Repository navigation
Conversation
(bot) (ai: vllm-qwen/qwen3.8-27b) 查重: 不重复 —— 「裸宣告式终止轮」(模型只说一句 "Running verify after formatting:" 就干净停轮、无工具调用) 这个症状在 open+closed issue 里无同题;与 #2303(draft 尾 但审核发现几个必须先处理的问题,按严重度列: 1. issue 引用错了(硬伤)。 标题和正文都写 2. 范围混装:把 #3906 的 streamMarkers 改动塞进了本 PR。 3. 正文描述与实现不符,且漏掉的正是精度边界。 正文说检测器要求「伴随常见动作宣告动词(running/checking/testing/let me 等)」且「无未闭合 markdown 代码栅栏」,实际实现只有「≤400 字符 + 以 4. 测试缺口 + 死导入。 测试文件 import 了 5. 行为变更未披露。 旧行为:这类短宣告轮正常结束;新行为:隐藏地多发一次 completion(计费 + 延迟)。PR 正文没有任何 old→new 披露,我会在对齐 commit 后补进 PR 描述。 方向判定:问题真实存在(生产 stub 证据 ses_eb6da0d60ffeYpnNQDoGz000Pc,deepseek-v4.1-flash,24 输出 token),机制归位正确(loop 门完整继承 #2303 的全部守卫:one-shot 预算、wire 边界、#2612 compaction 豁免、abort 检查),不是症状层打补丁。方向通过,按上述 1–5 做对齐重构。 接下来我会直接往你的分支推标注清晰的修复 commit(外部 PR 质量不达标走直接对齐,不拉锯),补立缺失 issue,然后看 CI。有问题随时喊停。 |
… PR ranxianglei#4934 Align-refactor (review of ranxianglei#4935): this PR is scoped to the bare-announcement terminal-turn retry (ranxianglei#3905 symptom), but it also carried the LoopCtx streamMarkers field and the emitMarker consumption gate that belong to the streamMarkers setting feature. That feature has its own PR (ranxianglei#4934, which wires the config surface but lacks these two hunks) — keeping them here makes ranxianglei#4934 unable to compile standalone and creates a merge-order dependency between two independent features. The hunks move to ranxianglei#4934 so each PR is a closed loop.
…he PR body describes Align-refactor (review of ranxianglei#4935): the PR body promises detection is 'accompanied by common action announcement verbs' and requires 'no markdown code fence open', but the shipped detector was tail-char + length only — any turn under 400 chars ending in ':' or ellipsis qualified, including legitimate short answers ('The answer is 42:') that would now cost an extra billed completion call each. Implement both promised gates: an ANNOUNCE_CUE word list (progressing forms + intent phrases only, mirroring the body's examples) and a reject on odd backtick-fence count (an unclosed fence is the ranxianglei#1755/ranxianglei#2190 truncation class, not a stall). English-first by design: non-English announcements keep the pre-ranxianglei#3905 no-retry behavior until traffic evidence justifies extending the cue list.
…ei#2303 sibling; pin the 400-char cap Align-refactor (review of ranxianglei#4935): the suite only covered unit + compress-loop lanes while the change also rewires pipePluginChatWithStrip and pipePluginResponsesWithStrip terminal classification — the ranxianglei#2303 sibling suite covers all three lanes, so match that bar: plugin chat positive (ranxianglei#3905 log line, one [DONE], single finish_reason stop), colon-answer and ellipsis-answer negatives, tool-call exclusion, retry-degenerates in-band error (ranxianglei#870); responses-pipe mirror set. Also fix at-l3, whose 'long answer' fixture did not actually end on the tail char and passed for the wrong reason — the rewritten fixture (441 chars, cue-bearing, colon tail) pins the length cap itself at loop level. Drop dead imports.
Align-refactor (review of ranxianglei#4935): this PR defines the streamMarkers field on LoopCtx and wires the config surface end-to-end, but nothing ever reads the field — the emission gate that honors it was shipped in PR ranxianglei#4935 (the announcement-retry PR), which makes this setting a dead knob on its own and couples the two PRs into a merge-order dependency. Move the consumption gate (and its failure-mode comment) here from ranxianglei#4935 so each PR is a closed loop. Behavior with the default (unset) is unchanged: markers still stream.
