decision request: retry truncated compression-draft turns (length-cut finish reasons) #6019

Activity

1. ranxianglei

   added

   p1

   scope/wiring

   on Nov 9, 2026
2. ranxianglei commented on Nov 9, 2026

   

   ranxianglei

   on Nov 9, 2026

   OwnerAuthor

   More actions

   🤖 Powered by ework · qwen3.8-27b

   (ai: vllm-qwen/qwen3.8-27b)

   查重：不重复。 Searched open+closed issues/PRs for a retry-gate extension to length-cut reasons — no existing item covers this gap. Related-but-distinct entries checked:

   • [Bug] Drifted open-tag names (acpt/acct/accent…) and mid-text draft envelopes reach the client verbatim — two thinking-leak shapes #5175 / PR [PR] fix(#5175): strip drifted open-tag names and mid-turn draft envelopes #5178 — detection + stripping of drifted tag names and mid-turn draft envelopes (different mechanism: payload hygiene, not retry eligibility); [PR] fix(#5175): strip drifted open-tag names and mid-turn draft envelopes #5178 is upstream of this decision — its isDraftDump predicate is what option C builds on;
   • Render-tag drift is deployment-dependent: 30% of adequately-exposed sessions on one self-hosted endpoint vs 1.1% elsewhere (p=8.6e-5) — measurement for #5175 #5349 — drift-rate measurement for [Bug] Drifted open-tag names (acpt/acct/accent…) and mid-text draft envelopes reach the client verbatim — two thinking-leak shapes #5175 (observability only);
   • OpenAI lane: the compression nudge never tells the model to use the compress tool, so it writes an <analysis>/<summary> draft as prose — and the draft-tail recovery nudge then steers it somewhere else #4764 / [PR] fix(openai-lane): name the compress tool in the draft-tail retry nudge #4785 — nudge wording never names the compress tool, so retries can produce prose drafts (retry-outcome quality, gate eligibility untouched);
   • [PR] Empty-turn continuation retry: off by default (opt-in switch) + neutral nudge wording #5897 / PR [PR] Fix 400 CONTEXT_WINDOW_EXCEEDED on anthropic-wire sessions bili believes fit #5899 — opt-in switch (degenerateRetry, default OFF) + neutral nudge wording for the degenerate-retry family (billing posture; sequencing interaction with this decision, see point 6);
   • decision request: plugin degenerate-turn retry failure mode — passthrough toggle design (#4913) #4919 (closed) — decision record for the double-degenerate fatal-error mode in the same subsystem (second-attempt handling, not entry eligibility).
     No open [PR] line targets this thread (verified against the mirror list).

   标签: bug / opencode / scope/wiring / p2 / needs-decision — real convergence defect seen live in OpenCode plugin mode; the gate lives in the client-integration layer (plugin pipes + core loop), same subsystem as #4919; p2 = edge-shape UX breakage with a manual-continue workaround (feel free to re-grade at decision time).

   Verification (live: master @ 406193b, PR #5178 @ c6eaf1f). All claims in the body check out against code, with one correction that sharpens the decision:

