Skip to content

[PR] feat: declarable custom-wire model endpoints + commandcode-cli wire family (#1295) - #1444

Closed
ranxianglei wants to merge 7 commits into
masterfrom
2026-09-26_model-endpoint-patterns
Closed

ranxianglei wants to merge 7 commits into
masterfrom
2026-09-26_model-endpoint-patterns

Conversation

@ranxianglei

@ranxianglei ranxianglei commented Sep 26, 2026 •

Copy link
Copy Markdown
Owner

Part of #1295

Follow-up families tracked under #1295 (same symptom: declared custom-wire endpoints not yet compressed): TRAE SOLO SSE, Qoder sash SSE, and #2975 (workbuddy reverse proxy) — pending captures + owner go-ahead.

What

Two commits implementing the owner-approved plan:

① f4862e8 — one config source for both claim gates

  • New global config modelEndpointPatterns: [{ match: <absolute URL prefix>, wire: "openai"|"anthropic"|"responses"|"google"|"commandcode" }], longest-prefix wins.
  • Proxy side (src/server.ts classifier): a declared pattern now decides protocol classification for its URL — including known families on non-standard paths (previously suffix-table only).
  • Client side (src/agent/native-intercept.ts): isModelApiUrl consults the same patterns (module-level registry, loaded best-effort from the global config file by every native entry: pi/omp/opencode V1+V2/dsh at apply time), so the fetch patch claims declared endpoints too. The two hardcoded tables can no longer drift for declared paths.
  • Body-based disambiguation stays ([Bug] DSH native fetch 拦截误伤非 LLM 端点:任何以 /messages 结尾的 URL 都被当成模型 API #1284 lesson): a declared endpoint whose body is not a convertible conversation is still relayed byte-for-byte with the existing once-per-endpoint warning.

② 1cd7e6a — first new wire family: commandcode

  • Motivating case: @mars-sea/dsh-commandcode-provider Go plan — POST https://api.commandcode.ai/alpha/generate, nested CLI envelope request ({config, threadId, params:{model, stream, messages, tools, system, …}}, CC block content), bare JSONL event-stream response ({"type":"text-delta",…}, {"type":"finish","finishReason","totalUsage"}) — not SSE.
  • src/commandcode-wire.ts: strict envelope unwrap (→ flat OpenAI-family body for the existing pipeline) and rewrap (final wire boundary stays the envelope; pipeline modifications ride inside params). Unsupported shapes → verbatim relay, never silent loss. Fidelity rules: params.system ↔ leading system message round-trip, tool results keep toolName (+error-text type via errorResultIds), unknown top-level keys preserved.
  • src/loop/adapter-commandcode.ts: OpenAI-family adapter spread with a JSONL codec (no SSE framing, no [DONE]); usage from finish.totalUsage; tag-echo filtering applies to prose deltas only; proxy tool calls settled into structured events.
  • Plugin mode: pipePluginJsonlWithStrip (usage accounting + degenerate-turn warning, no synthetic tails).
  • Error frames: single-line JSONL {"type":"error","error":…,"message":…} matching what the client throws on.
  • WIRE-CONTRACTS.md (+zh): per-endpoint upstream-constraint ledger WC-1…WC-6 (each constraint discovered → permanent entry, cited inline in code).
  • Docs: modelEndpointPatterns section added to CONFIGURATION.md + zh mirror; cross-references in Providers sections.

Behavior changes (old → new)

  1. New capability: custom-path model endpoints can now be compressed when declared (before: relayed byte-for-byte, zero compression — the dsh使用插件dsh-commandcode-provider反代go套餐的话,不进billion-context压缩 #1290 gap).
  2. Gate unification: client claim + proxy classification now share modelEndpointPatterns as an additional source (built-in suffix tables unchanged; declarations are additive and checked first).
  3. dsh unrouted log message: the "not a recognized model endpoint" line now also mentions modelEndpointPatterns as the fix path (log text only).
  4. Client-side declarations apply at agent entry start (pi/omp/opencode/dsh load the config file when the plugin applies; changing the list requires restarting that agent's session host). Proxy-side picks up restart as before.
  5. commandcode v1 scope: streaming envelopes only — a declared commandcode endpoint receiving a non-streaming envelope is demoted to verbatim relay (WC-6), not converted.
  6. Unknown JSONL lines: undecodable lines pass through byte-exact in round 1 of a compress loop and are dropped in later rounds (WC-3); unknown typed events pass through unchanged.

Pre-flight

  • npm run typecheck clean
  • npm test: identical failure set to clean master baseline (single pre-existing #393: --upstream mode … claude-opus-5(-thinking) to 1M failure, present on master before this branch)
  • npm run build OK
  • New tests: tests/commandcode-wire.test.ts (codec round-trips/fidelity), tests/adapter-commandcode.test.ts (JSONL parse/emit/usage/errors), tests/commandcode-endpoint.test.ts (server-level: declared endpoint unwrapped upstream & rewrapped back down, non-conversation relay, WC-6 demotion, declaration-driven classification)

No version bump (release flow owns that). No acp-kernel change.


Fixes #1295

(孤儿认领:标题本已指向 #1295,补上关闭关键字 —— 第三批清理)

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review started. Early finding: the PR base (897c9d77) is 224 commits behind master (825f1529, v0.1.158) and GitHub marks it unmergeable (dirty) — src/server.ts was rewritten by #1195 (overflow-refold pipeline) plus several other PRs landed on the hot files meanwhile. I'm rebasing the two commits onto current master, resolving the 10 conflict hunks (mostly import-list unions + one semantic integration in src/server.ts where the #1295 runPrepare wrapper must nest inside #1195's runPreparedPipeline closure), then re-running typecheck/tests/build and checking CI before posting the full verdict.

@ranxianglei
ranxianglei force-pushed the 2026-09-26_model-endpoint-patterns branch from 1cd7e6a to 56406c1 Compare September 26, 2026 17:40
@github-actions

github-actions Bot commented Sep 26, 2026 •

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-09-26_model-endpoint-patterns (1370426)

Latest build — install this exact version

npm install -g billion-context@0.0.9312-pr.1444.46f3cc9

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

Recent builds (commit → version)

commit version build
1370426 billion-context@0.0.9312-pr.1444.46f3cc9 build
d9963db billion-context@0.0.8862-pr.1444.6e7c15b build
04a7847 billion-context@0.0.8844-pr.1444.eabab04 build
ee19f22 billion-context@0.0.8838-pr.1444.50da3b5 build
793a488 billion-context@0.0.8821-pr.1444.ac78385 build
f03e892 billion-context@0.0.8820-pr.1444.d3fd0b0 build
a66487c billion-context@0.0.8818-pr.1444.f272c10 build
0f502f4 billion-context@0.0.8814-pr.1444.965e470 build
b7c9a02 billion-context@0.0.5527-pr.1444.64bb29f build

Download artifact

  1. Download the artifact from the Actions run
  2. Extract the tarball and install:
tar xzf billion-context-pr1444.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review complete — PR rebased onto current master, all gates green, mergeable_state: clean. Ready for human merge.

What I did

The PR base was 224 commits behind master and GitHub-marked unmergeable (dirty) — src/server.ts had been rewritten by #1195 (overflow-refold pipeline) plus other hot-file churn. I rebased the two feature commits onto origin/master (825f1529) and force-pushed 2026-09-26_model-endpoint-patterns. New head: 56406c11 (3 commits: the two original features + one doc fix I made).

Conflict resolutions (10 hunks): 8 were import-list/adjacent-line unions (master added noteRoutedOrigin/observeRoutedOrigin/waitForProxyVersion/envMillis/isLegacyBcpEntry/ContextOverflowInfo/resolveConversation etc. in the same regions the PR extended). One was semantic: the #1295 runPrepare stamping wrapper had to nest inside #1195's new runPreparedPipeline(respondFailFast, overflowWindow?) reusable closure — resolved by renaming master's inner fn to runPrepareInner so the wrapper (which stamps prepared.commandcode = { meta }) sits between it and every call site. Verified wiring at src/server.ts:2228 (inner), :2265 (wrapper), :2394 (initial), :2409 (overflow re-entry), and the preflightCompressIfNeeded(prepared, runPrepare, …) call site — initial prepare, preflight re-prepares, and #1195 overflow refold re-entry all flow through the stamped fn, which is exactly what the feature needs.

Doc fix (commit 56406c11): WC-4 in WIRE-CONTRACTS.md + zh had unwrap/rewrap directions swapped vs the code (code: unwrap does input obj → JSON.stringify → arguments, rejecting non-object input; rewrap does arguments string → parse → input, malformed → {} — src/commandcode-wire.ts:87,242). Fixed both files to match code.

① Direction vs original issue (#1295)

Implements exactly the owner-approved plan: modelEndpointPatterns as ONE config source consulted by both claim gates (proxy classifier + client isModelApiUrl), longest-prefix-wins, body-based disambiguation kept as the guard (#1284 lesson), plus the first new wire family (commandcode) with strict envelope codec. No scope drift. One design point worth flagging: src/agent/dsh-native.ts:576 adds a declaration-based takeover-gate bypass (matchModelEndpoint(getDeclaredModelEndpoints(), url) → claim). This goes slightly beyond the plan's literal text ("attribution would hit the dsh session naturally"), but is necessary because provider-plugin LLM calls can execute outside dsh's ALS initiator scope (the #1290 zero-traffic case); the proxy-side body check remains the real guard, so a declared-but-non-conversation endpoint still relays verbatim with the once-per-endpoint warning. Acceptable.

