Skip to content
/
Tier B A16 Cross-vendor router

PlanGate

plangate · s977043/PlanGate · ★ 2 · last commit 2026-05-26

Enforce a gated spec-driven AI development workflow with human approval at planning and PR stages, multi-model review, and full auditability via event logs.

Best whenAI development safety requires structural approval boundaries, not just guidelines — no code without a human-approved, hash-locked plan.
Skip ifScope expansion without explicit approval, Implementing without an approved plan
vs seeds
kiro(both enforce gated spec-first development), but implemented as a POSIX CLI + Claude Code plugin rather than a closed ID…
Primitive shape 60 total
Commands 4 Skills 21 Subagents 23 Hooks 12
00

Summary

PlanGate — Summary

PlanGate is a governance-first AI coding agent harness that enforces a 5-phase gated workflow: planning, requirement expansion, design, implementation, and handoff/verification. The core invariant is "no code without an approved plan" — human approval gates (C-3 for plan, C-4 for PR merge) bracket AI execution. It ships a plangate POSIX shell CLI (v0.2.0) with 30 subcommands, 4 Claude Code slash commands, 21 skills, and 24 specialized agents (orchestrator, implementation-agent, qa-reviewer, requirements-analyst, solution-architect, etc.), plus 12 enforcement hooks (EH-1 through EH-9 + EHS-1 through EHS-3). The workflow uses 5 "modes" (ultra-light through critical) with proportionally more verification steps. Events are logged to events.ndjson enabling sprint retrospective derivation. PlanGate v8.9.0 adds Reporting & Retrospective v1 on top of the governance harness.

differs_from_seeds: PlanGate is the closest to kiro (gated spec-first development with explicit approval gates) but implemented as a POSIX shell CLI + Claude Code plugin rather than a closed IDE. Where kiro enforces the pipeline via IDE hook events, PlanGate uses PreToolUse hooks for C-3/C-4 gate invariants. The 24 named personas (orchestrator, scrum-master, acceptance-tester, etc.) is the most extensive agent roster in this batch, more structured than BMAD's 6 personas.

01

Overview

PlanGate — Overview

Origin

Published by s977043 (Japanese developer), v8.9.0, MIT license. Primarily Japanese-language README and documentation, with an English version (README_en.md). Shell-based CLI. 4 contributors. Actively maintained — 68 CLI tests + 78 hook tests passing. Last commit 2026-05-26.

Philosophy

From the README: "承認なし、コードなし" ("No approval, no code") — a gate-based workflow for safe AI coding agents.

"PlanGate が配布するのは『AI 開発の安全な型』です。AI に何でも自動でやらせる枠組みではありません。" ("PlanGate distributes a 'safe pattern for AI development.' It is not a framework for having AI automatically do everything.")

PlanGate explicitly positions against full-autonomy approaches: "一般的なエージェントフレームワークが『自律性』を重視するのに対し、PlanGate は承認境界・監査可能性・スクラム親和性を重視します。" ("While general agent frameworks emphasize 'autonomy', PlanGate emphasizes approval boundaries, auditability, and Scrum compatibility.")

Manifesto-style statements

  • "No code without an approved plan" — the C-3 gate
  • "承認境界・監査可能性・スクラム親和性" (Approval boundaries, auditability, Scrum affinity)
  • "観測 (Steering Loop) ではなく評価 → 学習 → ガバナンス" (Not steering loop, but evaluation → learning → governance)
  • "PlanGate では失敗・成功を後から説明可能にする（観測・再現基盤）" (Make failure/success explainable afterward)

Staged adoption

5 levels: Level 1 (plan approval only) → Level 5 (full eval + timeline + dogfooding). "New users start from Level 1."

Target user

Development teams using Claude Code (with optional Codex CLI, Gemini CLI, Cursor) who want Scrum-compatible gated AI development with human approval at planning and PR stages.

