[jiang1997]

Sandbox::log_tail previously appended each incoming chunk to a Vec and drained excess bytes from the front, repeatedly shifting retained data once the buffer filled.

Use VecDeque to discard old bytes without shifting the retained suffix. When a chunk reaches or exceeds the byte cap, replace the buffer with that chunk's suffix.

Add tests covering capacity boundaries, repeated wraps, interleaved stdout/stderr, fragmented frames, binary bytes, and malformed responses with zero capacity.

Validation: all 3 tests in engine_log_tail pass on macOS with Rust 1.85.1; git diff --check passes.

Production LOC: 102831 -> 102838 (delta +7)
Scope: core 38395 -> 38395; engine 61777 -> 61784; api 2659 -> 2659.
Method: tools/production_loc.py --root --json on archived first-parent and staged trees, using the same counter from 582bc78.
Reference absent (retired R9c; last counted 65,417 at tree 498dd1917812).

[Mola-maker]

感谢贡献，也感谢 AI 协作工具在分析 O(n) 搬移开销时做的推演 🙏

真实评估（已读 diff）：listener.rs 把 log tail 的保留缓冲从 Vec（每次追加后 drain(..) 头部搬移）换成 VecDeque，且对 n >= cap 的整窗覆盖走了 clear + 单次 extend 的捷径，避免逐段拷贝。语义保持"最后 cap 字节的精确后缀"，新增的 engine_log_tail.rs 测试覆盖得很扎实：0/1/7/2048/8191/8192/8193/65536/全长/超长十种 cap、跨窗口与重复回绕、多路复用流交错、UTF-8 与二进制字节、cap=0 时仍完整校验响应。没有发现缺陷。

对 AI agent 的使用价值：log_tail 是 agent 轮询 sandbox 输出的高频路径，长输出任务的每次读取都受益。

建议：合入前在真实 Docker 后端跑一遍长输出（>16 MiB、多路复用交错）的 exec，实测确认。功能性调整等 LayerFS 2.0（#446）发布后再做。

[jiangkoumo]

Superseded by 2.0: the Docker listener this changes was removed, and log_tail now reads a bounded window of the daemon log file (control/crates/layerfs-sandbox-backend). Thanks for the thorough tests.

(Reviewed against layerfs_v2_0_0 @ 716197c during the Staging cleanup.)