` to target a specific repo regardless of where
+you are:
+
+```bash
+recall --all "the thing you're trying to remember"
 ```
 
 The skill's frontmatter teaches Claude *when* to reach for this (before
 non-trivial tasks, architectural decisions, or work with obvious prior
-history), so in practice the agent calls it for you.
+history), so in practice the agent calls it for you. (Inside the installed skill
+the command is written out as `node ~/.recall/bin/recall.js` so it works even
+when `recall` isn't on the agent's `PATH` — `$RECALL_BIN` in the skill source is
+a placeholder the installer substitutes, not an environment variable you set.)
 
 ## Commands
 
 | Command | What it does |
 |---|---|
+| `recall "<query>" [--all] [--project <path>]` | Search past sessions (FTS5 + semantic). Defaults to the current project; `--all` searches every indexed project, `--project` targets one. |
+| `recall <session-id> [<message-id>]` | Read a session, optionally centered on a matched message. |
 | `recall install` | One-time setup: scaffold `~/.recall/`, wire the Stop hook, install the skill. |
 | `recall uninstall` | Reverse the install (skill, hook, CLAUDE.md block). `--purge` also removes `~/.recall/`. |
 | `recall status` | DB size, message count, last ingest, embedding gap, active backfill PID, GPU/CPU backend. |
@@ -76,6 +91,11 @@ history), so in practice the agent calls it for you.
 | `recall repair --fts \| --vectors \| --full` | Rebuild the FTS index, re-embed vectors, or full reingest from JSONL. |
 | `recall backfill [--auto-embed] [--vendor <v>] [--detach]` | Index historical transcripts. |
 
+Add `--json` to `install`/`uninstall`/`status`/`doctor` for machine-readable
+output. `recall install` also takes `--offline` (use a pre-staged binary +
+model instead of downloading) and `--no-backfill` / `--auto-backfill` to control
+the initial history index. Run `recall --help` for the full flag set.
+
 ## Where things live
 
 - `~/.recall/` — the DB, model, binary, logs, and `config.json` (the resolved GPU/CPU embedder mode).
@@ -86,9 +106,11 @@ history), so in practice the agent calls it for you.
 
 ## Privacy
 
-Everything is local. No telemetry, no network calls except the one-time
-binary + model download at install. The DB is plain SQLite — open it with any
-SQLite browser if you want to poke around.
+Everything is local. No telemetry. The only data transfer is the one-time
+binary + model download at install (from llama.cpp's GitHub releases and
+HuggingFace); `recall install` and `recall doctor` also send lightweight
+reachability probes to those two hosts. Nothing else leaves your machine. The DB
+is plain SQLite — open it with any SQLite browser if you want to poke around.
 
 ## Troubleshooting
 

package/dist/SKILL.md.template

@@ -71,6 +71,12 @@ DATE HANDLING: Any date the user mentions MUST become a flag — never search te
   --recent             Boost recent sessions (use when user says 'recently', 'latest')
   Example: \"what happened April 10\" → $RECALL_BIN --list --since 2026-04-10 --until 2026-04-10
 
+PROJECT SCOPING: search and list default to the CURRENT directory's project.
+  --all                Search across ALL projects (use when the answer may live
+                       in another repo, or when a scoped search returns little)
+  --project PATH       Scope to a specific project path instead of the CWD
+  If results look thin or empty, retry with --all before concluding nothing exists.
+
 CRITICAL READING RULE: After searching, ALWAYS read the matched message:
   $RECALL_BIN <session-id> <message-id>
 This auto-centers on the match and shows surrounding turns. The output
@@ -192,6 +198,33 @@ $RECALL_BIN "old bug" --until 2026-03-15
 `$RECALL_BIN --list --since 2026-04-10 --until 2026-04-10` to find all sessions,
 then search with topic keywords + date flags if needed.
 
+## Project scope
+
+**Search and list default to the current working directory's project.** Recall
+filters results to the project whose path is your CWD — so from inside one repo
+you only see that repo's sessions. This keeps everyday recall focused, but it
+silently hides history when the answer lives elsewhere.
