diff --git a/docs/developer-guide/en/independent-review.md b/docs/developer-guide/en/independent-review.md
new file mode 100644
index 000000000..32b5617da
--- /dev/null
+++ b/docs/developer-guide/en/independent-review.md
@@ -0,0 +1,149 @@
+# Independent Review
+
+[中文版](../zh/independent-review.md)
+
+A recommended (optional, not a CI gate) checklist for reviewing anolisa code
+changes — self-checking your own change before a PR, or reviewing someone
+else's. It complements human CODEOWNERS review and CI; it does not replace them.
+
+## Core discipline
+
+The whole value is in reviewing **without contamination**:
+
+- Do it in a **fresh review pass that does not inherit the implementation
+  context** — a separate agent session (if you drive one) or a second reviewer.
+  It must not carry the author's intent, suspicions, or fix conclusions.
+- **Review-only**: read the code, `git diff`, tests and in-repo docs. Do not edit
+  code while reviewing.
+- **Zero-direction**: do not enter with "verify that known bug X is fixed". Look
+  for problems from scratch.
+- If several reviewers/angles run, **do not cross-feed** one's conclusions into
+  another — it biases them.
+
+## When to run how much
+
+- **Simple change → one holistic pass.** Simple = single file / local, no
+  interface-contract change, not cross-component, no concurrency or lifecycle, not
+  on a hot path, narrow blast radius.
+- **Risky change → a full set of review angles.** In anolisa, a change is **never
+  "simple"** if it touches any of: another component / crate, a public or FFI
+  interface, `src/*/probes/*.bpf.c`, sec-core sandbox/privilege, kernel / arch /
+  page-size code, or `*.spec.in` / `component.toml`. A small-but-dangerous
+  eBPF / FFI / sandbox diff is not simple.
+
+## Generic review axes (starting skeleton)
+
+architecture · correctness · stability · security · performance · holistic.
+
+Readability is **deprioritized** — clippy / rustfmt / ruff / eslint already gate
+style in CI. Only raise what mechanical linters cannot see (naming, dead
+abstraction, `mod.rs` layout per AGENTS.md). Apply the language-appropriate
+conventions per changed file (Rust / TypeScript / Python / eBPF).
+
+## anolisa-specific review angles (mount by changed path)
+
+These are the classes CI structurally cannot catch. Treat as a checklist keyed
+by what the diff touches, not as separate reviewer roles.
+
+1. **eBPF verifier & kernel safety** — `src/*/probes/*.bpf.c` has no automated
+   test gate, so review is the only net. Check: bounded loops; explicit size
+   clamps before `bpf_probe_read_*` (an unclamped size trips the verifier with
+   "R2 …negative"); no unbounded variable-length access; every producer
+   initializes newly added shared-header fields; stack under 512 B; per-CPU vs
+   shared-map concurrency. **State whether it was verifier-load-tested on a real
+   kernel** — an x86 CI load is not an aarch64 load.
+
+2. **Architecture / environment matrix** — kernel-parameter, `/proc`, `/sys`
+   code carries silent arch/kernel assumptions. Flag hardcoded page sizes (page
+   size is arch-dependent — use `sysconf(_SC_PAGESIZE)`), kernel-version gates,
+   container-vs-host `/proc` reads. Name which (arch × kernel × container) cells
+   go untested; recommend a run on the affected arch rather than asserting safety.
+
+3. **FFI / cbindgen ABI boundary** — anolisa crosses Rust ↔ C ↔ Python. Verify
+   the generated header == the Rust signature == the example in docs/headers; all
+   crossing types are `#[repr(C)]`; no panic unwinds across the FFI boundary
+   (`catch_unwind`); the cbindgen drift-guard was actually re-run, not assumed.
+   Header/example drift is silent and high-impact; the generic correctness axis
+   misses it.
+
+4. **Cross-component contract** — the seams between agentsight, cosh,
+   anolisa-cli, and the genai storage schema; SKILL.md auto-discovery;
+   `component.toml` fields; CLI/JSON shapes consumed across components. Trigger
+   when a diff changes a **producer without its consumer**; ask "does the other
+   side assume a stronger contract than this now provides?"
+
+5. **Agent security** — sec-core is a whole component, and agent workloads bring
+   specific threats: prompt injection, sandbox escape, deny-list bypass (a
+   code-execution deny-list must match on the resolved path, not the spelling),
+   command injection in generated shell, privilege escalation on root / sysctl
+   writes, secret leakage into logs or PR bodies. Give privileged-write and
+   untrusted-input paths a dedicated look.
+
+6. **Packaging / distribution correctness** — `*.spec.in` / `component.toml` /
+   manifests are not exercised by code-test CI. Check spec dependencies match
+   Cargo/npm dependencies, `requires_*` matches reality, versions are bumped
+   consistently, and there is no orphaned or missing registration.
+
+## Verification discipline (before claiming "verified")
+
+Hard-won on this repo's CI:
+
+- **Reproduce with the toolchain CI pins**, not just your local default — a lint
+  that passes on an older toolchain can fail on the pinned one. Run
+  `cargo +<CI-version> fmt --check`, `clippy --all-targets -- -D warnings`, `test`.
+- **New logic goes in a lib crate, not a bin** — otherwise incremental
+  coverage reads 0 %.
+- **commitlint**: `type(scope):` lowercase, scope in
+  `.github/commitlint.config.json`, header ≤ 120 chars, body wrapped ≤ 100.
+- **rustfmt is a build dependency** (libbpf-cargo's skeleton builder), not only a
+  linter.
+- **If it can be end-to-end tested, end-to-end test it.** Compiling or unit tests
+  alone is not "verified". In the PR body, split the testing status into
+  done-E2E / unit-only / not-verified — never a blanket "all passing".
+- **Tests must discriminate**: reverting the fix should make the test fail
+  (prove it, e.g. by a quick mutation). Cover both directions. Beware
+  short-circuits (`mtime` checks, `Ok(false)`) that give false confidence.
+- **Extract pure functions** so logic that doesn't need the kernel / root / I/O
+  is unit-testable off a privileged box.
+
+## Finding discipline (avoid false findings)
+
+- **"The code differs from what I expected" ≠ a bug.** First read `git blame` /
+  the commit message / surrounding design intent and answer "why is it this way".
+  Findings raised without checking design intent are frequently walked back.
+- **Don't trust the PR's abstract terms — verify the real mechanism.** Confirm
+  the claimed threat model / dependency / mechanism ("an X injection", "uses
+  library Y") actually exists in the code — the real reader/parser may differ
+  from the description, and a fix built on the description may not match the real
+  implementation. Review to the code, not to the PR's wording.
+- **Kernel / security bugs are not discounted by trigger frequency** — a
+  use-after-free, uninitialized read, info leak or race is a deterministic
+  property of the source; "not seen in a synthetic run" is about frequency, not
+  existence. Report it.
+- **Adversarially verify a conclusion before acting on it** — especially before
+  opening an issue, so it isn't based on a stale reading of the code.
+- **Withdraw a finding that doesn't hold up on verification** — don't keep it to
+  look thorough.
+- **"No change needed" is a valid outcome** — if an audit finds nothing, say so
+  and close; don't manufacture changes to look busy.
+
+## When reviewing someone else's PR
+
+- **State the problem, don't prescribe the fix**: give the fact, the evidence
+  (`file:line`), and the impact — let the author choose how to address it.
+- Read the author's design intent before raising a finding (same lesson as above).
+
+## Output contract
+
+- Findings first, ranked by severity, each with `file:line`. If clean, say so
+  explicitly and note residual risk. Don't lead with a long summary.
+- **Out of scope**: do not re-report what clippy / rustfmt / ruff / eslint /
+  commitlint / coverage already gate — that is noise and erodes trust. Focus only
+  on what mechanical CI and a single reviewer cannot see.
