PR review handler skill

This skill iterates a PR through review/comment cycles until there is nothing left to address.

Your role on a PR

When this skill is in use, the agent's job is to bring the PR into merge-ready shape: solve the original problem the PR was opened for, and address every legitimate finding the PR has accumulated, from every source. Reviewers, linters, and CI all matter. Comments are the loudest source but they are not the only source -- you must proactively pull findings from every channel that reports on the PR, not wait for something to surface as a chat message.

Sources of findings, in priority order:

1. Human review comments -- maintainers / devs / community.
2. AI bot review comments -- cubic-dev-ai, coderabbitai, etc.
3. SonarCloud PR findings -- new code-smell / vulnerability / security-hotspot issues introduced by this PR. SonarCloud does NOT post these as inline GitHub review-comments; only a QualityGate summary is posted to GitHub. The actual findings live behind the SonarCloud API and must be pulled explicitly.
4. CI failures relevant to this PR -- shellcheck, codeql, build / test failures caused by the PR's changes.
5. Anything else this repo configures (Codacy, custom workflows, ...).

A finding is "relevant to this PR" if its existence (or its line location) is plausibly caused by the PR's diff. CI failures unrelated to this PR (a flaky test on an unrelated module, an infra outage) are NOT in scope -- note them, surface to the user at the end, do not fix them here.

The bar is the project's performance, stability, and long-term maintainability. Don't dismiss findings because they look minor.

MANDATORY rules

These are non-negotiable. Skipping any of them will cost the user time.

1. Pagination paranoia. Do not stop at round numbers. If a fetch returns exactly 100 / 200 / 300 items, the round count is suspicious -- GitHub pagination defaults to 100, and round-multiples almost always mean there is a next page that the previous client missed. Always re-probe with an explicit page=N+1 request. fetch-all.sh does this automatically.
2. Accept all comments and address them all. No exceptions. No "this is minor, skip it." The bar is: Netdata's performance, stability, and long-term maintainability.
3. Verify every comment properly. No shortcuts. Read the code, follow the trace, confirm the claim. AI bots produce false positives -- judge each one on its merits.
4. Reply per-thread, one by one. No bulk replies. No mechanical "fixed" answers. Each thread gets a substantive reply that explains what you did or why the comment doesn't apply.
5. Don't dismiss comments because they look minor. Even style nits compound. The goal is for the project to thrive.
6. Help bots when they're confused. AI reviewers sometimes flag false positives because the surrounding code is ambiguous. Add a short comment in the source that clarifies the intent -- it helps the next reviewer (human or bot).
7. Check CI BEFORE every push, but never WAIT for CI between iterations. Waiting for CI between bot-review cycles destroys throughput -- a CI run can take 30+ minutes, and during that time the AI reviewers are idle. The right cadence is:
   • Before each push: run ci-status.sh. If there are FAILURES, fix them and bundle into the same push. If checks are still running, that's fine -- ignore them and push anyway. The next push triggers fresh CI on the new code, which is what we actually care about.
   • After each push: re-trigger the bots, then wait-for-activity.sh for new comments (NOT for CI).
   • If wait-for-activity.sh times out (30 min, no new comments): re-check ci-status.sh. If checks are still running, that's normal, surface to the user. If there are failures, fix and iterate.
8. Re-trigger AI reviewers explicitly. They do NOT react to thread replies or pushed commits the way humans do. Re-trigger every reviewer the exit condition waits on -- if such a reviewer is never re-triggered it can never post its "no new findings" pass, so the loop cannot end. A reviewer that is deliberately excluded from the loop (see Copilot below) must also be excluded from the exit condition, so the two lists always match.
   • cubic-dev-ai: post a new top-level comment mentioning it (trigger-cubic.sh).
   • coderabbitai: post a new top-level comment with a command (trigger-coderabbit.sh; @coderabbitai review is incremental, full review re-reads the whole PR).
   • Copilot: NOT re-triggered. Its billing model means the org normally has no credits for it, so a re-request produces nothing and only adds latency. trigger-copilot.sh is kept for the case where credits exist, but it is not part of the loop and Copilot never blocks the exit condition. Address any comments it does post like any other AI bot.
9. Don't loop forever on silent bots. Some assistants stop responding. That's fine. Use wait-for-activity.sh with the 30-min timeout and move on if nothing changes.
10. When a bot finds a legit issue, search the WHOLE PR for similar issues. This is the most expensive rule to ignore. AI reviewers surface their top 3-7 findings, not the full set. If you fix only the ones they pointed at, you'll spend dozens of round-trips discovering the rest one at a time. Every round-trip is 30+ minutes of bot review latency. The fix for one issue means a full re-audit of the PR for the same class of issue. Do that before pushing.
11. Don't trust linters alone -- smoke-test every fix. Static analyzers (shellcheck, etc.) verify a property of the code; they don't verify behavior. A "correct per the linter" fix can change runtime behavior in subtle ways (e.g. a printf format-string fix that stops escape-sequence interpretation, breaking colored output that the linter never knew about). After every fix, run the affected script (or the smallest invocation that exercises the change) and verify the output looks right. "Linter green" is not the same as "still works."
12. Before every push, spawn a subagent for a holistic PR review. See Step 4a in the workflow. The orchestrator's context is biased toward the fixes it just made; a clean-context subagent re-reviewing the WHOLE diff is what catches the issues the orchestrator and the AI reviewers missed. Skipping this turns each iteration into a 30-minute round-trip to discover issues that could have been found in 2 minutes locally.
13. Before every push, re-fetch all finding sources one last time. See Step 4-pre. Reviewers post in parallel; if findings arrive while you're addressing the current batch, they belong in THIS push, not the next one. Without this sync barrier, you and the reviewers stay one round out of sync forever -- the next iteration is always "fixing" issues that no longer apply.

