Repository navigation
[PR] fix: classify the dsh image-offload placeholder so it cannot steal the last-user anchor (#6021) - #6027
Conversation
…e last-user anchor (#6021) dsh compaction-image-offload rewrites an OLD message's inline image into placeholder prose when the route's request budget is exceeded. The rewrite changes content identity, so bili re-ingests the cell as a NEW row (fresh ref / storage seq / timestamp) — a two-day-old message now sitting at the tail of the resent history. Unclassified, the placeholder was a plain user-role text message: once everything after the newest human request had folded away, that stale prose was the last visible user message, won the Rule-3 last-user anchor (#5109) and pinned the protected zone on the stale topic — the model answered it for the whole turn (#6021 repro, bili 0.1.213 / DSH 0.2.2-rc.1). #5752 shipped the sibling live handle ("; request preview ") knowing this slot would next carry this placeholder and deliberately left it unclassified; this change lands that deferred family now carrying production traffic (KDD #9 evidence-permitlist, dsh lane only). - new dshImageOffloadPlaceholderPath classifier -> oneshot/image-offload-placeholder (substring match on purpose: [text,image] originals splice prose AFTER the "]" tail); both route variants (image-budget + text-only) share one path - wired into BOTH registries (slot-exclusion + drift accounting) and into dshCarrierRawMarkers (stored-row backfill pre-filter) - behavior change: placeholder cells lose last-user-anchor eligibility and join expected-rebase drift accounting (their own retirement is accounted instead of alarming); T2c pin updated 1 -> 2 accordingly - tests: new tests/issue6021-image-offload-anchor.test.ts (units pos/neg per #3617, BOTH-registries wiring, e2e anchor geometry through the real dsh classifier — THE repro data-loss pin + no-bloat pin), registry row in issue5173, count pin in issue5142
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g billion-context@0.0.8804-pr.6027.428eb80This exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr6027.tgz
npm install -g packageInstall instructions are refreshed on each push. |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Reviewed PR #6027 against live state, synced it to current master, and pushed. Status below. What I verified (live, just now):
Review judgment:
8问(结构化审核结论)
标签已在位且与 diff 相符(bug / p1 / dsh / scope/wiring)。判定:可合并——等同步后 head |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) CI follow-up on the synced head |
…the offload placeholder A [text, image] original keeps its prose before the '[' bracket, so the marker can sit mid-text; pin that the documented substring match classifies it (coverage folded in from the parallel analysis in ranxianglei#6027).
|
(ai: vllm-qwen/qwen3.8-27b) This PR fixes the same issue (#6021) as #6024, which was opened ~40 minutes earlier by an external contributor, and the src changes are functionally identical (same classifier Per repo rules (first-come-first-served; AGENTS.md §4 "External contributors' PRs — fix directly, don't replace"), #6024 is kept as the canonical vehicle. The one unique coverage point here that #6024 lacked (the words-before-marker splice-shape positive) has been folded into #6024 as commit c855d6f, so nothing substantive is lost by closing this one. Recommend closing this PR as a duplicate of #6024. (Closure left to the owner — agents do not close.) |
…fload placeholder (folded in from parallel PR ranxianglei#6027) The unit rows pin that the classifier resolves; this file pins the actual failure mode of ranxianglei#6021 through applyCompression with the REAL dsh registry: a fold covering the offloaded tail must keep the live request visible (the stale placeholder loses Rule-3 last-user anchor eligibility), and a placeholder ahead of the live request still folds normally (no bloat).
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Independently verified the duplicate determination before agreeing — all three claims check out against live state:
One gap the summary missed, now fixed: beyond the unit row, this PR also carried the end-to-end anchor-geometry pin ( Actions taken:
Net result for the owner: merge #6024 (it now carries the complete fix + full test coverage), close #6027 as its duplicate. |
📦 Merged Build PublishedThis merge to master was published as npm install -g billion-context@0.0.1122-master.c812a84 # this exact build
npm install -g billion-context@master # always the newest merged build
|
Closes #6021
Root cause
dsh compaction-image-offload rewrites an OLD message's inline image into placeholder prose when the route's request budget is exceeded. The rewrite changes content identity, so bili re-ingests the cell as a NEW row (fresh ref / storage seq / timestamp) — a two-day-old message now sitting at the TAIL of the resent history. Unclassified, the placeholder is plain user-role text: once everything after the newest human request has folded away, that stale prose is the last visible user message, wins the Rule-3 last-user anchor (#5109) and pins the protected zone on the stale topic — the model answers it for the whole turn (#6021 repro; bili 0.1.213 / DSH 0.2.2-rc.1, ~6900-event session).
#5752 shipped the sibling live handle ("; request preview ") knowing this slot would next carry this placeholder and deliberately left it unclassified; this change lands that deferred family now carrying production traffic (KDD #9 evidence-permitlist, dsh lane only).
Changes
dshImageOffloadPlaceholderPathclassifier ->oneshot/image-offload-placeholder; substring match on purpose ([text,image] originals splice prose AFTER the "]" tail — a prefix match misses the m03269 shape); both route variants (image-budget + text-only) share one pathdurableMessageToolPaths.dshslot exclusion +oneShotCarrierPaths.dshdrift accounting) and intodshCarrierRawMarkers(stored-row backfill pre-filter)Behavior change disclosure
Config surface
None. No new key / env var / CLI flag.
Kernel
Untouched (host-side classifier only).
Tests