+
+## Integration loop
+
+Dedupe all findings → fix the real ones first → re-run affected tests → if a fix
+materially widens the diff, review again (at least holistic / correctness /
+stability / performance) → report honestly: which tests ran, which angles were
+reviewed, what was found, what was fixed, what residual risk remains.
diff --git a/docs/developer-guide/zh/independent-review.md b/docs/developer-guide/zh/independent-review.md
new file mode 100644
index 000000000..3c62c0b77
--- /dev/null
+++ b/docs/developer-guide/zh/independent-review.md
@@ -0,0 +1,76 @@
+# 独立 Review
+
+[English](../en/independent-review.md)
+
+一套推荐的（可选、非 CI gate）anolisa 代码审查 checklist——提 PR 前自检自己的改动，或 review 别人的。它补充人工 CODEOWNERS review 与 CI，不替代它们。
+
+## 核心纪律
+
+全部价值在于**无污染**地审查：
+
+- 在一个**不继承实现上下文的全新 review pass** 里做——一个独立的 agent session（如果你用），或第二位 reviewer。它不能带着作者的意图、怀疑点或修复结论。
+- **只读**：读代码、`git diff`、测试和仓内文档。review 时不改代码。
+- **无方向**：不要带着"验证已知 bug X 修好了"进入，从 0 找问题。
+- 如果跑多个 reviewer / 角度，**不要把一个的结论交叉喂给另一个**——会带偏。
+
+## 何时跑多少
+
+- **简单改动 → 一个 holistic pass。** 简单 = 单文件 / 局部、不改接口契约、不跨组件、不涉并发或生命周期、不在热点路径、影响面窄。
+- **风险改动 → 起一组多角度审查。** 在 anolisa 里，只要动到以下之一就**永远不算"简单"**：另一个组件 / crate、public 或 FFI 接口、`src/*/probes/*.bpf.c`、sec-core 沙箱 / 权限、内核 / arch / 页大小相关代码、`*.spec.in` / `component.toml`。一个小但危险的 eBPF / FFI / 沙箱 diff 不算简单。
+
+## 通用审查 axis（起点骨架）
+
+architecture · correctness · stability · security · performance · holistic。
+
+readability **降级**——clippy / rustfmt / ruff / eslint 已在 CI gate 风格。只报机械 linter 看不到的（命名、死抽象、`mod.rs` 布局，见 AGENTS.md）。按改动文件的语言用对应约定（Rust / TypeScript / Python / eBPF）。
+
+## anolisa 特有审查角度（按改动路径挂载）
+
+这些是 CI 结构上抓不到的类别。当作一份按 diff 触及内容挂载的 checklist，不是一组独立 reviewer 角色。
+
+1. **eBPF verifier 与内核安全** —— `src/*/probes/*.bpf.c` 没有自动测试门禁，review 是唯一的网。检查：bounded loop；`bpf_probe_read_*` 前显式 size clamp（size 不 clamp 会让 verifier 报 "R2 …negative"）；无 unbounded 变长访问；每个 producer 都初始化新增的共享 header 字段；stack < 512 B；per-CPU vs 共享 map 的并发。**说明有没有在真内核上 verifier-load 过**——x86 CI 的 load 不等于 aarch64 的 load。
+
+2. **架构 / 环境矩阵** —— 内核参数、`/proc`、`/sys` 相关代码携带隐含的 arch/内核假设。排查硬编码页大小（页大小随 arch 不同——用 `sysconf(_SC_PAGESIZE)`）、内核版本 gate、容器 vs host 的 `/proc` 读取。点名 (arch × kernel × container) 哪些格子没测；建议在受影响 arch 上真跑，而不是断言安全。
+