Author classes -- different handling per class

• AI bots (cubic-dev-ai[bot], coderabbitai[bot], copilot[bot] and variants): handle autonomously. Verify the finding, fix or push back with reasoning, reply in-thread, resolve thread.
• Informational bots (sonarqubecloud[bot], github-actions[bot], netdata-bot[bot]): read for signal (e.g. quality gate status). They don't usually require a reply.
• Humans (developers, maintainers, community): consult the user. Maintainer comments matter most -- in this project, we are usually contributors, they are the project owners. Do not respond on the user's behalf without their direction. Surface human comments to the user with a recommendation, then act per their instruction.

Setup

gh CLI authenticated for the repo. Nothing else.

The skill reads upstream (or origin) from git remotes to derive the repo slug. Override with PR_REPO_SLUG=owner/repo if working cross-repo.

State for each PR is cached under <repo-root>/.local/audits/pr-reviews/pr-<N>/:

• pr.json -- top-level PR metadata
• issue-comments.json -- top-level PR comments (REST)
• review-comments.json -- inline review comments (REST)
• reviews.json -- review submissions with body (REST)
• review-threads.json -- per-thread, with isResolved (GraphQL)
• summary.txt -- human-readable triage summary

Workflow

The order is: gather all findings -> address them per-thread / per-finding -> check CI for failures the PR caused -> push -> retrigger -> wait -> loop.

1a. Fetch all comments (paranoid)

bash .agents/skills/pr-reviews/scripts/fetch-all.sh <PR_NUMBER>

Tail-prints a summary.txt that shows the per-author count and the list of open review threads. Use this as the input to the rest of the cycle.

1b. Fetch SonarCloud PR findings

bash .agents/skills/pr-reviews/scripts/fetch-sonar-findings.sh <PR_NUMBER>

SonarCloud findings are NOT delivered as inline GitHub comments -- only a QualityGate summary is. The actual issue list lives behind the SonarCloud API. This script writes:

• .local/audits/pr-reviews/pr-<N>/sonar-issues.json
• .local/audits/pr-reviews/pr-<N>/sonar-hotspots.json
• a brief summary to stdout (counts by rule and severity).

Requires the same .env config the sonarqube-audit skill uses (SONAR_TOKEN, SONAR_HOST_URL, SONAR_PROJECT). If .env is missing, the script prints what's needed and exits.