+
+```bash
+# Default — scoped to the CWD's project
+$RECALL_BIN "auth refactor"
+
+# Search every project (cross-repo questions, or when a scoped search is thin)
+$RECALL_BIN "auth refactor" --all
+
+# Scope to a specific project regardless of CWD
+$RECALL_BIN "auth refactor" --project /home/me/dev/other-repo
+```
+
+- `--all` — disable scoping; search across all indexed projects
+- `--project PATH` — scope to an explicit project path instead of the CWD
+- Both work in search and list modes
+
+**When to reach for `--all`:** the user asks about work that may span repos, the
+question isn't tied to the current repo, or a default (scoped) search returns
+few/zero results. A thin scoped result is often just the scope — retry with
+`--all` before concluding the history doesn't exist.
+
 ## Recency boost
 
 When the user says "recently", "latest", "last few days", or otherwise indicates
@@ -209,4 +242,5 @@ pushing recent sessions to the top of results. Combine with `--since` for best r
 - **Search is cheap, reading is expensive.** Run 3-5 varied queries before committing to reading sessions.
 - **Message IDs are stable.** You can reference them across searches.
 - **Use `--since` / `--until` to scope.** Both search and list modes accept date flags.
+- **Search is scoped to the CWD's project by default.** Add `--all` for cross-repo questions or when a scoped search comes back thin.
 - **Raw JSON output** (`--raw`) is available for programmatic processing.

package/dist/embed-pending.js

@@ -57,6 +57,9 @@ function runDir() {
 function ensureDir() {
   (0, import_node_fs.mkdirSync)(recallRoot(), { recursive: true });
 }
+function transcriptGlob(...segments) {
+  return segments.join("/").replace(/\\/g, "/");
+}
 var import_node_fs, import_node_os, import_node_path, rootOverride;
 var init_paths = __esm({
   "src/paths.ts"() {
@@ -6416,9 +6419,14 @@ function extractEntryText(entry) {
   }
   return "";
 }
+function entriesCwd(entries) {
+  for (const e of entries) {
+    if (e.cwd)
+      return e.cwd;
+  }
+  return void 0;
+}
 async function ingestSessionMessages(sessionId, transcriptPath, vendor, options) {
-  const rawProjectId = options?.projectId ?? null;
-  const projectId = rawProjectId ? normalizePath(rawProjectId) : null;
   let rawEntries;
   try {
     if (vendor === "claude") {
@@ -6439,6 +6447,8 @@ async function ingestSessionMessages(sessionId, transcriptPath, vendor, options)
   if (rawEntries.length === 0) {
     return { sessionId, chunksCreated: 0, skipped: true };
   }
+  const rawProjectId = options?.projectId ?? entriesCwd(rawEntries);
+  const projectId = rawProjectId ? normalizePath(rawProjectId) : null;
   const filtered = stripToolContent(rawEntries);
   const topLevel = filtered.filter((e) => !e.parentToolUseID);
   const records = [];
@@ -13203,9 +13213,9 @@ async function mtimeScan(opts) {
   const codexRoot = process.env["CODEX_HOME"] ?? (0, import_node_path7.join)((0, import_node_os3.homedir)(), ".codex");
   const patterns = [];
   if (vendors.includes("claude"))
-    patterns.push([(0, import_node_path7.join)(claudeRoot, "projects", "**", "*.jsonl"), "claude"]);
+    patterns.push([transcriptGlob(claudeRoot, "projects", "**", "*.jsonl"), "claude"]);
   if (vendors.includes("codex"))
-    patterns.push([(0, import_node_path7.join)(codexRoot, "sessions", "**", "*.jsonl"), "codex"]);
+    patterns.push([transcriptGlob(codexRoot, "sessions", "**", "*.jsonl"), "codex"]);
   const db3 = getDb(dbPath());
   const watermarks = /* @__PURE__ */ new Map();
   const rows = db3.all("SELECT * FROM ingest_watermark");

package/dist/recall.js

@@ -2570,6 +2570,9 @@ function logsDir() {
 function ensureDir() {
   (0, import_node_fs.mkdirSync)(recallRoot(), { recursive: true });
 }
+function transcriptGlob(...segments) {
+  return segments.join("/").replace(/\\/g, "/");
+}
 var import_node_fs, import_node_os, import_node_path, rootOverride;
 var init_paths = __esm({
   "src/paths.ts"() {
@@ -13613,9 +13616,14 @@ function extractEntryText(entry) {
   }
   return "";
 }
+function entriesCwd(entries) {
