[BUG] In some cases, with Codex's auto model + DeepSeek, it will suddenly terminate automatically.

log

2026-09-12T14:43:08.075Z [info] forward POST → https://api.deepseek.com/responses
2026-09-12T14:43:22.209Z [info] [01a09604-f526-7822-bd15-37389368b47d] [acp-usage] round 1 input=86426 cached=84864 (cache hit 98%)
2026-09-12T14:43:22.210Z [info] [01a09604-f526-7822-bd15-37389368b47d] [acp-proxy: compress requested 1 range(s): m00109–m00241]
2026-09-12T14:43:22.210Z [info] [01a09604-f526-7822-bd15-37389368b47d] [acp-proxy: ctx has 150 message(s), state has ? ref(s) mapped]
2026-09-12T14:43:22.210Z [info] [01a09604-f526-7822-bd15-37389368b47d] [acp-proxy: first msg ids: acp_summary_b1(4061c), h_04717dfa80527a27(874c), h_c409b04ed442fbde(57c), acp_summary_b2(3581c), acp_summary_b3(11663c), h_111d38449f61bea7(4469c), h_d15b859a5b434346(290c), h_7e894daa91504165(154c), h_acd8d70bbfb03b00(5254c), h_8ea31d5c56c4658a(154c)]
2026-09-12T14:43:22.215Z [info] [01a09604-f526-7822-bd15-37389368b47d] [acp-compress-obs] shrink 35% (~29987/86426 tok) foldPoint=m00109 blocks=1 anchor≈2479 tok (2 active blocks, sys excluded) postCtx≈56439 → next-request cache ceiling ≥4%
2026-09-12T14:43:22.215Z [info] [01a09604-f526-7822-bd15-37389368b47d] [acp-proxy: [Compressed m00109–m00241 → 1 block(s), ~29987 tokens saved.]]
2026-09-12T14:43:22.218Z [info] [01a09604-f526-7822-bd15-37389368b47d] [acp-loop] round 1: re-request view refreshed to post-compress fold
2026-09-12T14:43:22.218Z [info] [01a09604-f526-7822-bd15-37389368b47d] [acp-loop] round 1: proxy tool executed; re-requesting so the model sees the result
2026-09-12T14:43:32.513Z [warn] [degenerate-turn] responses: turn ended completed with zero visible text and zero tool calls (thinking present; no visible text emitted) — the agent receives an empty turn and may stall until nudged (#673)
2026-09-12T14:43:32.513Z [info] [01a09604-f526-7822-bd15-37389368b47d] [acp-usage] round 2 input=53587 cached=13184 (cache hit 25%) fold=new
2026-09-12T14:44:08.791Z [info] [update] checking npm registry for billion-context (last check 180s ago)…
2026-09-12T14:44:11.201Z [warn] [update] running v0.1.106 but v0.1.107 is installed — restart bili to activate (auto-update replaced the on-disk install; this process is still on the old code)
2026-09-12T14:46:30.533Z [info] [01a09604-f526-7822-bd15-37389368b47d] dropped 1 whitespace-only message item(s) before projection (flattened-turn artifact)
2026-09-12T14:46:30.538Z [info] [01a09604-f526-7822-bd15-37389368b47d] processTurn: 72 msgs, renderTags=text-only, 6 text tagged, 0 tool tagged (should be 0 with text-only)
2026-09-12T14:46:30.538Z [info] [01a09604-f526-7822-bd15-37389368b47d] nudge idle: usage=42% (53587/128000), growth=-2852/22500 (ref=56439, interval=50000), pendingT1=27079/50000 model=deepseek-flash, reason="max compressible 27079 < threshold 50000; growth -2852 < floor 22500"
2026-09-12T14:46:30.538Z [warn] [01a09604-f526-7822-bd1

check 2 timestamp
2026-09-12T14:43:32.513Z auto mode stop and don't know why
2026-09-12T14:44:08.791Z

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷

Thanks for the log — this is a real, reproducible failure mode and I can pinpoint both timestamps.