02

Architecture

PlanGate — Architecture

Distribution

  • Standalone repo (clone + plugin registration OR .claude/ copy)
  • Claude Code plugin: plugin/plangate/
  • No npm package

Install methods

# Option A: Clone + plugin registration (recommended)
git clone https://github.com/s977043/plangate.git
# Then follow plugin/plangate/README.md

# Option B: Copy .claude/ directly
git clone https://github.com/s977043/plangate.git
cp -r plangate/.claude/ your-project/.claude/

# Hook enforcement (REQUIRED after install):
bin/plangate doctor --fix

CLI binary: plangate (POSIX shell, v0.2.0)

Located at bin/plangate. Subcommands (from source analysis): help, version, init, doctor, status, validate, validate-schemas, eval, metrics, abort, timeline, resume, maintenance, plan-check, report, context, keep-rate, review, exec, brainstorm, plan, gate, verify (+ more not listed)

Directory structure

plangate/
├── .claude/
│   ├── commands/     # 4 slash commands
│   ├── skills/       # 21 skills
│   ├── agents/       # 24 agent personas
│   ├── rules/        # mode-classification.md, etc.
│   └── settings.example.json
├── bin/plangate       # POSIX shell CLI
├── scripts/hooks/     # 12 enforcement hooks (EH-1 through EH-9 + EHS)
├── docs/
│   ├── plangate.md
│   ├── philosophy.md
│   ├── staged-adoption-guide.md
│   ├── ai/eval-baselines/
│   └── working/discussions/  # 5 discussion logs (Claude × Codex × Gemini)
├── schemas/           # JSON schema validation
├── workflows/         # WF-01 through WF-05 workflow definitions
├── plugin/plangate/   # Claude Code plugin
├── pages/             # GitHub Pages documentation
└── tests/
    ├── run-tests.sh   # 68 CLI tests
    └── hooks/run-tests.sh  # 78 hook tests

Required runtime

  • git + POSIX shell (bash/zsh)
  • python3
  • Optional: gh CLI (for PR/issue operations)
  • Optional: Codex CLI (recommended for exec + C-2/V-3)
  • Optional: Gemini CLI (for parallel external review)
  • Optional: Cursor (partial support)
  • OS: macOS, Linux (POSIX). Windows: WSL recommended.
03

Components

PlanGate — Components

CLI binary: plangate (30+ subcommands)

Subcommand Purpose
init Initialize PlanGate in a project
doctor --fix Verify + fix hook wiring
status Show current task/phase status
validate Validate invariants
validate-schemas JSON schema validation
eval 8-dimension evaluation + release blocker detection
metrics Workflow event aggregation from events.ndjson
abort Abort current task
timeline Trace timeline visualization (experimental)
resume Resume interrupted task
plan-check Lightweight plan quality check
report Sprint retrospective from events.ndjson
context Context management
keep-rate Keep rate metric (v8.8.0)
review External review coordination
exec Execute implementation
brainstorm Brainstorming phase
plan Plan generation
gate Gate check
verify Verification
maintenance Maintenance mode

Claude Code Commands (.claude/commands/)

Command Purpose
plangate-setup Setup/onboarding
ai-dev-workflow AI development workflow
working-context Load working context
(others in README)

Claude Code Skills (.claude/skills/) — 21 skills

Skill Purpose
acceptance-criteria-build Build acceptance criteria
acceptance-review Review acceptance criteria
architecture-sketch Architecture design
brainstorming Brainstorming
codex-multi-agent Codex multi-agent orchestration
context-load Load context artifact (WF-01 entry)
edgecase-enumeration Enumerate edge cases
feature-implement TDD implementation (WF-04)
known-issues-log Log known issues/compromises
nonfunctional-check Non-functional requirements check
plan-quality-check Lightweight plan quality assessment
plangate-setup PlanGate setup skill
requirement-gap-scan Scan for requirement gaps
risk-assessment Risk assessment
self-review 17-item self-review checklist (C-1)
setup-team Team setup
skill-creator Create new skills
skill-ops-planner Skill operations planning
skill-optimizer Optimize existing skills
subagent-driven-development Subagent coordination
systematic-debugging Systematic debugging

