Repository navigation
[PR] fix(#5080): teach only the single-string compress form in proxy-lane receipts - #5105
ranxianglei wants to merge 0 commit into
Conversation
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g billion-context@0.0.7159-pr.5105.4f02f68This exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr5105.tgz
npm install -g packageInstall instructions are refreshed on each push. |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) 查重:不重复。 本 PR 是 #5080 的实现 PR( 标签: Review 结论(独立重审,非复读 issue 分析)方向正确、边界处理准确:只做 advertise-side 对齐,kernel parser 未动(legacy object-form salvage 仍有效,新测试钉住);plugin lanes 不受影响(KDD #7 抑制工具注入)。"外部模式每块需至少一行 hint"的说法我对照 kernel 实际行为核实过,准确: 但 PR 自身引入了 9 个红 shard——全部确定性、零 flake、本地全部复现。 三处遗漏:
证据链:master 在 base 已在分支直接修好,3 个标注 commit(新 head
本地验证:定向 7 套件共 78/78 通过 + 结构化八问
收工判定:可合入,待新 head |
(ai: vllm-qwen/qwen3.8-27b) 查重: 无需单独立项——这是 PR #5105(fix #5080)自身的 CI 红,按 #4599 规则直接在 PR 分支处理,不为 CI 红开新 issue。 结论先行: 9 个红 shard 全部源于本 PR 有意的措辞变更漏更了另外 4 处钉住旧文案的测试面;修复已推 PR 分支,head 八问
行为变更披露(必须项): 9 处模型可见文案由「对象/数组双形式并教」改为「只教 single-string 形式」( 过程备注: 两条 CI 修复车道并行处理同一红,各自独立定位到同一根因、产出等价修复;本车道的修复留在本地未推,避免覆盖对方已推提交(两者 golden 完全一致,仅一处测试断言写法取径不同)。绿 run 之后系统自动打标又触发了几次 label 事件 run,按 workflow 设计这类事件 scope=skip(重活跳过、报绿),不影响上述权威绿结论。 判定: CI 绿(head |
🔶 需要决策 · (ai: vllm-qwen/qwen3.8-27b) I picked up the "PR conflicts with base" blocker for PR #5105 (fix #5080) and stopped before pushing anything — the conflict is not mechanical, it's a head-on design fork that needs your call. Branch is untouched (still at its last head What I found
Not cosmetic — the two branches hand the model contradictory instructions about the compress Why I won't force a side
Options
Recommendation: lean B — #5555 is already merged on master and intentionally supports + teaches array-in-string, so #5105 looks superseded. But that only holds if #5080-class Default if undecided: branch left untouched, no push, stopping here. Label 结构化八问(决策请求)
判定:暂停等 owner 拍板(默认不动分支、不 push)。 |
Re-verified live first (as instructed) — the red is still live, not already fixed. Mirror cache for PR #5105 @ head What I inspected (one representative shard, ubuntu 22,3):
Why I'm not applying a mechanical fix here:
So I'm holding rather than speculatively editing an unfamiliar subsystem on a head that's about to be replaced. Both blockers (conflict + this red-CI) share the same unlock: your A/B/C decision. Once you pick, I'll integrate master cleanly (per-file 3-way vs the true base + regenerate the storage-render goldens), re-verify, and fix anything that still reds on the new head against fresh master. If you'd rather I dig into the #5140 409 in isolation on the stale head right now anyway, say the word — but I recommend resolving the direction first. 结构化八问(本轮 CI 诊断)
判定:维持 |
Re-verified live this run — unchanged: still 5 red test shards + New evidence this run (why the red is staleness, not a PR defect): since the head, master has reworked the exact subsystem that's failing —
So the To unblock (one pick): A = keep this PR's single-string-only wording (I override master's #5555 wording, keep its parser salvage) · B = adopt master/#5555 (this PR's core change superseded → I realign it) · C = reconcile to one canonical form. On your pick I'll do it in one pass: per-file 3-way vs the true base + regenerate the storage-render goldens + re-verify; the staleness red should clear, and I'll fix any residual red against fresh master then. 结构化八问(增量状态)
判定:维持 |
Re-verified live this run — still unchanged: Useful context I confirmed from the full thread this run — this is not new and not a quick mechanical fix:
I'm holding, not re-posting the diagnosis: the design call (A = keep this PR's single-string-only wording · B = adopt master's #5555 array-in-string · C = reconcile to one form) is owner-only per policy; I won't make it myself or do out-of-scope fork/snapshot surgery on a head that will be rewritten post-decision. Unblock = one word: reply 判定维持 |
(ai: vllm-qwen/qwen3.8-27b) Already posted the full status a moment ago — nothing has changed since. One-line state:
Unblock = one word: reply |
Re-verified live this run (mirror cache): unchanged — Why the red-CI is not a small mechanical fix I can apply in this auto-fix lane:
Still blocked on the owner's A/B/C (needs-decision label applied, decision posted). On the pick I integrate master in one pass, which clears BOTH the conflict and the 5 red shards; then re-verify and fix any residual red against fresh master. 判定维持挂起:5 分片红是 stale-head 的 fork/snapshot 409( |
(ai: vllm-qwen/qwen3.8-27b) Restarted after the 33-min hang — reporting state, nothing has moved. Re-verified live (this run):