② Regression inventory (existing behaviors on touched paths)

  • Proxy classification (classifyWireFamily): declared patterns checked after route.explicitProtocol, before built-in suffix tables. Purely additive — an undeclared URL takes byte-identical behavior to before; explicit route protocol still outranks declarations. ✅ no regression
  • Client-side isModelApiUrl: declarations checked before suffix tables but AFTER control-plane exclusions — /bili/ and /__bili/ are never claimed even when declared (tested). Built-in table behavior unchanged. ✅
  • 使用pi 压缩后就提示prompt is too long #1195 overflow-refold semantics: untouched except the fn rename above; respondFailFast/overflowWindow contract intact; suite passes. ✅
  • Non-commandcode endpoints: zero path changes beyond the additive classifier step; SSE/stream/error paths unchanged. Error emitters' new "commandcode" branch is gated on prepared.commandcode. ✅
  • Fake-completion retry ([bug] 假完成:模型把工具调用写成文字(无 tool_use 块)→ 轮次提前终止,agent 显示正常完成(#361 根因 follow-up) #371): disabled only for commandcode envelopes (no fake-completion shape exists in JSONL; finish IS the terminator). ✅
  • dsh log line now mentions modelEndpointPatterns — log text only, disclosed by author. ✅
  • _resetForTest gained opts.anchor on master; merged with the new declaredPatterns = [] reset; test suite green. ✅
  • version field untouched, acp-kernel pin untouched, no lockfile churn. ✅
  • Author-disclosed behavior changes (all 6 old→new items verified against code): declaration-driven compression for custom paths; additive gate unification; dsh log text; client declarations apply at agent entry start (session-host restart needed); commandcode v1 streaming-only (non-streaming demoted to verbatim relay, WC-6); unknown JSONL lines byte-exact in round 1 / dropped in later rounds (WC-3). All match the implementation.

③ Code

Strict unwrap/rewrap fidelity holds up under scrutiny: unknown top-level keys and params.* passthrough both directions, params.system ↔ leading system message round-trip, tool-result toolName/error-type preservation via id maps, documented normalizations are honest (multi-part user text join, assistant content always block array). Adapter reuses the OpenAI adapter spread and only overrides the transport codec; real tool calls replay verbatim from the upstream's original event line (byte-exact tool-args invariant #1039 preserved); proxy tools settled as structured events; tag-echo filtering prose deltas only (reasoning-delta unfiltered — consistent with adapter-openai.ts). firstRoundOnly/passthrough are established core contracts, used correctly. No as any/@ts-ignore; no console.error reachable from the proxy process (the loader warning lives on the client-side entry path only).

④ Tests

Real teeth, not ceremony: commandcode-wire.test.ts (round-trip deepEquals + rejection matrix), adapter-commandcode.test.ts (event sequences, verbatim replay, round gating, exact JSONL emit shapes), commandcode-endpoint.test.ts drives a real HTTP proxy against a fake upstream asserting wire bytes both directions (upstream sees flat OpenAI body, client gets rewrapped envelope; non-conversation relayed byte-exact; WC-6 demotion; declaration-driven classification with usage accounting), model-endpoints.test.ts (fail-loud parsing, segment-aligned matching, never-throwing client loader). These would catch regressions of exactly the reported bug class.

⑤ Diff & gates

After rebase: exactly 22 files, +1898/−34 — every line serves the PR's purpose; no whitespace churn, no generated artifacts, no lockfile drift.
Gates on pushed head 56406c11: local typecheck clean · npm test 2876 pass / 0 fail / 12 skip · build OK · GitHub CI fully green (CI, E2E real-codex, e2e-fake-codex, e2e-local-registry, e2e-native opencode/pi/dsh, PR Build Artifact, Windows codex overlay). Local E2E couldn't run here (no local sglang upstream), but CI's E2E-real-codex job covers it.


中文摘要:PR #1444(可声明自定义 model 端点 + commandcode wire 族)方向与 owner 批准的 #1295 方案一致、代码与测试质量过关;发现并直接修复了两处问题——基线落后 master 224 个提交导致的 10 处冲突(含一处 #1195 管线重构的语义性整合,已验证初始/preflight/overflow 三条路径全部经过 commandcode 打标点)和 WIRE-CONTRACTS 文档 WC-4 方向写反——已 rebase 到最新 master 并强推,本地+GitHub CI 全绿,可以合并。

@ranxianglei

Copy link
Copy Markdown
Owner Author

Deep review (on top of 56406c1 → f5461db)

Reviewed the full diff (two feat commits + docs) against #1295. Verdict: approve after the hardening commit below — architecture is right (single declaration source for both claim gates, strict-fidelity reject→verbatim-relay default, WC ledger, no pipeline changes for standard families).

Fixed on the branch (f5461db)

# finding fix
F1 normalizeMatch silently stripped userinfo → a typo'd https://u:p@host/… declaration mints a credentials-less origin claim that never matches reject loudly at startup (same treatment as search/hash)
F3 rewrap's user branch coerced non-string flat content to "" — latent WC-2 violation if any path ever emits array-content user messages mirror unwrap's text-part join
F4 route.explicitProtocol silently overrode a declaration one-time warn naming both the declaration and the pin
F6 GET/HEAD/empty-body probes against a declared endpoint logged "unrecognized path" noise declared endpoints are excluded from the unrecognized-path warn (carve-out scoped via the same longest-prefix matcher); undeclared paths still warn — both asserted
F7 dead meta.reasoningEffort field dropped (reasoning_effort round-trips via extraParams)
F8 bare console.error in a src-root module documented rationale (host-process loader, no logger sink)
F9 three coverage gaps new tests/commandcode-plugin-pipe.test.ts: plugin pipe (tag-echo strip, byte-verbatim non-prose, terminal fragment policy, in-band truncation on read failure), commandcode branches of all three stream-error emitters, duplicate-tool-name verbatim-replay pin; F6 assertions in the endpoint e2e
F2/F5 doc gaps WC-2 empty-message note (WIRE-CONTRACTS.md); two-gate URL semantics ("declare the URL the request actually lands on") + userinfo rejection in CONFIGURATION.md (+zh)

Non-blocking observations

  • The two gates match different URLs on path-rewriting providers (client claim = called URL, classifier = rewritten URL). Documented; a future follow-up could align them if it ever bites in practice.
  • WC entries are docs-only (enforcement lives in the codec test files, not tests/wire-contract.test.ts gates) — acceptable for codec contracts; upstream-rejection constraints stay in the gated ledger.

Gates

typecheck ✓ · targeted suites 107/107 (model-endpoints, commandcode wire/endpoint/adapter/plugin-pipe, plugin-agent, proxy-stream-error) · CI running remotely on f5461db.

@ranxianglei
ranxianglei force-pushed the 2026-09-26_model-endpoint-patterns branch from f5461db to d44e333 Compare September 27, 2026 02:08
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Hardening commit f5461db verified line-by-line — all nine findings land as described. One claim/code discrepancy found and fixed directly on the branch (d44e3330), plus a fresh rebase onto current master. All gates green, mergeable_state: clean.

Verification of f5461db (independent re-check)

  • F1 ✓ normalizeMatch now rejects userinfo (u.username || u.password → undefined → startup throw, same treatment as search/hash). A typo'd https://u:p@host/… can no longer mint a credentials-less origin claim.
  • F3 ✓ rewrap's user branch now mirrors unwrap's text-part join (parts.join("\n")) instead of coercing array content to "".
  • F4 ⚠️ warn present but not once-gated — see below.
  • F6 ✓ GET/HEAD/empty-body probes against a declared endpoint skip logUnrecognizedPath; carve-out derives the target URL exactly like buildForwardTarget and scopes via the same longest-prefix matcher; undeclared paths still warn — both directions asserted via getUnrecognizedPathStats().
  • F7 ✓ dead meta.reasoningEffort dropped from both interfaces and the unwrap site (reasoning_effort round-trips through extraParams).
  • F8 ✓ bare console.error documented as host-process loader path with no logger sink.
  • F9 ✓ tests/commandcode-plugin-pipe.test.ts (145 lines) adds the plugin pipe, all three stream-error commandcode branches, and the duplicate-tool-name verbatim-replay pin; F6 assertions added to the endpoint e2e. New tests have real assertions (byte comparisons, stat deltas), not ceremony.
  • F2/F5 ✓ WC-2 empty-message note + two-gate URL semantics ("declare the URL the request actually lands on") + userinfo rejection, both languages.

Fixed on the branch — d44e3330

F4 said "one-time warn" but the code warned on every request. The classifier hunk called loggerLog("warn", …) unconditionally inside handle() — a misconfigured client retrying in a tight loop would spam the proxy log per request, contradicting both the stated intent and the codebase convention (nonConversationRelayWarned, bounded FIFO sets). Fix: module-level declPinConflictWarned set keyed by (declaration, pin) with the same 4096-cap FIFO eviction, plus an exported getDeclPinConflictStats() following the getUnrecognizedPathStats pattern. While there I also corrected the message wording: explicitProtocol is a URL-embedded pin (/bili/<family>/https://…, parsed in resolveUpstream at src/server.ts:220-231), not a "provider config" entry — the old text pointed users at a config location that doesn't exist. Regression test in tests/commandcode-endpoint.test.ts: two conflicting pinned requests against the declared endpoint assert the counter advances by exactly 1.

