Grade what triage cannot see and add /codetrial-odds

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

@@ -0,0 +1,69 @@
+---
+name: codetrial-odds
+description: >
+  List every open CodeTrial pull request with its merge confidence, size and
+  who it waits on, and what moved since the last run. Read-only.
+disable-model-invocation: true
+---
+
+# How close each open pull request is to merging
+
+A shortcut for one question the triage skill already answers: "which open PRs
+could merge, and how sure is that". It runs only when typed as
+`/codetrial-odds`, so it never competes with
+[codetrial-triage](../codetrial-triage/SKILL.md) for a request described in
+words; the grading rules stay in one place, [Rank by size and merge
+confidence](../codetrial-triage/pulls.md#rank-by-size-and-merge-confidence),
+and this file only says how to present them. It posts, approves and resolves
+nothing.
+
+## Take the snapshot
+
+[snapshot.sh](snapshot.sh) runs the sweep against the last snapshot, which
+fills `moved` and `gone`, and replaces it only with a whole sweep. When it
+fails, stop and say so: a report from the old snapshot would read as today's.
+Then read the fields the report needs from the snapshot file:
+
+```sh
+snap=$(.claude/skills/codetrial-odds/snapshot.sh) &&
+jq '{viewer, maintainers_known, unknown, moved, gone, prerequisites,
+     overlap, collide, collide_unread,
+     status: [.status[] | {pr, title, author, author_is_maintainer, size,
+       mergeable, decision, grade, waits_on, waits_on_all, why,
+       files_truncated, checks_truncated: .checks.truncated}],
+     truncated: [.prs[] | select(.reviews_truncated or .comments_truncated
+       or .threads_truncated or .closes_truncated) | .number]}' "$snap"
+```
+
+Report the open count, and every PR in `unknown`, `truncated`, or with
+`files_truncated` or `checks_truncated` set, since its grade rests on less
+than the others'. When `maintainers_known` is false, `author_is_maintainer` is
+null and a read-only collaborator's comment may have counted as a request; say
+so.
+
+## Present it
+
+Group the PRs by `waits_on`, maintainer first, since those are the ones a merge
+decision can act on today; within a group keep the sweep's order, smallest
+first, and say the size is total lines split into source and tests. Give each
+PR its grade, title and the `why` in plain words, and when `waits_on_all` names
+more than `waits_on`, say who else still owes a move. Say that `high` with no
+review yet means ready for a first review, not approved. Mark a PR whose
+`author_is_maintainer` is true as a maintainer's own, and one by `viewer` as
+the reader's own, which needs another reviewer.
+
+End with the merge order the snapshot records, `prerequisites` and `collide`
+(with each collision's path and function), read as
+[pulls.md](../codetrial-triage/pulls.md#similar-pull-requests) says, and name
+`collide_unread` PRs as not compared. Say nothing about order the snapshot
+does not record. `overlap` sets no order: list its pairs apart, as two PRs
+closing one issue.
+
+Then what moved, read as the end of [Rank by size and merge
+confidence](../codetrial-triage/pulls.md#rank-by-size-and-merge-confidence)
+says. On a first run `moved` and `gone` are null: say this run is the first
+baseline.
+
+The grade reads only what GitHub records. Reading a PR's non-test diff can
+lower it, never raise it without a named reason; do that only for the PRs the
+user asks about, and say which grades rest on a diff read.

.claude/skills/codetrial-odds/snapshot.sh

@@ -0,0 +1,33 @@
+#!/bin/sh
+# Refresh the /codetrial-odds snapshot: run pr-sweep.sh against the last one,
+# and replace it only with a whole sweep, so a failed or interrupted run keeps
+# the old baseline. It lives in the git directory beside the triage review
+# ledger, so it survives sessions and is never tracked, one per repository,
+# since the sweep matches PRs by number alone.
+#
+# Usage: snapshot.sh [repo] (default sysprog21/codetrial)
+#        Prints the snapshot's path; exits 1, printing nothing, when the sweep
+#        failed. ODDS_SWEEP names the pr-sweep.sh to run, for tests.
+
+set -eu
+
+repo=${1:-sysprog21/codetrial}
+sweep=${ODDS_SWEEP:-$(dirname "$0")/../codetrial-triage/pr-sweep.sh}
+dir="$(git rev-parse --git-common-dir)/codetrial-triage"
+mkdir -p "$dir"
+last=$dir/pr-sweep-$(printf '%s' "$repo" | tr / -).json
+
+# An empty stale-days leaves pr-sweep.sh its own default.
+set -- "$repo" ""
+if [ -s "$last" ]; then
+    set -- "$@" "$last"
+fi
+tmp=$(mktemp "$dir/pr-sweep.XXXXXX")
+trap 'rm -f "$tmp"' EXIT
+if sh "$sweep" "$@" > "$tmp" && jq -e '.status | arrays' "$tmp" > /dev/null; then
+    mv "$tmp" "$last"
+    echo "$last"
+else
+    echo "snapshot.sh: the sweep failed; the last snapshot is kept" >&2
+    exit 1
+fi

