prjct-cli 2.49.1 → 2.50.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "prjct-cli",
-  "version": "2.49.1",
+  "version": "2.50.0",
   "description": "Context layer for AI agents. Project context for Claude Code, Gemini CLI, and more.",
   "main": "dist/bin/prjct.mjs",
   "bin": {

package/templates/skills/prjct/SKILL.md

@@ -66,6 +66,19 @@ Heavy quality workflows (`review`, `qa`, `security`, `investigate`, `audit`, `au
 
 **CONTENT LANGUAGE — author every stored memory in ENGLISH**, no matter what language the user speaks (Spanish, Japanese, German — any). When you `capture`/`remember`, translate the intent into a clean English entry; the persisted knowledge is always English. LLMs comprehend English better and embeddings stay high-quality in one canonical language — mixed-language content produces cross-language retrieval noise and extra token cost on every later recall.
 
+## Loop discipline — stop, delegate, or audit (keep the orchestrator thin)
+
+Concrete triggers that keep the main thread thin and stop you from working forward over a broken state. When one fires, do the action BEFORE continuing. These do NOT contradict `default DIRECT` — they mark WHEN direct stops being safe.
+
+| Trigger | Do this before continuing |
+|---|---|
+| Reading **4+ files** just to understand a flow | Delegate exploration to a fresh-context subagent (`general-purpose`, `model: "sonnet"`) — it returns the map; your context stays clean. |
+| Touching **2+ non-trivial files** | Keep ONE writer (no fan-out onto shared files), then a fresh `review` before you call it done. |
+| About to **commit / push / open a PR** after code changes | Run `review` first — skip ONLY for a trivial docs/text/version diff. |
+| **Wrong cwd, worktree/git accident, merge recovery, or a confusing test/env failure** | STOP. Re-orient or run `audit` before any more edits — never debug forward over a broken state. |
+
+Model on EVERY dispatch (quick-reference — never omit `model:`): the implementer that writes code → `model: "opus"`; any reviewer/judge (`review`/`security`/`investigate`/`audit-spec`) → `model: "sonnet"`; pure routing/orchestration → `model: "haiku"`. Rationale + fan-out rules in `workflows.md`.
+
 ## Verb intent map — you run the verb, the user never types it
 
 On every turn ask: "what is the user trying to accomplish?" and match to a verb below. The user may phrase intent in ANY language — the verbs are language-agnostic. These are *signals*, not phrase templates. The **Tier** column governs whether you auto-run or confirm first (see Routing).