chore(followup): after acp-kernel#205 ships — bump acp-kernel pin, drop duplicated SUMMARY FIDELITY RULES section

[ranxianglei]

来源: #310 分析 summary-fidelity 内核迁移时发现(跟踪项)。

CLOSED — SUPERSEDED (2026-09-06)

Per owner direction, the kernel migration was executed directly inside PR #310 (commit 11912fe) instead of a separate follow-up PR: acp-kernel 0.0.50 → exact 0.0.55, adapter SUMMARY FIDELITY RULES section removed (single-sourced from kernel #205 via ${prompts.howToCompressRules}), 3 unit tests adapted to post-#194/#198 semantics. Verified locally on a clean tree (565/0/3, typecheck+build clean).

Remaining item before #310 can merge (tracked in the #310 thread): CI E2E is red on head — scenario scripts/e2e/scenarios/03-nudge-triggered.json fails on all three e2e jobs (ubuntu/windows/docker): nudgeDetections: 0, nudgeCount >= 1 got 0, blockCount >= 1 got 0. Root cause: same #198 min-pressure-benefit gate (max(5000, limit×1%)) — the scenario's modelContextLimit: 1500 plus ~2K tokens of cumulative growthText puts effective pending below the 5K floor, so the pressure-band nudge is suppressed by design. Fix: scale scenario growthText above the floor (fixture-only change, mirrors the e2e-compress-config unit-test adaptation).

---

Original scope (kept for reference)

• Bump acp-kernel devDependency to exact 0.0.55; refresh package-lock.json in the same commit (bump alone breaks every workflow's npm ci with EUSAGE).
• Remove adapter SUMMARY FIDELITY RULES from src/system-prompt.ts; keep src/summary-sanitize.ts permanently.
• Adapt 3 pre-fix(compress): 'nothing to compress' is terminal — one-shot continue msg, not re-nudge #194/release v0.1.46 #198 tests: tests/e2e-compress-config.test.ts:119 (release v0.1.46 #198 gate, v0.0.52/PR compress 伪成功死循环:重复压缩已压缩区间报告 success,模型上下文自复位无限重试(33 次,真实会话复现) #199), tests/sent-view-arbitration.test.ts:100 and :120 (fix(compress): 'nothing to compress' is terminal — one-shot continue msg, not re-nudge #194 first-sight bypass, v0.0.51/PR feat: salvage malformed compress content strings (omp#121) #196). ✅ Done in 11912fe (fixtures only, assertions intact).

Version span reviewed during bump (0.0.50 → 0.0.55)

• v0.0.51: fix(compress): 'nothing to compress' is terminal — one-shot continue msg, not re-nudge #194 first-sight mass bypass + drain-path regression test
• v0.0.52: release v0.1.46 #198 min-pressure-benefit gate max(5000, limit×1%) — kills nudge→compress→reject zero-yield loops
• v0.0.53: retry whole write cycle on transient ENOENT/ENOTDIR
• v0.0.54: report tool accounting — tool-results attributed to their tool
• v0.0.55: ci: per-PR npm preview builds (pr-artifact workflow) #205 write-side quote fidelity merged into tuned compress prompts

Process note (footgun hit during #310 review)

A failed npm ci (EUSAGE sync error) exits non-zero without touching node_modules, so subsequent local runs silently execute against the stale installed version — during #310 review this produced false-green results (runs believed to be on 0.0.55 actually ran on the older installed kernel). Always confirm the npm ci exit status and the installed kernel version (node_modules/acp-kernel/package.json) before trusting local test results.

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 开始处理。第一步先核实 acp-kernel#205 是否已随正式版本发布(npm latest / 完整版本列表 / PR 状态),再决定 bump 目标版本;随后做区间 review、bump pin、删重复节、全量验证、开 PR。进展会持续同步在此。

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 结论:前置条件未满足,清理动作暂挂;准备工作(重复性核验 + 区间 review)已完成,#205 发版后可立即执行。

一、核实结果:触发条件未达成

按 issue 步骤 1 逐项核实(2026-09-06):

项	Issue 假设	实际
acp-kernel#205	待发版	仍 open、未合并(API:state=open, merged=false)。npm latest=0.0.54(发布于 09-06 04:55 UTC);pr-205 仅是 prerelease tag(0.0.54-pr.205.56,09-06 13:26 UTC,晚于 0.0.54 发布)。即:尚无任何正式版本包含 #205。
PR #310	"已按 owner 指示合入"	同样仍 open、未合并(head 分支 2026-09-06_issue309-summary-fidelity)。因此 master 上既没有 SUMMARY FIDELITY RULES 节,也没有 src/summary-sanitize.ts——两者都只存在于 #310 分支上,"删除该节"目前在 master 上没有对象。
master pin	0.0.48(package.json:63)	已是 0.0.50(package.json:64,v0.1.55 之后有人 bump 过),issue 描述已过时。

二、已完成的准备(执行时零额外调研)

1. 重复前提核验通过:逐条比对 fix(compress): normalize double-escaped summaries at ingest; flag unverifiable user quotes (closes #309) #310 新增的 SUMMARY FIDELITY RULES 4 条规则与 kernel#205 对 src/compression-rules.ts 的修订(HOW_TO_COMPRESS_RULES 开头段 + KEEP VERBATIM user-intent 条 + tier-2 规则)。语义完全对应:①任务状态标为历史("TASK AS OF THIS BLOCK")、②verbatim 引用必须带 mNNNNN ref、③引用是历史记录非当前指令、④禁抄 \uXXXX/JSON 转义。ci: per-PR npm preview builds (pr-artifact workflow) #205 发版后删适配器节无覆盖缺口(内核文本经 ${prompts.howToCompressRules} 等插值进同一 system prompt)。
2. 区间 review(master 0.0.50 → 目标版本前段 0.0.54):共 4 个修复版本——fix(compress): 'nothing to compress' is terminal — one-shot continue msg, not re-nudge #194(bypass growth floor for first nudge)、release v0.1.46 #198(pressure-band nudge 最小收益门槛)、persist 写周期 ENOENT/ENOTDIR 重试加固、report 工具 token 归属(#386 follow-up)。无破坏性 API 变更:仅新增可选配置 NudgeConfig.minPressureBenefitTokens?(additive,默认内部计算,设 0 恢复旧行为)与 isToolMessage 导出。行为漂移面 = nudge 时机 + status 报告展示,风险低;billion-context-pi 全量测试对着新 pin 跑即可端到端兜底。
3. src/summary-sanitize.ts(+其测试)确认与内核互补、保留不动,符合 issue 步骤 4。

三、建议的执行顺序(请 owner 拍板操作)

1. 合并 acp-kernel#205(CI 全绿,只差人工 merge——merge 权限在我这里被系统规则禁止,只能由您操作)。
2. 等 CI 发布正式版本(预计 0.0.55),npm view acp-kernel version 确认后我立即开工。
3. billion-context-pi#310 可随时合并,顺序不限:先合它则过渡期靠适配器节兜覆盖,后合则直接单源。
4. 我随后提交单一 PR:acp-kernel exact pin 0.0.50→新版本 + npm install 刷 lockfile + 删 SUMMARY FIDELITY RULES 节 + 保留 summary-sanitize.ts + typecheck && test && build 全量验证。一次发版完成单源切换,避免 0.0.54 与 ci: per-PR npm preview builds (pr-artifact workflow) #205 两个连续小版本的双份发布开销。

另:若您希望先行把 0.0.50→0.0.54 的版本差距补掉(与 #205 无关的独立维护),说一声我就单独开 PR——区间 review 已做完,风险如上表。但默认建议仍是 #205 落地后一个 PR 收口。

Issue 保持 open,待上述 merge 完成后我继续。