What actually stopped — 14:43:32.513Z

That timestamp IS the stop. The sequence:

14:43:22 (round 1): the model issued a compress call → bili executed it server-side (Compressed m00109–m00241 → ~29987 tokens saved) and re-requested so the model sees the result (round 1: proxy tool executed; re-requesting).
14:43:32 (round 2): the model returned a degenerate turn — it finished normally (completed) but emitted only a thinking/reasoning block, zero visible text, zero tool calls. That's exactly what the [degenerate-turn] … (#673) warning reports. The agent gets an empty completed turn and stalls until nudged.

So the "stop" is the upstream model (deepseek-flash) ending its turn with thinking only. bili detected it correctly and flagged it — it just had no recovery step, so the empty turn was passed straight through and the agent sat idle until the next request at 14:46:30.

The second timestamp — 14:44:08.791Z

Unrelated to the stop. It's just the routine npm auto-update check (checking npm registry … last check 180s ago). The following line (running v0.1.106 but v0.1.107 is installed — restart bili) likewise has nothing to do with the stall — it only means you should restart bili to pick up v0.1.107. The gap between 14:43:32 and 14:46:30 is simply the agent stalled by the empty turn.

Root cause / layer

This is an upstream reasoning-model behavior, not a bili data bug: a reasoning model occasionally returns a reasoning-only completion (no content, no tools). It's most likely right after a post-compress re-request, where the model sees the freshly-shrunk context and "wraps up" into a silent thought.
bili behaved correctly and even surfaced it via the [bug] #644 follow-up (post-#646, v0.1.97): typo'd render-tag echo (<acip …) bypasses the strip regex and passes through as a 43-char degenerate end_turn turn — agent stalls mid-orchestration until manual nudge #673 detector. The gap: [bug] #644 follow-up (post-#646, v0.1.97): typo'd render-tag echo (<acip …) bypasses the strip regex and passes through as a 43-char degenerate end_turn turn — agent stalls mid-orchestration until manual nudge #673 shipped detection only — the CHANGELOG literally calls it "the groundwork for auto-retry," but the retry was never implemented.
The existing retry backstop ([bug] 假完成：模型把工具调用写成文字（无 tool_use 块）→ 轮次提前终止，agent 显示正常完成（#361 根因 follow-up） #371, BILI_FAKE_COMPLETION_RETRIES) does not cover this: (a) it's off by default, (b) it detects a different shape (the model writing tool-call XML as plain text), and (c) it only guards the first fetch, not the compress loop's round-2+ re-requests — which is where this happens.

Fix I'm implementing

An invisible single-retry inside the compress loop (src/loop/core.ts): when a terminal turn is degenerate (normal completion, zero visible text, zero tool calls) and nothing has been forwarded to the client yet (forwardedAny === false — true for these post-compress rounds, since round>1 reasoning/meta are suppressed), re-fetch once with a light continuation nudge. A plain re-fetch would deterministically reproduce the same silent output, so the nudge is what breaks the pattern. It's one-shot per request and mirrors the existing invisible truncation-retry guard (#413), so it can't corrupt the client stream. This completes the auto-retry that #673/#674 set up.

Opening a PR with a regression test. If you'd rather I gate it behind an env flag instead of on-by-default, say so and I'll adjust before merge.

——
中文摘要：14:43:32 那次停止是模型（deepseek-flash）在压缩后重请求的 round 2 只返回了思考块、没有任何可见文本或工具调用（即 #673 的 degenerate-turn），agent 拿到空回合就卡住了；14:44:08 只是例行 npm 自动更新检查，与卡住无关。根因在上游推理模型的偶发行为，bili 检测正确但缺重试。我会在压缩循环里加一个“对客户端不可见的单次重试（带继续提示）”来补上 #673 预留的 auto-retry。

🤖 Powered by ework · qwen3.8-27b

Short answer on your two timestamps:

14:43:32.513Z is the stop. That's the [degenerate-turn] line: right after bili ran your round-1 compress (Compressed m00109–m00241) and re-requested, the model (deepseek-flash) returned a turn that ended completed but contained only a thinking block — no visible text, no tool call. The client got an empty completed turn, so the agent sat idle until the next request arrived at 14:46:30.
14:44:08.791Z is unrelated — that's just the routine npm auto-update check (checking npm registry … last check 180s ago). The follow-up "running v0.1.106 but v0.1.107 is installed — restart bili" is the same housekeeping, not the stall.

