@lbruton/specflow 3.5.13 → 3.5.14

package/README.md

@@ -152,7 +152,7 @@ Each agent reads the current spec state, advances the workflow, and writes the r
 
 ### Claude Code
 
-**Step 1 — Install the MCP server** (provides tools and prompts):
+**Step 1 — Install the MCP server** (provides tools, prompts, and templates):
 
 Add to your user-level settings (`~/.claude/settings.json`):
 
@@ -167,16 +167,25 @@ Add to your user-level settings (`~/.claude/settings.json`):
 }
 ```
 
-**Step 2 — Install the plugin** (optional — provides skills and slash commands):
+The npm package contains the MCP server, dashboard, and the 9 spec templates (requirements, design, tasks, implementer-prompt, spec-reviewer, code-quality-reviewer, product, tech, structure). Templates auto-populate `DocVault/specflow/templates/` on first run.
+
+**Step 2 — Install the skills and slash commands** (recommended — provides `/spec`, `/prime`, `/wrap`, etc.):
 
 ```bash
-git clone https://github.com/lbruton/specflow.git
-cp -r specflow/plugin/ ~/.claude/plugins/marketplaces/specflow-marketplace/
+git clone https://github.com/lbruton/specflow.git ~/specflow-source
+cp -r ~/specflow-source/skills/* ~/.claude/skills/
+cp -r ~/specflow-source/commands/* ~/.claude/commands/
 ```
 
-Skills like `/prime`, `/wrap`, `/spec` are available immediately after restarting Claude Code.
+Or, if you prefer not to clone the repo, [download the latest ZIP](https://github.com/lbruton/specflow/archive/refs/heads/main.zip), unpack it, and copy the `skills/` and `commands/` directories into your `~/.claude/` directory.
+
+After restarting Claude Code, the new skills (`/prime`, `/wrap`, `/spec`, `/audit`, `/publish-templates`, `/migrate-skill`) and the slash commands they invoke will be available.
 
-> **Note:** The plugin provides skills and commands only — it does not start the MCP server. The MCP server must be installed separately via Step 1.
+> **Architecture:** The two install steps are independent and serve different purposes:
+> - **The npm package** (`@lbruton/specflow`) ships the MCP server, dashboard, and bundled spec templates. It is the engine.
+> - **The GitHub repo** (`specflow/skills/` and `specflow/commands/`) ships the markdown skills and slash commands that Claude reads directly. They are the user-facing workflows.
+>
+> Skills are NEVER bundled into the npm package — they are plain markdown installed by copying. See [`CLAUDE.md`](./CLAUDE.md) § "Architecture & Distribution Model" for the full pipeline diagram.
 
 ### Gemini CLI — Manual Install
 

package/dist/markdown/templates/tasks-template.md

@@ -214,7 +214,7 @@ After ALL tasks are [x] and implementation logs are recorded:
   - Purpose: Cross-model peer review catches blind spots a single AI might miss; using a different model architecture surfaces different classes of bugs
   - _Leverage: `/codex:review` skill, `codex:codex-rescue` subagent, branch git diff_
   - _Requirements: All_
-  - _Prompt: Implement the task for spec {{spec-name}}, first run spec-workflow-guide to get the workflow guide then implement the task: Role: Code Review Coordinator | Task: First, check if this task should be skipped: (a) if running from Codex (check $CODEX_SESSION or session origin), skip and note "skipped: running from codex". (b) if Codex CLI not installed and codex:codex-rescue agent type unavailable, skip and note "skipped: codex not available". Otherwise, dispatch /codex:review (or codex:codex-rescue subagent) for an independent code review of the branch diff. Review the findings. Fix any Critical or Important issues. Minor issues are advisory. Record the review results (or skip reason) in the implementation log. | Restrictions: Do not skip this step without invoking the guards above. Do not dismiss findings without investigating. If skipped due to guards, the skip reason MUST be logged. | Success: Codex review completed with Critical/Important issues addressed, OR task is skipped with documented reason. Results logged in implementation log. BLOCKING: After review (or skip), you MUST call the log-implementation tool with review findings (or skip reason) before marking [x]. Do NOT mark [x] until the log-implementation tool call succeeds._
+  - _Prompt: Implement the task for spec {{spec-name}}, first run spec-workflow-guide to get the workflow guide then implement the task: Role: Code Review Coordinator | Task: First, check if this task should be skipped: (a) if running from Codex (check $CODEX_SESSION or session origin), skip and note "skipped: running from codex". (b) if Codex CLI not installed and codex:codex-rescue agent type unavailable, skip and note "skipped: codex not available". Otherwise, choose the Codex invocation path based on the current dispatch context: PREFERRED — if executing in main context (not inside a Claude Agent subagent dispatch), invoke /codex:review --base <main-or-dev> --background directly (purpose-built tool, review prompt baked in). AGENT-DISPATCH CONTEXT — if executing inside a Claude Agent subagent dispatch (slash commands are unreachable from there), pause the orchestration loop, surface the exact command /codex:review --base <ref> --background for the user to run manually in main context, and resume the loop only when the user reports results. DO NOT substitute codex:codex-rescue for a branch-diff review — that subagent is reserved for delegating investigation, bug-fix, or continuation tasks, not code review; substituting it reproduces the PAPiTA PAP-1 hang. If codex:codex-rescue must be dispatched for a legitimate delegation reason, use fully-qualified subagent_type: "codex:codex-rescue" with run_in_background: true per global CLAUDE.md rules. After dispatching /codex:review, poll progress via /codex:status <task-id> — do NOT tail the JSONL output file directly (opaque and lags behind). IF the task remains in phase: starting with no progress for more than ~3 minutes, cancel via /codex:cancel <task-id> and re-dispatch from main context via the slash-command path; NEVER fall back to codex:codex-rescue on retry. Review the findings. Fix any Critical or Important issues. Minor issues are advisory. Record the review results (or skip reason) in the implementation log. | Restrictions: Do not skip this step without invoking the guards above. Do not dismiss findings without investigating. If skipped due to guards, the skip reason MUST be logged. | Success: Codex review completed with Critical/Important issues addressed, OR task is skipped with documented reason. Results logged in implementation log. BLOCKING: After review (or skip), you MUST call the log-implementation tool with review findings (or skip reason) before marking [x]. Do NOT mark [x] until the log-implementation tool call succeeds._
 
 - [ ] C7. Loop or complete
   - **Recommended Agent:** Claude

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lbruton/specflow",
-  "version": "3.5.13",
+  "version": "3.5.14",
   "description": "MCP server for spec-driven development workflow with real-time web dashboard",
   "main": "dist/index.js",
   "type": "module",