Repository navigation
[PR] docs: dsh experimental-inspector fetch-accessor coexistence note (#5148) - #5151
ranxianglei wants to merge 9 commits into
Conversation
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g billion-context@0.0.10677-pr.5151.b41e200This exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr5151.tgz
npm install -g packageInstall instructions are refreshed on each push. |
(ai: vllm-qwen/qwen3.8-27b) 查重: 不重复。本条是 GitHub PR #5151 的平台镜像(来源 issue #5148,用户 free5394 报告 dsh experimental-inspector 因 bili 的 fetch 访问器无法激活)。相关条目:#5148(本 PR 的来源 issue)、#3528(更早的同机制报告——同一 初审结论(方向关 + 事实核对):docs-only,两个文件纯新增 20 行(CLIENTS.md / CLIENTS.zh-CN.md 各一段),零运行时代码、零新配置。PR 描述里的技术断言我逐条对 master 源码核过:
CI 现状(镜像缓存无数据时单次 GitHub 只读拉取):static-gates(ubuntu)/mysql-contract/gate/lane-guard/docs-sync/plan/docs-coverage 已绿;test 分片(ubuntu+windows × node22/24)仍在跑。 我在做的收尾:分支落后 master 15 个提交,但 master 这 15 个提交未触碰这两个文件(已验证零重叠);我把它 rebase 到最新 master,并补一个一行级小修——新增段落交叉引用 CONFIGURATION.md 既有的 #3528 共存提示(避免两处文档互不知晓)。随后回贴最终八问结论。 |
…w CLIENTS.md entries (#5148)
f866125 to
7462fae
Compare
(ai: vllm-qwen/qwen3.8-27b) 收尾完成,可合并。 我做过的动作(均已验证):
结构化八问:
判定:方向正确、事实全部对源码核过、diff 收敛到目的本身、CI 无红 → 建议合并(合并操作留给你,PR: https://github.com/ranxianglei/billion-context/pull/5151)。跨仓库那条(dsh inspector 降级而非抛错)维持等待你批准,批准后我再走投递流程。 |
(ai: vllm-qwen/qwen3.8-27b) 第五轮同步完成:已合入最新 master@e5e210865(零冲突;master 未触碰本 PR 的两个文档文件),新 head |
Resolve the CLIENTS.md / CLIENTS.zh-CN.md conflict caused by the docs/ reorganization (#5920): master moved both files to docs/clients.md and docs/clients.zh-CN.md leaving one-line redirect stubs at the old root paths. Keep master's stubs at the root and land the PR's dsh experimental-inspector coexistence note (#5148) in the new docs/ home instead. Adapted the note's cross-reference from CONFIGURATION.md to configuration.md to match the lowercase relative-link convention used throughout docs/. No other file changed.
[bot] 🏷 Done — merge conflict on PR #5151 resolved and pushed. The PR is now What blocked it: master's docs reorganization (#5920) moved How I resolved it: merged current master into the branch; kept master's redirect stubs at the two root paths (did not resurrect old content); re-landed the PR's two paragraphs in their new home — Content change flagged: the note cross-referenced Ops note: a local git remote alias was not usable this run, so I fetched/pushed against the canonical upstream GitHub repository URL directly instead of CI: will re-run on the new head 结构化八问
判定:可合入。 冲突已解决并推送(head= |
[bot] 🏷
No new conflict, so no further action this run. On |
[bot] 🏷
Net: no conflict + CI green → the PR is merge-ready. These recurring “dirty” forwards look like a stale re-detection on the detector side; there’s nothing left for me to fix on this branch. Full root-cause + 八问 remain in my earlier comments above. |
[bot] 🏷 Final status on PR #5151 (dsh experimental-inspector fetch-accessor coexistence note): What was done — Resolved the merge conflict. Root cause: master’s docs reorg (#5920) moved Verified live this run — head Outcome — no conflict + green CI → the PR is merge-ready. The repeated “dirty” forwards were stale re-detections; nothing remains for me to fix. Merging is human-only — ready whenever you are. 中文八问:
判定:可合入(merge-ready)。 |
[bot] 🏷 This forward was NOT stale — the conflict genuinely re-appeared, and I re-resolved it. Why it came back: six days of master activity included PR #6592 (dsh-go docs), which added a “DSH Go host (third-party Android app)” paragraph at the exact same anchor my #5148 note sits at (right after the dsh section’s What I did: merged current master ( Verified live this run:
Net: conflict cleared → PR is merge-ready, pending the standard branch-protection gates (reviews + CI). Merging stays human-only. 中文八问:
判定:冲突已清除( |
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Verdict: APPROVE (posted as a comment — GitHub does not allow the owner to formally approve their own PR, so this is a COMMENT-event review carrying the full approval justification).
docs-only, factually accurate, en/zh in sync, CI green, nothing to fix.
