shintaii/claude-powermode

Summary

claude-powermode is a Claude Code plugin that enforces a disciplined multi-agent engineering workflow: plan → explore → implement → verify → simplify. It works by installing slash commands (/powermode, /pm-plan, /pm-checkpoint, etc.), Python hook scripts that intercept Claude Code tool calls to enforce delegation rules (blocking direct edits unless routed through pm-implementer), and agent definitions that assign specific Claude model tiers (Haiku/Sonnet/Opus) to different roles. The net effect is a guardrailed CI-like loop running inside your Claude Code session.

Great for

people interested in building behavioral guardrails and multi-agent orchestration patterns for Claude Code — specifically enforcing workflow discipline through hooks that intercept and block tool calls at runtime

Easy wins

• +Add a test suite for the hook scripts — they're pure Python with stdin/stdout JSON contracts, making them trivially unit-testable without Claude Code (see comment-checker.py's regex logic or stop-validator.py's transcript parsing)
• +Add a CONTRIBUTING.md explaining how to test hooks locally (the CLAUDE.md already documents `echo '{...}' | python3 hooks/stop-validator.py` but this isn't surfaced to contributors)
• +Fix the incomplete README — the Automatic Behaviors table is cut off mid-row (ends with `| *`), which is a clear documentation gap
• +Add a hooks integration test that replays a sample JSONL transcript through stop-validator.py to verify blocking/approving behavior across edge cases

Red flags

• !Single commit repo — it's unclear if this is actively maintained or a one-shot upload; no commit history to assess development trajectory
• !References 'GPT-5.4' in /pm-codex-review documentation — this model does not exist; either a placeholder or hallucinated model name
• !The Claude Code plugin marketplace (`claude plugin marketplace add`) does not appear to be a publicly available feature, making the primary installation path non-functional for most developers
• !No tests whatsoever despite hooks having complex stateful logic (attempt counting, session file management, transcript parsing) — bugs in stop-validator.py's PRD tracking would silently mis-block or mis-approve sessions
• !context-monitor.py is listed in the file tree but not included in source samples — given it tracks token counts and writes state files, bugs here could silently corrupt session state
• !CLAUDE.md mentions `plugin-dev:plugin-validator` and `plugin-dev:skill-reviewer` as validation tools but these appear to be internal Anthropic tools not available to external contributors

Code quality

The hook scripts show solid defensive programming: every code path ends with `sys.exit(0)`, all `except` blocks fall back to safe JSON output, and the `try/except` wrapping in `__main__` is consistent across files. delegation-enforcer.py correctly handles multi-agent concurrency (implementer-sessions/ directory with per-agent files plus legacy single-file fallback). However, plan-checkpoint-validator.py's drift detection is naive keyword overlap that will produce noisy false positives in practice (e.g., 30% overlap threshold on stopwords-heavy todo text). The comment-checker.py regex patterns are extensive but the string-in-string detection heuristic (counting unescaped quotes) is fragile — it won't handle multiline strings or template literals correctly.

What makes it unique

The hook-based enforcement approach is genuinely interesting — using PreToolUse hooks to block direct edits and force routing through pm-implementer is a novel pattern for workflow enforcement rather than just prompting. However, this is essentially a sophisticated CLAUDE.md + hooks configuration, and the core concept (structured agentic workflows for Claude Code) is an emerging space with several similar repos. What distinguishes it is the depth of the hook system (17 hooks) and the hierarchical Project→Feature→Task planning artifact structure, not just prompt templates.