symphlo: Instructions file for Codex

AGENTS

默认用中文交流；公共 API、协议字段、代码标识符和对外开发者文档使用英文。

Read First

进入 workspace 后按顺序阅读：

1. PROJECT_KNOWLEDGE.md
2. PROJECT_SPEC.md
3. README.md
4. PUBLIC_SOURCE_MANIFEST.md
5. 当前 task 对应的代码和测试

Owner Model

• 主 Agent 是 owner / decider / implementer / verifier。
• 中等及以上任务先明确 goal、scope、affected contracts、validation、risks 和 non-goals，获得 checkpoint 后再实现。
• 一次只推进一个最小可验收 task；每次收敛都回写受影响的 public contract、tests 和 release evidence。
• 保留用户已有改动；不重置、不清理、不顺手修无关文件。

Product Boundaries

• Symphlo is the durable Flow and collaboration control plane in the active Symphlo + Clerklet product portfolio. The historical huisezhiyin/agent-flow project is retired and must not receive new product work.
• Symphlo coordinates multiple independent executors, not only fixed Agent chains. Agents, applications, MCP, HTTP, CLI, scripts, compute and Humans may collaborate through the same versioned Node and handoff contracts.
• Fixed and repeated orchestration is one strong Flow shape, not the complete product definition. Multi-Agent and multi-application collaboration is a core product scenario.
• Symphlo is not a general low-code platform and not an Agent UI shell.
• An Agent is a first-class Node executor. It may keep its internal loop, while Symphlo owns observable inputs, effects, executor identity, events, accepted results, Context and Artifacts.
• Symphlo guides Agent work through external task boundaries; it does not modify, expose or claim control over private Agent reasoning.
• Loop ownership slides with explicit Node granularity: broad Agent Nodes preserve more Agent autonomy, while finer semantic boundaries give the Flow more orchestration, evidence and recovery responsibility.
• Do not infer atomic work from a small task. Agent Node may contain an opaque loop; model.task + model_cli guarantees one adapter request, while tool.task guarantees one fixed semantic CLI, MCP or HTTP operation and records symphlo.tool-call-evidence.v1.
• Flow controls what / who / when / dependencies / authority / handoff; each Agent or application controls how inside its Node boundary.
• Flow definitions, Run state, Context, Artifacts and history are product truth. Agent sessions and conversations are not.
• The public product must remain agent-agnostic and capability/adapter-driven.
• Clerklet is the separate standalone Agent and an optional reference participant. Neither repository may become the other's hidden runtime, database or source-path dependency.
• Canvas is optional infrastructure and must not define product identity or become the source of truth.