Rebase note

Master moved while this was in flight (#1285, #1419 landed — PR state was behind). I checked the new master diff for semantic interaction with this PR's hunks: it touches compress-settings/registry price-profile regions only, disjoint from the classifier/wire/pipeline paths here. Rebased cleanly (no conflicts), so the branch is now on current master b371bd04.

Gates on final head d44e333

  • Local: typecheck clean · npm test 2913 pass / 0 fail / 12 skip · build OK
  • GitHub CI: all runs success — CI (ubuntu 22+24, windows 22+24), E2E real-codex, e2e-fake-codex, e2e-local-registry, e2e-native opencode/pi/dsh, Windows codex overlay, PR Build Artifact, version-guard (13/13 check-runs)
  • PR: mergeable: true, mergeable_state: clean

Your non-blocking observations (two-gate URL semantics on path-rewriting providers; WC entries docs-only with enforcement in codec tests) are correctly scoped — no action taken.


中文摘要:逐行核对了加固提交 f5461db,F1–F9 全部属实且实现正确;唯一出入是 F4 声称「一次性告警」但代码实际每请求都告警,已直接在分支上补了按(声明,pin)去重的有界 FIFO 门控并修正了误导性文案(该 pin 来自请求 URL 而非 provider 配置),附回归测试;同时把分支 rebase 到了最新 master(确认与新增的 #1285/#1419 无语义交互),本地+GitHub CI 全绿,可以合并。

ranxianglei pushed a commit that referenced this pull request Sep 30, 2026
…ap before chain stamp; port commandcode error shape into nativeErrorChunk/legacyCompletionChunk
ranxianglei pushed a commit that referenced this pull request Sep 30, 2026
…ng args (anthropicCacheMarks, streamErrorShape) around the commandcode ternary
ranxianglei pushed a commit that referenced this pull request Oct 18, 2026
ranxianglei pushed a commit that referenced this pull request Oct 18, 2026
…export internal ModelEndpointWire

Post-merge gate fixes on the merged tree:
- config-surface-guard (#2643, landed after this branch cut): the new
  top-level field needed a seed entry in website/config-reference/*.yaml;
  added to server.yaml (upstream group) and regenerated the marked doc
  blocks via tools/gen-config-docs.mjs (hand-written section prose kept).
- unused-exports: ModelEndpointWire is only used inside
  src/model-endpoints.ts — drop the leading export (INTERNAL_ONLY).
Local: both gates green, tsc clean.
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

查重: N/A —— 系统转发的「上游 PR 冲突」唤醒(本 PR 分支 vs master),不是新 issue 接单。
标签: scope/wiring opencode pi dsh p3(功能 PR,横跨三个原生客户端接线层 + 新 wire family;非缺陷不打 bug;功能类取 p3 档)

冲突已解决——结构化八问

1. 是什么问题:分支自 9-26 停更,期间 master 走了 2692 个提交(最大动作是 #1440 把请求管线从 src/server.ts 拆进 src/server/*.ts 模块群),两侧改写了同一片文件区域,GitHub 判定不可合并(dirty),PR 卡住。

2. 什么原因导致的:不是代码缺陷,是长命分支与快进主干的漂移。master 侧把分类器/forward() 搬进了 handle.ts/relay.ts、把各原生 lane 的 bootstrap 重构为 armNative* 函数、新增 #1909 providers 表声明与 #2924 dsh 门控机制、给 TagEchoFilterStats 加了必填 dropCount 字段、上了两个新门禁(config-surface-guard / unused-exports);我们分支还在旧位置编辑同一逻辑 → 9 个文件 20 个 hunk 文本冲突。

3. 影响范围与严重程度:只影响本 PR 自身的可合并性,冲突解决本身不改变任何用户可见行为;卡住期间阻塞 #1295 的功能落地。现已解除。

4. 怎么修复的:把 master 合入分支(merge 而非 rebase,保留 PR 历史与既有评审记录),按「意图并集」解 9 文件冲突:PR 功能的完整增量(commandcode wire family + modelEndpointPatterns 声明机制,共 20 个 hunk)逐段重新落位到新家园(src/server/handle.ts 的分类器/unwrap/prepared.commandcode/preflight 错误面,src/server/relay.ts 的 rewrap/wireTransform/流错误/plugin JSONL 直管/适配器选择);agent 各 lane 采用 master 的 armNative* 结构 + PR 的声明加载调用点;dsh 的声明判定并入 master #2924 takeoverGate 顶部。随后一个跟进提交 cd55447 补上两个新门禁的集成债(config-reference 种子条目 + ModelEndpointWire 降为内部类型)。最后又追平两批零文件交集的 master 小差量(#3398/#3412 billing 修复、#3721 Windows teardown 重试、kernel EOF 换行),当前 head 8549b43 已完全同步 master 尖端 c398859。

5. 是否有回归问题:对 diff 触碰路径的既有行为逐项清点——全部有意变更且已在下方清单披露,无无意回归。验证证据:本地 typecheck/build/定向套件全绿(214 pass / 0 fail,覆盖 commandcode 四套件、model-endpoints、native-intercept、proxy-stream-error、dsh/omp/pi/opencode native);前一 head cd55447 的远程 CI 除下述 2 个既有红道外全部绿(含所有 e2e lane 与其余测试分片);最终 head 8549b43 push 后静态门禁全绿、测试矩阵运行中。

6. 是否破坏缓存:不涉及。持久化格式、会话键、压缩缓存策略均未动(零 kernel 改动)。

7. 是否改了设计方向:没有。合并完全跟随 master 的拆分管线架构,PR 功能是重新落位而非重设计。唯一一处集成级决策:分类器同时遇到「#1909 providers 表按路由声明」与「#1295 全局 modelEndpointPatterns 声明」时的优先级定为 URL pin > #1909 > #1295 > 内置启发式(路由特化压全局),已在 handle.ts 决策点注释固化理由。

8. 是否引入了新的配置:无新增。modelEndpointPatterns 本身是 PR 内容(CONFIGURATION.md 中英散文节由 PR 自带);本次仅按新门禁要求补了 config-reference 索引条目(生成器标记块内,散文未动)。

行为变更清单(旧→新,均为有意变更)

  1. 分类器优先级(新集成决策):URL pin > [feature] configurable wire-path table (per-provider custom paths) - Claude Desktop's claude.ai private protocol as motivating case #1909 按路由声明 > [Feature] Declarable custom-wire model endpoints — route providers like commandcode's Go plan (/alpha/generate) into compression #1295 全局 pattern > 内置启发式(此前分支内只有后两级,master 侧只有 [feature] configurable wire-path table (per-provider custom paths) - Claude Desktop's claude.ai private protocol as motivating case #1909,二者合并后的相对次序是本次新定的)。
  2. ProxyOptions.modelEndpoints:必填 → 可选(仅类型放宽,运行时无变化,缺省按空数组处理;因 master 大量新测试字面量不含该字段而被迫放宽,两个消费点均走容忍 undefined 的 matchModelEndpoint)。
  3. omp/pi lane 声明加载位置:移入 NODE_TEST_CONTEXT 守卫块内、紧贴 installNativeFetchIntercept(opencode/dsh 保持 PR 原位置);生效时机不变(宿主启动时读一次配置)。
  4. dsh-native:声明检查置于 [dsh-native] takeover-gate refusal (threw state) fails open with the FULL uncompressed transcript; a >window local history dies upstream with 400 (maximum context length) #2924 归因机制之前直接返回 true(显式声明优先于启发式归因);刻意不重置 threwStreak/noteAttrRecovery,保持 压缩接管失效后完全静默:面板与 /__bili/status 看不出压缩已停,用户全价直连直到重启才恢复 #3581 健康上报诚实。
  5. PR 自身新文件的潜在类型错误随集成修掉:TagEchoFilterStats.dropCount 两处(master 新增必填字段)、tests/commandcode-endpoint.test.ts 8 处 address() 平台类型断言、parseStream 参数个数、emitReasoning 非空断言、wire 测试的类型收窄。
  6. ModelEndpointWire:导出 → 内部类型(unused-exports 门禁要求;无任何外部引用)。

两条红道说明(均为既有问题,与本 PR 无关)

收工判定

冲突清零:分支已与 master 尖端同步(head 8549b43,mergeable=True 无冲突),静态门禁全绿,测试矩阵运行中(若 Windows 分片再次命中已知 flake 见 #3725)。CI 转绿后此 PR 即可合并——合并操作在你这边(系统规则不允许我代合)。

@ranxianglei ranxianglei changed the title feat: declarable custom-wire model endpoints + commandcode-cli wire family (#1295) [PR] feat: declarable custom-wire model endpoints + commandcode-cli wire family (#1295) Oct 18, 2026
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

冲突已解决并推送:b7c9a0230(merge origin/master into 2026-09-26_model-endpoint-patterns)。GitHub 状态已从 dirty 转为可合并(当前显示 behind——推送期间 master 又前进了几笔,非冲突)。

怎么解的:8 个文件、10 处 hunk,全部按「双方意图取并集」处理,无设计决策分歧:

验证:tsc --noEmit 干净;定向测试 170/170 通过(native-intercept、model-endpoints、commandcode-wire/endpoint/adapter-commandcode、dsh-native)。全量回归交给 CI 的 test lanes。