This is a genuine deadlock only you can break. Both blockers share one unlock — integrating current master — and that integration is gated on the A/B/C compress-wording decision (this PR teaches single-string-only; master's #5555 teaches JSON-array-in-string; both are deliberate and test-pinned, so picking one reverts the other). That's a product-direction call, not a mechanical conflict, so per policy I won't make it myself. The 5 red shards are a stale-head fork/snapshot 409 whose fix already lives in current master (846+/92− across 7 files + ~14 sidecar/snapshot tests), so they clear in the same integration. There is no small-mechanical-fix path I can take in this auto-fix lane without either (a) making the A/B/C call for you, or (b) doing a half-integration of master that gets discarded the moment you pick. So I'm holding rather than thrash. Unblock — one of:
判定维持 |
(ai: vllm-qwen/qwen3.8-27b) Re-verified live (this run): still |
1 similar comment
(ai: vllm-qwen/qwen3.8-27b) Re-verified live (this run): still |
(ai: vllm-qwen/qwen3.8-27b) Conclusion for this run (re-verified live, not from memory): Done (this run): re-checked the live mirror cache for PR #5105 — still Verified: the conflict is a genuine A/B/C design fork (this PR teaches single-string-only; master's #5555 teaches JSON-array-in-string; both test-pinned, can't both pass). The 5 red shards are a stale-head fork/snapshot 409 whose fix already lives in current master (846+/92− across 7 files + ~14 sidecar/snapshot tests since the head). Remaining / risk: both blockers share one unlock — integrating current master — gated on your A/B/C call. I won't make that call myself (owner-only) or do out-of-scope fork/snapshot surgery on a head that will be rewritten post-decision. Nothing pushed; branch untouched; needs-decision label applied. Unblock = one word: |
(ai: vllm-qwen/qwen3.8-27b) Re-verified live (this run): still |
(ai: vllm-qwen/qwen3.8-27b) Re-verified live (this run): still |
(ai: vllm-qwen/qwen3.8-27b) Re-verified live (this run): still |
1 similar comment
(ai: vllm-qwen/qwen3.8-27b) Re-verified live (this run): still |
c8b319f to
a259e0b
Compare
What & why
#4252 narrowed the advertised compress schema to exactly ONE shape —
contentis a single plain string holding every range (mNNNNN–mNNNNN optional topicheader line + that block's summary markdown). But five proxy-lane model-facing texts still taught the retired object/array forms, so guided-decoding models blended the two advertised shapes into parser-invalid hybrids and hit theno-valid-rangesrejections reported in #5080.This aligns all nine model-facing strings to the single advertised form:
src/stream.ts— the fiveapplyRanges()rejection receipts: armed-refusal, empty-call, corrupt-args, retry line, and theno-valid-rangesreceipt (the exact text pasted in the issue).src/compress-tool.ts— standing notes appended to proxy-lane nudges:FIRST_SIGHT_DRAIN_NOTE,FIRST_SIGHT_DRAIN_NOTE_EXTERNAL,SUMMARY_BUDGET_NOTE.src/external-summary-surface.ts—EXTERNAL_SUMMARY_NOTE(now also requires ≥1 hint line, sinceparseLineEntryrejects empty summaries andwithOptionalSummariesonly fills object entries).Scope & non-changes
kernel/src/parse-compress-input.ts) still salvages legacy object/flat shapes and is untouched — matching the compress calls rejected when the summary text rides incontentnext to top-level startId/endId — plus U+2212 dash silently folds the wrong range #4252 design ("narrowing is advertise-side ONLY").COMPRESS_PARAMETERSstayscontent.type === "string"), no golden change.kernel-guarddeclaration needed.Behavior change disclosure
Nine model-facing prompt texts changed wording (object/array form → single-string form). The three byte-stable note constants in
compress-tool.ts+external-summary-surface.tsanchor prefix-cache segments, so this is a one-time cache invalidation on upgrade — same class as any prompt-text change. Receipts are ephemeral tool results (not cached). Acceptance set is unchanged.Config surface
None. No new/renamed/re-semantics'd config field, env var, or CLI flag.
Tests
New
tests/issue5080-compress-form-single.test.ts(6 tests) pins: schema staystype:stringwith no top-levelrequired; each of the five receipts teaches only the string form (including the exact #5080no-valid-rangessignature); and legacy object-form calls still parse (salvage intact). Updated three assertions intests/compress-observability.test.tsto the new wording. Local targeted run:tsc --noEmitgreen + 48/48 pass across the four touched test files. Full-suite verdict deferred to remote CI.Closes #5080