feat: add structured agent audit prompt pack

[huangrichao2020]

Summary

• add app/prompt/agent_audit.py, a reusable prompt pack for structured agent-runtime audits
• include playbooks, advanced playbooks, rubric guidance, report schema, and example report data in code
• add tests that validate the required audit artifacts, schema shape, playbook coverage, and prompt builder output
• mention the prompt pack in the README

What this adds

This PR adds a reusable prompt pack for diagnosing agent runtime failures in OpenManus.

The prompt pack is designed for cases such as:

• wrapper regression
• tool-discipline failures
• stale memory contamination
• hidden repair or retry layers
• rendering / transport mutation

It asks the agent to build structured audit artifacts before writing a final diagnosis:

1. agent_check_scope.json
2. evidence_pack.json
3. failure_map.json
4. agent_check_report.json

The module also includes:

• standard playbooks
• advanced playbooks
• an audit rubric
• a report schema
• an example structured report
• build_agent_audit_prompt() for composing the prompt with a selected playbook

How to use it

This PR does not add a new CLI command. Instead, it adds a reusable prompt pack that developers can attach to an agent, flow, evaluation script, or debugging workflow.

Example usage in code:

from app.prompt.agent_audit import build_agent_audit_prompt

audit_prompt = build_agent_audit_prompt("tool-discipline")

The returned prompt can be used as a system prompt or supplemental instruction for an OpenManus agent:

from app.prompt.agent_audit import build_agent_audit_prompt

agent.system_prompt = build_agent_audit_prompt("wrapper-regression")

Supported playbooks include:

• wrapper-regression
• memory-contamination
• tool-discipline
• rendering-transport
• hidden-agent-layers

A user-facing prompt could look like:

Use the agent-audit prompt pack to audit this agent runtime.
Focus on tool-discipline failures and stale evidence reuse.
Build agent_check_scope.json, evidence_pack.json, failure_map.json, and agent_check_report.json before giving the final diagnosis.

Expected final output includes:

• executive verdict
• severity-ranked findings
• conflict map
• contamination paths
• ordered fix plan

Why this fits OpenManus

OpenManus is a general agent framework, and runtime reliability issues often come from the orchestration around the model rather than the model itself. This prompt pack gives users and developers a reusable, evidence-first way to inspect those layers without changing the core agent loop.

I kept this as a prompt module and tests instead of adding a new runtime dependency or tool surface.

Verification

I ran the following locally:

cd /tmp/openmanus-agent-audit && uv venv .venv --python 3.12
cd /tmp/openmanus-agent-audit && uv pip install pytest black==23.1.0 isort==5.12.0 autoflake==2.0.1 pre-commit==4.4.0
cd /tmp/openmanus-agent-audit && .venv/bin/python -m isort app/prompt/agent_audit.py tests/test_agent_audit_prompt.py
cd /tmp/openmanus-agent-audit && .venv/bin/python -m black app/prompt/agent_audit.py tests/test_agent_audit_prompt.py
cd /tmp/openmanus-agent-audit && .venv/bin/python -m pytest tests/test_agent_audit_prompt.py -q
cd /tmp/openmanus-agent-audit && .venv/bin/python -m py_compile app/prompt/agent_audit.py tests/test_agent_audit_prompt.py

Results:

• 4 passed
• py_compile passed