   1. ✅ Detection claim holds. isDraftDump exists only in PR [PR] fix(#5175): strip drifted open-tag names and mid-turn draft envelopes #5178 (open; CI as of 09:19Z: 30 checks, 0 failed, 14 pending, mergeable_state blocked): true when the </analysis>→<summary> envelope is present without a later </summary> close — i.e. exactly while the draft is unfinished. Master today only has endsWithDraftClose (closed-tail form).
   2. ✅ Plugin chat pipe: CLEAN_TURN_REASONS = {"stop","end_turn","stop_sequence"} (src/plugin.ts:2529); retryEmptyTurn bails at src/plugin.ts:2794 on reason === undefined || !CLEAN_TURN_REASONS.has(reason) → length/max_tokens cuts are convicted but never retried.
   3. ✅ Plugin responses pipe: gate at src/plugin.ts:4309 requires status === "completed"; a length cut lands as incomplete → blocked.
   4. ⚠️ Correction — the core lane is a blacklist, not a clean-whitelist. The DSH native mode: assistant text ending with </summary> is treated as a completed step, so the turn silently stops (149 samples from session transcripts) #2303 draft-tail gate (src/loop/core.ts:1179-1199) excludes only {failed, incomplete, error} plus requires sawDone && !truncatedDone. The openai adapter emits done with the RAW finishReason and no truncated flag (src/loop/adapter-openai.ts:583, 588-592); anthropic likewise with raw stop_reason. So once [PR] fix(#5175): strip drifted open-tag names and mid-turn draft envelopes #5178 swaps in isDraftDump, proxy-mode openai/anthropic length-cut whole-turn dumps will already be retried — length/max_tokens are not in the exclusion set. Consequence: under option A the three lanes DIVERGE — proxy users get the retry for free while the reported OpenCode plugin-mode population stays broken. Option C would align all lanes.
   5. ✅ Budget claim holds: the one-shot degenerateRetried budget is shared per request across all degenerate paths (src/plugin.ts:2757), so C adds at most ONE extra upstream call per request.
   6. ⚠️ Sequencing note: PR [PR] Fix 400 CONTEXT_WINDOW_EXCEEDED on anthropic-wire sessions bili believes fit #5899 (from [PR] Empty-turn continuation retry: off by default (opt-in switch) + neutral nudge wording #5897, following closed decision record decision request: plugin degenerate-turn retry failure mode — passthrough toggle design (#4913) #4919) proposes making the whole degenerate-retry family OPT-IN behind a new degenerateRetry switch (default OFF, env BILI_DEGENERATE_RETRY=1) with neutral nudge wording. If [PR] Fix 400 CONTEXT_WINDOW_EXCEEDED on anthropic-wire sessions bili believes fit #5899 lands before option C, C's extended gate only fires for opted-in users; if C ships first, [PR] Fix 400 CONTEXT_WINDOW_EXCEEDED on anthropic-wire sessions bili believes fit #5899 will presumably cover it too. Worth deciding together.

   Independent option analysis:

   • A (status quo): zero code change, zero extra spend — but it cements the lane divergence from point 4. The reported population keeps seeing raw drafts until they manually continue.
   • B (any convicted dump): widens the gate for ALL convicted dumps including mixed turns (draft + real answer + tool call). Those turns carry visible work the host may act on — retrying discards it and can stack two drafts. Largest billing blast radius. Not recommended.
   • C (whole-turn dumps only): extend eligibility ONLY where conviction is isDraftDump ∧ no tool call ∧ the draft IS the entire visible text — which is exactly the existing draftTail predicate shape; the only missing piece is reason acceptance. Accept CLEAN_TURN_REASONS ∪ isMaxOutputReason (helper already exists at src/degenerate-turn.ts:29-32, wire vocabulary from [Bug][DSH] 长会话切换自定义模型后输出预算恢复到 249 token，连续 thinking-only / max-tokens（#1840 零上限来源场景） #2904) in the draft-dump branch of both plugin pipes (chat: finish reason; responses: also accept incomplete). Empty-turn branch keeps its clean-only whitelist; core lane needs no change for openai/anthropic (already passes), responses/google stay excluded. Worst case: model re-dumps on retry → two stacked drafts + the existing Plugin passthrough leaves the client an empty turn when the model echoes a render tag #870 in-band error (plugin) or pass-through (core) — same failure class as today's clean-stop DSH native mode: assistant text ending with </summary> is treated as a completed step, so the turn silently stops (149 samples from session transcripts) #2303 retry, not a new one. Per decision request: plugin degenerate-turn retry failure mode — passthrough toggle design (#4913) #4919's production dataset (~9/13 continuation retries succeed), expect most affected turns to recover.

