@clauderecallhq/cli 0.77.1 → 0.92.3

package/README.md

@@ -138,7 +138,7 @@ recall search "auth #auth-fix" -p Pest-Control
 
 Local **768-dimension embeddings** via `bge-base-en-v1.5` ONNX. Three-lane RRF fusion (BM25 + summary + vectors) finds sessions by *meaning*, not keywords. Your code never leaves your laptop.
 
-The embedder model (~110MB) auto-downloads on `npm install`. Skip with `RECALL_SKIP_MODEL_DOWNLOAD=1`.
+The embedder model (~110MB) is opt-in — install on demand with `recall semantic install`. Keeps `npm install -g` quick and avoids forcing the download on users who never touch vector search.
 
 **One-click vectorization from the web UI.** Each project header has a `🧠 Vectorize` button. Click → pre-flight dialog shows scope (X of Y eligible sessions), three-✓ cost guarantees (free, on-device CPU only, no network), depth selector (Quick / Standard / Full / Custom chunks-per-session), and a **depth-aware ETA** computed server-side per cap level so Standard and Full give honest, different numbers. **Counters are per-repository:** the Queued, Throughput, and ETA tiles only count chunks from the project you opened the dialog from, with a separate small disclosure line surfacing how many chunks are queued from other projects ahead of yours (the worker drains FIFO globally). While running, the dialog flips to monitor mode with a live progress bar, ETA tick countdown, observed-throughput (drained-in-this-session) feeding the estimate, and visible **warm-up** (yellow) + **wind-down** (red) progress bars with timing context for the parts that have no deterministic ETA (embedder load is 5-60s by hardware; worker batch wind-down is 5-15s). **Three explicit confirmation states close every run:** green `✓ Vectorization complete — N chunks embedded in Xs` banner when the queue drains naturally, green `✓ Already vectorized — N{project} is already done` pre-flight callout (replacing the gray "Nothing to do" text) when there's nothing to vectorize, and the existing green `✓ Stopped — worker halted and queue cleared` badge for Stop. Stop itself is a **global kill switch** — clears every project's queue and halts the worker. The vector worker **does not auto-resume on daemon restart** by default; enable `semantic.autoResumeWorker: true` in `~/.recall/config.json` if you want background draining without a click. Free tier doesn't see the button — endpoint is Pro-gated.
 
@@ -146,13 +146,15 @@ The embedder model (~110MB) auto-downloads on `npm install`. Skip with `RECALL_S
 
 ```bash
 recall similar <session-id>
-recall semantic install                  # one-time model download (skipped if `npm install` already pulled it)
+recall semantic install                  # one-time model download (~110MB, on-device)
 recall semantic reindex                  # vectorize on local CPU (idempotent, resumable)
 recall semantic verify-spawn             # diagnostic — confirms claude CLI honors the no-persistence flag
 ```
 
 **Two lanes.** Tier-2 (commands above — `install`, `reindex`, `verify-spawn`) is free, local, and never sends anything anywhere. Tier-1 (`recall semantic on / backfill / auto-extract`) shells out to your local `claude` CLI to summarize sessions and **costs plan tokens at ~30 sessions/min** — opt-in for users who want LLM-generated summaries on top of vectors. Default-off.
 
+**Switching backends in place.** Pro users running an older ONNX-based vector index who upgrade to the llama.cpp backend can run `recall semantic migrate` to re-embed the existing corpus under the new backend. The command is resumable (SIGINT-safe), atomic at switchover, and retains a 30-day rollback window via `recall semantic rollback-migration`.
+
 ### Cost Analytics
 
 Per-session and per-project token + dollar totals. Daily sparkline. Top-10 heaviest sessions. Know exactly where your Anthropic bill is going.
@@ -164,7 +166,7 @@ recall stats --project Tools --days 7
 
