Extend the skills and gate their scripts

.claude/skills/codetrial-contribute/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: codetrial-contribute
-description: Help CodeTrial contributors turn observations into clear English GitHub issues, small contribution plans, and pull request descriptions, and see them through. Use to draft, improve or file a CodeTrial issue, answer what a maintainer asked on it (a revision, a log, an initial analysis, a retest), prepare a PR description, respond to review requests such as rebase, squash or commit message fixes, check which of your own issues and PRs are waiting on you, ask a focused maintainer question, or choose a bounded first contribution with an AI agent. The deliverable is copy-ready text; judging an existing backlog for duplicates is codetrial-triage. Opening a PR goes through gh-submit when it is installed; this skill covers it otherwise.
+description: Help CodeTrial contributors turn observations into clear English GitHub issues, small contribution plans, and pull request descriptions, and see them through. Use to draft, improve or file a CodeTrial issue, answer what a maintainer asked on it (a revision, a log, an initial analysis, a retest), prepare a PR description, retrieve and address PR review comments, respond to requests such as rebase, squash or commit message fixes, check which of your own issues and PRs are waiting on you, ask a focused maintainer question, or choose a bounded first contribution with an AI agent. The deliverable is copy-ready text, or verified fixes when asked to address review feedback; judging an existing backlog for duplicates is codetrial-triage. Opening a PR goes through gh-submit when it is installed; this skill covers it otherwise.
 ---
 
 # Contribute to CodeTrial
@@ -235,6 +235,10 @@ everyone's list.
 
 ## Answer a review
 
+Fetching every thread, including long ones, and rechecking the head after a
+push, and where failed checks are diagnosed, are in
+[review-feedback.md](review-feedback.md).
+
 Git and commit-message requests recur alongside substantive code review. Read
 the current inline threads, unanswered cubic findings and failing checks as
 well as the review summary. The one-line requests are decoded in

.claude/skills/codetrial-contribute/review-feedback.md

@@ -0,0 +1,56 @@
+# Retrieve and close out review feedback
+
+What a request means, how to answer it and the local list of open requests are
+in [Answer a review](SKILL.md#answer-a-review); this file only covers fetching
+every thread and confirming the result. A request only to review or summarize a
+PR stays read-only. When the user names a subset of comments, address that
+subset; "address all the feedback" is the whole list, not a prompt to ask which.
+
+Run `gh auth status` first, so an access failure is reported as unreviewed
+feedback rather than as a PR with none. Top-level comments and reviews come from
+`gh pr view`; inline threads need their own query, with the ids and URLs that
+replying and resolving need:
+
+```sh
+gh pr view "$pr" --repo sysprog21/codetrial \
+  --json number,url,headRefOid,comments,reviews
+gh api graphql --paginate -F n="$pr" -f query='query($n: Int!, $endCursor: String) {
+  repository(owner: "sysprog21", name: "codetrial") { pullRequest(number: $n) {
+    reviewThreads(first: 100, after: $endCursor) {
+      pageInfo { hasNextPage endCursor }
+      nodes { id isResolved isOutdated path line originalLine
+        comments(first: 100) {
+          pageInfo { hasNextPage }
+          nodes { id author { login } body url } } } } } } }'
+```
+
+`--paginate` follows only the outer `pageInfo` and prints one JSON document per
+page. A thread whose `comments.pageInfo.hasNextPage` is true is incomplete;
+refetch all of its comments, where the thread's own `pageInfo` is the only one,
+before reading its last reply:
+
+```sh
+gh api graphql --paginate -F thread="$thread" -f query='query($thread: ID!, $endCursor: String) {
+  node(id: $thread) { ... on PullRequestReviewThread {
+    comments(first: 100, after: $endCursor) {
+      pageInfo { hasNextPage endCursor }
+      nodes { id author { login } body url } } } } }'
+```
+An outdated position does not mean fixed; compare the request with the current
+code. Verify a bot finding by the same standard as any other
+([collaboration.md](collaboration.md#review-evidence-including-bot-findings)).
+
+Fix each request with the smallest change and run the affected check, then the
+full gate with its skips read; browser behavior goes through
+[Browser investigation](../codetrial-verify/SKILL.md#browser-investigation)
+and red checks through
+[Failed GitHub Actions checks](../codetrial-verify/SKILL.md#failed-github-actions-checks).
+If the scope changed, update the PR description to describe the final result.
+
+A push does not prove every request addressed. After it, rerun both queries
+above and the checks, so a new top-level review or comment is read as well as
+a new thread, and reassess anything a new head or new comment touches.
+Replies and resolving follow
+[codetrial-conventions](../codetrial-conventions/SKILL.md#pull-request-branches-and-review-replies)
+and `CONTRIBUTING.md`, which closes an addressed thread with "Resolve
+conversation"; never resolve a disputed or unverified request.

