What: 把 build_prompt 拆成独立的 section builder（身份、工具、技能、记忆、安全、频道格式化、运行时提示等），支持 full/minimal/none 三种模式，加 prompt 总字符硬上限。
Why: 当前 prompt 构建是单一步骤，难以按场景裁剪。分段后可按模式选择装哪些段，也方便单独调优每个 section 的大小。
参考: elephant.ai 的 manager_prompt*.go，18 段组合 + 32K 字符上限 + 每技能体 1500 字符截断。
Effort: M Priority: P1
What: 给工具执行加装饰器链：至少包含熔断器（连续失败后自动熔断）和降级链（熔断后回退到备选工具）。可选：SLA 计时、参数校验、重试。
Why: 当前工具注册是扁平的，单个工具失败会直接暴露给 LLM。加熔断+降级后工具调用韧性大幅提升，LLM 不会卡在坏掉的工具上。
参考: elephant.ai 的 toolregistry/registry.go，5 层装饰器链 + DegradationExecutor + circuitBreakerStore。
Note: Per-tool timeout is now handled separately in executor.rs (Tool::with_timeout() + tokio::time::timeout() wrapper). This TODO should focus on circuit breaker + degradation chain only.
Effort: S-M Priority: P1
What: 在网关加一个快速分类步骤：用轻量 LLM 调用（或规则）把消息分成 direct（直接回复）、think（快确认+深度回复）、delegate（确认+后台 ReAct），简单消息不走完整 turn。
Why: 当前每条消息都跑完整 turn pipeline，简单问候或确认也要等全流程。分类路由后简单消息亚秒响应，体验质变。
参考: elephant.ai Lark 网关的「三模式大脑」，8s 快速 LLM 分类 → direct/think/delegate 三路由。
Effort: M Priority: P1
What: Publish conduit with FFI bindings so Python/JS developers can use the tape system and LLM toolkit.
Why: Expands addressable market from Rust-only to the entire agent ecosystem. Currently Eli only targets Rust developers building agents — a small intersection.
Pros:
- 100x larger audience
- Validates conduit as standalone value
- Enables the "SQLite for agent memory" positioning
Cons:
- Significant maintenance burden — two FFI surfaces to maintain
- API must be stable before binding to it
- Testing matrix explodes (Rust + Python + JS)
Context: The CEO review identified this as an "ocean" (too big to boil now) but high-value long-term. The Rust API needs to prove the decision layer thesis first. If persistent decisions don't matter to users, bindings don't help.
Effort: XL (human) → L with CC+gstack Priority: P3 Depends on: Conduit published as standalone crate, decision layer validated with real usage