1c. Note the CI signal as a third source

Run bash .agents/skills/pr-reviews/scripts/ci-status.sh <PR> once early to capture which checks are failing right now. You're looking for failures caused by the current PR (typo in a YAML file you added, a script that doesn't pass shellcheck, a build that breaks because of the diff). DO NOT fix CI yet -- just note the failures as input alongside review comments and Sonar findings. They all get addressed in the same iteration so a single push covers them.

2. List open threads (and Sonar findings)

bash .agents/skills/pr-reviews/scripts/list-open-threads.sh <PR_NUMBER>          # full bodies
bash .agents/skills/pr-reviews/scripts/list-open-threads.sh <PR_NUMBER> --short  # one line per thread

The "short" output is a table: thread-id | path:line | author. The full form prints every comment in each thread.

3. For each open thread, ONE AT A TIME

This is per-thread, not batched. Do not prepare a list of replies and fire them in a loop. Do not post all replies first and resolve all later. Walk one thread at a time:

For thread N:

1. Read the comment carefully. What is the bot/dev claiming?
2. Open the file at the line and verify. Does the claim hold against the current code? Is it valid in context?
3. Search the whole PR diff (and adjacent code) for the same class of issue. Rule #10 -- this is mandatory. (You only do this sweep once, on the first thread of a class -- subsequent threads in the same class share the same fix.)
4. Decide:
   • If valid -> fix it AND every similar instance you found.
   • If invalid -> understand why the bot got confused. Often a small source comment clarifying the intent will help the next reviewer.
5. Reply in the thread.
   Copy

   bash .agents/skills/pr-reviews/scripts/reply-thread.sh <PR> <comment-id> "<reply>"
   

   <comment-id> is the databaseId of the FIRST comment in the thread (from review-threads.json -> .[].comments.nodes[0].databaseId).
6. Resolve the thread immediately after the reply succeeds. "Succeeds" means you saw posted reply id=.... Resolving a thread whose reply failed hides it from the needs-attention view with nothing written in it, which reads to a human as a silently dismissed review.
   Copy

   bash .agents/skills/pr-reviews/scripts/resolve-thread.sh <thread-id>
   

   <thread-id> is the GraphQL node id (review-threads.json -> .[].id, starts with PRRT_). Resolving immediately after replying takes the thread out of the "needs attention" view; leaving threads open without resolution accumulates noise.

Then move to thread N+1. Reply-and-resolve, reply-and-resolve. Never queue them up.

The reason: the order makes intent visible to humans watching the PR -- they see "agent posted reply, agent resolved" as one motion per thread, not "agent dumped 14 replies, then dumped 14 resolves". Bulk operations look mechanical and erode trust in the address pass.

3b. Address each Sonar finding

For each issue in sonar-issues.json and each hotspot in sonar-hotspots.json:

1. Read the rule and the message. What is Sonar claiming?
2. Open the file at the line and verify. Does the claim hold against the current code?
3. Search the whole PR diff (and adjacent code) for the same class of issue. Same rule #10 as for review comments. If Sonar flagged one instance of S131 (case without default), sweep all case statements. If Sonar flagged S2245 (insecure RNG), sweep all random() callsites.
4. Decide:
   • If valid -> fix it AND every similar instance you found in the project where the same reasoning applies (within the diff, plus nearby code in files this PR already touches).
   • If invalid -> the corresponding sonarqube-audit skill provides sonar-mark.sh fp <KEY> "<reason>" to mark it False Positive directly in SonarCloud. Comments are ASCII-only (Cloudflare).
5. Repeat until sonar-issues.json has zero issues that we caused.

For Sonar there is no "thread reply" -- you address the issue with either a code fix or a sonar-mark.sh action. There's nothing to resolve in GitHub for Sonar findings.

4-pre. Before pushing -- MANDATORY final-fetch sync barrier

Reviewers run in parallel. Multiple bots and humans can be appending findings WHILE you're addressing the current batch. If you push the moment your queue is empty, the findings that arrived during this iteration get attributed to your fresh commit instead of the previous one -- and on the next round you end up "fixing" findings that no longer apply because you addressed them implicitly with the next push. The result: chronic desync, where your commit and the reviewers' findings are always one round apart.