八问:

  1. 是什么问题:PR 分支落后 master 约 6 天,双方同时改了同一批文件的 import 区与 dsh gate,产生文本冲突(merge state dirty)。
  2. 什么原因导致的:PR 长期在途,master 在此期间落地了 dsh 归属重构(refuseUnattributed 等)、Switching model mid-conversation restarts the ref ledger, but the rendered transcript still shows the old refs — the model keeps aiming compress at dead ranges until the circuit breaker fires | 切模型后消息编号整体换账本,但渲染历史仍显示旧编号,模型对着死区间反复压缩直至熔断 #4305/[Feedback] Noticed during testing that Zcode doesn't seem to trigger compression very often. | [反馈]在测试途中发现Zcode好像不怎么会触发压缩 #4344 缓存接缝、codexHealthBarApplies 等改动,与本 PR 的声明端点功能在同一行域交汇。
  3. 影响范围与严重程度:仅影响本 PR 分支自身的可合并性;不涉及任何用户行为——master 侧行为原样保留,PR 功能完整保留。
  4. 怎么修复的:merge(非 rebase,与此前几轮冲突修复一致),逐 hunk 取并集;语义冲突处按「新结构 + 新功能检查」合并。
  5. 是否有回归问题:无。既有行为逐项核对——master 的未归属拒绝路径、genBoundary/mcpDialect 嵌套顺序、codexHealthBarApplies 判定均未被触碰;PR 侧声明端点放行逻辑位置不变(gate 首语句)。
  6. 是否破坏缓存:不涉及——本轮只解冲突,未改变任何出站字节或 prompt 拼接顺序(relay round-2 的 systemPrompt 采用 master 的现行顺序,正是缓存接缝对齐后的版本)。
  7. 是否改了设计方向:无,双方各自方向原样合入。
  8. 是否引入了新的配置:无(配置面仍为本 PR 原有的 model endpoints 声明机制,未动)。

判定:冲突已清,等 CI 绿即可由 owner 合并。

@suyeku1

suyeku1 commented Nov 9, 2026

Copy link
Copy Markdown
Contributor

Rebased this branch onto current master in a cross-fork PR so it can move again: #6032 → #6032

Why: this PR has been 1725 commits behind master with 11 conflicted files (mergeable_state: dirty) since 2026-09-26. #6032 carries the identical feature set rebuilt on 4be21b08:

  • 28 files, +2158/−38, 0 behind master — same file list as here.
  • Merge commit 80e9d205eb has two parents — this branch's tip b7c9a02302 and master 4be21b08 — so merging [PR] Rebase #1444 onto master: declarable model endpoints + commandcode wire family #6032 back into 2026-09-26_model-endpoint-patterns is a fast-forward that leaves this PR mergeable against master as-is.
  • Conflicts resolved by keeping both sides: relay.ts (master's messageTimestampSections + the commandcode adapter), handle.ts (master's ingest-size guards + declared-endpoint matching/declPinConflictWarned), the six src/agent/* files, plugin.ts, native-intercept.test.ts.
  • CONFIGURATION.md / .zh-CN.md are pointers to docs/configuration.md on master now, so the prose moved into docs/configuration.* and the parameter table went through the generator (website/config-reference/server.yaml seed → tools/gen-config-docs.mjs generate, 254 i18n keys, check-coverage exit 0) rather than being hand-edited.

Verification on 80e9d205eb: tsc --noEmit exit 0 with no output; the six suites this PR adds or touches pass 124/124. CI on the fork PR: 10 success / 10 skipped-by-design / 0 failed.

No features were added beyond what is here — nothing from the later wire work is bundled.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

(ai: opencode/qwen3.8-27b)

Merge conflict resolved and pushed. I merged origin/master (4be21b0) into 2026-09-26_model-endpoint-patterns and pushed merge commit 0f502f4 (fast-forward, no force-push; remote head verified unchanged before push). The branch convention here is merging master (the previous tip b7c9a02 was itself such a merge), so rebase was ruled out.

Conflicts: 11 files, all resolved:

  • Code (union semantics — both sides kept): src/server/handle.ts (declared-endpoint matching from this PR and master's ingest-size guards from the 5 GiB body-cap work), src/server/relay.ts (master's messageTimestampSections wrapping and the commandcode adapter branch), src/plugin.ts (this PR's pipePluginJsonlWithStrip and master's new describeReadError helper from Replay a stream cut that lands after reasoning-only output #5951 — the two inserted at the same anchor; closing braces needed manual balancing), plus import unions in src/agent/dsh-native.ts, src/agent/native-intercept.ts, src/agent/opencode-native.ts, tests/native-intercept.test.ts.
  • Docs (followed master's new layout): root CONFIGURATION.md/.zh-CN.md are redirect stubs on master (2026-11-08 doc-layout ruling), so the PR's "Model Endpoint Patterns" prose was ported verbatim into docs/configuration.md/.zh-CN.md and the parameter row regenerated through tools/gen-config-docs.mjs (254 i18n keys, EN/ZH parity, coverage check exit 0). Same generator run refreshed website/main.js + website/docs/index.html.

Local verification (targeted; full suite is CI's job): tsc --noEmit clean · six touched suites (adapter-commandcode, commandcode-endpoint, commandcode-plugin-pipe, commandcode-wire, model-endpoints, native-intercept) 124/124 pass · tsup build OK. Post-merge diff vs master is 28 files, +2156/−38 — byte-identical stat to the original PR, i.e. the merge introduced no feature drift.

Re: @suyeku1's #6032 suggestion — thanks for the analysis; the described operation matches what this branch needed. I performed the merge independently on this PR's own branch rather than adopting the cross-fork tree (external work wasn't verifiable without consuming it), so #6032 is now redundant — up to the owner whether to close it.

Flag for reviewers (not fixed by me — layout decision, not a mechanical conflict fix): WIRE-CONTRACTS.md/.zh-CN.md sit at repo root, which contradicts the 2026-11-08 doc-layout ruling (docs belong under docs/ + index entries). No CI lane enforces root-doc placement, so I left them untouched.

Live status right now: GitHub reports mergeable: true, mergeable_state: blocked (required checks still running on 0f502f4). Mirror check cache: 30 checks, 14 in progress, 0 failed, e2e lanes skipped by design. If any lane goes red I'll pick it up in this lane (attempt 1/3) and coordinate via #4599 per policy.


结构化八问(中文)

  1. 是什么问题:PR 分支自 2026-09-26 落后 master 约 1725 个提交,11 个文件与 master 文本冲突(mergeable_state=dirty),无法合并也无法正常过审。
  2. 什么原因导致的:分支切出后 master 侧发生了大范围变更——文档布局迁移(根目录 CONFIGURATION.md → docs/configuration.md + website 生成器)、plugin.ts 截断诊断(Replay a stream cut that lands after reasoning-only output #5951)、native-intercept 的 pfChannel/respawn 冷却、handle.ts 的 ingest 尺寸守卫等——与本 PR 的改动区间重叠,产生冲突。属正常的长周期分支漂移,非任何一方代码错误。
  3. 影响范围与严重程度:仅影响本 PR 自身的可合并性;不影响任何用户可见行为。严重度:阻塞级(PR 卡住不能合),无运行时影响。
  4. 怎么修复的:按分支既有惯例把 master 合并进分支(而非 rebase,避免 force-push);全部 11 处冲突按「并集」语义解决——代码文件两侧的定义都保留,文档跟随 master 新布局(旧根目录 prose 移入 docs/,website 块用生成器重出);未删除、未改写任何一侧的功能。
  5. 是否有回归问题:未见回归证据。本地:tsc --noEmit 干净、本 PR 触及的 6 个测试文件 124/124 通过、tsup 构建通过、合并后 diff vs master 与原 PR 完全一致(28 文件 +2156/−38)。高危项逐一核对:plugin.ts 同时保住 PR 的 pipePluginJsonlWithStrip 与 master 的 describeReadError;handle.ts 同时保住 declared-endpoint 分类与 master 的 ingest 尺寸守卫(含 responses 协议 0 上限特判);relay.ts 同时保住 master 的 messageTimestampSections 包装与 commandcode 适配分支。全量套件判定以远程 CI 为准。
  6. 是否破坏缓存:不涉及——未触碰缓存/重试/增量游标/去重键逻辑。
  7. 是否改了设计方向:没有,纯冲突消解,无架构/数据流变化。一处留给评审判断:WIRE-CONTRACTS.md/.zh-CN.md 放在仓库根目录,不符合 2026-11-08 文档布局裁定,但无 CI 车道强制,我保持原样并在上文已点名。
  8. 是否引入了新的配置:我没有引入。本 PR 自带的 modelEndpointPatterns 配置项是 PR 作者的设计(非本次合并新增),随 PR 合并后生效。

收工判定:冲突已解并推送,PR 已恢复 mergeable;CI 在新 head 上运行中(当前 0 失败),转绿后即可由 owner 合并。

ranxianglei added a commit to suyeku1/billion-context that referenced this pull request Nov 9, 2026
…(relative link + index/architecture lines)

