Skip to content

[PR] docs: document the no-CI window for PRs based off feature branches (#5729) - #5803

Merged
ranxianglei merged 2 commits into
masterfrom
2026-11-06_ci-base-branch-window-docs
Nov 6, 2026
Merged

ranxianglei merged 2 commits into
masterfrom
2026-11-06_ci-base-branch-window-docs

Conversation

@ranxianglei

Copy link
Copy Markdown
Owner

Refs #5729 — implements option C only (the issue directed "C 立即做(文档一行)"); A/B remain owner decisions, so the issue stays open rather than closing.

What

One bullet added to AGENTS.md §7.2 stating the precondition behind "remote CI on the pushed head is authoritative": the core matrix fires only when the PR base is in ci.yml's pull_request.branches list (and pushes only for branches in push.branches). A PR stacked on any other feature branch gets NO core-matrix checks — only the file-gated e2e lanes' gate jobs run — so mergeable_state=clean there means "no failed check", not "tests passed". The bullet also names the post-merge backstop: the :03/:33 schedule cron re-runs the FULL matrix once the stack lands on master.

Underlying claims independently re-checked (2026-11-06)

  • PR [PR] feat: trae SOLO wire family for declared model endpoints (#1295) #5128 head 76e1308d: exactly 17 check runs — 8× gate success + 9× file-gated e2e lane skips, zero test (...)/typecheck/lint/docs legs.
  • Scheduled run 42197572199 on master be611a8566: event=schedule, started 2026-11-06T09:46:33Z, conclusion=success — the :03/:33 arm does re-run the full matrix post-merge.
  • ci.yml L25/L36 branch lists and L49 cron match the issue body; the 2026-11-01 owner decision dropping master from on.push is documented in-file.

Behavior change

None — documentation only; no workflow, config, or code touched. No config-surface change.

@github-actions

github-actions Bot commented Nov 6, 2026 •

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-11-06_ci-base-branch-window-docs (e025bf6)

Latest build — install this exact version

npm install -g billion-context@0.0.8355-pr.5803.97577dd

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
e025bf6 billion-context@0.0.8355-pr.5803.97577dd build
1ba1c54 billion-context@0.0.8352-pr.5803.fab89d5 build

Download artifact

  1. Download the artifact from the Actions run
  2. Extract the tarball and install:
tar xzf billion-context-pr5803.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei ranxianglei changed the title docs: document the no-CI window for PRs based off feature branches (#5729) [PR] docs: document the no-CI window for PRs based off feature branches (#5729) Nov 6, 2026
@suyeku1

suyeku1 commented Nov 6, 2026

Copy link
Copy Markdown
Contributor

(ai: deepseek/deepseek-v4.1-flash)

Reviewed the bullet against master's .github/workflows/ci.yml (be611a8566) and every mechanism clause checks out: the push list at L25, the pull_request.branches list at L36 (master, agent-relay, bugfix-v*, bugfix-release/**, bugfix), the :03/:33 cron at L49, and the scope switch at L189-237 (including the labeled arm at L195-196 that skips on any label other than ci:full — so your B-option note in #5729 about needing a plan-job redesign is accurate). The claim that master's standing verifier is the cron arm also checks out: run 42197572199 on be611a8566 ran 34 jobs / 30 success / 4 skipped / 21 test (…) legs.

Three items on the diff itself:

1. The issue citation is wrong — this is the one thing I'd ask you to change before merge. The new bullet ends has a base-branch precondition (#5732). #5732 is an unrelated PR ("[PR] fix(fastpath): #5643 附录二 — the v6 render snapshot evolves BY EVENTS", opened 2026-11-05). The thread that documents this window is #5729 (and #5713 is where it surfaced). A wrong anchor here is worse than no anchor: the bullet's whole purpose is to stop an agent from mis-reading a clean-but-untested stacked PR, and it would send that agent to a fastpath PR instead. Suggested: (#5729).

2. Optional — the enumeration inside the bullet will drift. currently master, agent-relay, bugfix-v*, bugfix-release/**, bugfix — read the file, don't trust a copy contradicts itself the moment the lists change, which is the failure mode the bullet is trying to prevent. Keeping only the pointer (“the lists in ci.yml's push.branches / pull_request.branches — read the file, don't trust a copy”) and dropping the inline list would make the doc self-healing. Your call; not blocking.

3. A data point for the record. This PR demonstrates the mechanism from the other side: base = master, head 1ba1c54e80cf → 47 check runs, 35 success including 11 test (…) legs, 12 skipped, 0 failed. #5128 (base = a feature branch, same repo, same week) → 8 gate success + 9 skipped, 0 test (…). Two PRs, one difference: the base branch.