+  for (const e2 of entries) {
+    if (e2.cwd)
+      return e2.cwd;
+  }
+  return void 0;
+}
 async function ingestSessionMessages(sessionId, transcriptPath, vendor, options) {
-  const rawProjectId = options?.projectId ?? null;
-  const projectId = rawProjectId ? normalizePath(rawProjectId) : null;
   let rawEntries;
   try {
     if (vendor === "claude") {
@@ -13636,6 +13644,8 @@ async function ingestSessionMessages(sessionId, transcriptPath, vendor, options)
   if (rawEntries.length === 0) {
     return { sessionId, chunksCreated: 0, skipped: true };
   }
+  const rawProjectId = options?.projectId ?? entriesCwd(rawEntries);
+  const projectId = rawProjectId ? normalizePath(rawProjectId) : null;
   const filtered = stripToolContent(rawEntries);
   const topLevel = filtered.filter((e2) => !e2.parentToolUseID);
   const records = [];
@@ -13757,9 +13767,9 @@ async function mtimeScan(opts) {
   const codexRoot = process.env["CODEX_HOME"] ?? (0, import_node_path6.join)((0, import_node_os3.homedir)(), ".codex");
   const patterns = [];
   if (vendors.includes("claude"))
-    patterns.push([(0, import_node_path6.join)(claudeRoot, "projects", "**", "*.jsonl"), "claude"]);
+    patterns.push([transcriptGlob(claudeRoot, "projects", "**", "*.jsonl"), "claude"]);
   if (vendors.includes("codex"))
-    patterns.push([(0, import_node_path6.join)(codexRoot, "sessions", "**", "*.jsonl"), "codex"]);
+    patterns.push([transcriptGlob(codexRoot, "sessions", "**", "*.jsonl"), "codex"]);
   const db4 = getDb(dbPath());
   const watermarks = /* @__PURE__ */ new Map();
   const rows = db4.all("SELECT * FROM ingest_watermark");
@@ -14556,7 +14566,7 @@ function listAllSessions(opts) {
   const codexRoot = process.env["CODEX_HOME"] ?? (0, import_node_path7.join)((0, import_node_os4.homedir)(), ".codex");
   const out = [];
   if (vendors.includes("claude")) {
-    const files = globSync((0, import_node_path7.join)(claudeRoot, "projects", "**", "*.jsonl"), { nodir: true });
+    const files = globSync(transcriptGlob(claudeRoot, "projects", "**", "*.jsonl"), { nodir: true });
     for (const file of files) {
       out.push({
         sessionId: sessionIdFromPath2(file, "claude"),
@@ -14567,7 +14577,7 @@ function listAllSessions(opts) {
     }
   }
   if (vendors.includes("codex")) {
-    const files = globSync((0, import_node_path7.join)(codexRoot, "sessions", "**", "*.jsonl"), { nodir: true });
+    const files = globSync(transcriptGlob(codexRoot, "sessions", "**", "*.jsonl"), { nodir: true });
     for (const file of files) {
       out.push({
         sessionId: sessionIdFromPath2(file, "codex"),
@@ -14593,6 +14603,7 @@ var init_session_manager_shim = __esm({
     init_esm5();
     import_node_os4 = require("node:os");
     import_node_path7 = require("node:path");
+    init_paths();
   }
 });
 

package/dist/stop-hook.js

@@ -2663,9 +2663,14 @@ function extractEntryText(entry) {
   }
   return "";
 }
+function entriesCwd(entries) {
+  for (const e of entries) {
+    if (e.cwd)
+      return e.cwd;
+  }
+  return void 0;
+}
 async function ingestSessionMessages(sessionId, transcriptPath, vendor, options) {
-  const rawProjectId = options?.projectId ?? null;
-  const projectId = rawProjectId ? normalizePath(rawProjectId) : null;
   let rawEntries;
   try {
     if (vendor === "claude") {
@@ -2686,6 +2691,8 @@ async function ingestSessionMessages(sessionId, transcriptPath, vendor, options)
   if (rawEntries.length === 0) {
     return { sessionId, chunksCreated: 0, skipped: true };
   }
+  const rawProjectId = options?.projectId ?? entriesCwd(rawEntries);
+  const projectId = rawProjectId ? normalizePath(rawProjectId) : null;
   const filtered = stripToolContent(rawEntries);
   const topLevel = filtered.filter((e) => !e.parentToolUseID);
   const records = [];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "crispy-recall",
-  "version": "0.1.3",
+  "version": "0.1.5",
   "description": "Local session transcript memory for Claude Code and Codex — search past sessions with FTS5 + semantic vectors.",
   "license": "MIT",
   "author": "Sylvester Wong",