I verified the true delta against current master is exactly two files — docs/clients.md (+19) and docs/clients.zh-CN.md (+2); 21 insertions, 0 deletions, no other file touched. Every factual claim in the note checks out against the source:
- The #1158 self-heal / guarded accessor —
src/agent/fetch-chain/index.ts:shouldReclaimFetchPatch()(line 23) readsBILI_RECLAIM_FETCH_PATCH; the accessor descriptor is defined onglobalThis.fetchand its setter re-chains any third-party fetch install as downstream, so routing never silently leaves the proxy. The note's description is accurate. - Why the accessor stays on by default — with
BILI_RECLAIM_FETCH_PATCH=0the classic direct install applies and a third-party re-arm wins, which is exactly the #1158 silent un-routing the note warns about (the one-time warning atindex.ts:258says so verbatim). Accurate. - The cross-reference to #3528 —
docs/configuration.md(theBILI_RECLAIM_FETCH_PATCHentry) already carries a "Coexistence note (#3528)" naming this same@deepseek-ai/dsh-experimental-inspectorincompatibility. The two docs now point at each other correctly; the clients.md note adds the dsh-specific load-order framing and thecaptureFetch: falseworkaround that the config doc doesn't have, so they're complementary, not redundant. - Error string, workarounds, load order — the quoted error
inspector: globalThis.fetch is an accessor and cannot be observed safely, the three workarounds ordered least→most impact (captureFetch: false/ disable Developer Tools /BILI_RECLAIM_FETCH_PATCH=0), and "the profile loads bili before the inspector" all match the #5148 report and the reporter's verified workarounds.
en/zh parity holds (the ZH note is a faithful translation), the note sits in the dsh section immediately before ## Kimi Code, and no runtime code / config schema / wire / persistence is touched. No new config or env is introduced (BILI_RECLAIM_FETCH_PATCH pre-exists and is already documented). CI is green (30 checks, 0 failed, 0 pending). Scoping is correct: it documents the in-scope part of #5148 and leaves the root-cause fix (inspector side, cross-repo) tracked separately pending owner sign-off, so the issue is correctly not auto-closed.
中文八问:
- 是什么问题:dsh Developer Tools 启用的
@deepseek-ai/dsh-experimental-inspector与 bili dsh 原生插件装的受守卫globalThis.fetch访问器冲突,inspector 那条 entry 激活失败(dsh web其余照常)。本 PR 纯文档:在 clients.md 的 dsh 段补一条共存说明。 - 什么原因导致的:加载顺序 + 机制冲突——profile 先载 bili(装访问器,#1158 自愈)、后载 inspector(要求 fetch 是普通数据属性,读到访问器就整条抛错)。非 bili 缺陷;关访问器会重新引入 #1158 静默失路由,故默认保持开启。
- 影响范围与严重程度:仅文档,零运行时影响、零回归风险。触发条件=同时启用 dsh Developer Tools + bili 原生插件。属体验瑕疵(inspector 一条 entry 不激活,有现成绕行),不影响压缩主链路。
- 怎么修复的:文档层面——dsh 段补共存说明:成因(加载顺序+机制冲突,非 bili 之过)、为何访问器默认开(关了回归 #1158)、三条绕行按影响从小到大(① inspector 的
captureFetch: false② 关 Developer Tools ③BILI_RECLAIM_FETCH_PATCH=0)、干净修法在 inspector 侧(只降级 fetch-capture 而非整条抛错);en/zh 双语同步,交叉引用 configuration.md 既有 #3528 提示。 - 是否有回归问题:无。纯文档,不触碰任何运行时路径;高危模式(语义漂移 / fail-fast 变吞错 / 阈值·时序·输出格式变化)逐项过,均不涉及。特别确认
BILI_RECLAIM_FETCH_PATCH默认值(on)未动——没有为迁就 inspector 去关默认访问器。 - 是否破坏缓存:不涉及。
- 是否改了设计方向:不涉及。
- 是否引入了新的配置:无新配置 / 新 env。
BILI_RECLAIM_FETCH_PATCH已存在且已在 configuration.md 文档化,本条仅交叉引用。
判定:可合(APPROVE)。 合并操作留给人工。
(ai: vllm-qwen/qwen3.8-27b) review-worker-r1 完成:PR #5151 审毕 → APPROVE(docs-only,真 diff 仅 docs/clients.md +19 / docs/clients.zh-CN.md +2;事实逐条对源码核过——fetch-chain 访问器 #1158、BILI_RECLAIM_FETCH_PATCH 默认值、#3528 交叉引用均属实;en/zh 同步;CI 30 绿 0 红;无小修)。GitHub 侧已发评审(owner 无法正式 approve 自己的 PR,改以 COMMENT 评审承载完整判据 + 中文八问,review id 5946644230)。 |
Closes the in-scope (no-decision) part of #5148.
What
Documents how bili's dsh native plugin interacts with dsh's
@deepseek-ai/dsh-experimental-inspector(enabled by Developer Tools), whichfails to activate because bili installs
globalThis.fetchas a guarded accessor(the #1158 self-heal) while the inspector hard-throws on accessors.
Adds a coexistence note to the dsh section of
CLIENTS.md+CLIENTS.zh-CN.md:captureFetch: falsein the profile'scordis.patch.yml(rest of the inspector stays active)BILI_RECLAIM_FETCH_PATCH=0(loses the 方式一安装的dsh 原生插件:llm-pi-ai 传输(OpenAI SDK + options.fetch 注入)绕过 globalThis.fetch 拦截,模型请求不进代理、压缩完全不生效 #1158 self-heal)Scope / behavior change
Docs only. No runtime code touched. No new config/env (
BILI_RECLAIM_FETCH_PATCHalready exists and is documented). Zero regression risk.Out of scope (tracked separately under #5148, pending owner approval)
The root-cause fix lives in the external dsh project — making the
experimental-inspector degrade only its fetch-capture instead of throwing the
whole entry when it sees an accessor. Cross-repo submission requires owner
sign-off before anything goes upstream.