Repository navigation
[PR] feat: add compress.preserveRecentTurns and protectUserMessages (#4593) - #4620
Open
ranxianglei wants to merge 21 commits into
Open
ranxianglei wants to merge 21 commits into
ranxianglei wants to merge 21 commits into
Conversation
ranxianglei
changed the title
Owner
Author
(ai: vllm-qwen/qwen3.8-27b) 查重: 不重复。 独立检索 open+closed issues/PRs（preserveRecentTurns / protectUserMessages / turn retention / user message protection / 「按轮保留」多组关键词）：唯一相关条目是功能请求本身（GitHub #4592 / ework #4593）与本 PR，无其他实现尝试或同机制在途工作。 标签: scope/kernel, dsh（主改动面在 kernel 保护层；DSH 是驱动场景，机制本身通用）。功能类 PR 不打 bug、不评缺陷严重度。 说明：这是本 daemon 另一会话（ework #4593 工作区）完成的实现 PR，正文无 先报一个已确认的硬伤：PR 正文写 review 进行中：正在拉分支核对代码、跑定向测试、查 CI。 |
ranxianglei
force-pushed
the
2026-10-25_turn-retention-user-protection
branch
from
October 24, 2026 21:29
d791f53 to
22f6e21
Compare
Contributor
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g [email protected]This exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr4620.tgz
npm install -g packageInstall instructions are refreshed on each push. |
Owner
Author
(ai: vllm-qwen/qwen3.8-27b) 冲突已解决。 PR #4620 分支落后 master 约 497 个提交，rebase 到当前 master（ 结构化八问
判定：冲突清零、分支已 rebase 到当前 master 并 force-push，本地定向验证绿；待 CI 全量 + 人工 review（kernel 变更保持人工把关）。 |
added 2 commits
October 25, 2026 06:01
kernel-change: per-turn retention window + hard protection for user messages requested in #4593 — the host side cannot express either: turn boundaries are positional (slide per user message, incompatible with permanent BLOCKED-ref identity decisions) and user-message hard exclusion needs to ride the shared isMessageProtected predicate so every consumer (refs, projection, report, absorb/ccr guards) stays in sync automatically. - turn-retention.ts: computeTurnRetention — visible user-message boundaries (consecutive users = one, summary carriers excluded), strict window starting AT the (N+1)-th-from-last boundary - recommend.ts: Rule 4 unions the retention suffix into the soft zone (partial overlap = existing #3953 carve + warning, full containment = zone fatal) - protected.ts: protectUserMessages hard-excludes user text messages (BLOCKED refs); compress.ts gate extended so the predicate actually runs - unit-projection.ts: warnings/fatal attribute causes (last-N-turns retention vs classic recent zone vs protectUserMessages) - config.ts: validateConfig checks; types.ts optional fields (defaultConfig untouched -> golden wire-contract stays stable) - version 0.0.151 -> 0.0.152
Config surface (owner-gated mapping, per-tier like siblings): - compress.preserveRecentTurns?: number (0/unset = off): last N turns (turn = user message -> next user message) join the soft-protected zone; submitted ranges crossing the boundary are carved with an auditable warning, fully-contained ranges are rejected - compress.protectUserMessages?: boolean (unset = false): every user text message hard-excluded (BLOCKED refs, never advertised), same semantics as protectedTools - parseCompressSettings: finite >= 0 number (explicit 0 = off) and strict boolean validation; COMPRESS_SETTING_FIELDS updated - mergeCompress/applyCompressSettings: per-field global -> provider -> model cascade with ?? (an explicit off at any tier wins) - preflight overflow relaxation deliberately does NOT touch preserveRecentTurns: folding user instructions defeats the feature, so unresolvable overflow fails loudly instead - CONFIGURATION.md en+zh: table rows + deterministic behavior spec (reject / filter+receipt / priority order)
Owner
Author
(ai: vllm-qwen/qwen3.8-27b) 查重:不重复 —— 本条为 CI 自动修复车道(attempt 1/2、2/2 合并交付),四条红已逐条分账并对照 #4599 总控核实,无新立项。 CI 四条红全部已在分支修复并推送,新 head 修复分账
八问
预期残留红:重触发后 11 个 test 分片仍会命中继承的 LIVE_FILES tripwire(#4599 盘点第 2 条: 结论:点名的四条门禁本地验证全绿、已推送,等 CI 重触发;剩余红均属 #4606 工作流,非本 PR 代码问题。本车道收工。 |
added 2 commits
October 25, 2026 06:28
…ior; floor fractional preserveRecentTurns Review fixes for the turn-retention feature (review of this PR): - types.ts / turn-retention.ts / recommend.ts: three comments still described the abandoned design (ANY overlap rejected). The implemented behavior is two-tier: partial overlap carves at the boundary with an auditable receipt, full containment rejects. Also corrected the window-start wording in types.ts (the boundary itself is included). - turn-retention.ts: floor fractional N. The host floors at parse time, but acp-kernel ships standalone — a direct-API fractional value would index boundaries[] fractionally and silently degrade to off. - unit-projection.ts: zone error wording uses the effective (floored) N. - tests: fractional-N case (1.9 -> window of 1 turn).
