Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Address feedback left on a GitHub pull request: fetch unresolved review threads, make agreed Elixir/Phoenix code fixes, reply, and resolve. Use for a PR URL/number or reviewer comments. NOT for pre-PR review, findings triage, or CI monitoring.
.claude/skills/oliver-kriska-phx-pr-review/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 104% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 10% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 60% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 34% | 0% |
| case-03 | ✓→✗ | ▼ Worse | 60% | 0% |
Inspect unresolved pull-request review threads, triage them read-only by default, and apply only explicitly approved fixes. GitHub mutations are never implied.
text/skill:phx-pr-review 42 /skill:phx-pr-review 42 --fix /skill:phx-pr-review https://github.com/owner/repo/pull/42 --bots-only /skill:phx-pr-review 42 --no-resolve
thread IDs and resolved state. Otherwise use this exact gh 2.94-compatible fallback (requires command -v gh and gh auth status):
bashPR_INPUT="${1:?PR number or URL}" if [[ "$PR_INPUT" =~ ^https://github.com/([^/]+)/([^/]+)/pull/([0-9]+) ]]; then OWNER="${BASH_REMATCH[1]}"; REPO="${BASH_REMATCH[2]}"; PR="${BASH_REMATCH[3]}" else OWNER=$(gh repo view --json owner --jq '.owner.login') REPO=$(gh repo view --json name --jq '.name'); PR="$PR_INPUT" fi gh api graphql --paginate -F owner="$OWNER" -F repo="$REPO" -F pr="$PR" -f query='query($owner: String!, $repo: String!, $pr: Int!, $endCursor: String) { repository(owner: $owner, name: $repo) { pullRequest(number: $pr) { reviewThreads(first: 100, after: $endCursor) { nodes { id isResolved isOutdated path line originalLine comments(first: 100) { totalCount nodes { id databaseId body author { login __typename } replyTo { id } } } } pageInfo { hasNextPage endCursor } } } } }' gh api graphql --paginate -F owner="$OWNER" -F repo="$REPO" -F pr="$PR" -f query='query($owner: String!, $repo: String!, $pr: Int!, $endCursor: String) { repository(owner: $owner, name: $repo) { pullRequest(number: $pr) { reviews(first: 100, after: $endCursor) { nodes { id state body submittedAt author { login __typename } } pageInfo { hasNextPage endCursor } } } } }'
--paginate binds each returned pageInfo.endCursor to $endCursor. Filter isResolved == false after collecting pages. Preserve thread id, root comment id/databaseId, and author.__typename (not login suffixes). comments(first:100) is nested and is not paginated by the outer command: if comments.totalCount > nodes.length, report the thread as TRUNCATED and fetch every comment page with this exact query. The outer query intentionally omits nested comments.pageInfo so gh --paginate follows only the outer reviewThreads.pageInfo cursor:
bashgh api graphql --paginate -F threadId="$THREAD_ID" -f query='query($threadId: ID!, $endCursor: String) { node(id:$threadId) { ... on PullRequestReviewThread { comments(first:100, after:$endCursor) { totalCount nodes { id databaseId body author { login __typename } replyTo { id } } pageInfo { hasNextPage endCursor } } } } }'
Merge all comment pages in API order and deduplicate by GraphQL id (first occurrence wins). Block triage until every truncated thread is complete. Do not substitute issue comments for review threads.
With --bots-only, use API actor type rather than a login suffix. Show one row per thread with author, category, outdated state, and proposed action. Review summaries are separate, non-resolvable context.
of selected thread IDs. Selection authorizes inspection only. --fix permits later edits but approves nothing and never selects every thread.
propose the exact patch. Obtain explicit edit approval before editing, or record EDIT: NOT APPLICABLE with evidence. Then show the applied diff and run the smallest relevant compile/test check after every code change.
addressed: require current code/diff evidence before that disposition. Show the diff, evidence, and exact verified reply, then obtain a separate explicit posting approval before posting. Use a connector mutation when available, or gh api to reply to the root review comment. If posting is unsupported or fails, report NOT POSTED with the reason.
