LearnNewsExamplesConnect
Frontmatter
assigneesneo-opus-vega
createdAtAug 15, 2026, 2:02 PM
updatedAtNov 8, 2026, 8:52 PM
githubUrlhttps://github.com/neomjs/neo/issues/17171
authorneo-opus-vega
commentsCount11
parentIssuenull
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]
closedAtNov 8, 2026, 8:33 PM

The repo-local lint workflows run on dev PRs; none is a required status context

Closed v13.4.0/archive-v13-4-0-chunk-2 bugaiarchitecturebuildai-generated
neo-opus-vega
neo-opus-vega commented on Aug 15, 2026, 2:02 PM

Current state — 2026-11-06

Re-measured from Emmy's census (comment 6676977305) and verified by the author against GET /repos/neomjs/neo/rules/branches/dev and git ls-tree on dev. The Problem below is the August state, kept as history.

  • Required set: 16 contexts: the ten baseline / … jobs, build-all, unit, components (1/3) to (3/3), and e2e-engine. The shared baseline is now a merge gate, so the "only integration-parity" premise is historical.
  • Repo-local lint workflows: three *-lint.yml files, none of them required:
    • guard-ci-parity-lint.yml: one lint job running the consolidated Engine guard family.
    • secrets-lint.yml: beside the required baseline / Secrets.
    • spec-retirement-lint.yml: its own lint job.
  • Name collision: all three workflows name their job lint, with no name: override (git show origin/dev:.github/workflows/<file>, 2026-11-06 21:5xZ). An enforcement design must give them distinct contexts before requiring any; requiring lint would be ambiguous.
  • Not merge enforcement: review-admission/mergeability is not in the required set, so its reviewer predicate does not gate merges by itself.
  • Baseline names may move: neomjs/neo-agent-skills#283 (merged) splits the producer into two jobs, but the Engine's caller still pins baseline v0.1.84 and emits the old ten names. A context-name migration follows the caller's adoption, not the producer's merge.
  • Stale headers: pr-baseline.yml and spec-retirement-lint.yml still carry the obsolete ruleset census; the source change reconciles them.

The enforcement decision stays the operator's (ruleset authority). The negative-control PR and the regression-monitoring ACs below are unchanged.

Decision — 2026-11-08 (operator, on the author's recommendation)

One aggregate context, served by a shared step. reusable-pr-baseline.yml in neo-agent-skills gains a step that runs the guard list each caller declares, so the rules stay in each repository and the mechanics are shared. The guards run pull-request code, so they run on the baseline's pull-request-code side: in the Skills materialized job, which already runs the caller's install, or in a sibling job beside it. The PR baseline job runs no pull-request code by design, because steps of one job share GITHUB_ENV, GITHUB_PATH and the filesystem (Grace, 2026-11-08). Either placement adds no slot per guard. The draft interface is one caller input naming the guard list (a JSON array of {name, run}) and one step per guard. It always reports exactly one context: baseline / guards here, Shared PR Baseline / guards in Brain.

The standalone guard workflows retire, and each dev ruleset gains that one context. Brain follows under its own ticket. Institution has no standalone lint workflows. This also settles the aggregate-required-contexts question #20328 was waiting on.

Rejected:

  • Per-lint contexts: a path-filtered required check that never runs blocks unrelated pull requests, and every new guard would need a ruleset edit.
  • A repo-local aggregate job: the same mechanism twice (Engine, Brain), and a runner slot each, on an org limited to 20.
  • Advisory by design: it leaves guards whose purpose is to block unenforced.

Sequence:

  1. A skills PR for the shared step, then its release.
  2. Here: the pin bump (this repo still pins v0.1.84), the guard list, retiring the standalone guard workflows, and the ruleset entry. That closes this ticket.
  3. Brain.

Context

Found by @neo-gpt during exact-head review of PR #17161, where I had claimed a new CI workflow closed a git push --no-verify hole. It does not, and the reason is not specific to that guard.

The Problem

Nineteen *-lint.yml workflows run on dev PRs. Zero of them are required status contexts.

ls .github/workflows/*-lint.yml | wc -l
→ 19

gh api repos/neomjs/neo/rules/branches/dev \
  --jq '[.[] | select(.type=="required_status_checks")] | .[].parameters.required_status_checks[]?.context'
→ integration-parity

integration-parity is the entire required set. The branch-protection endpoint 404s, so ruleset 19087298 is the whole story.

The consequence is mechanical: a failing lint moves a PR to mergeStateStatus: UNSTABLE, which means mergeable with failing non-required checks. It is a visible red mark, not a gate. Combined with git push --no-verify bypassing the local hooks, every lint in this repository is advisory at the moment of merge.

This is not a claim that lints are useless — they are loud, they are read, and authors fix them. It is a claim that the repo's own substrate describes them as gates when the ruleset does not back that. Sampled from the workflow headers currently on dev:

  • atomic-write-shape-lint.yml: "CI mirror of the .husky/pre-commit guard, so git commit --no-verify cannot bypass it."
  • The same sentence pattern recurs across the family.

Those comments are the enforcement claim, and the ruleset does not support any of them. An agent or human reading the header reasonably concludes the hole is closed.

Why this surfaced now

PR #17161 adds check-spec-retirement.mjs, whose entire purpose is catching a silent failure — a deleted unit spec, which no red build can report because the assertions leave with the file. I wired a CI mirror and wrote that it closed the --no-verify path. @neo-gpt checked the live ruleset instead of the workflow file and found the mirror is reachability, not enforcement.

That guard is a sharp instance because its defect class is silence: shipping it as advisory means a --no-verify push still deletes coverage silently, and the guard that exists to notice sits green-but-unconsulted beside it. But the gap is the family's, not that guard's.

The Architectural Reality

  • Ruleset 19087298 on dev; active rule types: code_scanning, copilot_code_review, deletion, non_fast_forward, required_status_checks. Only the last carries contexts: the 16 listed under Current state, none of them a repo-local lint.
  • The August reachability spec (lintWorkflowScanRootParity, scanned ⊆ watched) no longer exists on dev, and nothing asserts the enforcement half either.
  • buildScripts/util/check-guard-ci-parity-registry.json reasons about which lint-staged guards have CI mirrors (now one consolidated job in guard-ci-parity-lint.yml). That is again mirror existence, never required-context membership.
  • A reader of the ruleset does exist: get_conversation(projection: 'merge-readiness') returns the effective requiredSet and contextStates (Brain PullRequestService.mjs). It checks the membership and state of the required set, never whether every repo-local lint ought to be in it. That invariant is the gap.

The Fix (shape — the decision is the deliverable, not the code)

Three candidate shapes; this ticket exists to pick one, and picking is operator-owned because every option mutates repo settings.

  1. One aggregate required context. A single lint-gate job that fans out to the lint family and is the required context. One ruleset entry, no per-lint churn. Cost: a red aggregate does not name which lint failed without opening it.
  2. Per-lint required contexts. Precise, and it makes the ruleset a readable inventory. Cost: one entry per lint context (three today, once the two lint jobs carry distinct names), and every new lint needs a ruleset edit — which is exactly the drift lintWorkflowScanRootParity exists to prevent, now moved to a surface a spec cannot see.
  3. Fold the lint family into integration-parity's job. Retired 2026-11-06: integration-parity is no longer a required context, so folding into it enforces nothing. Folding into a required baseline / … job would instead change the shared producer (neomjs/neo-agent-skills), which is not this repository's decision.

Whichever is chosen, the mechanical follow-on is the same: a spec that reads the live ruleset and asserts the family's membership, so this gap cannot silently reopen. That is the piece an agent can build and the piece that makes the decision durable.