…carrier-exclusion wording and kernel version refs Review fixes for the turn-retention feature (review of this PR): - The two new compress keys were added to COMPRESS_SETTING_FIELDS and to the generated index tables by hand, but their seeds were missing from website/config-reference/*.yaml — both the config-surface-guard and the docs-sync CI jobs fail on that. Add the seeds (cw.prturns / cw.pum) and regenerate all derived artifacts through tools/gen-config-docs.mjs. Regeneration also corrects the hand-edited table rows: the Env column is for BILI_* env vars (these keys have none), so it renders '—'. - Detail sections claimed summary carriers are role system; the actual exclusion is by SUMMARY_HEADER text prefix on user-role messages (tool results never match on role). Fix en+zh wording to match the code (kernel/src/protected.ts isUserTextMessage). - Kernel version refs 0.0.152 -> 0.0.164: the branch was rebased onto master, where acp-kernel is at 0.0.163, so this change ships as 0.0.164 (CI-enforced version discipline).
ranxianglei
force-pushed
the
2026-10-25_turn-retention-user-protection
branch
from
October 24, 2026 22:36
e204c60 to
9df2264
Compare
Owner
Author
(ai: vllm-qwen/qwen3.8-27b) Review 结论：修复 + rebase 后可合并（等新 head 先说我在 PR 分支上直接做的四类事（均已 push，独立标注 commit）：
结构化八问
测试有效性（④）：kernel 13 个测试有牙（精确块构成 deepEqual、精确 ref 值如 BLOCKED/m00001..m00005、warning/error 正则、defaults-off 负守护），能抓住 filter↔reject 回归与审计归因丢失；host 侧 4 个钉住 cascade/显式 false 胜出/parse 边界（0=off、3.7→3 floor、strict boolean 拒 "yes"/1）。本地定向全绿：kernel 保护族 76 + recommend 族 41 + turn-retention 13、bili compress-settings 47、verif-subset-guard（compress.ts 在 Tier-A 守护清单内）2/2、config-surface-guard ✓、docs-sync 幂等 ✓、typecheck ✓、build ✓。全量套件按纪律交远程 CI——新 head 的 CI 正在跑，转绿后我会补一行确认。 判定：可合并（CI 绿后）。合并请 owner 执行：#4620 |
added 2 commits
October 25, 2026 06:53
…ync gate) The rebase bumped kernel/package.json to 0.0.164 but the kernel's own lockfile (standalone acp-kernel identity, #2092) still recorded 0.0.163 — the lock-sync CI job diffs the regenerated kernel/package-lock.json and fails on that drift. (The original branch had the same latent drift from its 0.0.151 -> 0.0.152 bump.)
….164 (lock-sync gate) Follow-up to the kernel/package-lock.json sync: the root lockfile also carries the file:-link's recorded version under packages."kernel", which the rebase bump to 0.0.164 left at 0.0.163 — the lock-sync job's root regen diffs exactly that line.
Owner
Author
…ment #4316 PR3/3 (restampRefsToLedger) re-homes citation numbers once on every full-mode cold load; goldens pinning exact ref numbers are born-red. Re-pin the 7 storage-lane tests (engine-proofs x6 + orphan-recovery x1) modulo that renumbering — zero production code touched. - golden-storage-drive.ts: per-turn citation-normalized digest (shaCiteNorm) and state digest modulo stripOneTimeAlignment (stateDigestAligned); the deep-sort helper is hoisted into golden-canonical.ts for sharing. - corrupt-index-child.ts: new stripOneTimeAlignment normalizer — strips the refLedgerAligned stamp and re-homes refs onto canonical raw-id order so a clean full-mode read and the legacy fallback load compare equal. - storage-engine-proofs.test.ts: wire legs assert shaCiteNorm equality vs the legacy restart leg and the committed goldens; state legs assert stateDigestAligned; rollback leg now spells BILI_STORAGE_UNIFIED=legacy explicitly (#3818 flip: absent resolves to full, deleting the env would vacate the legacy path). - storage-orphan-recovery.test.ts: pins the content-addressable byRaw->byRef round-trip instead of a counter-era m00003 literal. - golden/storage-render/*.json regenerated with shaCiteNorm/stateDigestAligned. Converges the master-red storage-lane family (#4599 plan A, owner directive).
…dger parity masks The #4316 one-time renumbering restamped messageRefs but left state.tokenSnapshot keyed by the OLD refs, so post-restart bodies carried stale per-message token counts (1.3K vs 1.4K) and the full-mode-restart wire drifted off the legacy baseline (#4599). persist.ts: two-pass remap (stage moves, delete old keys, then write new keys) so entries whose new ref collides with another move's old ref survive; fires only when superseded refs exist. Aligned normalizer (corrupt-index-child.ts): the cacheLedger's segment bookkeeping (be/bl/bs, unfinalized tail line), the derived agg block, and per-line miss-attribution (prov/seam) are restart experience, not content — no-restart and restart legs legitimately differ there. Masked in the aligned view only; raw digests keep full accounting. Goldens regenerated (stateDigest + aligned). storage-engine-proofs 12/12, golden-storage-render 5/5 locally.