The ranxianglei#1444 feature branch polished these five lines after this PR's tip
was cut: the wire-contracts ledger reference becomes a clickable
relative link (the file lives next to configuration.md), the docs
index row states the ledger's permanence rule, and the architecture
source map carries the longer comments with adapter-commandcode.ts
in position. Taking the ranxianglei#1444 versions verbatim makes this PR a
superset of the feature branch, so either merge path offered in the
PR body (merge tip into the feature branch, or replace its head)
lands identical docs.
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

合并前复核完成:对 review 期间顺手修的两个提交逐行核对,均按声称落地,本轮无新发现。审核结论维持:APPROVE(human merge tier)。

两个修复提交的核验结果

  • f03e8925c4(JSONL pipe 标签过滤状态):确认旧实现确实在每条 JSONL 行后 flush 过滤器状态机,摧毁跨 delta 的半标签/半标记暂存;修法与 SSE pipe(pipePluginChatWithStrip)语义对齐——每个 prose 字段一个独立过滤器实例(text/reasoning,避免字段交叉污染),仅 finish/error/EOF/读失败前 flush,暂存尾巴以原事件类型合成行写回(旧代码把 reasoning 尾巴写成 text-delta 的附带 bug 也一并修正),degenerate-turn 统计改为双字段聚合。三条回归测试分别钉住三个面:跨 delta 拆分标签整体剥离 / 暂存前缀存活到终端帧 / 跨 delta 伪造标记完整剥离。
  • 793a488b2b(文档布局):WIRE-CONTRACTS{,.zh-CN} 移入 docs/、docs/index{,.zh-CN} 补条目、5 处代码/测试注释 + 2 处配置文档引用改指(共 7 处),纯移动无内容改动。

当前状态

结构化八问

  1. 是什么问题:本轮无新缺陷;是对 review 期间两个顺手修的合并前复核——①plugin-mode commandcode JSONL pipe 逐行 flush 导致标签过滤状态丢失;②新文档落在仓库根目录违反文档布局裁决([PR] docs: batch-migrate 14 root-level documents into docs/ (consolidation + master index + redirect placeholders) | docs: 14 个根级文档一次性迁入 docs/(集中 + 总目录 + 重定向占位) #5920)。
  2. 什么原因导致的:①plugin-mode JSONL pipe 误按「每行 flush」实现过滤器生命周期,而过滤器设计上依赖跨 delta 持有半标签/半标记(SSE pipe 语义本就正确,仅 JSONL 车道实现错);②原 PR 新增文档置于仓库根目录,违反 2026-11-08 文档布局裁决。
  3. 影响范围与严重程度:①影响所有走 plugin mode 的 commandcode 端点——跨 delta 拆分的渲染标签泄漏进客户端可见文本、终端边界暂存前缀丢失、跨 delta 伪造 [ACP] 标记无法剥离(正确性缺陷);已在合并前修复,未发布版本不受影响。②仅文档位置,无运行时影响。
  4. 怎么修复的:JSONL pipe 对齐 SSE pipe 既有语义——每 prose 字段独立过滤器实例,仅终端帧/EOF/读失败前 flush,暂存尾巴按原事件类型合成行写回;文档移入 docs/ 并登记 index、引用批量改指。
  5. 是否有回归问题:无。修改变动仅限 JSONL plugin pipe 的 flush 时机,SSE pipe、proxy mode 路径、usage 结算、错误面均未触碰;三个缺陷面均有回归测试钉住,全量 CI 绿;高危模式(语义漂移/fail-fast 变吞错/阈值时序/输出格式变化)逐项核对均不涉及。文档移动为纯移动+引用改指,内容零改动。
  6. 是否破坏缓存:不涉及——未触碰任何缓存/重试/增量游标/去重键逻辑。
  7. 是否改了设计方向:没有。修复是把 JSONL 车道收敛到既有 SSE 车道语义,属对齐而非转向。
  8. 是否引入了新的配置:没有(modelEndpointPatterns 是 PR 自带功能,PR 描述行为变更清单已披露)。

收工判定:review 收口,verdict APPROVE(新配置面 + 新 wire family,属 §7.4 human merge tier,agent 不能代合);最终 head CI 绿、无冲突,等你合并。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

(ai: opencode/qwen3.8-27b)

Update — one red lane found and fixed, branch now fully up-to-date with master.

After my merge push, the full check-run list on the new head revealed one failure the mirror summary had not surfaced yet: architecture-docs. Root cause: this PR adds three src/ modules (src/model-endpoints.ts, src/commandcode-wire.ts, src/loop/adapter-commandcode.ts) but the pre-existing branch predates the architecture-docs guard, so reference/architecture.md's file map never listed them. Fixed in a66487c — three file-map rows added (the guard's mechanical requirement); verified green locally before pushing.

On the head that followed (793a488, after two additional commits landed on the branch from other work): all 56 check runs green or skipped-by-design, zero failures, architecture-docs included. Two things worth noting from that head:

  • My earlier flag about WIRE-CONTRACTS.md sitting at repo root was already actioned by a follow-up commit (793a488b2b moves it to docs/wire-contracts.{md,zh-CN.md} + index entries + reference updates) — no longer open.
  • A plugin-pipe filter-state fix (f03e8925c4) also landed with its regression test.

While I was watching CI, master advanced by two commits (#5913, stale-epoch overlay fix touching only src/session.ts). I merged them into the branch cleanly (no conflicts, disjoint files) and pushed final head ee19f22 — the branch is no longer "behind". CI is running on it right now: 0 failures at time of writing; the merged-in commits were already green on master itself, and the rest of the tree was green at 793a488. If any lane goes red I'll pick it up here (lane attempts 2/3 remain).


结构化八问(中文,覆盖本次 auto-fix 全程)

  1. 是什么问题:PR 分支落后 master 约 6 周 / 1725 提交,11 个文件冲突(dirty)无法合并;合并后又暴露一条红 CI(architecture-docs 文件清单缺 3 个新模块);最后 master 又前进 2 提交导致分支 behind。
  2. 什么原因导致的:①冲突 = 长周期分支漂移,master 侧文档布局迁移、plugin.ts 截断诊断(Replay a stream cut that lands after reasoning-only output #5951)、native-intercept/handle.ts 新守卫与本 PR 改动区间重叠;②CI 红 = PR 切出时 architecture-docs 守卫尚不存在,新增 src 模块未登记进 reference/architecture.md——是「先有守卫后补文档」的时序差,不是代码缺陷;③behind = master 正常推进。
  3. 影响范围与严重程度:仅本 PR 自身的可合并性/可过审状态,无用户可见行为影响。严重度:阻塞级(不修则 PR 永远卡住),运行时零影响。
  4. 怎么修复的:按分支既有惯例合并 master(非 rebase、无 force-push);11 处冲突全部并集语义解决,两侧功能无一丢失;CI 红按最小机械修复处理(只补文件清单 3 行,不动架构页/图——那是评审判断项);behind 用一次干净的小合并追平。每步推送前 fetch 核对远端 head,均为 fast-forward。
  5. 是否有回归问题:未见。合并后 diff vs master 与原 PR 完全一致(28 文件 +2156/−38);本地 tsc 干净、触及的 6 个测试文件 124/124、tsup 构建通过;高危点逐一核对(plugin.ts 双函数并存、handle.ts 分类+尺寸守卫并存、relay.ts 时间戳包装+commandcode 分支并存)。793a488b2b 上全部 56 个 check run 绿/跳过;最终 head ee19f22 仅并入 master 已绿的 [PR] fix: stale-epoch config overlay loses the named-provider recipe table (#5912) #5913,CI 运行中、当前 0 失败。
  6. 是否破坏缓存:不涉及——未触碰缓存/重试/游标/去重键逻辑。
  7. 是否改了设计方向:没有,纯冲突消解 + 机械修复 + 追平 master,无架构/数据流变化。
  8. 是否引入了新的配置:我没有引入。PR 自带的 modelEndpointPatterns 配置项是作者设计,随 PR 合并生效。

收工判定:冲突、红 CI、behind 三个阻塞全部解除,分支处于最新 master 基线上且上一 head 全绿;等 ee19f22 的 CI 转绿即可由 owner 合并。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[ework] CI fix verified green ✓

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Conflict resolved and pushed (auto-fix lane, attempt 1/3).

What I verified live before touching anything

  • Local head == remote PR head ee19f22c06; mirror cache (fetched ~7 min before this run) still reported dirty, and all 30 checks on that head were green/skipped — so the conflict was the only blocker.
  • Root cause of the stale dirty: the previous merge commit (ee19f22c06, 2026-11-09 23:36 +0800) had merged an older master — origin/master had moved 109 commits ahead of the PR head by the time of this run, so GitHub's dirty state was real and fresh, not stale metadata.