Claude Code Agents (.claude/agents/) — 24 agents

acceptance-tester, agile-coach, claude-code-reviewer, code-optimizer, documentation-writer, explorer-agent, implementation-agent, implementer, linter-fixer, migration-agent, orchestrator, project-planner, prompt-engineer, qa-reviewer, requirements-analyst, research-analyst, retrospective-analyst, scrum-master, setup-coordinator, skill-designer, solution-architect, spec-writer, workflow-conductor

Enforcement hooks (12 total: EH-1 through EH-9 + EHS-1 through EHS-3)

Hook ID Event Trigger Purpose
EH-1 PreToolUse Edit|Write Block editing without plan.md
EH-2 PreToolUse Edit|Write Block editing without C-3 approval
EH-3 PreToolUse Edit|Write Detect plan_hash tampering
EH-6 PreToolUse Edit|Write Detect scope-out-of-bounds file edits
EH-9 PreToolUse Bash Detect delegation commit boundary violations
SessionStart SessionStart — Pin gh CLI account
(EH-4, EH-5, EH-7, EH-8, EHS-1, EHS-2, EHS-3 inferred from README)

State files

  • docs/working/<TASK-XXXX>/ — per-task working directory
  • docs/working/<TASK-XXXX>/plan.md, todo.md, test-cases.md, review-self.md, review-external.md
  • approvals/c3.json — C-3 approval record with plan_hash
  • events.ndjson — event log for metrics + retrospective derivation
  • docs/ai/eval-baselines/ — frozen evaluation baselines
05

Prompts

PlanGate — Prompts

Verbatim excerpt 1: orchestrator agent (.claude/agents/orchestrator.md)

---
name: orchestrator
description: PlanGate × Workflow/Skill/Agent ハイブリッドアーキテクチャの実行層総責任者。WF-01〜WF-05 の
  phase 遷移を制御し、各 phase の完了条件判定・Agent への委譲・handoff 発行を行う。汎用マルチエージェント調整も兼ねる。
tools: Read, Grep, Glob, Bash, Write, Edit, Agent
model: inherit
---

# Orchestrator

PlanGate ハイブリッドアーキテクチャの実行層（WF-01〜WF-05）を統括する総責任者エージェント。
phase 遷移管理、Agent 間の委譲、完了条件判定、handoff 発行を担う。

## 委譲関係（PlanGate hybrid）

orchestrator (WF-01) ├→ requirements-analyst (WF-01 / WF-02) ├→ qa-reviewer (WF-02 締め / WF-05) ├→ solution-architect (WF-03) ├→ implementation-agent (WF-04) └→ （handoff 統合 / 発行）


## PHASE 遷移プロトコル

### WF-01 Context Bootstrap
1. `CLAUDE.md` と依頼文を読む
2. `requirements-analyst` に `context-load` Skill 実行を委譲
3. 出力（context artifact）が完了条件を満たすか検証
4. PASS なら WF-02 へ遷移、FAIL なら再委譲

Prompting technique: Japanese-language persona with delegation tree. Uses model: inherit to pass through whatever model the parent session uses. Explicit DAG of agent-to-agent delegation. Phase-based state machine with pass/fail gate checks at each transition.

Verbatim excerpt 2: feature-implement skill

---
name: feature-implement
description: "design artifact に従って、最小単位で実装・テスト・自己レビューを繰り返し、動作するコード差分を生成する."
---

# Feature Implement

design artifact に従って、機能を最小単位で実装する Skill。差分ごとに自己レビューと明示的な既知課題の記録を行う。

## 出力

known-issues artifact + コード差分:
- 動作するコード（テスト付き）
- 自己レビュー結果
- 明示的な既知課題（妥協点 / 未着手項目）
- 実装単位ごとのコミット履歴

## 使い方

- WF-04 で implementation-agent が呼び出す
- TDD サイクル（RED → GREEN → REFACTOR）を最小単位で回す
- 出力は WF-05 の入力となる

Prompting technique: Phase-scoped skill with explicit artifact contract. TDD cycle mandate (RED → GREEN → REFACTOR). Known-issues logging as a required output field — forces explicit acknowledgment of compromises rather than hiding them.

