skill
planning-with-files-governed
Governance-oriented planning with verifiable checkpoints, recoverable handoffs, and cross-platform continuity for long-running AI tasks.
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/xueleihu-labs/planning-with-files-governed itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 23, 2026planning-with-files-governed v2.0.0-rc.4
A governance-oriented, file-based planning and checkpoint system for AI coding agents — with immutable checkpoints, cross-platform handoff, and L0–L3 governance profiles.
Non-official derivative disclaimer: planning-with-files-governed is an independent, community-maintained derivative of planning-with-files. It is not affiliated with, endorsed by, or maintained by the upstream author.
The Python package version is 2.0.0rc4; the GitHub release/tag is v2.0.0-rc.4. This is a pre-release candidate, not the final v2.0.0 release.
Table of Contents
- Overview
- Upstream Attribution
- Key Differences from Upstream
- Compatibility Policy
- Installation
- Quick Start
- Governance Levels (L0–L3)
- Cross-Platform Handoff
- Threat Model
- Security
- Current Limitations
- 中文摘要
Overview
planning-with-files-governed (PWF-Governed) provides a file-based, single-source-of-truth planning framework for AI coding agents. It replaces in-memory or chat-based task tracking with durable, verifiable plan files, immutable checkpoints, and a deterministic handoff protocol that works across macOS, Windows, and WSL.
The core principle is simple: the plan file on disk is the only authority. Every phase transition, checkpoint, and handoff is recorded as a file with a SHA-256 attestation, so any agent or device can resume from a known-good state without trusting chat history.
Upstream Attribution
This project is derived from OthmanAdi/planning-with-files, tag v3.8.1, commit 117dfae83eefb0d4f0f5824252fd833dcde13459.
Dual Copyright
MIT License
Copyright (c) 2026 Ahmad Adi
Copyright (c) 2026 xueleihu52-arch
The upstream project's original work is gratefully acknowledged. This derivative adds governance profiles, checkpoint integrity, cross-platform handoff, and a selective-backport compatibility model while preserving the file-based planning philosophy of the original.
See UPSTREAM.md for full derivation details.
Key Differences from Upstream
- Single authoritative source of truth — no second state system; the plan file on disk is the only authority for task state.
- Immutable state chain — checkpoint → result → commit → head state forms a tamper-evident chain; each link references its predecessor by SHA-256.
- Cross-device handoff — deterministic protocol for handing off a task between macOS, Windows, and WSL with hash verification.
- L0–L3 governance profiles — four governance levels (
LIGHT_FAST,LIGHT_CONTROLLED,STANDARD,STRICT) that control checkpoint frequency, attestation requirements, and validation strictness. - Layout v3 — project-local task planning at
00.项目规划与治理/<task-id>/with deterministic single/multi-task discovery. - Task index —
task-index.yamlprovides deterministic task discovery; single-task auto-discovery, multi-task requires explicittask_id. - SHA-256 plan attestation — every plan file can be attested with a SHA-256 hash, enabling integrity verification at any time.
- Smart hook injection — selective injection into
userprompt,pretool, andprecompacthooks without blanket overrides. - Session catchup — recovery after
/clearor session restart by reading the latest checkpoint and plan state from disk. - Plan doctor diagnostics —
plan-doctorvalidates plan structure, checkpoint integrity, and handoff consistency. - Selective backport strategy — upstream changes are evaluated individually; this is not a drop-in mirror.
Compatibility Policy
planning-with-files-governed is an independent governance-oriented derivative of OthmanAdi/planning-with-files. It is not an upstream mirror and does not promise version parity with the latest upstream release.
Upstream changes are evaluated selectively. A change may be backported only when it preserves:
- the single-source-of-truth model;
- deterministic plan and checkpoint contracts;
- cross-platform path safety;
- recoverable and verifiable handoffs;
- backward-compatible data migration where practical; and
- the Community Edition security and test gates.
Compatibility is maintained at the workflow and data-contract level where documented. Drop-in compatibility with every upstream command, hook, plugin, or release is not guaranteed.
Every accepted upstream backport must record its source commit, affected files, compatibility impact, and validation evidence.
Installation
Prerequisites
- Python ≥ 3.10
- No external dependencies (standard library only)
- Git (recommended for version control of plan files)
Steps
# Clone the repository
git clone https://github.com/xueleihu-labs/planning-with-files-governed.git
cd planning-with-files-governed
# Optionally add to PATH for global access
export PATH="$PWD/scripts:$PATH"
Verification
python3 scripts/plan-doctor.py --version
python3 -m pytest tests/ -q
No pip install is required. All scripts use the Python standard library only.
On native Windows, use the .ps1 wrappers in scripts/ with PowerShell. They validate the selected Python executable and prefer a working python or py launcher over a broken python3 alias. The .sh wrappers are for macOS, Linux, WSL, or Git Bash environments that provide sh/bash.