.claude/skills/codetrial-triage/post-inline.sh

@@ -1,13 +1,14 @@
 #!/bin/sh
-# Post a sweep review's inline comments, one per finding, at the head they were
-# read at. review-prs.md describes the draft and why a sweep posts it without a
-# confirmation.
+# Post a sweep review's inline comments at the head they were read at, as one
+# review so the author gets one notification, not one per finding. review-prs.md
+# describes the draft and why a sweep posts it without a confirmation.
 #
 # Usage: post-inline.sh <pr> <draft.json> [repo]
-#        Prints each posted comment's URL, and on stderr each finding it
-#        dropped and why. Exits 3 when the head moved, checked before anything
-#        is read and again before every post, so a push during the reads posts
-#        nothing; 1 when GitHub refused a comment; 2 on bad arguments.
+#        Prints the review's URL, or each comment's when it fell back to one
+#        post per finding, and on stderr each finding it dropped and why.
+#        Exits 3 when the head moved, checked before anything is read and
+#        again before every post, so a push during the reads posts nothing; 1
+#        when GitHub refused a comment; 2 on bad arguments.
 #
 # The draft is {sha, comments: [{path, line, start_line?, body}], summary}. Two
 # kinds of finding are dropped rather than posted. One outside the diff, which
@@ -66,17 +67,11 @@ sh "$(dirname "$0")/../codetrial-verify/review-anchors.sh" "$pr" \
     exit 1
 }
 
-# Each hunk's new-side span, one tab-separated "path first last" per row. Header
-# handling is review-anchors.sh's; a hunk adding nothing on the right has no
-# span.
+# Each hunk's new-side span, one tab-separated "path first last" per row; a hunk
+# adding nothing on the right has no span.
 diff=$(gh pr diff "$pr" --repo "$repo")
-printf '%s\n' "$diff" | awk '
-    function header_path(line,    p) {
-        p = substr(line, 5)
-        sub(/\t.*$/, "", p)
-        if (p ~ /^".*"$/) p = substr(p, 2, length(p) - 2)
-        return p == "/dev/null" ? p : substr(p, 3)
-    }
+printf '%s\n' "$diff" \
+    | awk "$(cat "$(dirname "$0")/../codetrial-verify/diff-path.awk")"'
     /^diff --git / { header = 1; path = ""; next }
     header && /^\+\+\+ / { path = header_path($0); next }
     /^@@ / && path != "/dev/null" {
@@ -87,46 +82,68 @@ printf '%s\n' "$diff" | awk '
     }
 ' > "$work/hunks"
 