09

Uniqueness

PlanGate — Uniqueness & Positioning

differs_from_seeds

PlanGate is the most governance-heavy framework in this batch. The closest seed is kiro (explicit spec-driven pipeline with approval gates), but PlanGate is implemented as a POSIX CLI + Claude Code plugin rather than a closed IDE. Like kiro, it mandates a specific artifact progression (plan → review → approval → implementation → verification). Unlike kiro, PlanGate: (a) has 24 named persona agents with explicit delegation trees, (b) uses multi-model review (Claude + Codex + Gemini at different phases), (c) enforces plan integrity via plan_hash SHA-256 tamper detection, and (d) derives sprint retrospectives from events.ndjson event logs. The BMAD-METHOD seed is the closest in agent count (6 personas vs 24), but BMAD personas are free-form creative characters while PlanGate personas are phase-scoped specialists with explicit delegation contracts.

Positioning

  • Category: Governance-first AI coding harness with human approval gates
  • User: Japanese/international development teams using Claude Code + Codex + Gemini who want Scrum-compatible, auditable AI development
  • Integration: Layered on Claude Code; complements Codex CLI and Gemini CLI for external reviews
  • Value: The only framework that enforces plan integrity via hash tamper detection + derives sprint retrospectives from event logs + supports 5 task modes with proportional verification

Observable failure modes

  1. bin/plangate doctor --fix required: Forgetting this step means hooks are not wired — the core gate invariants silently don't apply
  2. Japanese-first documentation: Non-Japanese developers face a translation barrier despite README_en.md
  3. 24 agents complexity: The delegation tree depth (orchestrator → requirements-analyst → qa-reviewer → ...) adds cognitive overhead
  4. codex-guarded.sh recommended for Codex: Without the guarded wrapper, Codex exec may commit outside intended boundaries
  5. ABORT on 5× V-1 failure: Teams relying on automatic loop recovery may hit abort and require manual intervention more than expected
  6. plan_hash tamper detection: Hash is of plan.md at C-3 approval time — any valid update to plan after C-3 (e.g., CONDITIONAL approval requiring plan changes) requires a new C-3 cycle

Cross-references

  • Targets: Claude Code (primary), Codex CLI (C-2/V-3), Gemini CLI (V-3), Cursor (partial)
  • docs/working/discussions/ contains 5 design discussion logs with Claude, Codex, and Gemini — unusual transparency about AI-assisted framework development
04

Workflow

PlanGate — Workflow

5-phase gate flow (from docs/plangate.md)

Phase Actor Action Artifact
A: PBI Input Human Write requirements, scope, acceptance criteria pbi-input.md
B: Plan generation AI Generate plan + todos + test cases simultaneously plan.md, todo.md, test-cases.md
C-1: Self-review AI 17-item PASS/WARN/FAIL checklist review-self.md
C-2: External AI review AI (Codex/Gemini) Independent check by different model review-external.md
C-3: Human review GATE Three-value judgment (APPROVE / CONDITIONAL / REJECT) C-3 approval record
D: Agent execution AI TDD implementation (tests must all pass) Implementation code
L-0: Linter auto-fix AI autofix → AI fix up to 3 rounds → suppress + note Linter-clean code
V-1: Acceptance check AI Mechanically check test-cases.md conditions PASS / FAIL (FAIL → fix loop ≤5×)
V-2: Code optimization AI Reduce redundancy, improve readability (high-risk/critical) Optimized code
V-3: External model review AI (Gemini) Design quality check by external model Review result
V-4: Pre-release check AI Final quality gate (critical mode only) Check result
PR creation AI Create GitHub PR Pull Request
C-4: Human review GATE Final PR review on GitHub APPROVE / REQUEST CHANGES / REJECT
Merge → Done Human Merge Done