.claude/skills/codetrial-triage/SKILL.md

@@ -6,8 +6,9 @@ description: >
   review, and draft the reminders, acceptances and closes that follow. Use for
   backlog triage, checking new issues or PRs, asking what needs attention,
   deciding whether a symptom already has an issue, explicit PR sweeps,
-  approving GitHub Actions runs held for first-time contributors, and asking
-  conflicting PRs to rebase. Default to a read-only report; a PR sweep has
+  approving GitHub Actions runs held for first-time contributors, asking
+  conflicting PRs to rebase, and applying requested label, assignee, issue type,
+  sub-issue or blocked-by changes. Default to a read-only report; a PR sweep has
   narrowly defined actions in pulls.md. Use codetrial-contribute for drafts on
   a contributor's behalf, new issues or PRs, and focused maintainer questions.
 ---
@@ -403,6 +404,17 @@ that error, comment `Duplicate of #N` and close with `--reason "not planned"`,
 which on its own reads as rejected and would take a reopen to correct, notifying
 the reporter a second time for nothing.
 
+A sub-issue, a blocked-by dependency and a duplicate are three different
+records, and a cross-link in prose establishes none of them. `gh issue edit`
+covers labels and assignees; for issue types, sub-issues and dependencies read
+GitHub's current API documentation for that operation rather than guessing an
+endpoint, check which issue is the parent or the blocker, and pass a node or
+database ID where the endpoint wants one instead of an issue number. Put the
+request body in a JSON file and send it with
+`gh api <endpoint> --method <method> --input "$SCRATCH/payload.json"`, then read
+the relationship back. If the operation is unavailable, say so and list the
+intended relationship as a proposal, never as applied.
+
 Post any preserved reproduction evidence to the canonical issue before closing.
 Do not pair `--duplicate-of` with a `Duplicate of #N` comment; GitHub's marker
 already says it. Report applied actions separately from proposals, with links.

.claude/skills/codetrial-triage/issue-sweep.sh

@@ -34,8 +34,8 @@
 #               is the newest comment; stale is null when the history is cut,
 #               and idle_days too when none of their acts is in what was read
 #   truncated   issues with more than 100 comments or cross-references, 10
-#               assignees or 20 assignments, whose lists above may miss what
-#               was cut
+#               assignees, 20 assignments or 20 body edits, whose lists above
+#               may miss what was cut
 
