Repository navigation
[SCR-239] Check the issue fields - #241
Conversation
The checker validated labels and ignored every other field GitHub carries. Measured on content-rabbit: zero milestones exist, zero of 40 sampled issues carry a milestone or assignee, zero sub-issues, zero dependencies. Extend the prose-relationship rule to Depends on, Blocks and Related to. Ten phase epics say Depends on Phase 10 and passed the old rule while carrying no real dependency, which is the exact failure the rule exists to stop. A sentence merely containing the word depends is not a finding: the rule wants a line-leading label and a reference. Read the allowed project and milestone sets from the repo .agents/issues.md, beside the high-stakes paths, because a project name is repo-specific and cannot be derived from outside the repo. A field with no set defined is not checked. Zero milestones exist across the fleet, so a rule that always required one would fail every issue on day one and get the standard switched off. requireProof in pr-standards already works this way. An assignee is never checked. An agent does not decide who does the work.
|
Warning Review limit reachedNext included review available in 16 minutes. View limit detailsLimit details: You’ve used the included review currently available. Your 142 included PR review attempts over the past 7 days set your current allowance at 1 review per hour. Your organization has reached its usage spending cap. Adjust your spending cap in the billing tab. Review configuration: ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Team Run ID: 📒 Files selected for processing (2)
Comment |
There was a problem hiding this comment.
Reviewed head: fe55a8d7093a7cd38a600c00cd944070f2ca9fb8
I read pr.diff, council-findings.md, the PR body/verification, AGENTS.md, and traced every flagged line against the actual code and the repo's .agents/ config. No confirmed bugs survive verification, so I pushed nothing.
Council claims, checked against the code
"An issue with two projects passes when both are allowed" (correctness + scope lenses)
Rejected. This assumes an undocumented "at most one project" rule. I grepped issue-standards.md, .agents/*.md, and the PR body for any such constraint — none exists. GitHub issues can legitimately belong to multiple Projects v2 boards, and validateFields (issue-standards.mjs:290) is written to validate membership, not cardinality. No concrete requirement is violated.
"An issue with no project/milestone silently passes when a set is configured" (correctness + scope lenses)
Rejected as intentional, not a defect. The code comment directly above parseAllowedSets (issue-standards.mjs:254-257) and the PR body both state the design explicitly: "A field with no set defined is not checked... a rule that always required one would fail every issue on day one." The PR further states the check is deliberately biased against false positives: "rejecting honest work is the direction this check must not be wrong in." Requiring presence once a set is configured is a plausible future rule, but it isn't the rule this PR implements or claims to implement, and it isn't stated anywhere in the repo.
"gh issue view ApiError (e.g. missing read:project) is swallowed into [], letting a disallowed project pass" (correctness + scope lenses)
Rejected as intentional. fetchIssueProjects (issue-standards.mjs:575) has an inline comment saying exactly this: "Token may lack read:project; then this is empty rather than a check failure." The PR body confirms the token here lacks that scope today. This is the same fail-open convention fetchRepoFile already uses for a 404. Worth knowing about if Allowed projects is ever configured with a token that still lacks the scope, but it's a disclosed tradeoff, not a diff-introduced bug.
Gemini (performance) and Grok (maintainability) timed out — no claims to check. Kimi (security) found nothing.
My own pass
- Traced the merge-conflict story in the PR body: confirmed there's exactly one
fetchRepoFiledefinition now (issue-standards.mjs:557), the duplicate is gone, and both call sites (runCheck,runLint) use it. - Checked
NATIVE_LINK_REagainst every case in "How I verified" by hand (Depends on #12,Depends on Phase 10.,Blocks #9, and the negative cases) — all match the intended line-leading-label semantics, no false positives introduced. - PR-standard lens: branch
scr-239-check-the-issue-fields, title[SCR-239] Check the issue fields, body has exactly oneCloses #239,## What/## Why/## How I verified, ends inAssisted-by:. 2 files / ~196 counted lines — well under the 40-file/500-line caps. - Proof: the "How I verified" section runs the full test suite (73/73) and selfcheck, and hand-exercises the regex/validator against the exact strings named in the PR body. I could not re-run
node issue-standards.test.mjsmyself in this sandbox (command execution was blocked), so I verified by static trace instead. One gap worth naming: the evidence exercisesvalidateFields/parseAllowedSetsdirectly but doesn't demonstrate the livegh issue view --json projectItemswiring end-to-end — reasonable, since the PR discloses the token can't currently read project data, but a human should keep that in mind before configuringAllowed projects.
No fixes pushed — nothing survived verification as a real defect.
🧑⚖️ Council
| Model | Lens | Result |
|---|---|---|
| GPT-5.6 (Codex) | correctness | rejected |
| Gemini 3 Pro | performance | no findings (timed out) |
| Kimi K3 | security | no findings |
| Grok 4.5 | maintainability | no findings (timed out) |
| GPT-5.6 (scope) | scope | rejected |
Closes #239
What
The checker now validates the issue fields it was ignoring: project, milestone, and relationships written as prose. Labels were the only field it looked at.
Why
GitHub carries five settable fields on an issue and the standard specified one of them. Measured on content-rabbit:
The phase epics are the sharpest case. They say
Depends on Phase 10.in prose while carrying zero real dependencies, and the old rule passed them because it matched onlyParent:andBlocked by:. A dependency written in prose is invisible to every query, filter and board, so nothing shows the order the phases must run in.A field with no set defined is not checked. Zero milestones exist across the fleet, so a rule that always required one would fail every issue on day one and get the standard switched off.
requireProofinpr-standardsalready works this way, and this follows it deliberately.The allowed sets live in the repo's
.agents/issues.md, beside the high-stakes paths. A project name is repo-specific and cannot be derived from outside the repo. The token here also lacksread:project, so it could not be read from the API even if that were wanted.An assignee is never checked. An agent does not decide who does the work.
How I verified
The rules, exercised directly rather than trusted:
The two false-positive cases matter most. The rule targets a line-leading label followed by a reference, not any prose containing the word, because rejecting honest work is the direction this check must not be wrong in.
A merge conflict worth naming
This branched before #237 landed. Both that work and this one independently wrote a byte-identical
fetchRepoFilehelper, so the rebase produced a duplicate declaration that would not parse. The duplicate is removed and the exported one from #237 is used.That is the cost of two agents working the same file from the same base at the same time, and it was caught by
node --checkrather than by review.How it was built
Delegated to
cursor-agent. The prose-relationship extension,parseAllowedSets,validateFields, and the tests are its work, and it correctly implemented the "not checked when no set is defined" rule, which was the part most likely to be got backwards.Review found no defect in the delegated code. One apparent false positive turned out to be my own test string replacing the wrong character, not a bug in the rule.
Assisted-by: claude-personal-1:claude-opus-5