5 task modes

Mode Scope Verification steps
ultra-light typo/comment/README minor L-0 → V-1 → PR → C-4
light Bug fix, 1-2 files L-0 → V-1 → PR → C-4
standard Small feature, 3-5 files L-0 → V-1 → V-3 → PR → C-4
high-risk — L-0 → V-1 → V-2 → V-3 → PR → C-4
critical — L-0 → V-1 → V-2 → V-3 → V-4 → PR → C-4

Approval gates

  • C-3: Human three-value gate before implementation (APPROVE / CONDITIONAL / REJECT). plan_hash stored in approvals/c3.json — tamper-detected by EH-3 hook.
  • C-4: Human PR review on GitHub after implementation.

Phase-to-artifact map

Phase Artifact
B (plan) plan.md, todo.md, test-cases.md
C-1 review-self.md
C-2 review-external.md
C-3 approvals/c3.json with plan_hash
D Code changes
V-1 Acceptance check result
All phases events.ndjson entries
06

Memory Context

PlanGate — Memory & Context

State storage (multi-layer)

Per-task working directory

docs/working/<TASK-XXXX>/ — all artifacts for a specific task:

  • plan.md — the approved plan
  • todo.md — task list
  • test-cases.md — acceptance criteria / test cases
  • review-self.md — C-1 self-review result
  • review-external.md — C-2 external AI review result
  • pbi-input.md — input requirements

Approval records

  • approvals/c3.json — C-3 approval with plan_hash (SHA-256 of plan.md at time of approval). Tamper detection via EH-3 hook.

Event log

  • events.ndjson — NDJSON event log for all control points. Used for:
    • plangate metrics — workflow event aggregation
    • plangate report — sprint retrospective derivation
    • plangate timeline — trace timeline (experimental)
    • plangate eval — 8-dimension evaluation

Eval baselines

  • docs/ai/eval-baselines/ — frozen baselines from v8.6.0; used as comparison reference for plangate eval

Agent learnings

  • AGENT_LEARNINGS.md — accumulated learnings from sessions

Cross-session handoff

The context-load skill (WF-01) reads CLAUDE.md and current working state to restore context at session start. Per-task working directories persist all decision artifacts.

Context compaction

Dynamic Context Engine v1 (v8.8.0): model profile + context size management. Working discussions are in docs/working/discussions/ (5 discussion logs with Claude, Codex, Gemini).

Persistence

All task artifacts, approvals, and event logs persist in the project directory. The events.ndjson is the audit-safe, replay-capable record of what happened.

07

Orchestration

PlanGate — Orchestration

Multi-agent support

Yes — 24 named agent personas with explicit delegation trees. The orchestrator agent dispatches to: requirements-analyst, qa-reviewer, solution-architect, implementation-agent. Each agent has specific phase assignments (WF-01 through WF-05).

Orchestration pattern

Hierarchical: orchestrator → specialized agents per phase. The delegation tree is explicit: "orchestrator (WF-01) ├→ requirements-analyst ├→ qa-reviewer ├→ solution-architect └→ implementation-agent"

Isolation mechanism

Git worktree per task implied by Codex multi-agent skill (codex-multi-agent). The delegation_commit_boundary concept (EH-9 hook) shows delegation can be set to no-commit to prevent delegated tasks from committing — suggesting worktree or branch isolation.

Multi-model support

Yes — the framework is designed for multi-model use:

  • Primary: Claude Code (plan generation, orchestration)
  • C-2 external review: Codex CLI (via scripts/codex-guarded.sh)
  • V-3 external review: Gemini CLI (parallel external review)
  • Cursor: Partial support (PLANGATE_IMPL_AGENT=cursor)

Execution mode

Interactive-loop (human triggers gate phases) + event-driven (hooks fire on tool use).