 set -eu
 
@@ -48,8 +48,11 @@ query='query($owner: String!, $name: String!, $endCursor: String) {
     issues(states: OPEN, first: 50, after: $endCursor) {
       pageInfo { hasNextPage endCursor }
       nodes {
-        number title body createdAt lastEditedAt authorAssociation
-        author { login } editor { login } issueType { name }
+        number title body createdAt authorAssociation
+        author { login } issueType { name }
+        userContentEdits(last: 20) {
+          totalCount nodes { editedAt editor { login } }
+        }
         assignees(first: 10) { totalCount nodes { login } }
         assigned: timelineItems(last: 20, itemTypes: [ASSIGNED_EVENT]) {
           totalCount
@@ -94,7 +97,9 @@ pages=$(gh api graphql --paginate -F owner="${repo%/*}" -F name="${repo#*/}" \
 # collaborators with read or triage access, students among them, so the badge
 # counts only when the push listing could not be read. A bot's comment never
 # asks the reporter for anything. The reporter answering by editing the body
-# counts, since the bug form and maintainers both ask for that.
+# counts, since the bug form and maintainers both ask for that; every edit is
+# read rather than only the last, which a maintainer tidying the body would
+# otherwise hide.
 printf '%s\n' "$pages" | jq -s --argjson days "$days" --arg repo "$repo" \
     --argjson pushers "$pushers" --arg me "$me" '
 def ts: if . == null then 0 else fromdateiso8601 end;
@@ -119,28 +124,48 @@ def ink: [limit(40; split("\n")[] | select(startswith("#") | not)
 [.[].data.repository.issues.nodes[]] as $issues |
 def maint: [.comments.nodes[] | select(human and staff)];
 def cut: .comments.totalCount > 100 or .timelineItems.totalCount > 100
-  or .assignees.totalCount > 10 or .assigned.totalCount > 20;
+  or .assignees.totalCount > 10 or .assigned.totalCount > 20
+  or .userContentEdits.totalCount > 20;
 def prs: [.timelineItems.nodes[] | select(.source.number != null
   and .source.repository.nameWithOwner == $repo)];
 def split: {open: [.[] | select(.state == "OPEN") | .number] | unique,
   merged: [.[] | select(.mergedAt != null) | .number] | unique};
 def mine: .author.login as $a | [.comments.nodes[]
   | select($a != null and .author.login == $a) | .createdAt | ts]
-  + [if .editor.login == $a then .lastEditedAt | ts else 0 end] | max;
+  + [.userContentEdits.nodes[] | select($a != null and .editor.login == $a)
+      | .editedAt | ts] + [0] | max;
+# Per line rather than one regex over the body, since "the same or a higher
+# level" is not a backreference. A fence counts as content and what is inside
+# it is skipped, so the comments in a pasted shell session are not headings; it
+# closes only on its own character, at least as long, with nothing after it.
+def empty_section: [foreach (split("\n")[] | select(test("\\S"))) as $l
+    ({fence: null, level: null};
+     ([$l | capture("^ {0,3}(?<m>`{3,}|~{3,})(?<rest>.*)$")] | first) as $f
+     | if .fence == null then
+         if $f != null and ($f.m[0:1] == "~" or ($f.rest | contains("`") | not))
+         then {fence: $f.m, level: 0}
+         else .level = (($l | capture("^(?<h>#{1,6}) +\\S") | .h | length) // 0)
+         end
+       elif $f != null and $f.m[0:1] == .fence[0:1]
+           and ($f.m | length) >= (.fence | length) and ($f.rest | test("^\\s*$"))
+       then {fence: null, level: 0}
+       else .level = null end;
+     .level | select(. != null))] as $lv
+  | any(range($lv | length); . as $k | $lv[$k] > 0
+      and ($lv[$k + 1] // -1 | . == -1 or (. > 0 and . <= $lv[$k])));
 def gaps: (.body // "") as $b
   | ($b + " " + ([.author.login as $a | .comments.nodes[]
       | select($a != null and .author.login == $a) | .body] | join(" "))) as $said
   | [ (select($b | ink < 40) | "empty_body"),
       (select($b | test("\\[([Pp]lease )?[Ff]ill in|<[A-Z][a-z][^<>\\n]{15,}>"))
         | "placeholder"),
-      (select($b | test("(?m)^(#{1,6}) [^\\n]+\\n\\s*(^\\1 |\\z)"))
-        | "empty_section"),
+      (select($b | empty_section) | "empty_section"),
       (select(.title | test("^\\s*(\\[[^\\]]+\\]|[A-Za-z]+(\\([^)]*\\))?:)\\s"))
         | "title_prefix"),
       (select((.title | cjk > latin) or (($b | cjk) > ($b | latin)))
         | "not_english"),
       (select(.issueType.name == "Bug"
-          and ($said | test("codetrial [0-9.]+ \\([0-9a-f]{7,40}\\)|\\b(?=[0-9a-f]*[0-9])(?=[0-9a-f]*[a-f])[0-9a-f]{7,40}\\b") | not))
+          and ($said | test("codetrial [0-9.]+ \\([0-9a-f]{7,40}\\)|\\b(?=[0-9a-f]*[0-9])(?=[0-9a-f]*[a-f])[0-9a-f]{7,40}\\b"; "i") | not))
         | "no_revision") ];
 def claim: . as $i | maint as $m
   | (($pushers // []) + [$m[].author.login]
@@ -150,8 +175,9 @@ def claim: . as $i | maint as $m
   | ([$i.assigned.nodes[] | select(.assignee.login == $a) | .createdAt]
       + [$i.comments.nodes[] | select(.author.login == $a) | .createdAt]
       + [$xrefs[].createdAt] | map(ts) | max) as $last
-  | ([$m[] | select((.createdAt | ts) > ($last // 0)) | .createdAt]
-      | max) as $check
+  | (if $last == null then null
+      else [$m[] | select((.createdAt | ts) > $last) | .createdAt] | max end)
+      as $check
   | (if $check then $check | ts else $last end) as $since
   | ([$i.comments.nodes[] | select(human)] | max_by(.createdAt | ts)
       | .author.login == $a) as $asked

