manifesto-hci

Implement "Explicit State & Continuous Consensus" HCI pattern (v3.0). Combat information entropy, prevent intent drift, and maintain a shared source of truth...

Security Scan

The skill's code, instructions, and requirements are coherent with its stated purpose (Git-backed manifesto/state management) and there are no obvious requests for unrelated credentials or network access, but it will persist conversation content to disk and perform git commits which you should understand before use.

Purpose & Capability

Name/description ask for Git-backed state, tri-track architecture, and continuous consensus; the skill includes an on-disk project layout, a script (scripts/core.py) that initializes project dirs, creates a manifesto file, runs git init/add/commit, and logs turns — all of which are consistent and expected for this purpose.

Instruction Scope

SKILL.md instructs the agent to run scripts/core.py start/log and to append both user inputs and assistant outputs to persistent JSONL logs and to update/commit a Manifesto. That behavior is coherent with the manifesto purpose but is broad: every turn is persisted to disk (history JSONL and Git commits). The SKILL.md also specifies a background 'Diff Sub-Agent' that emits structured audit messages and performs commits; that auditing protocol is documented but not implemented in the provided code (the script only provides init and logging).

Install Mechanism

Instruction-only with a small included Python script; there is no install spec, no downloads, and no third-party package installs. Risk from installation is minimal.

Credentials

The skill declares no required environment variables, credentials, or config paths. It does require file-system write access in the agent's working directory (creates projects/prj_<id>/, state/, logs/, workspace/) and runs local git; those are proportional to the declared goal but do persist data locally (including full user/assistant messages).

Persistence & Privilege

The skill is not 'always' enabled and does not request elevated privileges, but it will create files and initialize a Git repository under projects/prj_<id>/ and make commits. This is normal for state-management but means the agent will retain conversation history on disk and in git history until you remove it.

Assessment

This skill appears to do what it says: it creates a per-project directory (projects/prj_<project_id>/), writes a Manifesto markdown file, logs every user and assistant turn to a JSONL file, and performs git init/add/commit. Before installing or using it, be aware that: (1) all conversation content and manifesto data will be stored on disk and recorded in git history (un-encrypted) — do not use with secrets or sensitive data unless you isolate or encrypt the project directory; (2) the SKILL.md describes an asynchronous 'Diff Sub-Agent' and socket push messages for audit results, but the provided script does not implement that background service — some auditing behavior is purely a protocol expectation the agent must follow; (3) the script invokes local git via subprocess and writes files relative to the agent's working directory, so confirm the agent has permission to write there and that created files won't inadvertently include other sensitive files; (4) no network endpoints or credentials are requested, and there are no external downloads. If you accept local persistent logging and git commits, the skill is consistent with its stated purpose. If you need guarantees about encryption, remote hosting, or removal of logs, request those features or review/modify the script before use.

Manifesto HCI - Explicit State & Continuous Consensus (v3.0)

This skill enforces a rigorous state management and consensus-building workflow using Git-backed physical decoupling and asynchronous auditing.

Core Interaction Logic

You MUST follow the "Tri-Track Architecture" and "Execution Hooks" defined in specification.md.

1. User Lifecycle Control (Slash Commands)

• /manifesto start [project_id]:
  • Run python3 scripts/core.py start <project_id>.
  • Perform git init and first git commit silently.
  • Output: [System: Manifesto 已挂载。后台 Diff 守护进程已就绪，当前分支: master。]
• /manifesto pause:
  • Pause the Diff Sub-Agent. Stop Manifesto updates.
  • Output: [System: Manifesto 暂停更新。进入自由发散模式。]
• /manifesto stop:
  • Final git commit and unmount.
  • Output: [System: Manifesto 已解绑。共识状态已安全固化至本地硬盘。]

2. Physical Decoupling (Tri-Track)

• All project data is stored in projects/prj_[project_id]/.
• Track 1 (Raw Log): Append-only JSONL in logs/.
• Track 2 (The State): YAML + Markdown Manifesto in state/. Single Source of Truth.
• Track 3 (Working Memory): workspace/ and LLM context.

3. Main Agent Protocol (Foreground)

Every foreground response MUST follow this rule:

• Pre-hook (Intent Sync): Every turn starts with [意图同步: -m "Short summary"].
• Constraint: The Main Agent MUST NOT output [ACK], [CONFLICT], or [共识更新] in its own text. It only handles the user's task.

4. Background Auditing (Diff Sub-Agent & Post-hook)

• On turn completion, a zero-temperature (T=0) Sub-Agent must audit history asynchronously using the Diff Agent Prompt.
• Stateful Write: Update Manifesto following the Manifesto Template structure and run git commit.
• Ephemeral/Noise: Discard.
• Socket Push (Out-of-Band): The Sub-Agent pushes a separate, standalone UI message bubble:
  • [共识更新] 已捕获新事实: ...
  • [逻辑冲突] 审计警报: ...
  • [ACK] 流程对齐中，无新共识产生。

5. Automated History Logging

• After generating a response, log both the user input and your response.
• Execute: python3 scripts/core.py log <project_id> user "<user_message>" and python3 scripts/core.py log <project_id> assistant "<assistant_response>".

Manifesto Organization

Follow Topic-based Classification in specification.md:

1. Core Anchors
2. Side-talk / Low Relevance
3. Multi-topic Parallelism

Constraints

• Explicit Commits: All changes must be visible via commit -m.
• Zero Data Loss: Never discard core arguments.
• Physics Separation: Manifesto vs History per project.
• Agent Decoupling: Main Agent (Interaction) vs Sub-Agent (Auditing).