Subagent definition format

persona-md — each agent is a Markdown file with YAML frontmatter including tools, model: inherit, and phase assignments.

Consensus mechanism

None. External AI reviews (C-2, V-3) are advisory, not consensus-based.

Prompt chaining

Yes — each phase's output becomes the next phase's input: plan.md → test-cases.md → implementation → review-external.md → PR.

Crash recovery

plangate resume subcommand. Linter fix loop has a maximum (3 rounds). V-1 fix loop max 5×, then ABORT → human judgment.

08

Ui Cli Surface

PlanGate — UI & CLI Surface

CLI binary: plangate (POSIX shell, v0.2.0)

  • Entry: bin/plangate
  • Subcommands: 30+ (init, doctor, status, validate, validate-schemas, eval, metrics, abort, timeline, resume, maintenance, plan-check, report, context, keep-rate, review, exec, brainstorm, plan, gate, verify, ...)
  • Is thin wrapper: No — the binary implements substantial workflow logic

Key subcommands:

bin/plangate doctor --fix      # Verify + fix hook wiring
bin/plangate metrics           # Aggregate workflow events
bin/plangate eval              # 8-dimension evaluation
bin/plangate report            # Sprint retrospective
bin/plangate timeline          # Trace timeline (experimental)
bin/plangate keep-rate         # Keep rate metric

Claude Code slash commands

  • /plangate-setup — onboarding
  • /ai-dev-workflow — AI development workflow
  • /working-context — load working context

Local UI surface

None. No web dashboard.

IDE integration

None beyond Claude Code support. Cursor has partial support (docs/rfc/provider-cursor.md).

Observability

  • events.ndjson: NDJSON event log — all gate events, approvals, phase transitions
  • plangate metrics: Aggregate events.ndjson data
  • plangate eval: 8-dimension evaluation with release blocker detection
  • plangate timeline (experimental): Trace timeline visualization
  • plangate report: Derive sprint retrospective from events.ndjson
  • docs/ai/eval-baselines/: Frozen baselines for comparison
  • Override logging: approvals/c3.json hash detection via EH-3

Test coverage

  • 68 PASS (CLI tests: sh tests/run-tests.sh)
  • 78 PASS (Hook tests: sh tests/hooks/run-tests.sh)
  • 8 plangate eval dimensions
  • JSON schema validation for all artifacts

Documentation

GitHub Pages documentation site (pages/). Extensive docs including Japanese philosophy, staged adoption guide, when-NOT-to-use, glossary, troubleshooting.

Related frameworks

same archetype · same primary tool · same memory type

CodeMachine CLI ★ 2.5k
A16 Cross-vendor router

JavaScript-DSL workflow orchestration engine that captures repeatable AI coding agent workflows with tracks, condition groups,…

Codexia ★ 690
A16 Cross-vendor router

Tauri desktop app providing visual control plane, task scheduler, git worktree manager, and headless REST API for Codex CLI +…

Kagan ★ 88
A16 Cross-vendor router

Kanban TUI for AI coding agents with a structurally enforced human review gate (REVIEW → DONE cannot be automated) — one git…

oh-my-claudecode (Yeachan-Heo) ★ 35k
A16 Cross-vendor router

Zero-learning-curve teams-first multi-agent orchestration for Claude Code with autopilot (6-phase lifecycle), ralph (PRD-driven…

Paseo ★ 6.8k
A16 Cross-vendor router

Multi-provider AI coding agent orchestration daemon with cross-device access (phone/desktop/CLI) and git worktree isolation.

CCG Workflow ★ 5.4k
A16 Cross-vendor router

Routes Claude + Codex + Gemini to task-appropriate collaboration strategies (direct-fix through full-collaborate) with hook-based…