[sky11023]

问题

品悟 Agent（约 25K token 系统提示词）冷 prefill 时，本地模型在首字节前静默约 30 秒（TensorFold 双卡 854 tok/s 实测，2026-10-29 现场）。TUI 三个流式适配器（chat / responses / anthropic）共用 45 秒的流打开（响应头）超时——这个预算是为「云主机死掉要快速失败」设计的；对本地端点，长提示词冷启动合法地超过 45s，用户侧表现为「回答被中断，当前请求超时请重试」。dspark 等卡同场景可复现，非单卡特有。

修复

crates/tui/src/client/stream_entry.rs：

• 新增 LOCAL_STREAM_OPEN_TIMEOUT = 180s 与 stream_open_timeout_for(base_url)：endpoint_is_local()（loopback / localhost / RFC1918 / link-local / ULA）命中的 base URL 走 180s 打开预算，其余维持 45s。
• chat / responses / anthropic 三个调用点全部切换到 stream_open_timeout_for。
• 显式环境变量覆盖对两类端点仍然最高优先。
• 块间空闲超时维持 900s 不变（已高于本地 120s 契约），改动只放宽「首个响应头」等待。

验证

• 本地 cargo test -p codewhale-tui --lib（精确选择）3/3 通过：local_endpoints_get_the_wider_open_budget、env_override_wins_over_the_local_budget、stream_open_timeout_defaults_and_clamps_env_values。
• 该提交（cb577ba55）已被 pinvou3 双机从机集成分支以 gitlink 引用并完成过整包构建装机验证（integrated.20261029.2，.10 设备 worker CLI 实测 exit 0）。请按原样合并，勿 rebase 改写哈希。

DCO：Signed-off-by 已带。

[asto18089]

First formal review — #130 fix(tui): widen local-endpoint SSE open timeout to 180s

Thanks for the PR and for the unusually well-documented reasoning (field data, per-class budget, and an env escape hatch preserved for both classes). Reviewed file by file against pinvou3-clean (fe90f72). No blockers found — this is a COMMENT review with minor items.

Per-file findings

crates/tui/Cargo.toml + Cargo.lock — url = "2"

• OK / necessary. url 2.5.8 is already in Cargo.lock transitively (via reqwest), so this adds no new build nodes. Correct tool for the job: hand-parsing IPv6 bracket hosts is error-prone. Spec "2" resolves to the locked 2.5.8 and matches the neighboring direct-spec style (urlencoding = "2.1", regex = "1.11"). There is no workspace.dependencies entry for url, so a direct spec is consistent with convention.

crates/tui/src/client/stream_entry.rs — stream_open_timeout_for

• Env override semantics are consistent with the existing path: presence check → delegate to stream_open_timeout() → same parse + clamp (5..=300, garbage falls back to 45s). The escape hatch still exceeds the local budget (up to 300s), so it works in both directions. Correct.
• nit: with env set to a invalid value (e.g. abc), a local endpoint silently drops from 180s to 45s (presence-check wins, parse falls back to default). Edge case, conservative direction — fine to leave.

crates/tui/src/client/stream_entry.rs — endpoint_is_local

• IPv4: loopback / RFC1918 / link-local via std predicates — correct.
• IPv6: ::1 + ULA fc00::/7 — correct as far as it goes.
• DNS names: only localhost / *.localhost count; a hostname that resolves to a private IP is deliberately not local (documented, no DNS on this path). This is the right conservative default (a cloud gateway behind a resolvable name would be misclassified the other way), but note it means LAN engines reachable only by hostname get 45s — only raw private IPs benefit.
• Conservative-direction gaps, each falls back to today's 45s (no regression), listed so the author can decide whether to widen: (1) IPv4-mapped IPv6 ([::ffff:127.0.0.1]) is Ipv6 and not loopback → remote; (2) IPv6 link-local fe80::/10 → remote (v4 link-local is covered, v6 is not); (3) CGNAT 100.64.0.0/10 — i.e. Tailscale-style endpoints, a very common way to reach a home GPU box, → remote; (4) 0.0.0.0 unspecified → remote. Given validate_base_url_security explicitly invites "llama.cpp on a private IP" over HTTP with the insecure-http env, item (3) is the one I'd most consider following up on (non-blocking).

