replen 1.0.4 → 1.0.6

package/extras/skills/replen-match/SKILL.md

@@ -207,14 +207,32 @@ that actually fit."* If the user agrees, run this checklist:
    what it is · stack · niche/domain · active areas · constraints/non-goals ·
    anti-patterns · integration preferences). Use the project's real domain
    vocabulary, not abstractions.
-3. **Register + tag.** Add the repo to Replen at **app.replen.dev/projects**
-   (toggle it watched), or it auto-registers on the next
-   `npx replen sync-projects`. Set **tags** that narrow the firehose to the
-   project's domain — derive them from the code (e.g. for a Python CCXT
-   trading engine: `crypto, trading, ccxt, python, quant, backtesting`) — at
-   /projects.
-4. **Re-run.** Once it has a remote + docs + tags, call `replen_match` again —
-   now it scopes to the project and matches against the real code.
+3. **Register + tag.** Register the repo: `npx replen sync-projects` (scans the
+   local repos and pushes them to Replen). Then **set the domain tags yourself
+   with the `replen_set_tags` tool** — derive them from the code you just read
+   (e.g. for a Python CCXT market-making engine:
+   `["crypto","trading","market-making","ccxt","quant","backtesting"]`).
+   **Do NOT tell the user to set tags on the web** — that's the sticky step this
+   replaces; set them with the tool. (They can still fine-tune later at
+   app.replen.dev/projects.) Tags matter most right after onboarding, before the
+   project has an embedding — without them a fresh project falls back to
+   language-only matching and surfaces noise.
+4. **Embed it now (don't wait for the daily run).** A freshly-registered
+   project has no embedding yet, so matching falls back to language/tags only
+   (noise) until the next scheduled run. Trigger an immediate run with the
+   `replen_run` tool — it builds the project's summary + embedding + initial
+   candidates from the README/CLAUDE.md you just pushed. It's async: poll
+   `replen_status` until the phase reports inventory ready (~1–3 min). Tell the
+   user it's processing.
+5. **Re-run.** Once the run finishes, call `replen_match` again — now scoped
+   AND embedded, matching against the real code (a dev-tool that only shared the
+   project's language now scores low on cosine and gets floored out).
+
+Note on recording actions: always use the MCP tools — `replen_state` (star /
+hide / handoff), `replen_record_triage` (your verdict), `replen_set_tags` — for
+any write back to Replen. Don't hand-roll `curl` to the API for these; the MCP
+path is the intended mechanism and avoids tripping host permission classifiers
+on the candidate repo name in a curl payload.
 
 **Candidate's README is unreachable (WebFetch 404)**. Note it in the
 writeup (`Caveats: README unreachable; verdict based on description

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "replen",
-  "version": "1.0.4",
+  "version": "1.0.6",
   "description": "Make your AI coding tools smarter. One command, no API keys. Replen scouts the OSS firehose against your projects and surfaces drop-in libraries, ideas to port, and dead deps to swap — the match decision happens inside your AI tool's session on your subscription tokens. 1-3 actionable matches a month, by design.",
   "type": "module",
   "bin": {