[Suggestion] renameWithRetry 与 packages/core/src/utils/atomicFileWrite.ts:50-70 中的实现完全相同——相同的重试逻辑、错误码（EPERM/EACCES）、指数退避和重试次数。两者均为私有函数，形成了代码重复。对重试逻辑的 Bug 修复需要在两处同步应用，存在代码分化风险。

— deepseek-v4-pro via Qwen Code /review

[Suggestion] config.storage 为公开 readonly 字段，此处直接访问 config.storage.getRuntimeStatusPath(sessionId) 暴露了内部 Storage 对象。而在 Config.startNewSession() 内部，同一操作使用的是 this.storage。如果将来 Storage 重构改变了 getRuntimeStatusPath 的语义，此处不会收到编译时反馈。

— deepseek-v4-pro via Qwen Code /review

[Suggestion] Both guards here are dead code:

• typeof v === 'number' already excludes booleans (typeof true === 'boolean'), so the trailing typeof v !== 'boolean' is unreachable.
• Number.isInteger(NaN/±Infinity) returns false, so Number.isFinite(v) after Number.isInteger(v) is also redundant.

Reduce to typeof v === 'number' && Number.isInteger(v) and rename the helper to isInteger for clarity.

[Suggestion] This branch is effectively dead code. Node's fs.readFile(path, 'utf-8') does not throw on invalid UTF-8 bytes — it substitutes U+FFFD and returns the decoded string. The returns null on invalid UTF-8 bytes test actually exercises the downstream JSON.parse failure path, not this err.message.includes('utf-8') check.

Suggested cleanup:

1. Drop this branch and let the unconditional return null handle the residual case.
2. Rename the test case to something like on bytes that decode to non-JSON so future readers aren't misled about what's being verified.

[Issue] The header says “External consumers should treat unknown fields as forward-compatible additions”, but readRuntimeStatus (line 168) uses strict equality on schema_version and rejects anything ≥ 2 outright — a v2 producer that only adds fields would still be rejected by every v1 reader. Doc and code disagree; pick one:

• Tighten the doc: declare schema_version a breaking version — bumping it is an ecosystem-wide break.
• Loosen the reader: accept >= 1 and rely on the per-field type checks (which already exist) to absorb additive fields.

Most schema_version conventions imply the latter; either is fine, but the contract should be explicit.

[Nit] This block is alphabetical (fileUtils → filesearch/fileSearch → formatters). runtimeStatus belongs in the r section (next to retry.js); inserting it between fileUtils and filesearch breaks the existing convention.

[Issue] sessionId is normally a randomUUID(), but both --resume <id> and Config.startNewSession(sessionId) accept user-supplied strings without validation. A value like ../../foo would resolve outside the chats/ directory via path.join.

Impact is low (the user already owns their filesystem), but a one-liner reject for ids containing /, \, or .. is cheap defense-in-depth and lines up with how peer CLIs (kimi-cli, Codex) sanitize session ids.

[Suggestion] Fire-and-forget IIFE here. If a user hits /clear twice within ~50 ms, two IIFEs interleave: unlink(A) → write(B) → unlink(B) → write(C), and the final on-disk state depends on kernel rename serialization rather than any guarantee in this code.

Two options:

1. Hang a _runtimeStatusTask promise chain off this so swaps strictly serialize:

   this._runtimeStatusTask = (this._runtimeStatusTask ?? Promise.resolve())
     .then(() => clearRuntimeStatus(oldPath))
     .then(() => writeRuntimeStatus(newPath, { ... }))
     .catch(() => {});

2. Add a one-line comment that documents “last writer wins, relies on rename atomicity” — at least the intent is recorded.

Unlikely to fire in practice, but spelling it out prevents the next contributor from re-discovering the same edge.

[Issue] The existing same-PID session swap test calls the low-level helpers directly — it does not exercise Config.startNewSession(). The real integration glue is exactly where rounds 3–5 kept finding regressions:

• whether the runtimeStatusEnabled gate fires;
• whether the previousSessionId !== this.sessionId short-circuit holds;
• whether the IIFE actually finishes deleting the old sidecar and writing the new one.

Suggested test in config.test.ts: build a Config, call markRuntimeStatusEnabled(), call startNewSession(newId), await new Promise(r => setImmediate(r)), then assert the old sidecar is gone and the new sidecar's fields are correct. That would catch the round 3–5 regressions in CI rather than in review.