What I did

  1. Merged current origin/master into pr1444-head (merge strategy, consistent with this branch's history — preserves the contributor's original commits; no rebase).
  2. Exactly one conflict hunk: src/server/relay.ts:2594–2606 — the PR's commandcode adapter ternary ([Feature] Declarable custom-wire model endpoints — route providers like commandcode's Go plan (/alpha/generate) into compression #1295) vs master's new loopSeparatorSlots declaration (Compression breaks Responses assistant history ordering, causing repeated invalid_assistant_history 400s #564/[Bug][Responses] acp_turn_sep 合成分隔符进入持久压缩覆盖集合,导致重复历史漂移误报 #2627 final-wire provenance). Resolved as the union: kept the PR's prepared.commandcode ? createCommandcodeAdapter(...) : pickAdapter(...) ternary AND added master's let loopSeparatorSlots: readonly SeparatorSlot[] = [] right after it. Verified the pickAdapter argument lists are byte-identical on both sides (no signature drift), and the SeparatorSlot import plus all downstream wiring landed intact (refreshFolded slots refresh at src/server/relay.ts:2672/2675, runCompressLoop gets separatorSlots: () => loopSeparatorSlots at src/server/relay.ts:2687).
  3. Reviewed every auto-merged region for semantic (not just textual) correctness: master's Compression breaks Responses assistant history ordering, causing repeated invalid_assistant_history 400s #564/[Bug][Responses] acp_turn_sep 合成分隔符进入持久压缩覆盖集合,导致重复历史漂移误报 #2627 object-graph rewrite machinery (compat roles / output steering / dropFields / chain-stamp / egress shed) auto-merged cleanly around the PR's commandcode work — the CLI-envelope rewrap still sits after every flat-body mutation and before the chain stamp (src/server/relay.ts:649–656 vs :669), exactly matching the prior merge round's established resolution. Commandcode requests carry no separator manifest, so all of master's new graph branches take their legacy path for them.
  4. Local gates on the merged head 04a7847a21: npm run typecheck exit 0; npm run build success (dist-import-annotations: OK, dist-external: OK). Full regression suite intentionally left to remote CI per local-testing discipline.
  5. Push discipline: git fetch origin immediately before push — remote head unchanged (no concurrent activity on the branch) — then a plain fast-forward push, no force-push: ee19f22c06..04a7847a21 on 2026-09-26_model-endpoint-patterns.

Post-push verification (live, mirror cache)


结构化八问(本次运行)

  1. 是什么问题:PR 分支与 master 存在合并冲突(GitHub merge state = dirty),合并被卡住。
  2. 什么原因导致的:上一轮合并提交(ee19f22c06,11-09 23:36)合入的是较旧的 master;此后 master 又前进了 109 个提交,PR head 未包含最新 master,冲突因此真实存在。
  3. 影响范围与严重程度:仅影响本 PR 的可合并性,不涉及任何用户可见行为;严重程度 = 合并被阻塞(该 head 上 CI 本身全绿)。
  4. 怎么修复的:再次把当前 origin/master merge 进 PR 分支(保留贡献者原始提交,不 rebase);唯一冲突在 src/server/relay.ts:2594–2606(PR 的 commandcode 适配器三分支表达式 vs master 新增的 loopSeparatorSlots 声明),按并集保留双方解决;并核对了 master 侧 Compression breaks Responses assistant history ordering, causing repeated invalid_assistant_history 400s #564/[Bug][Responses] acp_turn_sep 合成分隔符进入持久压缩覆盖集合,导致重复历史漂移误报 #2627 对象图重构与 PR 侧 commandcode 逻辑的全部自动合并区域语义正确(rewrap 仍在所有 flat-body 变更之后、chain stamp 之前,与上一轮决议一致)。
  5. 是否有回归问题:审过的既有行为——pickAdapter 两侧参数列表完全一致(无漂移);commandcode rewrap 顺序未变;loopSeparatorSlots 声明及下游接线完整;typecheck + build 通过。master 侧变更均为 master 上已发布行为,经 merge 整体带入,未做语义改动。
  6. 是否破坏缓存:否——本次运行未改动任何缓存/重试/增量游标/去重键逻辑,仅带入 master 既有变更。
  7. 是否改了设计方向:否。
  8. 是否引入了新的配置:否。

判定:冲突已解除并推送(fast-forward,无 force-push);CI 在新 head 上重跑中(截至写作时刻零失败)。转绿后由 owner 合并(merge 仅限人工)。

@ranxianglei ranxianglei left a comment

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

End-to-end review at final head 04a7847a2 (branch fully synced to master tip 4446301708; git merge-tree shows zero conflicts).

Independently re-verified this round:

  • Feature footprint vs master: exactly 31 files (+2236/−38) — the declared PR scope, nothing more; the two master-sync merges introduced no feature drift.
  • All 9 files touched by BOTH master's window and this PR, re-read at the final head: every PR hunk intact and correctly integrated — dsh takeover-gate declaration check remains the first statement ahead of master's refuseUnattributed path; classifier priority chain URL-pin > #1909 providers-table > modelEndpointPatterns > built-ins, with the bounded-FIFO once-warn on pin/declaration conflict; client-side claim keeps /bili/ + /__bili/ control-plane exclusions BEFORE declaration matching; relay rewraps the envelope at the final boundary AND on every compress-loop resend (after v1 wrapping), fake-completion retry disabled for JSONL (its finish IS the terminator), commandcode stream-error shapes gated on prepared.commandcode; the plugin JSONL pipe carries the per-field tag/marker filter fix (flush only at terminal frame / EOF / read failure, tails written back as synthetic lines of the original event type).
  • Docs: modelEndpointPatterns table row + prose section in docs/configuration{,.zh-CN}.md, wire-contracts under docs/ with index entries, architecture-map rows for the three new modules.
  • Boundaries: no kernel/, no pi-subagents/, version untouched, no lockfile churn.
  • CI on 04a7847a2: 30 checks, 0 failed / 0 pending (success/skipped only).

Tests have real teeth: codec round-trips + rejection matrix; a real-HTTP-proxy e2e asserting wire bytes in BOTH directions (upstream sees the flat OpenAI body, client gets the rewrapped envelope); plugin-pipe facet pins (cross-delta tag stripped as one unit, held prefix survives to the terminal frame, forged marker completed across deltas fully stripped); declaration-matching truth table including segment-boundary negatives and control-plane exclusion. These would catch regressions of exactly this bug class.

Behavior changes (the six old→new items disclosed in the PR body) all match the implementation; the declaration-driven compression of custom paths is additive — undeclared URLs take byte-identical behavior to before.

Verdict: APPROVE. New config surface (modelEndpointPatterns) + new wire family (commandcode) place this PR in the human merge tier per §7.4 — merging stays with the owner.

中文摘要:最终 head 已完全同步 master 尖端且零冲突;本 PR 特征集相对 master 恰为 31 个文件,两次 master 同步合并未引入任何功能漂移;master 窗口与本 PR 双触的 9 个文件逐一复核,PR 增量全部完整且集成正确(dsh 声明放行仍为 gate 首语句、分类器优先级链与有界一次性告警、客户端 claim 的 control-plane 排除前置、relay 终边界+压缩循环 resend 双重包、JSONL 关闭假完成重试、plugin pipe 逐字段过滤器修复在位);无 kernel/pi-subagents 改动、版本未动;CI 全绿(30 项检查 0 失败)。测试有牙:编解码往返+拒绝矩阵、真实代理双向线字节断言、plugin pipe 三个缺陷面钉死、声明匹配真值表。PR 描述披露的六项行为变更与实现一致。判定 APPROVE(新配置面+新 wire family,属 §7.4 人工合并档,合并操作归 owner)。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[ework] CI fix verified green ✓

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

head 04a7847a2 的端到端 review 已完成并提交到 GitHub(GitHub 不允许作者 token 发 APPROVE 事件,按前几轮先例以 review-comment 形式贴出完整审核,verdict 不变):APPROVE(human merge tier)。

本轮做了什么

结构化八问

  1. 是什么问题:本轮是合并前端到端复审(分支较上轮新增两笔 master-sync merge),本身不针对新缺陷。
  2. 什么原因导致的:不适用——本轮是合并状态核验,非缺陷排查。
  3. 影响范围与严重程度:无行为变化;核验覆盖 PR 触碰的全部路径(分类器 / 客户端 claim / relay / plugin pipe / 文档)。
  4. 怎么修复的:不适用(本轮无修复);审核本身已作为 GitHub review-comment 提交。
  5. 是否有回归问题:无。高危模式逐项核对(语义漂移 / fail-fast 变吞错 / 阈值时序 / 输出格式 / 缓存策略):优先级链集成决策已披露、body-based disambiguation 守卫完好、未声明 URL 行为与从前逐字节一致、JSONL pipe 修复仅改 flush 时机;PR 披露的六项行为变更与实现逐一相符。
  6. 是否破坏缓存:不涉及——缓存/重试/增量游标/去重键均未触碰;唯一重试路径变化是 commandcode 关闭 [bug] 假完成:模型把工具调用写成文字(无 tool_use 块)→ 轮次提前终止,agent 显示正常完成(#361 根因 follow-up) #371 假完成重试(该 wire 无假完成形态,PR 描述已披露)。
  7. 是否改了设计方向:没有。方向与 owner 批准的 [Feature] Declarable custom-wire model endpoints — route providers like commandcode's Go plan (/alpha/generate) into compression #1295 方案一致(单一声明源 + 严格保真 codec + WC 台账),本轮未改动任何架构决策。
  8. 是否引入了新的配置:有一项——modelEndpointPatterns({match, wire} 数组,默认 []),为本 PR 自带功能,PR 描述行为变更清单与 docs/configuration{,.zh-CN}.md 均已披露;进程启动时读取,修改后需重启生效。

收工判定:review 收口,verdict APPROVE(新配置面 + 新 wire family,属 §7.4 human merge tier,agent 不能代合);最终 head CI 绿、无冲突、无遗留发现——等你合并。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Conflict resolved and pushed again (auto-fix lane, attempt 1/3 of this cycle).

Why the trigger was real this time
Last cycle I pushed 04a7847a21 and the mirror showed blocked. Since then origin/master advanced another 43 commits (#6065 responses-acp-interception, #6068 egress-transform-pipeline, #6064 responses-head-role, …), which rewrote src/server/handle.ts (~106 lines) and src/server/relay.ts (~170 lines) again. A dry-run 3-way merge (git merge-tree) of current master vs the PR head proved a fresh conflict in src/server/handle.ts before I touched anything — the mirror cache snapshot I first read was ~7h stale (fetched 18:18Z Nov 9), so I did not trust its blocked label and verified by merge simulation.

What I did

  1. Merged current origin/master into pr1444-head (merge strategy, consistent with this branch's history — contributor commits preserved, no rebase).
  2. Exactly one conflict hunk: the import block at src/server/handle.ts:94–99 — master moved logRequestCost out of ../server.js into a new ./request-cost.js module, while the PR side had added declPinConflictWarned to the server.js import ([Feature] Declarable custom-wire model endpoints — route providers like commandcode's Go plan (/alpha/generate) into compression #1295 F4 once-gate). Resolved as union: kept declPinConflictWarned in the server.js import AND took master's split (import { logRequestCost } from "./request-cost.js"). Verified: declPinConflictWarned still exported from src/server.ts:209, request-cost.ts exists with the export, all 8 call sites intact, and no stale logRequestCost imports from server.js remain anywhere in src/.
  3. Semantic review of every auto-merged region (master's second rewrite pass): all commandcode/model-endpoint feature symbols survived — Prepared.commandcode wiring in src/server.ts, modelEndpointPatterns in src/config.ts, pipePluginJsonlWithStrip in src/plugin.ts:3934, the adapter ternary + both envelope rewaps + loopSeparatorSlots wiring in src/server/relay.ts (rewrap still sits after every flat-body mutation and before the chain stamp, same ordering as established last cycle), and all three feature modules present.
  4. Local gates on merged head d9963dbcc3: npm run typecheck exit 0, npm run build success (dist-import-annotations: OK, dist-external: OK). Full suite left to remote CI per local-testing discipline. (Environment note: the workspace had been re-provisioned between runs — all three node_modules trees were gone; ran npm install first, 611 packages, no lockfile changes.)
  5. Push discipline: git fetch origin immediately before push — remote PR head unchanged (no concurrent activity) — then a final dry-run merge against the newest master tip (c8dce581f5, which landed mid-flight) to confirm zero conflicts, then plain fast-forward push, no force-push: 04a7847a21..d9963dbcc3.

Post-push verification (live)


结构化八问(本次运行)

  1. 是什么问题:上一轮合并后 master 又前进 43 个提交,PR 分支与 master 再次产生合并冲突(src/server/handle.ts),合并被卡住。
  2. 什么原因导致的:仓库合入节奏极快(几分钟一个 PR);master 本轮重写了 handle.ts(把 logRequestCost 拆到新的 ./request-cost.js 模块)并在同一 import 行上与 PR 侧新增的 declPinConflictWarned 导入相撞。
  3. 影响范围与严重程度:仅影响本 PR 的可合并性,不涉及用户可见行为;严重程度 = 合并被阻塞(此前该 head 上 CI 全绿)。
  4. 怎么修复的:再次 merge 当前 origin/master(保留贡献者原始提交,不 rebase);唯一冲突是 src/server/handle.ts:94–99 的 import 块,按并集解决——保留 PR 的 declPinConflictWarned、采纳 master 的 logRequestCost 模块化拆分;随后核对了全部自动合并区域中 commandcode/model-endpoints 功能符号的完整性与既有顺序不变量。
  5. 是否有回归问题:审过的既有行为——declPinConflictWarned 导出仍在(src/server.ts:209)、logRequestCost 8 处调用点完整且无残留旧导入、commandcode 信封 rewrap 相对 chain stamp 的顺序未变、loopSeparatorSlots 接线完整;typecheck + build 通过。master 侧变更均为已发布行为整体带入,未做语义改动。
  6. 是否破坏缓存:否——未触碰任何缓存/重试/增量游标/去重键逻辑。
  7. 是否改了设计方向:否。
  8. 是否引入了新的配置:否。

判定:冲突已解除并推送(fast-forward,无 force-push);GitHub 现报 behind(无冲突,仅落后于持续前进的 master);CI 在新 head 上重跑中(截至写作时刻零失败)。转绿后由 owner 合并(merge 仅限人工)。若 master 后续再引入新冲突,下一轮触发时按同样流程处理。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Confirmation (this cycle's report is already on the thread as the previous comment, 01:30Z):

  • Conflict resolved: merged current origin/master into the PR branch, single hunk in src/server/handle.ts (import block — union of master's logRequestCost module split and the PR's declPinConflictWarned import), pushed fast-forward 04a7847a21..d9963dbcc3 (no force-push).
  • Local gates on d9963dbcc3: typecheck exit 0, build OK.
  • Live state at verification: mirror sha d9963dbcc3, mergeable_state = behind (conflict cleared — branch merely trails a fast-moving master); fresh CI running on the new head, zero failures so far.

Nothing further to add this run; will act again only if CI turns red or master introduces another conflict.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[ework] CI fix verified green ✓

1 similar comment
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[ework] CI fix verified green ✓

@suyeku1

suyeku1 commented Nov 10, 2026

Copy link
Copy Markdown
Contributor

Equivalence check, recorded so nobody has to redo it:

#1444's base is already master, and it is MERGEABLE / BEHIND on its own — no cross-branch sync is needed before merging it.

Proof that the two heads carry the same content despite diverging histories:

GET /git/trees/d9963dbcc3?recursive=1   (this PR's head)
GET /git/trees/9914dcae91?recursive=1   (#6032's head, suyeku1 fork)
  -> 2255 blobs each; only-in-A 0, only-in-B 0, differing blob shas 0
compare d9963dbc...9914dcae -> diverged, 6 ahead / 6 behind, files 0 content diff

So the three files #6032 picked up past this branch's tip (1a928c3ac plugin-pipe filter state across deltas, 48dc0d5ff wire-contracts wording, plus the two origin/master merge commits) are byte-for-byte what this branch already carries from its own commits (f03e8925c, 793a488b2). Either head can be merged; neither loses anything.

Nothing is outstanding on the code side here. The one remaining step is a single human merge, on either PR.

@ranxianglei ranxianglei mentioned this pull request Nov 10, 2026
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Equivalence record independently verified — and the branch is still clean as of this run.

What I did (verification only, nothing pushed)

  1. Fetched refs/pull/6032/head (9914dcae9126) directly from the base repo and ran git diff --shortstat d9963dbcc3 9914dcae91: empty output — zero content difference. Your tree-equivalence record holds right now, confirmed by direct tree comparison rather than blob counts.
  2. Live state of [PR] feat: declarable custom-wire model endpoints + commandcode-cli wire family (#1295) #1444 (mirror cache, fetched minutes ago): head d9963dbcc3, mergeable_state = behind, all 30 checks completed: 14 success / 16 skipped / 0 failed — CI fully green on this head.
  3. Master has advanced another 68 commits since my last sync (tip 4dea18eabf). Dry-run git merge-tree against that tip: exit 0, zero conflicts — so "no cross-branch sync needed before merging" still holds at this moment. Per this branch's established practice I am not re-syncing while the state is merely behind (every push re-runs the full CI matrix against a master that moves every few minutes); if a later sync introduces a real conflict, the next trigger handles it in-lane.

Where things stand: both PRs carry byte-identical content, so either can be merged with nothing lost. The remaining step — a single human merge on either #1444 or #6032, plus the call on closing the now-redundant twin — belongs to @ranxianglei (merging is human-only under system rules, and agents don't close PRs/issues).

ranxianglei added a commit that referenced this pull request Nov 10, 2026
[PR] Rebase #1444 onto master: declarable model endpoints + commandcode wire family
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Conflict resolved and pushed (auto-fix lane, attempt 1/3 of this cycle).

Environment note first: the workspace had been re-provisioned since the last run (fresh empty git repo, no node_modules). I re-fetched master + the PR branch and reinstalled dependencies (611 packages, no lockfile changes) before doing anything else.

Why the trigger was real
origin/master had advanced 179 commits past the PR head (tip 9674cfafbc). A dry-run 3-way merge (git merge-tree) proved a fresh conflict in src/server/relay.ts before I touched anything — the only conflicted file; handle.ts, plugin.ts, docs, and the rest auto-merged.

What I did

  1. Merged current origin/master into the PR branch (merge strategy, consistent with this branch's history — contributor commits preserved, no rebase).
  2. Exactly one conflict hunk: the round-2 systemPrompt line at src/server/relay.ts:2649–2653. Three-way stage comparison (git show :1/:2/:3) proved the PR side never modified this line (base == ours); master's Request two capabilities: compressed block "retirement" and summary "external rewrite". | 请求两个能力:压缩块「退役」与摘要「外部重写」 #5765 block-ops work changed it, adding the blockOpsActive argument to withMcpDialectNote(...) and the + blockOpsSection suffix. Resolved by taking master's line verbatim — there was nothing PR-specific in the ours-side copy to preserve. Verified blockOpsActive (:2615) and blockOpsSection (:2627) are defined in the merged tree and no conflict markers remain.
  3. Semantic spot-check of the auto-merged regions: all feature symbols intact — commandcode ×3 in src/server.ts, modelEndpointPatterns in src/config.ts (:1691/:2001), pipePluginJsonlWithStrip in src/plugin.ts:3960, the adapter ternary at src/server/relay.ts:2653, both envelope rewaps (:686–688 final wire boundary, :759 resend path — still after every flat-body mutation, before the chain stamp), declPinConflictWarned gate in src/server/handle.ts:94/:433, and all three feature modules present.
  4. Local gates on merged head 1370426c16: npm run typecheck exit 0, npm run build success (dist-import-annotations: OK, dist-external: OK). Full suite left to remote CI per local-testing discipline.
  5. Push discipline: git fetch origin immediately before push — remote PR head unchanged at d9963dbcc3 (no concurrent activity) — final dry-run merge against the newest master tip (unchanged at 9674cfafbc) exit 0, then plain fast-forward push, no force-push: d9963dbcc..1370426c1.

