<?xml version="1.0" encoding="utf-8"?>
<entry xmlns="http://www.w3.org/2005/Atom" xml:base="https://docs.spec-kitty.ai/adr/3.x/2026-07-17-1-red-main-is-honest-ci-is-release-authority/">
  <id>https://docs.spec-kitty.ai/adr/3.x/2026-07-17-1-red-main-is-honest-ci-is-release-authority/</id>
  <title>Red main is honest signal; CI status is the release authority</title>
  <updated>2026-11-18T00:00:00.000Z</updated>
  <author><name>Spec Kitty</name></author>
  <link rel="alternate" type="text/html" href="https://docs.spec-kitty.ai/adr/3.x/2026-07-17-1-red-main-is-honest-ci-is-release-authority/"/>
  <summary>A red mainline is honest signal and P0 reproductions may land red-first, but mainline CI status is the release authority: nothing ships on red, with no exceptions.</summary>
  <content type="html">&lt;span id=&quot;red-main-is-honest-signal-ci-status-is-the-release-authority&quot;&gt;&lt;/span&gt;
&lt;p&gt;&lt;strong&gt;Status:&lt;/strong&gt; Accepted (by consensus) · &lt;strong&gt;Deciders:&lt;/strong&gt; Robert Douglass (CEO / product owner), Kent (QA lead), maintainers · &lt;strong&gt;Date:&lt;/strong&gt; 2026-07-17 · &lt;strong&gt;Related:&lt;/strong&gt; &lt;a href=&quot;/development/reference/red-main-and-release-readiness/&quot;&gt;&lt;code dir=&quot;auto&quot;&gt;docs/development/red-main-and-release-readiness.md&lt;/code&gt;&lt;/a&gt;, &lt;a href=&quot;/development/testing/testing-flakiness/&quot;&gt;&lt;code dir=&quot;auto&quot;&gt;docs/development/testing-flakiness.md&lt;/code&gt;&lt;/a&gt;, &lt;a href=&quot;/development/how-to/pr-landing/&quot;&gt;&lt;code dir=&quot;auto&quot;&gt;docs/development/pr-landing.md&lt;/code&gt;&lt;/a&gt;, the charter’s red-first / test-remediation standing order.&lt;/p&gt;
&lt;h2 id=&quot;context-and-problem-statement&quot;&gt;Context and Problem Statement&lt;/h2&gt;
&lt;p&gt;Our &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; branch runs a large CI suite whose result the team reads as the truth about the product. Two pressures pull against each other:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;The convenience pull&lt;/strong&gt; — a green &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; feels good, so there is a standing temptation to keep it green by reverting anything that goes red, or by declining to merge a test that reproduces a known bug. That produces a &lt;em&gt;green-washed&lt;/em&gt; mainline: green because the failures are hidden, not because the product works.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The honesty pull&lt;/strong&gt; — the whole point of CI is to be a trustworthy signal. A P0 defect that exists in the product but is invisible on &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; is a lie the whole team then builds on: releases ship on a false green, expensive QA runs validate a base that is secretly broken, and the same bug gets rediscovered downstream.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The precipitating events: #2752 was merged with known reds (a CI-guard change whose companion arch-gate fix, #2753, landed right after), and #2754 (a pre-existing &lt;code dir=&quot;auto&quot;&gt;setup-plan&lt;/code&gt; error-translation regression) was filed &lt;em&gt;because&lt;/em&gt; it was surfaced honestly during a landing pass rather than papered over. Both cases raised the same question — &lt;strong&gt;is a red &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; a process failure to be hidden, or a true signal to be honoured?&lt;/strong&gt; — and the team needed a single, written answer.&lt;/p&gt;
&lt;h2 id=&quot;decision-drivers&quot;&gt;Decision Drivers&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Honest signal over convenience.&lt;/strong&gt; A false green is worse than an honest red: it moves the cost of a defect downstream and destroys trust in CI.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;CI must mean one thing.&lt;/strong&gt; If CI status is sometimes decorative (kept green by hiding failures) and sometimes authoritative, it is authoritative for nothing.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Protect expensive verification.&lt;/strong&gt; Manual QA and long/expensive automated runs are wasted if they execute against a base that is already known-broken.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Reproduce, don’t just assert.&lt;/strong&gt; A P0 that carries a failing test is unambiguous, self-documenting, and un-loseable; a P0 described only in prose drifts and gets forgotten.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Recovery is the priority.&lt;/strong&gt; Red &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; is not acceptable as a resting state; it is an alarm that pulls maintainer attention to the front of the queue.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;considered-options&quot;&gt;Considered Options&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Option A — Always-green mainline (green-washing).&lt;/strong&gt; Keep &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; green at all times: revert anything that reds it, and refuse to merge failing reproduction tests. &lt;em&gt;Rejected&lt;/em&gt; — it hides P0s, lets releases ship on a false green, wastes QA on a secretly-broken base, and contradicts the transparency value.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Option B — Honest mainline; CI is the release authority (chosen).&lt;/strong&gt; &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; is allowed to be red when it carries accepted release-blocking (P0) defects; CI status is the single authoritative release gate; failing reproduction tests for P0s are encouraged; expensive QA is gated on green; maintainers prioritize red recovery.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Option C — Separate always-green release branch, &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; allowed red.&lt;/strong&gt; Maintain a parallel release line that is kept green while &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; may be red. &lt;em&gt;Rejected&lt;/em&gt; — it splits the source of truth into two CI signals, adds branch-management overhead, and dilutes exactly the “CI means one thing” property Option B is designed to protect.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;decision-outcome&quot;&gt;Decision Outcome&lt;/h2&gt;
&lt;p&gt;&lt;strong&gt;Chosen option: “Option B — Honest mainline; CI is the release authority.”&lt;/strong&gt; The five binding rules:&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;&lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; may be red.&lt;/strong&gt; A red &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; is legitimate and expected when it reflects known release-blocking (P0) defects. It is a &lt;em&gt;true signal&lt;/em&gt;, not a workflow violation. (This is orthogonal to the no-direct-push / Protect-Main policy, which is unchanged — see &lt;a href=&quot;#consequences&quot;&gt;Consequences&lt;/a&gt;.)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Mainline CI status is authoritative for release readiness.&lt;/strong&gt; &lt;strong&gt;Red CI means no release.&lt;/strong&gt; No release proceeds on a red mainline, full stop. There is no “known-red, ship anyway” path.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;P0 filing/acceptance may (and should) add a failing test.&lt;/strong&gt; When a P0 bug is filed or accepted, the reporter/accepter is free and &lt;strong&gt;encouraged&lt;/strong&gt; to land a test that reproduces the defect — a red-first reproduction at mainline scope. Turning &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; red this way is honouring the process, not breaking it.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Expensive QA and internal manual testing do not run on a red mainline.&lt;/strong&gt; There is no value in validating a base that is already known-broken; QA effort waits for green.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Maintainers prioritize red builds.&lt;/strong&gt; Green &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; is the goal; a red &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; is not an acceptable resting state and pulls to the front of the maintainer queue. We favour transparency and honest reporting over convenience: &lt;em&gt;it’s broken, we hate it, but it’s the truth, so we do not hide it.&lt;/em&gt;&lt;/li&gt;
&lt;/ol&gt;
&lt;h3 id=&quot;consequences&quot;&gt;Consequences&lt;/h3&gt;
&lt;h4 id=&quot;positive&quot;&gt;Positive&lt;/h4&gt;
&lt;ul&gt;
&lt;li&gt;Release readiness has a single, trustworthy gate — CI — that cannot be green-washed.&lt;/li&gt;
&lt;li&gt;P0 defects are visible on &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; and, when reproduced by a test, self-documenting and un-loseable.&lt;/li&gt;
&lt;li&gt;Expensive QA and manual-test effort is never spent validating a known-broken base.&lt;/li&gt;
&lt;li&gt;The policy reinforces the existing red-first / never-retry-to-green discipline (see &lt;a href=&quot;/development/testing/testing-flakiness/&quot;&gt;testing-flakiness&lt;/a&gt; and the landing-pass bin classification in &lt;a href=&quot;/development/how-to/pr-landing/&quot;&gt;pr-landing&lt;/a&gt;) by extending it from the PR scope to the mainline scope.&lt;/li&gt;
&lt;/ul&gt;
&lt;h4 id=&quot;negative&quot;&gt;Negative&lt;/h4&gt;
&lt;ul&gt;
&lt;li&gt;A red &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; is psychologically uncomfortable and can alarm contributors who read it as “the project is broken” rather than “a known P0 is being worked.” Mitigation: the accompanying guide states the meaning plainly, and P0s that red &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; carry a filed issue.&lt;/li&gt;
&lt;li&gt;The policy depends on maintainer discipline to actually prioritize recovery; without it, “red is allowed” could erode into “red is tolerated.” The “not an acceptable resting state” rule and the QA-gating consequence are the counter-pressures.&lt;/li&gt;
&lt;/ul&gt;
&lt;h4 id=&quot;neutral&quot;&gt;Neutral&lt;/h4&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;No-direct-push is unchanged.&lt;/strong&gt; Changes still reach &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; only through pull requests and the Protect-Main workflow; this ADR governs what a red CI &lt;em&gt;means&lt;/em&gt;, not how commits arrive.&lt;/li&gt;
&lt;li&gt;A P0 reproduction test that reds &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; is expected to be tracked by its P0 issue, so the red is always traceable to an accepted defect.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;confirmation&quot;&gt;Confirmation&lt;/h3&gt;
&lt;p&gt;The decision is confirmed in practice by two observable invariants: (a) no release is cut while mainline CI is red, and (b) an accepted P0 either carries a failing reproduction test or a filed issue (ideally both). A release cut on red, or an accepted P0 silently reverted to keep &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; green, is a violation of this ADR. The operational runbook lives in &lt;a href=&quot;/development/reference/red-main-and-release-readiness/&quot;&gt;&lt;code dir=&quot;auto&quot;&gt;docs/development/red-main-and-release-readiness.md&lt;/code&gt;&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id=&quot;amendment-2026-10-04--the-red-lives-in-the-nightly-never-in-a-pr&quot;&gt;Amendment 2026-10-04 — the red lives in the nightly, never in a PR&lt;/h2&gt;
&lt;p&gt;Rule 3’s “red-first reproduction at mainline scope” now has one fixed home. A reproduction of an open P0 carries &lt;code dir=&quot;auto&quot;&gt;@pytest.mark.p0_repro(issue=N)&lt;/code&gt; and runs &lt;strong&gt;only&lt;/strong&gt; in the &lt;code dir=&quot;auto&quot;&gt;p0-repro&lt;/code&gt; job of &lt;code dir=&quot;auto&quot;&gt;.github/workflows/ci-nightly.yml&lt;/code&gt;. &lt;code dir=&quot;auto&quot;&gt;tests/_support/p0_repro.py&lt;/code&gt; deselects it from every other run: every per-PR lane, &lt;code dir=&quot;auto&quot;&gt;make test-fast&lt;/code&gt; / &lt;code dir=&quot;auto&quot;&gt;make test-full&lt;/code&gt;, and every other nightly lane. So a pull request always lands on a green per-PR gate, while an open P0 keeps the nightly honestly red, and each failure carries an &lt;code dir=&quot;auto&quot;&gt;[OPEN P0 #N]&lt;/code&gt; banner naming its issue.&lt;/p&gt;
&lt;p&gt;Rules 1, 2 and 5 are unchanged in substance. “Mainline CI” includes the nightly, and a release requires a green nightly run for the exact release SHA (&lt;code dir=&quot;auto&quot;&gt;scripts/ci/release_nightly_gate.py&lt;/code&gt;), so an open P0 reproduction still blocks every release. The only bypass is the &lt;code dir=&quot;auto&quot;&gt;skip_nightly&lt;/code&gt; maintainer waiver on a manually dispatched &lt;code dir=&quot;auto&quot;&gt;.github/workflows/release.yml&lt;/code&gt; run (never on a tag push), which skips the nightly gate and says so in the run summary. The &lt;code dir=&quot;auto&quot;&gt;regression&lt;/code&gt; marker no longer carries P0 reproductions. It stays the marker for issue-pinned guards of bugs that are already fixed, and those guards run per PR. The fix PR for a P0 removes the &lt;code dir=&quot;auto&quot;&gt;p0_repro&lt;/code&gt; marker in the same change, which turns the reproduction into such a guard.&lt;/p&gt;
&lt;h2 id=&quot;amendment-2026-11-18--the-performance-verdict-is-weekly-7704&quot;&gt;Amendment 2026-11-18 — the performance verdict is weekly (#7704)&lt;/h2&gt;
&lt;p&gt;Operator ruling on #7704. The &lt;code dir=&quot;auto&quot;&gt;performance&lt;/code&gt; job of &lt;code dir=&quot;auto&quot;&gt;.github/workflows/ci-nightly.yml&lt;/code&gt; now runs weekly, on the Sunday run (cron &lt;code dir=&quot;auto&quot;&gt;17 3 * * 0&lt;/code&gt;), or on any &lt;code dir=&quot;auto&quot;&gt;workflow_dispatch&lt;/code&gt;. Every other nightly job still runs every day. A green nightly on the release SHA therefore no longer proves that the performance suite ran, so the release gate (&lt;code dir=&quot;auto&quot;&gt;scripts/ci/release_nightly_gate.py&lt;/code&gt;) has two conditions:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Exact SHA, unchanged.&lt;/strong&gt; A green &lt;code dir=&quot;auto&quot;&gt;ci-nightly&lt;/code&gt; run for the exact release SHA.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Performance verdict, new.&lt;/strong&gt; In the newest completed &lt;code dir=&quot;auto&quot;&gt;ci-nightly&lt;/code&gt; run created in the last 7 days that executed the &lt;code dir=&quot;auto&quot;&gt;performance&lt;/code&gt; job, that job must conclude &lt;code dir=&quot;auto&quot;&gt;success&lt;/code&gt;. The performance job alone decides: a red stress or e2e job in the same run does not make the verdict red (the exact-SHA condition still judges the whole run). It must be on the release SHA, or on a &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; commit the release SHA descends from. A skipped job is no verdict, and neither is a covering run still in flight. A newer red verdict outvotes an older green one, so a release cannot publish past an open performance red. Without a green verdict, the gate’s &lt;code dir=&quot;auto&quot;&gt;--dispatch&lt;/code&gt; (which &lt;code dir=&quot;auto&quot;&gt;release.yml&lt;/code&gt; passes) runs the nightly by the release tag, and a dispatched run always executes the job.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Residual risk, accepted by the ruling.&lt;/strong&gt; A performance regression that lands between the weekly run and a release whose verdict comes from an ancestor commit ships unless the next weekly run catches it. The 7-day window bounds that exposure. Every other suite keeps the exact-SHA rule. Rules 1-5 are otherwise unchanged, and the &lt;code dir=&quot;auto&quot;&gt;skip_nightly&lt;/code&gt; waiver still skips the whole gate.&lt;/p&gt;</content>
</entry>