 <div align="center">
 
-<img src="https://cdn.clauderecall.com/marketing/cost-analytics.gif?v=hd3" alt="Cost analytics dashboard — daily spend bars, top sessions ranked by cost, primary model breakdown" width="540">
+<img src="https://cdn.clauderecall.com/marketing/cost-analytics.gif?v=hd4" alt="Cost analytics dashboard — daily spend bars, top sessions ranked by cost, primary model breakdown" width="540">
 
 <sub>Hover any day. Exact cost. Exact tokens.</sub>
 
@@ -278,7 +280,7 @@ recall install-extension
 
 <div align="center">
 
-<img src="https://cdn.clauderecall.com/marketing/vs-obsidian.gif?v=hd3" alt="Side-by-side: Obsidian vault on the left, Claude Recall TUI on the right showing 247 sessions across 8 projects with live preview" width="900">
+<img src="https://cdn.clauderecall.com/marketing/vs-obsidian.gif?v=hd4" alt="Side-by-side: Obsidian vault on the left, Claude Recall TUI on the right showing 247 sessions across 8 projects with live preview" width="900">
 
 <sub>People sometimes ask if this is "Obsidian for Claude Code." It is not. A general note-taking app cannot watch your filesystem for new sessions, index JSONLs into FTS5 + vector search, expose those sessions as MCP tools, or pipe a session back into Claude with one command. Claude Recall is built specifically for the loop you actually run.</sub>
 
@@ -504,7 +506,7 @@ All writes are rate-limited (default 60/min), zod-validated, and audited to `~/.
 
 <div align="center">
 
-<img src="https://cdn.clauderecall.com/marketing/privacy-trust.gif?v=hd3" alt="Local-first design — daemon binds to 127.0.0.1, session content stays on your machine" width="640">
+<img src="https://cdn.clauderecall.com/marketing/privacy-trust.gif?v=hd4" alt="Local-first design — daemon binds to 127.0.0.1, session content stays on your machine" width="640">
 
 </div>
 
@@ -593,10 +595,13 @@ recall correlate               # link sessions to commits
 recall blame <sha>             # commit -> session reverse lookup
 
 # Semantic / vector search (Pro)
-# Embedder model auto-installs on `npm install`; manual install also available.
-recall semantic install        # download on-device embedding model (auto-runs on npm install)
+# Embedder model is opt-in — `npm install` does NOT auto-download it.
+recall semantic install        # download on-device embedding model (~110MB)
 recall semantic status         # model + backfill progress + auto-extract state
 recall semantic reindex        # re-embed everything
+recall semantic migrate        # re-embed existing corpus under a different backend (Pro)
+recall semantic rollback-migration --force  # restore prior corpus within 30-day window
+recall semantic prune-rollback --force      # drop backup table early
 recall semantic auto-extract on  # daemon nibbles un-extracted sessions through Claude (Pro)
 recall similar <id>            # cosine kNN over session chunks
 
@@ -676,7 +681,7 @@ Claude Recall is a Node.js CLI with a few native dependencies (`better-sqlite3`,
 
 **Node:** 22 LTS or 24 LTS. Node 20 and earlier are unsupported (declared in `engines.node`).
 
-**Semantic search is optional.** The on-device embedder is the only feature that loads the native ONNX runtime. The model auto-downloads (~110MB) on `npm install`; skip with `RECALL_SKIP_MODEL_DOWNLOAD=1` and install later via `recall semantic install`. Core CLI features (search, list, context, daemon, MCP) work on every supported platform regardless of whether the embedder is installed. If the embedder fails to load on your platform, you get a clear error pointing here, and the rest of Claude Recall keeps working.
+**Semantic search is optional.** The on-device embedder is the only feature that loads the native ONNX runtime. The model (~110MB) is opt-in — install on demand with `recall semantic install`. Core CLI features (search, list, context, daemon, MCP) work on every supported platform regardless of whether the embedder is installed. If the embedder fails to load on your platform, you get a clear error pointing here, and the rest of Claude Recall keeps working.
 
 <br />