Post-push verification (live)


结构化八问(本次运行)

  1. 是什么问题:上次同步后 master 又前进 179 个提交,PR 分支与 master 再次产生合并冲突(src/server/relay.ts),合并被卡住。
  2. 什么原因导致的:master 的 Request two capabilities: compressed block "retirement" and summary "external rewrite". | 请求两个能力:压缩块「退役」与摘要「外部重写」 #5765 block-ops 改动修改了 round-2 重建的 systemPrompt 行(加 blockOpsActive 参数与 blockOpsSection 后缀),与 PR 侧的旧基线同一行相撞。三方 stage 比对证明 PR 侧从未改过这行——纯长命分支漂移,非代码缺陷。
  3. 影响范围与严重程度:仅影响本 PR 自身的可合并性,不涉及用户可见行为;严重程度 = 合并被阻塞(此前该 head 上 CI 全绿)。
  4. 怎么修复的:再次 merge 当前 origin/master(保留贡献者原始提交,不 rebase);唯一冲突按三方比对结果原样采纳 master 版本(PR 侧该行无独有内容);随后核对了全部自动合并区域中 commandcode/model-endpoints 功能符号的完整性与既有顺序不变量(rewrap 仍在 flat-body 变更之后、chain stamp 之前)。
  5. 是否有回归问题:审过的既有行为——blockOpsActive/blockOpsSection 定义在位(:2615/:2627)、commandcode 双 rewrap 位置与顺序未变、adapter 三元在位、declPinConflictWarned 门控完整、无残留冲突标记;typecheck + build 通过。master 侧变更均为已发布行为整体带入,未做语义改动。
  6. 是否破坏缓存:否——未触碰任何缓存/重试/增量游标/去重键逻辑。
  7. 是否改了设计方向:否。
  8. 是否引入了新的配置:否(blockOps 配置面是 master Request two capabilities: compressed block "retirement" and summary "external rewrite". | 请求两个能力:压缩块「退役」与摘要「外部重写」 #5765 自带,随合并带入,非本次引入)。

判定:冲突已解除并推送(fast-forward,无 force-push);GitHub 现报 mergeable: true / blocked(无冲突,等待 CI 与人工合并);CI 在新 head 上重跑中(截至写作时刻尚未出结论,不预断绿)。转绿后由 owner 合并(merge 仅限人工)。若 master 后续再引入新冲突,下一轮触发时按同样流程处理。

@suyeku1

suyeku1 commented Nov 12, 2026

Copy link
Copy Markdown
Contributor

#1444 is now an empty PR — the close decision, with evidence

Filed here rather than in #5128 because the decision belongs to this PR. Verified live just now, no re-derivation needed on your side:

GET /repos/ranxianglei/billion-context/pulls/1444
  -> state=open, merged=false, mergeable=true, mergeStateStatus=BLOCKED
     changed_files=0, additions=0, deletions=0, commits=7
     head=1370426c162a1f3f1ef431256d3b12ad9a8974a0
     base=master @ 9674cfafbc78c50245ae0d887483d3d6ce880c15
GET .../pulls/1444/files                     -> [] (0 files)
GraphQL pullRequest(1444).changedFiles       -> 0

That is not a rendering glitch. The head tree and the base tree are identical:

GET /git/trees/1370426c16?recursive=1   -> 2272 blobs
GET /git/trees/9674cfafbc?recursive=1   -> 2272 blobs
  only-in-head 0 / only-in-base 0 / differing blob shas 0
GET /repos/.../compare/9674cfafbc...1370426c16 -> ahead 7, behind 0, files 0

1370426c16 is Merge remote-tracking branch 'origin/master' into pr1444-head (parents d9963dbc + 9674cfaf), i.e. the auto-fix lane's catch-up merged current master in, and because this PR's own content had already reached master through #6032 (e1e33a974c), the merge absorbed every remaining delta. Net result: merging this PR would write zero bytes to master.

Suggested disposition: close it (optionally with Part of #1295 — content landed via #6032), rather than merging it. Merging an empty PR would only add a merge commit with no content.

One caveat on permissions: this head lives in the base repo (isCrossRepository: false, author: ranxianglei), and my token reports viewerCanUpdate: false / viewerCanClose: false there — so the close has to come from you. I am flagging the state, not asking for a change of plan on #5128.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

@suyeku1 的空 PR 判定已核实属实;唯一红 lane 是 master 级测试问题,不是本 PR 的回归。

本次实跑验证(非转述)

  1. 空 PR — 确认。 git diff 9674cfafbc 1370426c16(master tip → PR head)零输出:head tree 与 base tree 逐字节相同。ahead 的 7 个提交全是 master 同步 merge + 已落地的 follow-up。合并此 PR 对 master 写入 0 字节。
  2. 内容已经 [PR] Rebase #1444 onto master: declarable model endpoints + commandcode wire family #6032 落地 — 确认。 git merge-base --is-ancestor e1e33a974c origin/master 成立;特征文件 src/commandcode-wire.ts、src/model-endpoints.ts、src/loop/adapter-commandcode.ts、docs/wire-contracts.md 均在 master 上。
  3. 红 lane 不是本 PR 回归。 test (windows-latest, 22, 5)(run 43194814188)失败于 tests/graceful-close.test.ts:25(#1982 graceful client-side close,step2 backstop):timeout waiting for: linger backstop warn(waitFor 预算 3000ms,duration_ms 10395)。测试要求客户端持 socket 不发 FIN,等服务端 400ms linger 预算到期打出 backstop warn——慢速 Windows runner 上 3s 窗口没等到。因 head tree 与 master 逐字节相同,跑的就是 master 的代码;该测试文件最后改动 2026-10-29(Host-launched proxy can wedge on Windows when the parent stops draining stderr (pipe backpressure unguarded) #5012),早于本 PR base,且同文件 RESPONSE_BUDGET_MS 注释自述「was 8_000 — floored at the observed 11.6s flake window」——有已知 flake 史。这是 master 级测试问题(flakes 或预算偏紧),归 CI Control Center: Centralized CI Failure Handling & Status Sync (Comment all CI issues here; do not open new issues) | CI 总控: CI 失败集中处理与状态同步(所有 CI 问题在此评论, 勿开新 issue) #4599 哨兵 triage,不在本空分支修。

处置(需要你一步操作)

结构化八问

  1. 是什么问题:PR 已空(head tree == master tree),内容经 [PR] Rebase #1444 onto master: declarable model endpoints + commandcode wire family #6032 已在 master;唯一剩余阻塞是一个非本 PR 回归的红测试 lane。
  2. 什么原因导致的:空——自动修复 lane 的 catch-up merge 吸收了当前 master,而内容已经 [PR] Rebase #1444 onto master: declarable model endpoints + commandcode wire family #6032 落地,merge 后无剩余 delta。红 lane——timing 敏感测试(400ms linger 预算 / 3s 等待)在慢速 Windows runner 上超时,跑的是 master 代码。
  3. 影响范围与严重程度:无用户面影响(内容已在 master 正常发布)。PR 本身仅是「无法有意义地合并」;红 lane 是 CI 信号噪音(p3)。
  4. 怎么修复的:本分支无东西可修(无独有内容)→ 关 PR;graceful-close lane 走 CI Control Center: Centralized CI Failure Handling & Status Sync (Comment all CI issues here; do not open new issues) | CI 总控: CI 失败集中处理与状态同步(所有 CI 问题在此评论, 勿开新 issue) #4599 哨兵 triage。
  5. 是否有回归问题:无——PR 贡献 0 字节,不可能引入回归;红 lane 跑的是 master 代码。
  6. 是否破坏缓存:不涉及——无代码变更。
  7. 是否改了设计方向:不涉及。
  8. 是否引入了新的配置:不涉及。

收工判定:空 PR 已核实,建议 owner 关闭 #1444;红 lane 定性为 master 级测试问题,已在 #4599 留痕,本分支无需进一步动作。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[ework] auto-fix stalled: test (windows-latest, 22, 5) — stopping for human review

@ranxianglei

Copy link
Copy Markdown
Owner Author

test-connection

@ranxianglei

Copy link
Copy Markdown
Owner Author

Closing: the branch is fully in sync with master — the three-dot diff origin/master...head is empty, so merging would apply zero changes. The feature is already on master (src/commandcode-wire.ts, src/loop/adapter-commandcode.ts; registered in the source map by c7ac7b5).

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[Feature] Declarable custom-wire model endpoints — route providers like commandcode's Go plan (/alpha/generate) into compression

2 participants