Acceptance Criteria

  • A decision is recorded (aggregate context / per-lint contexts), with the rejected alternatives and why. (Decision section, 2026-11-08)
  • The chosen enforcement exists on the live dev ruleset, verified by gh api repos/neomjs/neo/rules/branches/dev output pasted into the PR — not by workflow-file inspection. (#20598 comment 6726039256: 7 contexts, including baseline / guards)
  • A falsifier: a PR that violates one lint is demonstrated to be merge-INELIGIBLE, not merely UNSTABLE. (PR #20604 was BLOCKED, receipt 6726301145, which states the isolation bound)
  • A spec holds every CI-mirrored lint-staged guard to the guards list in pr-baseline.yml (baseline / guards, the one required home), failing when any other workflow mirrors one. Revised 2026-11-08, on Grace's proposal (6725099960): the merge-readiness reader lives in the Brain, which Engine CI no longer checks out (#20451), so the live ruleset is read once, for AC 2, and pasted into the PR.
  • Every workflow header claiming --no-verify closure either becomes true or is corrected. spec-retirement-lint.yml is already corrected in PR #17161 and is the model. (#20598 retires that workflow with its header, and pr-baseline.yml names the ruleset as the authority)
  • #17151's enforcement half is discharged by this ticket, and its PR's narrowed claim is reconciled. (The spec-retirement guard runs inside the required baseline / guards; its "visibility, not enforcement" header retired with the workflow)

Out of Scope

  • Changing what any individual lint checks.
  • The pre-commit / pre-push hook layer itself — hooks are bypassable by design; this is about the gate behind them.
  • integration-parity's own scope.

Avoided Traps

  • Fixing it for one guard. Adding only spec-retirement-lint as a required context would close my instance and leave eighteen open, while making the substrate more misleading — one honest header among nineteen false ones.
  • Treating the workflow file as the source of truth. That is the exact error this ticket documents. Every AC above demands live ruleset output, because a workflow header is a claim and the ruleset is the fact.
  • Assuming advisory means broken. Nineteen lints run and authors fix them; the loop works socially. The defect is the gap between that and what the substrate says, which is what a --no-verify push exploits without anyone lying.

Evidence class

L2 — live API reads against the deployed ruleset, reproducible by any seat with repo read access. The enforcement change itself needs repo-admin and is therefore operator-gated.

Related

#17151 (the guard whose PR surfaced this; its enforcement AC lands here) · PR #17161 · lintWorkflowScanRootParity.spec.mjs (the reachability half, already built) · guard-ci-parity-registry.json

Live latest-open sweep: checked the latest 20 open issues at 2026-08-15T12:01Z; no equivalent found — nearest is #17151 itself, which owns the guard rather than the gate. A2A in-flight claim sweep over the last 12 messages: no competing [lane-claim] on ruleset or CI-enforcement scope.

Origin Session ID: 5cd926fa-77e1-4309-8bbf-ca563ab07403

Retrieval Hint: query_raw_memories("lint workflows advisory not required status context ruleset dev") · falsification anchor: gh api repos/neomjs/neo/rules/branches/dev returns integration-parity as the sole required context.

tobiu referenced in commit acfb498 - "feat(build): a deleted unit spec must say where its coverage went (#17151) (#17161) on Aug 15, 2026, 2:57 PM
tobiu referenced in commit 319c477 - "feat(ai): derive an ai/scripts entrypoint's execution plane from what it reaches (#16929) (#17191) on Aug 16, 2026, 9:15 PM
dawesi referenced in commit efba884 - "feat(build): a deleted unit spec must say where its coverage went (#17151) (#17161) on Aug 25, 2026, 5:41 PM
dawesi referenced in commit 273cf1e - "feat(ai): derive an ai/scripts entrypoint's execution plane from what it reaches (#16929) (#17191) on Aug 25, 2026, 5:41 PM
tobiu referenced in commit b6758e9 - "ci(baseline): the repository's guards run as one required guards context (#17171) (#20598) on Nov 8, 2026, 8:33 PM
tobiu closed this issue on Nov 8, 2026, 8:33 PM