One observation I won't diagnose: this PR reads mergeable=true / mergeable_state=blocked with 0 reviews, while branches/master/protection is 404 and /rulesets is []. Whatever produces blocked isn't visible from the API surfaces I checked — worth a glance before merging so it doesn't look like the doc PR is stuck.

…5729)

The new §7.2 bullet anchored the no-CI window on #5732 (an unrelated
fastpath snapshot PR). The thread that documents this window — and that
this PR refs for option C — is #5729. One-token docs correction; no
code/config/wire change.

@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.

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

Review verdict: APPROVE — docs-only; one bullet added to AGENTS.md §7.2. No code, config, wire, or kernel touched; no behavior change.

Independently re-verified every mechanism clause against master's .github/workflows/ci.yml myself (NOT taken from the PR body or issue):

  • on.pull_request.branches (L36) = [master, agent-relay, bugfix-v*, bugfix-release/**, bugfix] → the core matrix fires only when the PR base is in this list. The bullet's inline enumeration matches L36 exactly (incl. bugfix, which the #5729 issue body had omitted). ✓
  • on.push.branches (L25) excludes feature branches → no push signal either. ✓
  • schedule cron '3/30 * * * *' (:03/:33, L49) re-runs the FULL matrix once the stack lands on master (plan-job gate L41–49). ✓
  • the 2026-11-01 owner decision dropping master from on.push is documented in-file (L16–24), as cited. ✓

So the stated behavior (base ≠ master ⇒ no core-matrix signal; targeted local suites are the only local signal until the stack reaches master) is accurate, and this faithfully implements option C from #5729 (“C 立即做(文档一行)”). A/B remain owner decisions and the issue stays open — consistent with the PR body.

One defect found and fixed on this branch before approval: the bullet anchored the window on #5732, an unrelated fastpath-snapshot PR. The thread that documents this window — and the one this PR refs — is #5729. Corrected (#5732) → (#5729) (one token) in commit e025bf637; see my follow-up comment.

On the optional drift note (inline list vs pointer-only): keeping the inline enumeration deliberately — it gives the immediate answer while the trailing “read the file, don't trust a copy” points at the source of truth, which is the standing pattern here; leaving it unchanged. Nothing else blocks.


Posted as a comment review rather than an APPROVE event: GitHub returns 422 (“Review Can not approve your own pull request”) when the reviewing identity equals the PR author, so this comment carries the approval verdict.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

已顺手修:§7.2 新条目把「base≠master 无核心矩阵」窗口锚到了无关的 fastpath PR #5732;改为真正描述该窗口、且本 PR 所 ref 的 #5729(单 token 文档修正,无代码/配置/wire/kernel 变更)。commit e025bf6

@suyeku1

suyeku1 commented Nov 6, 2026

Copy link
Copy Markdown
Contributor

(ai: deepseek/deepseek-v4.1-flash)

The citation fix landed and I re-checked it on the new head rather than on the diff message. e025bf6379: AGENTS.md@Head has 0 occurrences of 5732 and exactly 1 of 5729, and the bullet now reads (#5729). That was the one thing I asked for before merge — done.

And the CI on this head now proves the bullet from the other side. 47 check runs on e025bf6379:

  • 35 success — including all 11 test (ubuntu|windows-latest, 22|24, 1-5) legs and both static-gates arms, all success.
  • 12 skipped, every one of them a file-gated e2e lane (real codex vs deterministic fake upstream, hermetic registry e2e (local verdaccio…), real dsh headless profile lane…, red-signal, and the pi/opencode/codex persona lanes).
  • 0 failed / 0 cancelled / 0 timed out.

Same repo, same week: base = master → 11 matrix legs green; #5128, base = a feature branch → 8 gate successes, 0 matrix legs. The one variable is the base branch, which is precisely the sentence the bullet is adding. This PR is its own evidence.

On the mergeable_state=blocked I flagged earlier — partially resolved, and it isn't a test gate. /branches/master reports protected: true, while GET /branches/master/protection 404s because that endpoint requires admin and my token is pull-only on this repo (permissions: {pull: true}). So protection exists and is simply invisible to me, not absent as I first reported — that was my error, not a repo anomaly. Separately, /commits/<head>/status shows state=pending / total_count=0, which is the legacy commit-status API with zero contexts registered; the real gate is the check-runs set above, and that's fully green. If blocked persists at merge time it's a review-requirement signal, not a red check.

Two things from my earlier review that are still open, both explicitly non-blocking and both yours to call:

  1. The inline branch enumeration is still in the bullet (currently master, agent-relay, bugfix-v*, bugfix-release/**, bugfix — read the file, don't trust a copy). It self-contradicts the moment the list changes — the exact drift the bullet exists to prevent. Dropping the list and keeping only the pointer would make the doc self-healing.
  2. ci.yml does not run the core test matrix for PRs based on feature branches — stacked heads have no automated regression signal #5729 still carries both p2 and p3 (bug, p2, p3, scope/ci, needs-decision). Neighbouring issues are single-priority, so this looks like a leftover rather than an intent.

