DIKWP AgentTrace OS

English documentation · 中文速览 · Download the release package

Release layout / 发布结构 — The runnable source is distributed in dikwp_agenttrace_os_open_source_app.zip. Download and extract the package first, enter the extracted application directory, and then run the commands in Quick start. / 可运行源码位于上述 ZIP 发布包中；请先下载并解压，进入解压后的应用目录，再执行下方命令。

中文速览

• 它解决什么问题：把一次 AI Agent 运行转换为防篡改的 DIKWP 语义追踪账本，记录观察、决策、工具调用、记忆写入、证据与风险事件。
• 适合谁：Agent 开发者、AI 安全与治理团队、模型评测人员，以及需要事故回放和证据保全的研究者。
• 如何开始：下载并解压发布包，在应用目录中完成本地安装，然后使用示例运行生成时间线、风险事件、证据账本与回放包。
• 使用边界：本项目用于经过授权的 AI 系统审计，不用于隐蔽监控人员；实际部署应包含告知、目的限制、脱敏、访问控制、保留期限和人工复核。

Open-source AI Agent black-box recorder, semantic trace ledger, and incident replay toolkit.

DIKWP AgentTrace OS turns an AI agent run into a tamper-evident semantic trace ledger. It helps teams answer:

• What did the agent see?
• What did it decide?
• What tool did it call?
• What memory did it write?
• Which evidence supported the final output?
• Where did intention drift, prompt injection, tool overreach, or evidence gaps appear?
• Can the run be replayed and reviewed without trusting the model's story after the fact?

The project is designed as part of the DIKWP open-source ecosystem. It complements:

• DIKWP AnswerGraph Studio: AI answer asset readiness.
• DIKWP IntentGuard OS: pre-execution intent firewall.
• DIKWP MemoryLedger OS: long-term memory governance.
• DIKWP SemanticEnergy OS: semantic cost and energy accounting.
• DIKWP ProofLedger OS: claim-to-evidence auditing.

AgentTrace OS fills the missing runtime layer: agent observability, semantic forensics, and replayable audit custody.

Why this matters

Agentic AI systems are no longer simple chatbots. They call tools, retrieve context, write memory, delegate tasks, and act across multiple steps. Ordinary logs show raw events; they do not preserve DIKWP semantics, intention continuity, evidence custody, or replayable accountability.

AgentTrace OS provides a local, dependency-light prototype for:

• AI agent flight recording.
• Run-level semantic hash chains.
• Tool-call incident timelines.
• P-layer intention drift detection.
• Memory-write and prompt-injection risk detection.
• Evidence gap and unsupported claim detection.
• Replay bundle generation.
• OpenTelemetry-style span export.

Quick start

pip install -e .
dikwp-agenttrace analyze examples/sample_agent_run.json --policy configs/default_policy.json --out outputs/demo

Run static boundary audit:

dikwp-agenttrace static-audit src --out outputs/demo/static_boundary_audit_report.json

Output files

The demo command generates:

• agent_trace_report.json
• trace_timeline.csv
• risk_events.csv