-gh api "repos/$repo/pulls/$pr/comments" --paginate > "$work/threads.json"
-
-jq -n --slurpfile anchored "$work/anchored.json" \
-    --slurpfile threads "$work/threads.json" --arg sha "$sha" --arg me "$me" \
-    --rawfile hunks "$work/hunks" --slurpfile payload "$work/payload.json" '
-  [$hunks | split("\n")[] | select(length > 0) | split("\t")
-    | {path: .[0], first: (.[1] | tonumber), last: (.[2] | tonumber)}] as $spans
-  | def span: {path, first: (.start_line // .line), last: .line};
-    def overlaps($a; $b): $a.path == $b.path
-      and $a.first <= $b.last and $b.first <= $a.last;
-  ($threads | add // [] | map(select(.commit_id == $sha))) as $here
-  | [$here[] | select(.user.login == "cubic-dev-ai[bot]" and .line != null
-      and .side == "RIGHT") | span] as $cubic
-  | def posted: . as $c | any($here[]; .user.login == $me
-      and .path == $c.path and .line == $c.line and .body == $c.body);
-  def claimed: span as $s
-      | any($spans[]; . as $h | overlaps($h; $s) and any($cubic[]; overlaps($h; .)));
-  ($anchored[0].comments) as $inline
-  | def anchored: . as $c | any($inline[]; .path == $c.path and .line == $c.line
-      and .start_line == $c.start_line);
-  $payload[0].comments[]
-  | if anchored | not then {drop: "outside the diff"} + .
-    elif posted then {drop: "already posted at this head"} + .
-    elif claimed then {drop: "in a hunk cubic already commented on"} + .
-    else . end
-' > "$work/sorted.json"
+# Sorts the findings against the review threads as they stand, leaving those to
+# post in $work/post. Run again after a refused review, since a review that
+# landed with its answer lost reads as refused, and its findings are then
+# already posted at this head; a second sort only ever drops more.
+sort_findings()
+{
+    gh api "repos/$repo/pulls/$pr/comments" --paginate \
+        > "$work/threads.json" < /dev/null
+    jq -n --slurpfile anchored "$work/anchored.json" \
+        --slurpfile threads "$work/threads.json" --arg sha "$sha" --arg me "$me" \
+        --rawfile hunks "$work/hunks" --slurpfile payload "$work/payload.json" '
+      [$hunks | split("\n")[] | select(length > 0) | split("\t")
+        | {path: .[0], first: (.[1] | tonumber), last: (.[2] | tonumber)}] as $spans
+      | def span: {path, first: (.start_line // .line), last: .line};
+        def overlaps($a; $b): $a.path == $b.path
+          and $a.first <= $b.last and $b.first <= $a.last;
+      ($threads | add // [] | map(select(.commit_id == $sha))) as $here
+      | [$here[] | select(.user.login == "cubic-dev-ai[bot]" and .line != null
+          and .side == "RIGHT") | span] as $cubic
+      | def posted: . as $c | any($here[]; .user.login == $me
+          and .path == $c.path and .line == $c.line and .body == $c.body);
+      def claimed: span as $s
+          | any($spans[]; . as $h | overlaps($h; $s) and any($cubic[]; overlaps($h; .)));
+      ($anchored[0].comments) as $inline
+      | def anchored: . as $c | any($inline[]; .path == $c.path and .line == $c.line
+          and .start_line == $c.start_line);
+      $payload[0].comments[]
+      | if anchored | not then {drop: "outside the diff"} + .
+        elif posted then {drop: "already posted at this head"} + .
+        elif claimed then {drop: "in a hunk cubic already commented on"} + .
+        else . end
+    ' > "$work/sorted.json"
+    jq -c 'select(.drop | not)' "$work/sorted.json" > "$work/post"
+}
+sort_findings
 
+# GitHub creates a review whole or not at all, so one refused comment refuses
+# them all; only then is each posted on its own, so the rest still land.
+status=0
+if [ -s "$work/post" ]; then
+    same_head
+    jq -s --arg sha "$sha" '{commit_id: $sha, event: "COMMENT", body: "",
+        comments: .}' "$work/post" > "$work/review.json"
+    if gh api "repos/$repo/pulls/$pr/reviews" --input "$work/review.json" \
+        --jq .html_url < /dev/null 2> "$work/refused"; then
+        : > "$work/post"
+    else
+        why=$(head -1 "$work/refused")
+        echo "post-inline.sh: #$pr refused the review${why:+ ($why)};" \
+            "posting each finding" >&2
+        sort_findings
+    fi
+fi
 jq -r 'select(.drop) | "post-inline.sh: #'"$pr"' dropped \(.path):\(.line), \(.drop)"' \
     "$work/sorted.json" >&2
-jq -c --arg sha "$sha" 'select(.drop | not) | . + {commit_id: $sha}' \
-    "$work/sorted.json" > "$work/post"
-
-# One at a time, so a refused comment does not stop the rest.
-status=0
 while IFS= read -r comment; do
 
     # Before every write: a push mid-way leaves what already landed, which a
     # rerun at the new head reads as not current, and posts nothing more.
     same_head
-    printf '%s\n' "$comment" > "$work/one.json"
+    printf '%s\n' "$comment" | jq --arg sha "$sha" '. + {commit_id: $sha}' \
+        > "$work/one.json"
     gh api "repos/$repo/pulls/$pr/comments" --input "$work/one.json" \
         --jq .html_url < /dev/null || {
         echo "post-inline.sh: #$pr refused $(jq -r '"\(.path):\(.line)"' "$work/one.json")" >&2