…rs (ranxianglei#2107) Align-refactor (review of ranxianglei#4935): the PR's website/config-reference/compress.yaml seed edit was committed without re-running tools/gen-config-docs.mjs, so the ranxianglei#2107 docs-sync drift guard goes red on this head. Regenerate the four derived artifacts (CONFIGURATION.md / CONFIGURATION.zh-CN.md / website/main.js / website/docs/index.html) — content is exactly the new streamMarkers row plus its BILI_STREAM_MARKERS env index entry.
…roseAcc (ranxianglei#3905) proseAcc accumulates every streamed text field including the thinking/ reasoning channel. endsWithAnnouncement(proseAcc) therefore saw the tail of the model's internal monologue: a thought ending mid-sentence on ':' or an ellipsis followed by a whitespace-only answer was reclassified as a bare announcement (the new ranxianglei#3905 branch) instead of the ranxianglei#2756 degenerate shape — the retry still fired but under the wrong log line, and a whitespace-only turn on an unclean finish lost its 'not retried ... not a clean stop' refusal log entirely (stalledTail suppressed it). Two subtests in tests/issue2756-degenerate-gate.test.ts went red on every shard-3 CI leg. Operand swap to answerProse in both plugin pipes (chat + responses). The compress-loop twin already tests text-only assistantText and is unaffected.
(ai: vllm-qwen/qwen3.8-27b) CI auto-fix lane (attempt 1/2) — 根因定位、已修复、已推送。 新头 查重: 不重复 —— #4599 记录的 master 侧红是 时间线说明: 分析期间作者推了 3 个新 commit( 结构化八问:
验证与状态: 本地定向复现(两条腿的失败子测试)+ 全量 blast-radius(337 断言)+ typecheck 均在新头通过;已推送 |
(ai: vllm-qwen/qwen3.8-27b) Merge guard sync: merged origin/master into this branch (new head 90db83f, fast-forward, zero conflicts — both sides' changes verified intact in src/degenerate-turn.ts / src/loop/core.ts / src/plugin.ts; PR regression suite 19/19 green + typecheck clean on the merged tree locally). State after push: no longer behind, mergeable, CI lanes re-running (queued) on the new head. |
Resolve the single src/loop/core.ts import conflict (union of the ranxianglei#3905 endsWithAnnouncement addition with master's scrubLoneSurrogateEscapesOnJson rename) and align the ranxianglei#3905 ranxianglei#870 in-band-error test assertion to master's ranxianglei#5958 protocol-native error-frame default (was the legacy [ACP] stream error completion shape).
Conflict resolved and pushed. I merged What I did this run:
CI checks: 24 queued, 0 failed (just started after the push). Full-suite verdict comes from CI. 八问
结论:冲突已解决并推送,PR 现为可合并( |
Status check (no new work needed) — this forward's I verified the LIVE state this run (not from memory):
So the PR is mergeable and CI-clean — this forward was a duplicate/stale conflict detection that landed after my fix. Nothing new to do; it's ready for a human to merge. 八问(本轮仅做状态核实,无新改动)
结论:冲突已解(上轮)、CI 全绿(本轮核实),PR 处于可合并状态,等待人工合并。 |
[ework] CI fix verified green ✓ |
[bot] (ai: vllm-qwen/qwen3.8-27b) PR review 结论(独立审核,非复读 issue 侧)—— 可合。 查重: 不重复。 本 PR 修「裸宣告式终止轮」(模型只说一句 "Running verify after formatting:" 就干净停轮、无工具调用)。与同族 #2303(未闭合草稿标签 本轮独立动作(不复读 floor 3/4):重取当前 head 结构化八问
行为变更披露(旧→新,Deliverables 第 4 条):旧 = 以冒号/省略号收尾的短轮正常完成;新 = 其中 ≤400 字符、带动作线索词、代码围栏闭合的轮会触发一次额外计费的 completion 调用(一次性预算封顶,重试再退化走 #870 错误而非循环)。残留误报面:含 "let me"/"i'll" 等线索词的合法收尾句若恰以省略号结尾仍会触发(有界、低频)。English-first:非英文宣告保持修复前行为,直到有流量证据扩线索表。 已知非阻塞项(需 owner/系统侧,agent 无权改 PR 标题/正文):PR 标题/正文引用 判定:可合。 方向对(修根因非打补丁)、无回归、测试有牙(三车道 × 正反 × 一次性预算 × #870 × 400 上限 × 工具调用排除 × 围栏排除 × #2303 同侪对齐)、diff 干净(4 文件全相关,无跑题/无空白 churn/无 lockfile)、CI 绿。等 owner 合并。 |
[ework] CI fix verified green ✓ |
Third identical
The Net: nothing to fix. The PR is mergeable and CI-clean; it's ready for a human to merge. The repeated 八问(本轮仅实时核实,无代码改动)
结论:无冲突、CI 全绿,PR 处于可合并( |
[ework] CI fix verified green ✓ |
|
Landed via batch PR #6482 (merged) — closing as superseded. |
