What's Changed

Synced with upstream v4.3.1 (obra/superpowers):

• Windows hook fix: POSIX-safe path resolution, quoted CLAUDE_PLUGIN_ROOT for paths with spaces
• Restored polyglot wrapper for Windows hook window spawning
• .DS_Store added to .gitignore

Fork maintenance:

• Updated skill cross-references from superpowers: to superpowers-extended-cc:
• Verified native task integrations intact across all customized skills

Changes

Plan Mode Defense

Adds explicit EnterPlanMode and ExitPlanMode prohibitions to skills that operate in normal mode, preventing Claude Code's trained auto-entry behavior from conflicting with the skill workflow.

• skills/writing-plans — CRITICAL CONSTRAINTS extended to block both tools
• skills/executing-plans — CRITICAL CONSTRAINTS section added
• skills/brainstorming — prohibition note added
• skills/using-superpowers — flow diagram updated (EnterPlanMode? DON'T)
• commands/test-writing-plans — compliance check updated to verify neither tool is called
• README — added "Recommended: Disable Auto Plan Mode" section with deny list config

Upstream sync (obra/superpowers)

• .cursor-plugin/plugin.json — Cursor plugin manifest added and rebranded for this fork
• hooks/session-start.sh — hook response compatibility
• README — updated Verify Installation section

Fixes

• HARD-GATE on ExitPlanMode: Added <HARD-GATE> at the Execution Handoff section in writing-plans skill, right at the decision point where the model goes wrong.

• Remove disable-model-invocation from commands: Commands (write-plan, brainstorm, execute-plan) no longer block model invocation. Fixes skill chaining failures.

Cleanup

• Remove maintainer-only files from published module: Deleted skills/release/SKILL.md (maintainer workflow, moved to .claude/commands/), development plan files, .claude symlink, .idea/ config.

• Restore finishing-a-development-branch to upstream: Removed local /worktree-cleanup reference, restored upstream inline bash.

Upstream Sync (v4.3.0)

• Brainstorming enforcement: HARD-GATE prevents implementation before design approval, new checklist and process flow diagram
• Synchronous SessionStart hook: using-superpowers loads before first turn instead of async
• Writing-plans fix: 4-backtick fence for nested code blocks in Task Structure template
• Test fix: --verbose flag for stream-json output

Conflict Resolutions

• .claude-plugin/*.json: kept fork name/version
• RELEASE-NOTES.md: removed (standard)
• All native task sections preserved in auto-merge

Fix: Command skill references after upstream merge

Commands in commands/*.md (write-plan, brainstorm, execute-plan) referenced the upstream plugin name (superpowers:*) instead of our fork's name (superpowers-extended-cc:*). This caused disable-model-invocation errors when invoking these commands via the Skill tool.

Changes

• commands/write-plan.md — superpowers:writing-plans → superpowers-extended-cc:writing-plans
• commands/brainstorm.md — superpowers:brainstorming → superpowers-extended-cc:brainstorming
• commands/execute-plan.md — superpowers:executing-plans → superpowers-extended-cc:executing-plans

Maintenance

• Documented command reference check as post-merge step in fork-maintenance.md

Fix: writing-plans skill compliance

Problem: The writing-plans skill was ignored by the model in two key areas:

1. TaskList not called as required first step
2. ExitPlanMode called instead of AskUserQuestion for execution handoff

Root cause: Critical constraints were buried in the middle of a long reference document. The model skipped procedural requirements.

Changes:

• skills/writing-plans/SKILL.md: Added CRITICAL CONSTRAINTS section at the top with explicit ExitPlanMode prohibition. Fixed TaskCreate description: removed fake placeholder paths, replaced with instruction to copy from the plan.
• commands/test-writing-plans.md: New command that spawns a Sonnet subagent to verify writing-plans skill compliance.

Verification:

• Cross-tested original vs fixed SKILL.md with parallel Sonnet subagents
• Original: TaskList NOT called first (failed)
• Fixed: TaskList called first (passed)

What's Changed

Merge upstream/main (16 commits) including:

• Windows fixes and PowerShell invocation improvements
• Codex native skill discovery and installer rewrite
• Documentation polish from 5-agent review
• Performance fix: O(n²) escape_for_json replaced with parameter substitution
• SessionStart hook async fix for Windows

Fork maintenance

• All native task integrations verified intact
• No new legacy TodoWrite/TodoRead references

Fix: writing-plans Execution Handoff

Problem: The execution handoff at the end of planning was ambiguous about which tool to use, causing the assistant to call ExitPlanMode instead of presenting the choice to the user. The user never saw the "Subagent-Driven vs Parallel Session" question.

Fix:

• Explicit AskUserQuestion YAML format with both options
• Added warning against using ExitPlanMode or plain text
• Ensures the user always sees and can answer the execution choice

Sync with upstream (77 commits) + Native task conversion + Release workflow

Core changes:

• Eliminated all TodoWrite → uses TaskCreate/TaskUpdate throughout
• Added /release skill for full release workflow

Skills updated:

• subagent-driven-development: native tasks in flowchart
• using-superpowers: TaskCreate in flowchart
• writing-skills: TaskCreate in checklist instruction
• release: NEW - full release workflow with Opus subagent research

Upstream merged:

• Worktree requirement for subagent-driven-development and executing-plans
• Main branch protection warnings

Add native task integration section to dispatching-parallel-agents skill