selfforge-avatar-reviewer

name	selfforge-avatar-reviewer
description	Use when complex work needs a SelfForge avatar to supervise Codex, review another agent, catch preference drift, or make preference-based decisions with retrieved local evidence.

SelfForge Avatar Reviewer

This is a portable reviewer workflow. It installs the trigger skill and recall scripts, but it does not ship the original private avatar_index.json.

Why: avatar_index.json can contain personal chat history, secrets, and preference evidence. Teammates should build their own local index from their own approved training data.

When To Use

Use when:

• the user asks for 分身, SelfForge avatar, avatar reviewer, or supervision;
• another Codex/agent produced work that needs acceptance;
• a complex task needs preference-based review;
• work is about to be called complete and preference drift is possible.

Skip for tiny local fixes unless the user explicitly asks for the avatar.

Step 1 - Retrieve Evidence First

Before reviewing, retrieve avatar evidence:

python <this-skill-dir>\scripts\recall_avatar.py `
  --query "<task, diff summary, screenshot notes, or decision question>" `
  --top-k 8 `
  --json

If the index is missing, stop and report that private avatar evidence is not installed. Do not invent user preference evidence.

Build an index only from user-approved JSONL sources:

python <this-skill-dir>\scripts\build_avatar_index.py `
  --source <approved-training-corpus.jsonl> `
  --source <approved-post-training.jsonl>

Records should contain text, or a combination of question, label, and tags.

Step 2 - Decision Interface

If a project-local decision interface exists, call it after retrieval:

python <project-or-skill-decision-interface>\decide.py `
  --task "<task, diff summary, screenshot notes, or decision question>" `
  --top-k 3

This package does not include a private decide.py. If it is unavailable, continue only as an evidence-based reviewer when retrieval succeeded. If retrieval also failed, stop.

Report missing decision interface like this:

SelfForge Decision_Interface 不可用

错误：<exact error message or exit code>

继续方式：只使用 recall_avatar.py 召回证据做审查；不代表用户做高风险最终决策。

Step 3 - Review Workflow

1. Summarize the worker task and retrieved evidence.
2. Inspect changed files, screenshots, command output, logs, API responses, and persistence behavior locally, read-only.
3. If a read-only subagent capability is available and the task is non-trivial, spawn one read-only reviewer subagent. Pass the retrieved avatar evidence into the subagent prompt.
4. Never grant the reviewer or subagent write access to the supervised workspace.
5. Merge findings with local evidence.
6. Output review result and next worker instructions.

Output Format

结论：通过 / 需修改 / 阻塞

关键问题：
- [P1/P2/P3] 问题。证据：...

召回证据匹配：
- ...

缺失闭环：
- ...

给 worker 的下一步：
1. ...
2. ...

Write Boundary

This skill must never write to the supervised workspace. Allowed write destinations:

• <this-skill-dir>\tmp\
• $env:TEMP\selfforge-reviewer\

Do not store new personality text in this skill. Add avatar facts to local training data and rebuild the vector index.