Quick Start
1. Initialize a session
python3 scripts/init-session.py \
--project-root /path/to/project \
--task-id my-first-task \
--governance-level L2
This creates the Layout v3 structure:
/path/to/project/
00.项目规划与治理/
task-index.yaml
my-first-task/
task_plan.md
1_master_plan.md
3_status_update.md
4_handoff.md
WORKFLOW_CHECKLIST.md
2. Run plan doctor
python3 scripts/plan-doctor.py \
--project-root /path/to/project \
--task-id my-first-task
The plan doctor validates plan structure, checkpoint integrity, and handoff consistency.
3. Attest a plan
python3 scripts/attest-plan.py \
--project-root /path/to/project \
--task-id my-first-task \
--plan-file task_plan.md
This computes a SHA-256 hash of the plan file and records the attestation.
4. Inject plan context
python3 scripts/inject-plan.py \
--project-root /path/to/project \
--task-id my-first-task \
--hook userprompt
This injects the current plan state into the specified hook context.
Governance Levels (L0–L3)
| Level | Name | Checkpoint Freq | Attestation | Use Case |
|---|---|---|---|---|
| L0 | LIGHT_FAST | Per-phase | Optional | Quick prototyping, low-risk tasks |
| L1 | LIGHT_CONTROLLED | Per-phase | Required | Standard development tasks |
| L2 | STANDARD | Per-phase + on-demand | Required + verified | Production-bound features |
| L3 | STRICT | Every significant action | Required + verified + signed | Critical/irreversible operations |
See docs/governance-levels.md for detailed specifications.
Cross-Platform Handoff
planning-with-files-governed supports deterministic handoff between macOS, Windows, and WSL:
- Stable checkpoint — the current device writes a checkpoint with SHA-256 attestation.
- Handoff record — a
4_handoff.mdfile records the handoff with predecessor hash, device info, and outstanding work. - Write lock release — the current device releases the task write lock.
- Verification — the receiving device verifies the checkpoint hash, predecessor chain, and write lock state before resuming.
Example: macOS → Windows
# On macOS: create checkpoint and handoff
python3 scripts/init-session.py --project-root /path/to/project --task-id shared-task
# ... work ...
python3 scripts/checkpoint.py --project-root /path/to/project --task-id shared-task --handoff
# On Windows: verify and resume
python3 scripts/init-session.py --project-root C:\Users\<username>\projects\example --task-id shared-task --resume
See examples/cross-platform-handoff/ for a complete example.
Threat Model
The primary threats addressed by this system:
- Untrusted external checkpoints — checkpoint data from external sources is treated as untrusted input; SHA-256 verification and schema validation are required.
- Path traversal — all file operations are confined to expected root directories.
- Stale plans — plan staleness is detected via checkpoint comparison and plan-doctor diagnostics.
- Concurrent writers — write lock enforcement prevents simultaneous modifications.
- Credential leakage — plan files must never contain credentials; runtime state is excluded from Git by default.
See docs/threat-model.md for the full threat model.
Security
- Plan files may contain sensitive task information; do not commit runtime state to Git by default.
- External checkpoint and read-head data is treated as untrusted input.
- Schema validation does not guarantee content trustworthiness.
- File paths must be confined to expected root directories to prevent path traversal.
- No credentials should be stored in plan files.
See SECURITY.md for vulnerability reporting and security guidelines.
Current Limitations
- No remote state sync — plan files are synchronized via user-managed file sync (e.g., Syncthing, Git); no built-in remote state server.
- No automatic conflict resolution — concurrent writes from two devices without proper handoff will produce a conflict that must be resolved manually.
- Single-language scripts — all runtime scripts are Python; no shell-only fallback is provided.
- No GUI — all interaction is via CLI and file editing.
- Schema validation is structural — JSON schema validation ensures structure but cannot verify semantic correctness of plan content.
- Windows path handling — backslash paths are normalized internally, but users should prefer forward slashes in configuration files.
中文摘要
planning-with-files-governed 是一个面向 AI 编码智能体的文件级规划与检查点治理系统,衍生自 OthmanAdi/planning-with-files(v3.8.1)。
核心特性:
- 单一事实源:磁盘上的计划文件是唯一权威状态
- 不可变状态链:检查点 → 结果 → 提交 → 头状态,通过 SHA-256 链接
- 跨设备交接:支持 macOS / Windows / WSL 之间的确定性交接
- L0–L3 治理分级:从轻量快速到严格签名,四级治理配置
- Layout v3:项目本地任务规划,确定性单/多任务发现
- SHA-256 计划认证、智能 Hook 注入、会话恢复、计划诊断
重要声明: 本项目是 planning-with-files 的独立社区衍生版本,与上游作者无隶属关系,未经上游作者背书或维护。
许可证: MIT(双重版权:Copyright (c) 2026 Ahmad Adi + Copyright (c) 2026 xueleihu52-arch)
安装: Python ≥ 3.10,无需外部依赖,克隆后即可使用。