feat: add worktree isolation to /implement, /validate, and /release

plugin/skills/implement/SKILL.md

@@ -13,9 +13,31 @@ Load plan, dispatch subagents per task in wave order, verify completion.
 - **Session metadata is the source of truth** — the session-start hook injects a metadata block with `epic-id`, `epic-slug`, `feature-id`, `feature-name`, `feature-slug`, parent artifacts, and `output-target`. Use these values verbatim — do NOT re-derive, re-extract, or generate alternatives.
 - **Wave ordering drives sequencing** — foundation before consumers; parallel-safe waves dispatch all tasks concurrently; reviews run sequentially after all implementations complete
 - **Model escalation** — start cheap (haiku), escalate on failure (sonnet, then opus). See Reference > Model Escalation.
-- **Working directory isolation** — the CLI provides the working directory; skills don't manage worktrees; each phase commits to the feature branch at checkpoint; merge happens only at /release
+- **Working directory isolation** — `/implement` always creates a dedicated git worktree so the main repo stays clean on its current branch and multiple features can run in parallel. The worktree persists until `/release` completes the squash-merge.
 - **Subagent safety** — one agent per task; parallel-safe waves dispatch all tasks simultaneously, non-parallel-safe waves dispatch sequentially; agents commit per task via `git add <files>` + `git commit`; agents must NOT read the plan file, modify files outside their task's file list, or push/switch branches
 
+## Phase 0: Worktree Setup
+
+Before dispatching the taskplanner, ensure work happens in an isolated worktree.
+
+### 0a. Check for existing worktree
+
+```bash
+git worktree list
+```
+
+If `.claude/worktrees/<epic-slug>` already appears (interrupted prior session), use `EnterWorktree` with `path: ".claude/worktrees/<epic-slug>"` to re-enter it, then skip to Phase 1.
+
+### 0b. Create worktree (first run)
+
+```bash
+git worktree add .claude/worktrees/<epic-slug> -b feat/<epic-slug>
+```
+
+Then use `EnterWorktree` with `path: ".claude/worktrees/<epic-slug>"` to switch the session into it.
+
+All subsequent work (taskplanner, implementers, reviewers, commits, tests) runs inside this worktree. The main repo working directory is never touched.
+
 ## Phase 1: Execute
 
 Dispatch taskplanner, loop over waves dispatching implementers and reviewers.

plugin/skills/release/SKILL.md

@@ -16,6 +16,27 @@ No release without passing validation.
 - **Bump type auto-detected, not user-prompted** — commit message conventions determine major/minor/patch automatically
 - **Warn-and-continue for non-blocking failures** — report problems, attempt fixes, only hard-stop on critical validation failures
 
+## Phase 0: Pre-Execute
+
+### 0. Enter Worktree
+
+Release must run inside the implement worktree so that release-notes generation, the retro context walker, and the feature-branch checkpoint commit all operate on the feature branch's state. The squash-merge in Phase 3 still happens from the main repo via explicit `cd "$main_repo"`; this Phase 0 step only governs the parent skill's cwd and the cwd that subagents (notably the retro) inherit.
+
+```bash
+git worktree list
+```
+
+If already inside `.claude/worktrees/<epic-slug>` (current directory matches): skip to Phase 1.
+
+If at repo root, look for `.claude/worktrees/<epic-slug>` in the worktree list. If found, use `EnterWorktree` with `path: ".claude/worktrees/<epic-slug>"` to switch into it.
+
+If no matching worktree exists, STOP:
+
+```
+BLOCKED — no worktree found for <epic-slug>.
+Run /beastmode:implement and /beastmode:validate first, then retry.
+```
+
 ## Phase 1: Execute
 
 ### 1. Stage Uncommitted Changes
@@ -136,8 +157,6 @@ Run a context reconciliation pass across all phase artifacts before releasing.
 
 4. **Apply BEASTMODE.md Updates** — If no L0 changes proposed, skip this step. Apply L0 changes and log.
 
-> **TRANSITION BOUNDARY — Steps below operate from main repo, NOT the feature branch working directory.**
-
 ### 2. Commit to Feature Branch
 
 Before merging to main, commit all release artifacts to the feature branch:
@@ -147,6 +166,8 @@ git add -A
 git commit -m "release(<epic-name>): checkpoint"
 ```
 
+> **TRANSITION BOUNDARY — Steps below operate from main repo, NOT the feature branch working directory.**
+
 ### 3. Squash Merge to Main
 
 ```bash

plugin/skills/validate/SKILL.md

@@ -20,6 +20,25 @@ No release without passing validation.
 
 ## Phase 0: Pre-Execute
 
+### 0. Enter Worktree
+
+Validate must run inside the implement worktree where the feature code lives.
+
+```bash
+git worktree list
+```
+
+If already inside `.claude/worktrees/<epic-slug>` (current directory matches): skip to step 1.
+
+If at repo root, look for `.claude/worktrees/<epic-slug>` in the worktree list. If found, use `EnterWorktree` with `path: ".claude/worktrees/<epic-slug>"` to switch into it.
+
+If no matching worktree exists, STOP:
+
+```
+BLOCKED — no worktree found for <epic-slug>.
+Run /beastmode:implement first to create the worktree and build the feature.
+```
+
 ### 1. Check Feature Completion
 
 Scan for implementation artifacts to verify all features have been implemented: