Repository navigation
[PR] kernel: add receipt stop-signal clause to the lean acpTags (#6930) - #6948
ranxianglei wants to merge 6 commits into
Conversation
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g billion-context@0.0.11012-pr.6948.1f22deaThis exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr6948.tgz
npm install -g packageInstall instructions are refreshed on each push. |
…acpTags kernel: add receipt stop-signal clause to the lean acpTags (#6930) The standing compress bullet orders the model to compress consumed history but never tells it to stop, so after a "No compressible ranges remain" receipt the model keeps issuing small fragment folds (wasted summarize rounds + clamped/guessed end refs). Add one counter-clause bullet directly after the standing bullet: compress receipts are authoritative stop signals — issue no further compress calls until a receipt/nudge names ranges to fold or the user explicitly asks. No numbers in the text (nothing to propagate, §7.3). kernel-change: prompt-surface counter-clause for the standing compress bullet; the prompt text is kernel-owned (packs.ts LEAN acpTags), the host cannot add it. Regression test asserts the bullet is present and positioned after the standing bullet (kernel/tests/packs.test.ts).
37b2c6d to
5ee4f7f
Compare
(ai: vllm-qwen/qwen3.8-27b) 查重: 不重复。 This PR is the only open [PR] targeting #6930 (checked the mirror's open [PR] list). Related #6928 / PR #6929 is the status-line half (already shipped) — different mechanism (per-turn status line re-ordering folds), complementary, not a dup. Review of PR #6948 (head Code review — PASS. Every factual claim in the PR body verified against the tree (not taken on faith):
Consumption chain (model-facing text, §7.3): the model will record "stop folding after a dead-end receipt; resume only when a receipt/nudge names ranges or the user asks." No numbers/refs/positions in the text — nothing to propagate past expiry. Over-application risk is bounded by the two named resume conditions. One residual tension: the already-compressed receipt's own "run acp_status and target one of the CURRENT compressible ranges" advice is a third resume path the clause doesn't name — but over-stopping costs at most one missed fold, while under-stopping is the #6930 pathology (wasted rounds + clamped refs). Net improvement. CI (live, mirror check-cache @ head
八问(结构化结论):
判定: 代码本身审过、可合;3 个 CI 红里 2 个是机械性的,我正在分支上修(一次 push,CI 重跑待出); |
🔶 需要决策 (ai: vllm-qwen/qwen3.8-27b) 背景: PR #6948(kernel: 给 lean acpTags 加 receipt 停止信号条款,Fixes #6930)被 待决问题: 这个 kernel 变更在线程里没有 owner 的明确批准(#6930 与 PR #6948 的评论全是 bot 帖)。PR body 里写的是 选项:
不决定的默认行为: PR 停在 kernel-guard 红、不可合并;test lane 因门禁链保持 skipped,代码进不了 CI 全量回归,直到声明就位。 |
(ai: vllm-qwen/qwen3.8-27b) CI 状态更新(新 head |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Closing summary for this review pass: Done & verified:
Remaining — your call (🔶 需要决策, floor above): the only red is
Risk if no decision: PR stays blocked on kernel-guard, test lanes stay skipped; nothing else degrades. Also on record: kernel version collision with PR #6926 (both bump 0.0.198→0.0.199 — the second one to merge must rebase to 0.0.200). |
🔶 需要决策 (ai: vllm-qwen/qwen3.8-27b) Re-verification at the current head
Per the 2026-11-01 / 2026-10-26 authorization-gate policy I will not append the declaration myself: the gated diff (the lean acpTags stop-signal bullet) is not a mechanical kernel bug I fixed, and there is still no explicit owner approval in this thread. The decision request from floor 5 stands unchanged — PR #6948, gated diff =
On record again: kernel version collision with PR #6926 — both bump 0.0.198→0.0.199; whichever merges second must rebase and bump to 0.0.200 (+ both lockfiles). 八问(结构化结论):
判定: 机械项已全部转绿且在新 head 上复验保持;唯一红 = kernel-guard 授权门禁,等 owner 拍板(A/B),我停手不再动这个 PR。 |
🔶 需要决策(pending — 复核确认,无新请求) (ai: vllm-qwen/qwen3.8-27b) Re-verification for this wake (all live, this run):
I am stopping here per policy — no commits pushed, no PR-body edits, no duplicate decision comment. Waiting on your A/B; once you approve (A), the one-line canonical 八问(结构化结论):
判定: 无新工作;唯一红 = kernel-guard 授权门禁,等 owner 拍板(A/B),停手。 |
…eceipt-stop-clause
(ai: vllm-qwen/qwen3.8-27b) Merge conflict resolved and pushed — branch head is now What I did this run
Known remaining red (not re-escalating): 八问(结构化结论):
判定: 冲突已解决并推送(7509a95,mergeable: true);本 lane 机械工作完成。唯一遗留红 = kernel-guard 授权门禁,等 owner A/B 拍板(已挂 needs-decision,不重复立帖)。 |
(ai: vllm-qwen/qwen3.8-27b) Re-verified live — no new work to do this run.
No commits, no pushes, no body edits, no duplicate decision comment this run. Waiting on the owner's A/B. 八问(结构化结论):
判定: 无新工作,维持等待 owner A/B 拍板(needs-decision 已挂)。 |
Summary
The standing compress bullet in the lean acpTags orders the model to compress
consumed history but never tells it to stop, so after a "No compressible
ranges remain" receipt the model keeps issuing small fragment folds (wasted
summarize rounds + clamped/guessed end refs). This adds one counter-clause
bullet directly after the standing bullet: compress receipts are authoritative
stop signals — issue no further compress calls until a receipt/nudge names
ranges to fold or the user explicitly asks.
kernel-change
Prompt-surface counter-clause for the standing compress bullet. The prompt text
is kernel-owned (
kernel/src/packs.tsLEAN acpTags — one object shared verbatimby the top-level surface and the pi adapter); the host cannot add it. The fix is
scoped to the lean lane only: the default acpTags (
kernel/src/compress-tools.ts)is a tag-format section with no standing compress instruction, so it needs no
counter-clause.
Changes
kernel/src/packs.ts: one new bullet directly after the standing compress bullet.kernel/tests/packs.test.ts: regression test — the bullet is present, positionedafter the standing bullet, and names the stop triggers + the two resume conditions.
kernel/package.json(+ lock): version 0.0.198 → 0.0.199.Regression inference (§7.3 — model-facing text ADDITION)
when ranges are named or the user asks." The clause cites receipt fragments the
system actually emits:
No compressible ranges remainis verbatimNO_RANGES_REMAIN_TEXT(host-emitted,src/stream.ts:637);already compressed/ no-op are the kernel applyRanges no-op wordings (
kernel/src/compress.ts).No phantom values introduced.
nudge names ranges or the user asks — the clause names exactly those two resume
conditions. No ceiling/position/count fields added, so there is no new anchor to
carry past expiry.
folds (the banner names the ranges), first-sight mass folds.
Tests
kernel/tests/packs.test.ts: 27/27 pass (new test asserts presence + position +stop/resume markers).
kernel/tests/surface-config.test.ts: 18/18 pass.tsc --noEmit: clean.tests/golden/wire-contract/kernel.json) does not include the leanacpTags — no regeneration needed.
Related: #6928 / PR #6929 (status-line drip removal — the status-line half, already
shipped). This PR is the prompt-side half that remains.
Fixes #6930