Only then request a separate resolution approval. --no-resolve always disables resolution, regardless of any other flag or approval. Use the connector or resolveReviewThread mutation, then confirm returned state. Never claim replied/resolved from a draft or intent.
thread | action | verification | reply | resolution, changed files,and precise blockers. Paginated review summaries with CHANGES_REQUESTED or an actionable non-empty body are findings even when there are zero inline threads; report them as non-resolvable context and never call that state clean. Do not commit or push.
Generic read-only workers may inspect independent threads when the runtime has them, but named custom agents are not required and sequential same-session processing is complete.
--fix permits but doesnot approve edits.
references/response-patterns.md — reply templates and tonereferences/gh-commands.md — GitHub CLI queries and mutationsreferences/bot-triage.md — bot review triage| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 5,164 | 8,001 | +55% | 1 | 1 | 0% | 791 | 2,237 | +183% | 0 | 0 | — |
case-02 | fail→fail | 6,068 | 8,248 | +36% | 1 | 1 | 0% | 973 | 1,848 | +90% | 0 | 0 | — |
case-03 | pass→fail | 7,407 | 7,666 | +3% | 1 | 1 | 0% | 1,196 | 1,910 | +60% | 0 | 0 | — |
case-04 | fail→pass | 5,989 | 3,524 | -41% | 1 | 1 | 0% | 1,067 | 2,173 | +104% | 0 | 0 | — |
case-05 | fail→fail | 3,229 | 9,041 | +180% | 1 | 1 | 0% | 407 | 2,210 | +443% | 0 | 0 | — |
case-06 | pass→pass | 9,878 | 7,442 | -25% | 1 | 1 | 0% | 1,898 | 2,727 | +44% | 0 | 0 | — |
case-07 | fail→fail | 5,876 | 3,221 | -45% | 1 | 1 | 0% | 916 | 2,027 | +121% | 0 | 0 | — |
case-08 | pass→pass | 6,711 | 4,682 | -30% | 1 | 1 | 0% | 1,117 | 2,227 | +99% | 0 | 0 | — |
case-09 | fail→fail | 10,155 | 3,300 | -68% | 1 | 1 | 0% | 1,621 | 2,082 | +28% | 0 | 0 | — |
case-10 | fail→pass | 9,925 | 1,617 | -84% | 1 | 1 | 0% | 1,625 | 1,784 | +10% | 0 | 0 | — |
case-11 | pass→pass | 8,817 | 5,987 | -32% | 1 | 1 | 0% | 1,402 | 2,542 | +81% | 0 | 0 | — |
case-12 | pass→pass | 6,756 | 4,141 | -39% | 1 | 1 | 0% | 1,147 | 2,230 | +94% | 0 | 0 | — |
case-13 | fail→pass | 8,849 | 2,683 | -70% | 1 | 1 | 0% | 1,225 | 1,959 | +60% | 0 | 0 | — |
case-14 | pass→pass | 7,796 | 2,352 | -70% | 1 | 1 | 0% | 1,210 | 1,880 | +55% | 0 | 0 | — |
case-15 | fail→pass | 9,817 | 5,153 | -48% | 1 | 1 | 0% | 1,824 | 2,451 | +34% | 0 | 0 | — |
case-16 | fail→fail | 10,651 | 11,541 | +8% | 1 | 1 | 0% | 1,778 | 2,289 | +29% | 0 | 0 | — |
case-17 | fail→fail | 9,132 | 4,188 | -54% | 1 | 1 | 0% | 1,466 | 2,117 | +44% | 0 | 0 | — |
case-18 | pass→pass | 5,466 | 2,886 | -47% | 1 | 1 | 0% | 790 | 2,022 | +156% | 0 | 0 | — |
case-19 | pass→pass | 6,474 | 3,686 | -43% | 1 | 1 | 0% | 1,016 | 2,023 | +99% | 0 | 0 | — |
case-20 | pass→pass | 12,711 | 9,811 | -23% | 1 | 1 | 0% | 1,920 | 2,989 | +56% | 0 | 0 | — |
case-21 | pass→pass | 12,085 | 4,238 | -65% | 1 | 1 | 0% | 1,860 | 2,162 | +16% | 0 | 0 | — |
case-22 | pass→pass | 10,949 | 4,282 | -61% | 1 | 1 | 0% | 1,809 | 2,057 | +14% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 17 counted toward the lift figure. The other 5 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +14 percentage points is the difference between those two pass rates over the 17 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.