crates/tui/src/client/{anthropic,chat,responses}.rs — wiring

• Complete. Verified by grep on base: exactly three stream_open_timeout() call sites exist (chat.rs:1346, anthropic.rs:354, responses.rs:269) — all three rewired with the identical comment; the only other match is doc text in client.rs:3028. The H1-fallback retry reuses the same StreamOpenRequest, so the widened budget carries into the retry. Per-chunk idle (900s) unchanged and already ≥ the local contract, as the commit message states.
• nit: classification reads self.base_url while the anthropic open goes through messages_transport_url; in production the transport URL derives from base_url (the override fields are test-only), so this is fine today.

Tests

• Not flaky: no sleeps; 180s is a constant comparison, not a timing assertion.
• ENV_LOCK correctly serializes the two env-touching tests, and no pre-existing test in this module touches the env, so no cross-test race.
• MINOR (hermeticity): local_endpoints_get_the_wider_open_budget calls stream_open_timeout_for, which reads ambient env. On a machine where CODEWHALE_STREAM_OPEN_TIMEOUT_SECS is exported, the == LOCAL_STREAM_OPEN_TIMEOUT assertions fail. Please clear the var under the lock at test start (or assert on endpoint_is_local only for the budget pair).
• nit: in env_override_wins_over_the_local_budget, if anything panics between set_var and remove_var, the var leaks into the process for other tests. A drop-guard (or catch_unwind) would make cleanup panic-safe.

Process notes

• DCO: Signed-off-by matches author/committer (l28756 <li.jiawen@h3c.com>) — DCO satisfied. The git name l28756 vs GitHub account sky11023 mismatch is noted; not a blocker, but aligning them avoids future verification confusion.
• Behind by 2 (fe90f72f #114, f1eb10ef #125): zero file overlap (engine/skills/docs vs client/*, no Cargo manifests), GitHub reports MERGEABLE. Per CONTRIBUTING, no rebase requested — the Merge Queue will gate against latest main.
• CI: no workflow runs exist for head cb577ba5 (nothing in the runs list, no check runs, combined status pending with zero statuses — the first-contribution approval gate). Nothing to approve via the API from this side; maintainers, please approve/trigger the workflow run so checks actually execute.

Conclusion

No BLOCKER, no MAJOR. MINOR items: test hermeticity vs ambient env; the enumerated conservative-direction classification gaps (CGNAT/Tailscale the most notable); optional panic-safe env cleanup. Comfortable to merge after a maintainer eyes the MINORs; comments-only, changes not requested.

[sky11023]

感谢首轮正式评审的细致核对（逐文件、突变与 CI 归因都做了）。三个 MINOR 已全部处置（ca079796c）：

1. 测试密封性：local_endpoints_get_the_wider_open_budget 现在在 ENV_LOCK 下先 remove_var 两个超时变量再断言——宿主机 export 过该变量也不再误红。
2. panic 安全清理：env_override_wins_over_the_local_budget 改用 Drop 守卫（EnvReset），断言 panic 也会清掉变量，不再泄进进程。
3. 分类缺口：endpoint_is_local 补齐三档——CGNAT 100.64.0.0/10（手写第二字节 64..=127 判定，is_shared() 在当前工具链还是 unstable feature ip）、IPv6 链路本地 fe80::/10、IPv4-mapped v6（按 v4 载荷重判）。0.0.0.0 作为目的地刻意维持远端（见注释：不该有人拨它，放宽只会掩盖调用方 bug）——若您倾向也纳入可以说一声，一行的事。

验证：三测试精确名单 3/3 绿；cargo fmt --check 绿。（跑宽过滤时 config::tests 那条已知编排型测试偶发红，与本次改动无关——不碰任何 config 变量。）

DCO 备注：git 身份 l28756 与 GitHub 账号 sky11023 的不一致已知，后续提交会对齐。

[github-actions]

Thanks @sky11023 for taking the time to contribute.

This repository is observing a maintainer-managed PR intake gate in dry-run mode, so this pull request is staying open. This note helps maintainers prepare the allowlist before any enforcement is considered.

Please read CONTRIBUTING.md for the expected contribution shape. A maintainer can grant recurring PR access by commenting /lgtm on a pull request.