…persisted state (#4599) The fork-survival test asserted immediately after the compress tool response, but the fold lands in the live session state asynchronously (#3573 replay semantics). Slow CI runners persisted the pre-fold state ('0 block(s) (compressions=0)' cold load) and the deepEqual diverged. Bounded 5s wait on the in-memory fold; fast machines exit on the first iteration. Product semantics unchanged.
The #4316 cold-load alignment reads the unified sidecar; its ingest of just-flushed records completes asynchronously. Slow runners compared the pre-ingest state and saw refLedgerAligned missing while the fold itself had landed. Poll (bounded 5s) until the loaded state carries the flag; fast machines exit on the first probe.
Git-for-Windows autocrlf checks the bundled .md posts out with CRLF; the exact-match '---' fences then failed to parse and every post lost its title/excerpt/date frontmatter (Windows CI: 'bundled meet-bili post found' assertion). Normalize CRLF to LF inside parseFrontmatter when the fence is CRLF-shaped; LF content is byte-identical as before. Regression test with a CRLF fixture.
…help-merge sweep)
… storage mode (#4599) The #4316 cold-load alignment sets refLedgerAligned only when the unified store is available; config-clean machines (CI runners) run legacy mode where it is an honest skip. The test asserted the flag unconditionally, so every shard carrying this file was born-red on CI while unified-config machines passed. Expectation now mirrors store.replicasAvailable(); verified 49/49 in BOTH modes locally.
…help-merge sweep)
…mspec wrap on win32 (#4599)
The ci-check lane's expectation change (c496cfb) exposed a real Windows bug: stripInheritedProxy builds inlineEnv as a spread copy of process.env, which keeps the original key casing (ComSpec) — Windows env vars are case-insensitive but plain objects are not. planClientSpawn's exact env.COMSPEC probe then missed it and silently degraded every wrapped spawn to the bare-name 'cmd.exe' fallback (verified by repro on a real win32 box: childEnv.COMSPEC=undefined while process.env.COMSPEC was set). Fix: probe COMSPEC/ComSpec/comspec in planClientSpawn and planDshSpawn. Verified on real Windows (Node 22): launcher-opencode-inline + fix-679 + dsh-refresh 37/37; host tsc clean, 35 pass / 2 win32-skips.
…OMSPEC case-probe c496cfb re-derived the win32 expectation from ambient process.env, which still mismatches production: Node's process.env member access is case-insensitive on Windows while the child env handed to the planner is a plain object whose key keeps the OS casing (ComSpec), so the exact env.COMSPEC probe fell through to the bare-name fallback. The assertion now derives from the child's OWN captured env/options and probes COMSPEC/ComSpec/comspec exactly like planClientSpawn does post-dfeea8a11 — self-contained under any concurrent-test env churn in single-process CI shards — and additionally pins windowsVerbatimArguments=true on the wrap.
…help-merge sweep)
This was referenced Oct 26, 2026
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
kernel-change: #4593 requests two protection mechanisms the host side cannot express: (1) per-turn retention is POSITIONAL — the window slides forward with every new user message, so it cannot ride permanent BLOCKED-ref identity decisions (aged-out messages would stay blocked forever, or force per-turn BLOCKED→numeric migration = prefix-cache churn); it belongs in the kernel's per-turn-recomputed soft zone; (2) user-message hard exclusion must ride the shared isMessageProtected predicate so every consumer (ref assignment, unit projection, report, absorb/ccr guards) stays in sync automatically.
Closes #4592
What
Two opt-in keys under
compress, both off by default:preserveRecentTurns?: number— the last N turns (turn = user message → next user message) join the soft-protected zone alongside preserveRecentMessages/Tokens/latest-user pin. Strict window: starts AT a user-message boundary, never mid-turn. Consecutive user messages count as ONE boundary; summary carriers are not boundaries.protectUserMessages?: boolean— every user text message hard-excluded (BLOCKED refs, never advertised as compressible), same semantics asprotectedTools. Host-injected runtime-context user messages are covered too (deliberate conservative default: over-protection costs ratio, under-protection loses instructions).Submitted-range behavior (per the author's in-thread refinement — filter, don't reject):
Range is entirely within the protected zone (…), cause named)Config surface
compress.preserveRecentTurns(number, default off/0),compress.protectUserMessages(boolean, default off/unset)??merge)Behavior decisions (flagged per owner policy)
Tests
查重: 不重复（open+closed issues/PRs searched, EN+CN keywords, no match）
Kernel version: acp-kernel 0.0.163 → 0.0.164 (branch rebased onto master, where acp-kernel is at 0.0.163; single bump for this change set, CI-enforced)