   八问（结构化结论）

   1. 是什么问题：整轮可见文本全是压缩草稿（<analysis>…</analysis>+<summary>…）时被上游长度上限截断的回合——检测能定罪，但重试门只认干净结束，所以永远不会重试；用户看到裸草稿、回合不收敛。
   2. 什么原因导致的：两条 plugin 车道的重试资格门是白名单式「干净结束」，而长度截断的结束原因是 length/max_tokens（responses 线是 incomplete），不在名单内。层次上：门的原始注释担心「截断的重试只会再截断一次」——对空回合成立，对「整轮即草稿」不成立：那里重试是继续写草稿，不是重放截断。另有一处纠正（见上第 4 点）：core 车道是黑名单而非白名单，[PR] fix(#5175): strip drifted open-tag names and mid-turn draft envelopes #5178 合并后 proxy 模式的 openai/anthropic 长度截断整轮草稿已经在被自动重试。
   3. 影响范围与严重程度（按选项）：
      • A：零代码变更；但 [PR] fix(#5175): strip drifted open-tag names and mid-turn draft envelopes #5178 合并后三车道行为分裂——proxy 用户免费获得重试，报告的 OpenCode plugin 群体继续看到垃圾草稿直到人工续。严重度：特定条件下有问题（整轮草稿+长度截断），可人工续，p2。
      • B：所有被定罪草稿（含混合回合）都允许长度截断重试；最坏每请求多一次全价调用，客户端可能看到两份叠放草稿，且混合回合里已有的真实内容被丢弃重来。
      • C：仅「整轮即草稿 ∧ 无工具调用 ∧ 长度原因」形态多至多一次有预算上限的继续调用（复用一次性 degenerateRetried）；其余路径逐字不动。
   4. 怎么修复的（推荐 C）：两条 plugin 车道的 retryEmptyTurn 中，把草稿定罪分支的合格结束从 CLEAN_TURN_REASONS 扩为 ∪ isMaxOutputReason（现成 helper，src/degenerate-turn.ts:29-32；responses 线同分支接受 incomplete）；空回合分支保持干净白名单不变；core 车道 openai/anthropic 已放行无需改动，responses/google 保持排除。不改任何既有判决，只补一个入口。
   5. 回归风险（按选项）：A=无代码风险，但引入车道间不对称（新状态）。B=把空回合/混合回合拉进长度截断重试，扩大 Plugin passthrough leaves the client an empty turn when the model echoes a render tag #870 双退化语义与计费面，回归面最大。C=新增面严格限定在「草稿即整轮 ∧ 长度原因」；既有干净结束行为逐字不变；已识别的最坏情形=模型重试时再次倾泻草稿（两份叠放+Plugin passthrough leaves the client an empty turn when the model echoes a render tag #870 或透传），与 DSH native mode: assistant text ending with </summary> is treated as a completed step, so the turn silently stops (149 samples from session transcripts) #2303 现有失败模式同类而非新增类。
   6. 是否破坏缓存：不涉及——重试走既有 refetch 机制（原请求体+nudge 尾随 user 消息，前缀不变，前缀缓存命中不受影响），与现有 DSH native mode: assistant text ending with </summary> is treated as a completed step, so the turn silently stops (149 samples from session transcripts) #2303 重试完全同路。
   7. 是否改了设计方向：不改架构与边界，只是把既有的「草稿即停滞」判决在长度截断入口补齐。需一并定的排序问题：[PR] Fix 400 CONTEXT_WINDOW_EXCEEDED on anthropic-wire sessions bili believes fit #5899（degenerateRetry 默认关）若在 C 之前合入，C 的效果会被该开关罩住（仅 opted-in 用户生效）。
   8. 是否引入了新的配置：C 本身不引入新配置（复用 isMaxOutputReason 与既有预算）；若叠加 [PR] Fix 400 CONTEXT_WINDOW_EXCEEDED on anthropic-wire sessions bili believes fit #5899 则受其 degenerateRetry 开关管辖。

   Awaiting your pick: A / B / C (my recommendation: C). Once decided I'll open the single fix PR for this thread (no [PR] exists for it yet). If you'd rather sequence with #5899 first, say so and I'll hold.
3. ranxianglei

   added

   bugSomething isn't workingSomething isn't working

   needs-decision

   opencode

   p2

   on Nov 9, 2026