+3. **FFI / cbindgen ABI 边界** —— anolisa 跨 Rust ↔ C ↔ Python。核对生成的 header == Rust 签名 == docs/header 里的示例；所有跨界类型 `#[repr(C)]`；FFI 边界不 panic unwind（`catch_unwind`）；cbindgen drift-guard **真跑过**、不是假设。header / 示例漂移是静默高危，通用 correctness axis 会漏。
+
+4. **跨组件契约** —— agentsight、cosh、anolisa-cli、genai storage schema 之间的接缝；SKILL.md 自动发现；`component.toml` 字段；跨组件消费的 CLI / JSON 形状。当一个 diff **改了生产方却没改消费方**时触发；问"另一端是不是假设了比现在更强的契约"。
+
+5. **agent 安全** —— sec-core 本身是一个组件，且 agent 负载带来特有威胁：prompt 注入、沙箱逃逸、deny-list 绕过（代码执行 deny-list 必须匹配解析后的路径，而非拼写）、生成 shell 的命令注入、root / sysctl 写入的提权、secret 泄漏进日志或 PR body。给特权写入和不可信输入路径专门看一眼。
+
+6. **打包 / 分发正确性** —— `*.spec.in` / `component.toml` / manifests 不被 code-test CI 覆盖。检查 spec 依赖对齐 Cargo/npm 依赖、`requires_*` 符合现实、版本一致 bump、无孤儿或缺失的注册。
+
+## 验证纪律（report "已验证" 前）
+
+本仓 CI 上的血泪：
+
+- **用 CI 锁定的 toolchain 复现**，别只用本地默认——旧 toolchain 上过的 lint 可能在锁定版上挂。跑 `cargo +<CI-version> fmt --check`、`clippy --all-targets -- -D warnings`、`test`。
+- **新逻辑放 lib crate 不放 bin**——否则增量覆盖率读 0%。
+- **commitlint**：`type(scope):` 小写开头、scope 在 `.github/commitlint.config.json`、header ≤ 120 字符、body 换行 ≤ 100。
+- **rustfmt 是编译依赖**（libbpf-cargo 的 skeleton builder），不只是 linter。
+- **能端到端测就端到端测。** 只编译或只单测不算"已验证"。PR body 里把测试状态拆成 已 E2E / 仅单测 / 未验证——绝不整篇"全部通过"。
+- **测试要有判别力**：撤回修复应让测试失败（要证实，比如快速 mutation）。正反双向都覆盖。警惕给假信心的短路（`mtime` 检查、`Ok(false)`）。
+- **抽取纯函数**，让不需要内核 / root / I/O 的逻辑能在非特权机器上单测。
+
+## finding 纪律（避免假 finding）
+
+- **"代码和我预期的不同" ≠ bug。** 先读 `git blame` / commit message / 周边设计意图，回答"为什么现在是这样"。不查设计意图就报的 finding 经常被撤回。
+- **别信 PR 描述的抽象名词，查真实机制。** PR 声称的威胁模型 / 依赖 / 机制（"某某注入""用了某某库"）都去代码里核实是否真实存在——真正的读取/解析方可能和描述不符，据此的修复也可能与真实实现不匹配。按代码审，别照 PR 措辞审。
+- **内核 / 安全 bug 不用触发频率打折** —— UAF、未初始化读、信息泄漏、race 是源码的确定性属性；"合成场景里没见到"只关乎频率，不关乎是否存在。发现就报。
+- **下结论前对抗式验证** —— 尤其开 issue 前，别基于对代码的过期理解。
+- **核实后不成立的 finding 主动撤回** —— 别为显得审得深而硬留。
+- **"无需改动"是合法产出** —— 审计查完没问题就明说并收尾；别为显得做了事而制造改动。
+
+## review 别人 PR 时
+
+- **只陈述问题，不替作者开药方**：给事实、证据（`file:line`）、影响——让作者自己选怎么处理。
+- 报 finding 前先读作者的设计意图（同上教训）。
+
+## 输出契约
+
+- findings 优先，按严重度排序，每条带 `file:line`。若干净，明说无 finding 并注明残余风险。不要先写大段总结。
+- **不在范围内**：不重报 clippy / rustfmt / ruff / eslint / commitlint / coverage 已经 gate 的东西——那是噪音、蚀信任。只报机械 CI 和单个 reviewer 看不到的。
+
+## 集成循环
+
+去重所有 findings → 先修真问题 → 重跑受影响测试 → 若某个修复明显扩大 diff，再 review 一轮（至少 holistic / correctness / stability / performance）→ 如实报告：跑了哪些测试、review 了哪些角度、发现了什么、修了什么、还剩哪些残余风险。