… PR ranxianglei#4934 Align-refactor (review of ranxianglei#4935): this PR is scoped to the bare-announcement terminal-turn retry (ranxianglei#3905 symptom), but it also carried the LoopCtx streamMarkers field and the emitMarker consumption gate that belong to the streamMarkers setting feature. That feature has its own PR (ranxianglei#4934, which wires the config surface but lacks these two hunks) — keeping them here makes between two independent features. The hunks move to ranxianglei#4934 so each PR is a closed loop. (cherry picked from commit fa60f8f)
…he PR body describes Align-refactor (review of ranxianglei#4935): the PR body promises detection is 'accompanied by common action announcement verbs' and requires 'no markdown code fence open', but the shipped detector was tail-char + length only — any turn under 400 chars ending in ':' or ellipsis qualified, including legitimate short answers ('The answer is 42:') that would now cost an extra billed completion call each. Implement both promised gates: an ANNOUNCE_CUE word list (progressing forms + intent phrases only, mirroring the body's examples) and a reject on odd backtick-fence count (an unclosed fence is the ranxianglei#1755/ranxianglei#2190 truncation class, not a stall). English-first by design: non-English announcements keep the pre-ranxianglei#3905 no-retry behavior until traffic evidence justifies extending the cue list. (cherry picked from commit aec1e1b)
…ei#2303 sibling; pin the 400-char cap Align-refactor (review of ranxianglei#4935): the suite only covered unit + compress-loop lanes while the change also rewires pipePluginChatWithStrip and pipePluginResponsesWithStrip terminal classification — the ranxianglei#2303 sibling suite covers all three lanes, so match that bar: plugin chat positive (ranxianglei#3905 log line, one [DONE], single finish_reason stop), colon-answer and ellipsis-answer negatives, tool-call exclusion, retry-degenerates in-band error (ranxianglei#870); responses-pipe mirror set. Also fix at-l3, whose 'long answer' fixture did not actually end on the tail char and passed for the wrong reason — the rewritten fixture (441 chars, cue-bearing, colon tail) pins the length cap itself at loop level. Drop dead imports. (cherry picked from commit 206ddc9)
Acknowledged — superseded, no action taken. I verified the supersession is real (live, not from memory):
So I am not resolving the conflict (no point on a superseded PR) and not closing it (close is your action, not the agent's). The 八问(本轮仅核实 supersede,无代码改动)
结论:supersede 成立(#6482 已合并、内容已入 master),本 PR 冲突无意义,等待你关闭 #4935。 |
[ework] CI fix verified green ✓ |
PR 2: fix(loop): retry stalled bare-announcement terminal turns (#4944)
Fixes #4944
Problem / Background
Certain chat-optimized models (notably DeepSeek and other open-weights instruction models) occasionally emit a short single-line announcement promise (e.g.
"Let me run the tests now:","Checking the logs...","Running verify after formatting:") and stop cleanly without emitting any tool calls (tool_calls: [],finish_reason: "stop").In Billion-Context proxy environments, this failure mode is exacerbated when prompt guidance or ACP tool markers introduce slight turn distraction. The model intends to run a tool, emits an introductory colon-terminated statement, but hits an empty completion turn. The harness/user is left waiting on an unfinished sentence with no follow-up action executed.
Existing mitigations handle empty turns (#821) and incomplete draft closes (#2303 /
endsWithDraftClose), but bare announcements terminating with a colon or ellipsis fell through and ended the loop prematurely.Mechanism & Changes
src/degenerate-turn.ts:endsWithAnnouncement(text: string): boolean.:or.../…, accompanied by common action announcement verbs (running,checking,testing,inspecting,let me, etc.).src/loop/core.ts:endsWithAnnouncement(), injects a single-shot continuation retry (acp_draft_tail_retry_r${round}) nudging the model to proceed with execution.src/plugin.ts:endsWithAnnouncementcheck into plugin terminal tail validation.tests/issue3905-announcement-tail.test.ts:Verification
tests/issue3905-announcement-tail.test.tsverifying positive matches ("Running tests:","Checking the output...") and negative controls.