Repository navigation
fix(acp-status): aggregate pre-compaction archive section by reason in default view (#3105) - #3111
Conversation
📦 Built Package ArtifactBranch: Option A — Install this exact build (recommended)npm install -g billion-context@0.0.3557-pr.3111.c62be19This exact version is immutable and reproducible. The rolling Option B — Download artifact
tar xzf billion-context-pr3111.tgz
npm install -g packageInstall instructions are refreshed on each push. |
d8e70c7 to
cce5073
Compare
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) PR #3111(修 #3105)review 完成。独立核对了 diff、下游消费者代码、CI 全量日志(6820 pass / 4 fail)。结论:代码可合,CI 红是 master 既有噪音,与本 PR 无关。 结构化八问
CI 状态(独立核实):静态门全绿;test 四平台(ubuntu/windows × node22/24)全红,但失败恰好 4 个,全部是 判定:按 §7.4 属人工合并档(model 可见输出格式变更 + 新增 opt-in 工具参数,虽单模块纯 fix)。建议:#3119/#3120 落地后 rebase 一下再合(让 CI 全绿留痕);若 owner 认可该红为已知无关噪音,现在也可直接合并。代码侧我这边没有任何需要补的东西。 |
cce5073 to
0093618
Compare
📦 Merged Build PublishedThis merge to master was published as npm install -g billion-context@0.0.428-master.2f22bc6 # this exact build
npm install -g billion-context@master # always the newest merged build
|
Behavior change (intended, #3105): the default
acp_statusPRE-COMPACTION ARCHIVE section changed from one line per archived block (b1 — <reason>×N) to per-reason counts (<count>: <reason>) plus a hint pointing at an opt-in detail view. For the issue's exact scenario (237 blocks, one reason): ~20.2 KB of identical lines → 3 lines (~0.3 KB); the whole default receipt measured 740 B after vs ~20.9 KB before. Per-block ids + reasons stay reachable viaacp_status({ archiveDetail: true }), capped at 50 lines with an explicit omitted count (+187 more; showing first 50 of 237), so a 1,000-block archive still costs one bounded section.What changed
src/acp-status.ts: newformatPreCompactionArchive(archive, detail)— default aggregates by reason (count-desc), detail mode lists per-block lines capped atPRE_COMPACT_ARCHIVE_DETAIL_CAP = 50. The header was also generalized: it previously asserted "content was replaced by the client's native compaction summary" for every entry, which mislabels [Bug][proxy] unframed host-side history deletion leaves zombie fold blocks rendering carriers forever — one carrier flaps in/out every turn (persistent prefix-cache churn) #2695 substrate-destruction archives; it now names both mechanisms and each reason keeps its own wording on its count line.tests/compaction-archive.test.ts— updated the acp_status assertion (default no longer names individual blocks; detail mode does) + 3 new regression tests (300 same-reason blocks → constant-size 3-line section with the reason stated exactly once; multi-reason → accurate separate counts; detail cap 50 + omitted-count footer).tests/marker-render-contract.test.ts— new case proving both shapes are consumed whole by the marker body consumer (the acp_status visibility markers: stripMarkerLines removes the head line but leaves the multi-line report body in incoming history (#1029 follow-up) #2789/Inbound markers: one declarative strip engine for all four wires (#2789, #2798, #2811 plan B) #2815 drift guard).Downstream consumers checked: both text parsers anchor on the unchanged
PRE-COMPACTION ARCHIVE —head prefix and treat indented lines as continuation (src/acp-panel.ts,src/marker/status-body.ts) — all new lines stay indented, render-contract test green for both modes.Config surface: none. No config-file key, env var, or CLI flag.
archiveDetailis an optional tool-call argument parsed host-side inhandleAcpStatus; it is deliberately NOT added to the kernel tool schema (kernel changes are gated) — same pattern as the existing Tip-line-taught args (scope/view/tool) that the host already accepts whileACP_STATUS_TOOL_OPENAI.parameters.propertiesstays{}with noadditionalProperties:false, so providers pass the extra key through.Verification: targeted suites green locally (
tests/compaction-archive.test.ts,tests/marker-render-contract.test.ts— 15 tests), typecheck clean; full-suite verdict = CI on this head. Rendered the issue's scenario end-to-end: default receipt constant-size regardless of block count; detail bounded by the cap.Closes #3105