.claude/skills/codetrial-triage/pr-sweep.sh

@@ -62,7 +62,8 @@ known='["Rebase latest `main` and resolve conflicts",
 # first check suite, created when the push reached a workflow or app, or the
 # latest force push. The commit date, which the contributor's clock sets and
 # which can run ahead and make every later request look stale, is used only for
-# a head with no check suite.
+# a head with no check suite. A PR force-pushed back to its base has no commit
+# at all, and dates only its force push.
 # shellcheck disable=SC2016
 query='query($owner: String!, $name: String!, $endCursor: String) {
   repository(owner: $owner, name: $name) {
@@ -131,7 +132,7 @@ printf '%s\n%s\n' "$prs" "$runs" | jq -s --argjson days "$days" --arg me "$me" \
     --argjson pushers "$pushers" '
 def human: .author != null and .author.__typename != "Bot";
 def ts: if . == null then 0 else fromdateiso8601 end;
-def pushed: .commits.nodes[0].commit as $c
+def pushed: (.commits.nodes[0].commit // {checkSuites: {nodes: []}}) as $c
   | [([$c.checkSuites.nodes[].createdAt | ts] | min) // ($c.committedDate | ts),
      (.timelineItems.nodes[0].createdAt | ts)] | max;
 def trim: gsub("^\\s+|\\s+$"; "");
@@ -155,11 +156,10 @@ def requests: .author.login as $a | .headRefOid as $head | pushed as $p |
 def fixed_only: .inline == 0 and (.body | norm | IN($wordings[]));
 def asks_rebase: [requests[] | select(fixed_only)] | sort_by(.at | ts);
 def asked_here: any(requests[]; .current);
-# A conflicting PR whose only request on its current head is hand-written:
-# that request is the standing one, so no fixed one is kept beside it and it
-# is never offered for hiding.
-def hand_asked: .mergeable == "CONFLICTING"
-  and asked_here and all(asks_rebase[]; .current | not);
+# A PR whose only request on its current head is hand-written, whether or not
+# it still conflicts: that request is the standing one, so no fixed one is kept
+# beside it and it is never offered for hiding.
+def hand_asked: asked_here and all(asks_rebase[]; .current | not);
 def needs_rebase: .mergeable == "CONFLICTING" and (.isDraft | not)
   and .baseRefName == $base and $base == "main"
   and nudgeable and whole

.claude/skills/codetrial-triage/pulls.md

@@ -180,14 +180,15 @@ gh pr view "$pr" --repo "$repo" --json mergeable,headRefOid \
 Post it before hiding anything, so a failed post never leaves a conflicting PR
 with no visible request. `hide` lists the earlier requests, reviews and comments
 alike, in a known wording and with no inline comments: all of them on a PR in
-`rebase`, or on a conflicting PR whose current request is hand-written (that
-request stays and is kept out of `hide_ask`); on any other PR, all but one: the
-newest on the current head, or the newest overall when none is, since a review
-left on an older head can carry a later date. The entries for a PR in `rebase`
-assume its new request landed, so hide them only after that PR's post succeeded;
-when the check above skipped the post, or the post failed, keep the one that
-rule would keep, which is still the current request. Hiding a review does not
-dismiss it; the PR stays at changes requested until a reviewer re-reviews.
+`rebase`, or on a PR, conflicting or not, whose current request is hand-written
+(that request stays and is kept out of `hide_ask`); on any other PR, all but
+one: the newest on the current head, or the newest overall when none is, since a
+review left on an older head can carry a later date. The entries for a PR in
+`rebase` assume its new request landed, so hide them only after that PR's post
+succeeded; when the check above skipped the post, or the post failed, keep the
+one that rule would keep, which is still the current request. Hiding a review
+does not dismiss it; the PR stays at changes requested until a reviewer
+re-reviews.
 
 `hide_ask` lists the other requests that mention rebase on a PR that has, or is
 about to get, the fixed request: hand-written ones, and known-wording reviews

.claude/skills/codetrial-triage/waiting.md

@@ -38,10 +38,9 @@ The four sweep lists are the queue, and each one has a fixed next move:
   disposition first. When a linked PR merges or the facts change while a
   reminder is pending, rescope the request and keep its deadline. The closing
   comment names the fact again and says a maintainer reopens the issue once it
-is posted, since a comment alone does not reopen it and the sweep reads only
-open issues. An
-  issue a maintainer opened is left out of `waiting`, because its author already
-  knows the rule.
+  is posted, since a comment alone does not reopen it and the sweep reads only
+  open issues. An issue a maintainer opened is left out of `waiting`, because
+  its author already knows the rule.
 - `linked`, with `merged` set: a merged pull request referenced the issue after
   it was filed, and GitHub did not close it, so that PR only touched it, or its
   `Closes` line named another issue. Read the PR and the issue. When the PR does