Root cause. This is upstream reasoning-model behavior, not a bili data bug: deepseek-flash occasionally emits a reasoning-only completion, and the post-compress re-request makes it more likely (the model sees freshly-shrunk context and "wraps up" into a silent thought). bili already detected this correctly (#673) — but only as a warning. The detector was described as "the groundwork for auto-retry," and that retry was never implemented. (The separate #371 fake-completion backstop catches a different shape, is off by default, and only guards the first fetch, not later loop rounds.)

Fix (PR open): #734. It implements the missing retry. When a loop round ends completed with zero text and zero tool calls, thinking present, and nothing forwarded to you yet, bili re-fetches once with a short continuation nudge. It's invisible to the client, the nudge is never persisted into session state, it's one-shot per request, and it's guarded so a genuinely-empty (no-reasoning) terminal turn is left untouched. Pre-flight: typecheck ✅ / test ✅ (1431 pass, 0 fail) / build ✅.

Not released yet — it ships in the next release and your running proxy will pick it up via auto-update after that lands. Until then, if you hit a stall like this, sending any message (even a bare nudge) resumes it, because the trigger is the model's silent turn, not lost state.

中文摘要：你贴的两个时间戳里，14:43:32.513Z 才是卡住的原因——压缩后重新请求时 deepseek-flash 只返回了一个「纯思考」回合（状态 completed，但没有可见文本、没有工具调用），客户端拿到空回合就停住了；14:44:08.791Z 只是例行 npm 自动更新检查，跟卡住无关。根因是上游推理模型偶尔只输出思考块（bili 在 #673 已经能检测到，但当时只做了告警，自动重试一直没实现）。我已在 PR #734 补上缺失的重试：当某个循环回合以 completed 结束、零文本零工具调用、有思考、且还没向你转发任何内容时，带一条临时续写提示重取一次（对客户端不可见、也不落盘），每请求最多一次；真正为空的回合不受影响。typecheck / test / build 全绿。合并发布后你正在跑的代理会自动更新拿到它。

ework简述本issue在最新版本是否还有问题,哪个pr需要合并

fix: auto-retry invisible degenerate terminal turns (#732)

🤖 Powered by ework · qwen3.8-27b

结论：最新版仍有问题；需要合并的是 PR #734。

1) 最新版是否还有问题 —— 有。
npm 上现已发布到 v0.1.110，但本 issue 的 bug（压缩后重新请求时，模型只返回一个「纯思考」回合——状态 completed 却无可见文本、无工具调用 → 客户端拿到空 completed 回合而卡住）在 v0.1.110 里依然存在。我核对了 master 自 v0.1.107 以来的全部提交：没有任何一个修这个问题。其间的 reasoning-guard（#739/#740）是另一回事——针对 gpt-5.x/gpt-6.x 的推理截断保护，和本次「静默思考回合导致空回合」无关。所以最新版尚未修复。

2) 需要合并哪个 PR —— #734。
我刚把 #734 rebase 到当前 master（v0.1.110），重同步了依赖并重跑校验：

typecheck ✅ ／ test 1491 pass · 0 fail ✅ ／ build ✅
无冲突；CI 的非测试门禁已通过（version-guard、build-artifact、codex 零 token E2E 预检均为 success），4 个必需 test 矩阵（ubuntu/windows × node22/24）正在跑，转绿后 mergeable_state 即变 clean、可合并。

合并后 CI 会自动发布下一个版本，正在运行的代理会通过自动更新拿到该修复。（合并动作按仓库规则由人工执行。）

Merge pull request #734 from ranxianglei/2026-09-12_auto-mode-degener…