@kraftapps-ai/kai 1.3.4 → 1.4.1

package/CLAUDE.md

@@ -7,16 +7,26 @@ Autonomous AI Shopify expert, product manager, and developer loop for Claude Cod
 - `kai` — the entire tool; a single bash script that bootstraps the PM agent and dev loop
 - `package.json` — npm metadata and version
 - `.kai/PROMPT.md` — per-project context (generated at runtime, not in repo)
+- `.kai/memory.md` — persistent memory across sessions (generated at runtime, not in repo)
 - `.kai/loop.sh` — dev loop script (generated at runtime, not in repo)
 
 ## How it works
 
-1. `kai` script auto-inits project files (`kai.json`, `kai-progress.txt`, `.kai/PROMPT.md`)
+1. `kai` script auto-inits project files (`kai.json`, `kai-progress.txt`, `.kai/PROMPT.md`, `.kai/memory.md`)
 2. Auto-installs MCP servers if not already configured:
    - **Playwright** — browser testing inside Shopify Admin
    - **@shopify/dev-mcp** — Shopify docs, GraphQL schema introspection, Liquid/GraphQL/component validation
 3. Generates `.kai/loop.sh` (the autonomous dev loop)
-4. Launches Claude as a Shopify-expert PM agent
+4. Launches Claude with `--continue` to resume the last conversation (per-repo)
+5. Falls back to `.kai/memory.md` for context if conversation history is unavailable
+
+## Session persistence
+
+Kai uses a two-layer persistence strategy:
+- **`claude --continue`** — resumes the exact last conversation in this directory (primary)
+- **`.kai/memory.md`** — Kai-maintained memory file with key decisions, current focus, and session log (fallback)
+
+Use `kai --new` to force a fresh conversation (memory.md is still loaded for context).
 
 The PM creates user stories in `kai.json`, then kicks off `.kai/loop.sh` which runs Claude in a loop — one story per iteration with self-review. Both the PM and the dev loop worker use Shopify Dev MCP tools to introspect APIs, search docs, and validate code.
 

package/kai

@@ -15,9 +15,18 @@ if [ "${1:-}" = "--version" ] || [ "${1:-}" = "-v" ]; then
     exit 0
 fi
 
+# ── Handle --new flag ─────────────────────────────────
+
+USE_CONTINUE=true
+if [ "${1:-}" = "--new" ]; then
+    USE_CONTINUE=false
+    shift
+fi
+
 PRD_FILE="kai.json"
 PROGRESS_FILE="kai-progress.txt"
 PROMPT_FILE=".kai/PROMPT.md"
+MEMORY_FILE=".kai/memory.md"
 LOOP_SCRIPT=".kai/loop.sh"
 
 # ── Auto-init if needed ───────────────────────────────
@@ -65,6 +74,29 @@ Add your project-specific context here. The more you provide, the better Kai per
 EOF
 fi
 
+# Init memory file if it doesn't exist
+if [ ! -f "$MEMORY_FILE" ]; then
+    mkdir -p .kai
+    cat > "$MEMORY_FILE" << 'EOF'
+# Kai Memory
+
+This file is automatically maintained by Kai to preserve context between sessions.
+Do not edit manually unless you want to correct or add context for Kai.
+
+## Current Focus
+<!-- What we're currently working on -->
+
+## Key Decisions
+<!-- Important decisions made during our conversations -->
+
+## Open Questions / Blockers
+<!-- Things that need to be resolved -->
+
+## Session Log
+<!-- Brief log of what happened in each session -->
+EOF
+fi
+
 # ── Ensure MCP servers are available ─────────────────
 
 # Track what's already configured
@@ -318,9 +350,53 @@ progress=""
 Recent progress (kai-progress.txt):
 $(tail -30 "$PROGRESS_FILE")"
 
+memory=""
+[ -f "$MEMORY_FILE" ] && memory="
+
+## Kai Memory (from previous sessions)
+$(cat "$MEMORY_FILE")"
+
+dev_name=$(git config user.name 2>/dev/null || echo "")
+dev_greeting=""
+[ -n "$dev_name" ] && dev_greeting="
+The developer's name is: ${dev_name}"
+
+# ── Build Claude args ─────────────────────────────────
+
+CLAUDE_ARGS="--dangerously-skip-permissions"
+if [ "$USE_CONTINUE" = true ]; then
+    CLAUDE_ARGS="$CLAUDE_ARGS --continue"
+fi
+
 # ── Launch the PM ─────────────────────────────────────
 
-exec claude --dangerously-skip-permissions --system-prompt "You are Kai — an AI product manager, tech lead, and **Shopify expert**. You work directly with the developer to plan and ship Shopify apps, themes, and integrations.
+exec claude $CLAUDE_ARGS --system-prompt "You are Kai — an AI product manager, tech lead, and **Shopify expert**. You work directly with the developer to plan and ship Shopify apps, themes, and integrations.
+
+## Session persistence
+
+You maintain a memory file at \`.kai/memory.md\` that preserves context between sessions. This is critical — it's how you remember what happened in previous conversations.
+
+### Reading memory
+- **All context is already loaded in this system prompt** — story status, project context, progress log, and memory are all below. Do NOT read files on startup.
+- Use the loaded memory to greet the developer with instant awareness of where you left off
+- If memory has content, reference it naturally: \"Hey {name} — picking up where we left off on X\"
+
+### Writing memory
+You MUST update \`.kai/memory.md\` in these situations:
+- **After creating or modifying stories** — record what was planned and why
+- **After the dev loop completes** — record what was built and any issues
+- **When the developer makes a key decision** — record the decision and reasoning
+- **When there are open questions or blockers** — record them so you remember next time
+- **Before the developer leaves** — if they say \"bye\", \"done for today\", \"gotta go\", etc., immediately update memory with a session summary
+
+### Memory format
+Keep \`.kai/memory.md\` organized with these sections:
+- **Current Focus** — what we're actively working on
+- **Key Decisions** — important choices and their reasoning
+- **Open Questions / Blockers** — unresolved issues
+- **Session Log** — brief dated entries of what happened (newest first, keep last 10)
+
+Keep entries concise. Memory is context, not a novel.
 
 ## Your Shopify expertise
 
@@ -432,5 +508,5 @@ Shopify apps are **embedded inside Shopify Admin**. Pages are NOT directly acces
 - When the developer says \"ship it\" or \"go\" or \"do it\", start the loop
 - Don't ask unnecessary questions — use good defaults
 - When unsure about an API, introspect the schema instead of guessing
-- On startup, greet the developer briefly. If there are existing stories, give a quick status. If not, ask what they want to build. Keep it to 2-3 lines max.
-${status}${project_context}${progress}" "${@:-Hey Kai}"
+- **On startup: DO NOT read any files. All context is already loaded below.** Greet the developer by name immediately. If memory has context from previous sessions, acknowledge where you left off in 1-2 lines. If fresh, ask what they want to build. Keep greeting to 2-3 lines max. Be instant — no tool calls before your first response.
+${status}${project_context}${progress}${memory}${dev_greeting}" "${@:-Hey Kai}"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kraftapps-ai/kai",
-  "version": "1.3.4",
+  "version": "1.4.1",
   "description": "Autonomous AI developer loop for Claude Code",
   "bin": {
     "kai": "kai"