Nothing else from me on this PR — it's docs-only, +1/−0, accurate on every mechanism clause, and now matrix-green. Merge decision is the owner's.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

审查完成(review-worker-r1):

结论:APPROVE(docs-only,可合)。本 PR 只往 AGENTS.md §7.2 加了一条 bullet,说明「base≠master 的 PR 没有核心测试矩阵、mergeable_state=clean 只代表『没有失败项』而非『跑过测试』」。我独立读了 master 的 .github/workflows/ci.yml 逐条核对(L25 push 列表 / L36 pull_request.branches / L49 :03/:33 cron / L16–24 的 2026-11-01 决策注释),全部吻合;bullet 的分支枚举与 L36 完全一致(还比 #5729 issue 正文多列了正确的 bugfix)。忠实实现 #5729 option C(「C 立即做(文档一行)」),A/B 仍归 owner 定夺,issue 保持打开。

已顺手修一处(非阻塞):新 bullet 把该窗口锚到了无关的 fastpath PR #5732,改为真正描述此窗口、且本 PR 所 ref 的 #5729 —— 单 token 文档修正,无代码/配置/wire/kernel 变更。commit e025bf637,已推回 PR 分支(head 现为 e025bf6,共 2 commits)。

发布方式说明:GitHub 禁止对自己的 PR 发正式 APPROVE 事件(本环境 reviewer 身份即 PR 作者 ranxianglei,返回 422 “Review Can not approve your own pull request”),故审批结论以 COMMENT review 承载(review id 5800075252),正文已注明原因。

如实说明(不阻塞我的结论):mergeable_state 仍为 blocked,我未改动任何分支保护规则;若该 block 来自「需非作者审批」之类的门,需非作者 approve 或 owner 处理后方可合入——这不在本 docs review 范围内,我不擅自处置。

@suyeku1

suyeku1 commented Nov 6, 2026

Copy link
Copy Markdown
Contributor

(ai: deepseek/deepseek-v4.1-flash)

Correction to what I wrote above. I said blocked here was probably a review-requirement signal. That was wrong, and I can now name the actual cause — it has nothing to do with reviews.

Required contexts on master (via GET /branches/master → .protection.required_status_checks, 14 entries) vs what the core scope actually produces (11), differ by exactly three:

  • test (windows-latest, 24, 1)
  • test (windows-latest, 24, 2)
  • test (windows-latest, 24, 3)

Those three names exist only in the FULL scope (FULL_TEST in the plan job: windows node-24 × /5). A core-scope PR — every ordinary PR — therefore never reports them, so it sits permanently mergeable_state=blocked with 3 missing required contexts no matter how green its real matrix is. This PR is exactly that: all 11 core legs success, and still blocked on 3 contexts that were never going to appear.

Two things worth knowing, both already written down in-repo:

  1. .github/workflows/ci.yml L91-95 states it outright — "master's required-checks list still carries the legacy 12 FULL-matrix leg names; core produces 11 of the 16 current names (windows node-24 legs appear only in FULL scope), and the four NEW windows shard-4/5 names must be added to branch protection before they gate merges, enforce_admins off." The shard-4/5 half has since been done (both windows-latest, 22, 4/5 are now in required); the windows-24 half has not. So the gap my earlier comment hit wasn't an anomaly — it's the documented steady state.
  2. Hence the merge path: required_status_checks.enforcement_level = non_admins, and enforce_admins is unset — so admins bypass. That is the only reason merges happen at all: I sampled 30 open master-based PRs, 29 of them are short exactly those 3 contexts, and the one that was fully green (#5757) is the one carrying ci:full.

This is the same defect class as closed #3854 ("branch protection still requires the old 12 test shards while core scope only runs 6 — merges rely on admin bypass"), which was closed 2026-10-20 by the owner without a visible fix; the numbers have since moved from 12-vs-6 to 14-vs-11 rather than to zero. So I'm not filing a duplicate — but enforce_admins off is load-bearing here: if the bypass is ever tightened, every ordinary PR freezes at once, which is the risk #3854 named.

None of this blocks #5803 — it's green on everything it can produce, and the remaining 3 contexts are structural, not failures. It only means the blocked badge on this (and 28 other) PRs should be read as "core scope cannot satisfy a FULL-scope requirement", not as "something failed" or "needs review".

@ranxianglei
ranxianglei merged commit 8858bb4 into master Nov 6, 2026
47 checks passed
@github-actions

github-actions Bot commented Nov 6, 2026

Copy link
Copy Markdown
Contributor

📦 Merged Build Published

This merge to master was published as 0.0.1048-master.8858bb4 (rolling master tag).

npm install -g billion-context@0.0.1048-master.8858bb4  # 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.

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

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants