[ranxianglei]

What this fixes

A model that serializes its compress args as text escapes the header/summary separator twice, so the documented line-form content string carries a literal backslash+n (two characters) instead of a real newline. Undecoded, the kernel finds no newline → empty summary → rejects with kind=content-not-array, which hides the real reason (missing summary after the refs header line) and sends the model to retry the same broken shape. This is the next rung of the compress-arg shape-drift ladder (#1494 / #1559).

Repro from the issue: {"content": "m00003–m00107 example topic\\nTASK AS OF THIS BLOCK: ..."} fails, while the identical payload with a real newline succeeds.

Root cause

parseCompressInput → kernel parseObjectValue: a content string with no real newline goes through splitLineEntries (single chunk = whole string) → parseLineEntry (nl === -1) → summary === "" → rejected. The bare branch then sets diag.kind = "content-not-array" and returns without surfacing the per-entry reasons it did collect.

Fix (host-side, no kernel change)

In src/compress-tool.ts parseCompressInput, when the first parse is not ok, decode \n / \r\n once from a double-escaped content string and retry, adopting the result only if it yields strictly more ranges. Same pre-normalization discipline as #1494.

Safety invariant: a string with no real newline always parses to zero ranges undecoded, so adopting only on strictly-more-ranges can never degrade a working parse. A payload that already carries a real newline is left untouched (a literal backslash-n inside its summary body survives verbatim).

Boundary note: the shared home for any future multi-lane fix is the kernel parser (kernel/src/parse-compress-input.ts parseLineEntry). Kept host-side here on purpose — no kernel release cycle is needed for a proxy-lane arg-shape salvage (#1494 precedent), and every execution lane already converges on this single entry point.

Behavior change (disclosed)

• Before: a double-escaped line-form string was rejected with [Compression FAILED: ... kind=content-not-array ...]; the model retried the same shape.
• After: it is decoded once and recovered as a valid range, then folded normally (when it decodes into refs + a non-empty summary); it is still rejected unchanged when it does not decode into that form (e.g. no refs header).

No wire-shape change; both compression modes are unaffected (arg parsing precedes mode-specific handling); no message-ref/id reissue.

Config surface

None. No new env var, config key, or CLI flag.

Tests

tests/issue1494-compress-params.test.ts (+4):

• positive — double-escaped string recovered at parse level (startRef/endRef/topic/summary correct)
• positive — folds end-to-end through applyRanges (no FAILED, no content-not-array)
• negative — a payload with a real newline is NOT re-decoded (literal backslash-n in the body preserved)
• negative — a ref-less escaped string is NOT salvaged into a fake range

Targeted suite green locally (13/13 in the file); the full-suite verdict is remote CI.

Closes #3780

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-19_compress-double-escape-salvage (c824436)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.4499-pr.3785.4a7aade

This exact version is immutable and reproducible. The rolling pr tag tracks the newest test build across all PRs.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr3785.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

查重: 不重复——按失败签名（kind=content-not-array / double-escaped）检索 open+closed issue 与 PR：仅本 PR #3785 与其原始 issue #3780 在途；同家族 #1494/#1495/#1497/#1559 均已关闭，无并行重复修复。

标签: bug + dsh + scope/kernel + p2（缺陷修复；复现环境为 dsh desktop 插件 lane，机制本身是模型参数序列化漂移、非客户端专属；属压缩管线参数解析；能用但特定条件下有明显问题→p2）。

审核进行中（agent 提交的 PR，按规则不重做、只核对代码与 CI）：

• 当前 head (123301b) 的 CI 全绿（test ubuntu/windows × node 22/24 各 shard + static gates 全部 success）；
• 安全不变量已对 master 上的内核解析器源码逐条核验：kernel/src/parse-compress-input.ts 自 PR base 以来零改动，且「无真实换行的 content 字符串在未解码时必解析为 0 ranges」在 master 代码上成立（line-form 回退路径 parseLineEntry nl===-1 → summary 为空 → 拒绝；JSON-array 路径若成功则 ok、根本不触发 salvage），strictly-more-ranges 采纳条件不可能劣化任何可工作解析；
• master 自 base 前进 99 个文件，与本 PR 触碰的 2 个文件零重叠；
• 正在 rebase 到最新 master 并跑本地定向测试，rebased head 的 CI 绿后按结构化八问给最终审核结论。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

PR #3785 审核结论（独立重审）

已执行：rebase 到最新 master（123301b → c82443600，零冲突——master 自 base 前进 99 文件，与本 PR 的 2 个文件零重叠；kernel/src/parse-compress-input.ts 自 base 以来零改动）→ 本地 build + typecheck + 定向测试 13/13 全过 → rebased head 的远程 CI 全绿（18 success / 12 skipped 为门控 e2e lane / 0 failure）。以下结论基于我自己的 diff 与源码核验，未沿用 issue 侧分析。

八问：

1. 是什么问题：模型把 compress 参数当文本序列化时，行形式的头/摘要分隔符被转义两次，content 里是字面 backslash+n（两个字符）而非真换行；内核找不到换行 → summary 为空 → 拒绝，且报出的 kind=content-not-array 掩盖了真实原因（refs 头行后缺 summary），模型因此重试同一个坏形状。
2. 什么原因导致的：内核 parseLineEntry 以真实 \n 为头/摘要边界，无真实换行的单 chunk 必得空 summary 被拒；bare 分支只置 kind=content-not-array，不暴露已收集的 per-entry 原因。属 compress 参数形状漂移梯度的下一阶（compress failed may be caused by params #1494/fix: surface parse-dropped compress entries; normalize single-object content (#1494) #1497/[kernel] parse ladder: string in the content slot ({"content": "m1–m5 …"}) is rejected as content-not-array — next salvage candidate for weak-model drift #1559 同族）。
3. 影响范围与严重程度：任何「模型输出双转义行形式」的场景都会触发（复现环境 dsh desktop lane + DeepSeek；机制是模型序列化行为，不限客户端）。p2：该次压缩调用失败 + 浪费一轮重试，会话本身可用、可自愈。
4. 怎么修复的：host 侧预归一化（compress failed may be caused by params #1494 同一纪律）——首次解析失败且 content 是「无真实换行但含字面 \n」的字符串时，解码一次后重试，仅当 range 数严格更多才采纳。我逐条对 master 内核源码核验了安全不变量：无真实换行的字符串未解码时必解析为 0 ranges（line-form 回退路径 nl===-1 → 空 summary → 拒；JSON-array 路径若成功则 ok、根本不进 salvage），所以采纳条件不可能劣化任何可工作解析；含真实换行的 payload 提前 return 不碰，正文里的字面 backslash-n 逐字保留。多 lane 共享归宿（内核 parseLineEntry）以注释 boundary note 留档，本次刻意不动内核——省一个 kernel release 周期，且所有执行 lane 都收敛在这个入口。
5. 是否有回归问题：无。触碰路径既有行为逐项清点：①所有成功解析不受影响（salvage 由 !ok 门控）；②失败解析中仅该形状改变结局，其余拒因/日志不变；③日志仅新增一条 rescue warn，其余行原样；④wire 形状不变（参数解析先于模式处理，plugin/proxy 两模式共用此入口）；⑤无阈值/时序/输出格式漂移，无 fail-fast→吞错（解不出来的 payload 仍按原样拒绝）。rebase 后在 master 上重新验证成立（见首段）。
6. 是否破坏缓存：不涉及——未触碰缓存/重试/增量游标/去重键；无 message-ref/id 再发号。
7. 是否改了设计方向：没有——延续 compress failed may be caused by params #1494 的 host 侧预归一化路线，边界注释明确未来多 lane 修复的内核归宿（届时走 human-gated kernel 变更）。
8. 是否引入了新的配置：无（与 PR 声明一致）。

测试有效性（④）：4 个新测试有牙——正例钉死 startRef/endRef/topic/summary 映射 + applyRanges 端到端 receipt（[Compressed m00001–m00002 → 1 block(s)、无 FAILED/content-not-array、active blocks==1）；两个负例分别钉死「含真实换行不再解码」与「无 refs 不救成假 range」，正负成对符合 gate 纪律（#3617）。本 bug 回归时 test 10/11 会抓住。

已知残留（不阻塞）：同一字符串内真实换行与字面 \n 混合转义的 payload 不救（设计如此，避免动到有效载荷）；OP 建议 2（model-facing FAILED 消息暴露 invalidReasons / 独立 kind）属内核诊断面，留待另行决策——rescue 落地后该误导 kind 场景已大幅收窄。

§7.4 auto-merge 判定：全部满足（单模块、回归测试复现原 bug、rebased head 绿、无 config/persistence/wire/kernel 触碰、纯 fix:、diff 干净、Fixes #3780、不涉 load-bearing infra）→ 可合入，等待人工合并。

[github-actions]

📦 Merged Build Published

This merge to master was published as 0.0.622-master.f74f266 (rolling master tag).

npm install -g billion-context@0.0.622-master.f74f266  # this exact build
npm install -g billion-context@master  # always the newest merged build

npm view billion-context@master prints the current version at any time.