The fix: a sync barrier immediately before push. Re-fetch ALL sources (comments, Sonar, CI) one more time. If ANY new finding has arrived since you last looked, loop back to step 2 -- address those new findings in the SAME upcoming push -- then re-fetch again. Only push when a fresh fetch comes back with no new findings against the current HEAD. This guarantees you and the reviewers are synchronized.

bash .agents/skills/pr-reviews/scripts/fetch-all.sh <PR_NUMBER>
bash .agents/skills/pr-reviews/scripts/fetch-sonar-findings.sh <PR_NUMBER>
bash .agents/skills/pr-reviews/scripts/ci-status.sh <PR_NUMBER>

The ci-status.sh line is the third source: a CI failure that is CAUSED by this PR's changes (added a script that doesn't pass shellcheck, broke a YAML parse, etc.) is in scope and must be folded in. CI failures unrelated to this PR are noted, surfaced to the user at the end, but not fixed here.

If summary.txt shows any new open thread or sonar-issues.json shows any new issue you haven't addressed yet, do NOT push. Loop back to step 2 and address them first. Then re-run the fetch. Only push when the fetch is clean.

The same loop applies during the iteration: if you re-fetched while addressing the previous batch and saw new findings drop in, fold them into the same push rather than dispatching a half-done batch.

4a. Before pushing -- MANDATORY holistic PR review via subagent

This is the most important pre-push step. Skipping it is what makes review cycles last for hours.

After you have made all the fixes for the current iteration's findings but BEFORE running git push, spawn a subagent to re-review the WHOLE PR diff (not the small change you just made). The orchestrator's context is already loaded with the recent fixes; the subagent's clean context is what gives an honest second look.

Why this is non-negotiable:

• AI reviewers (cubic-dev-ai, coderabbitai, sonarqube) only surface their top 3-7 findings. The full set of similar issues remains hidden.
• Each fix can introduce its own new problems (a printf format-string fix that breaks color rendering, a portability fix that drops a feature, an input-validation fix that rejects valid inputs).
• Without a holistic pre-push review, every iteration takes ~30 min of bot review latency just to discover problems the orchestrator could have spotted in 2 minutes by re-reading the diff.

How to invoke:

Use the orchestrator's Agent / subagent tool (whatever the harness provides). Pass the subagent the PR diff (or the list of touched files) and ask it to:

• Verify each fix in this iteration solves the original finding without side effects.
• Sweep the touched files for similar patterns the original findings did not point at, but which the same reasoning would flag.
• Sweep the touched files for NEW issues the fixes themselves may have introduced (broken behavior, lost features, regressions).
• Report findings as a flat list -- file:line + class + suggested fix.

Then the orchestrator addresses every finding the subagent returns BEFORE push. Loop the subagent if necessary until it returns a clean review. Only then proceed to step 4b.

A good subagent prompt template:

Re-review PR <N> end-to-end. The current diff is on branch <X>; the base is <master|...>. Recent fixes addressed: <list>. For the WHOLE diff (not just the recent fixes), find:

1. Similar patterns to the ones recently fixed that were NOT pointed at by reviewers but where the same reasoning applies.
2. Issues the recent fixes may have introduced (regressions, broken behavior, dropped features).
3. Anything in the diff that does not match the project's conventions (AGENTS.md, sibling files, the rest of the repo). Report a flat list of file:line + class + suggested fix. Be exhaustive; do not stop at 3-7 findings.

4b. Before pushing -- check CI for FAILURES (don't wait)

bash .agents/skills/pr-reviews/scripts/ci-status.sh <PR_NUMBER>

Exit codes:

• 0 -- all green, safe to push
• 2 -- runs in progress -- IGNORE this; push anyway. Waiting for CI between iterations destroys throughput. The new push triggers fresh CI on the new code, which is what matters.
• 3 -- runs failing -- fix the failures and bundle them into the push.

CI failures unrelated to this PR (a flaky test on a different module, an infra outage) are NOT in scope for this PR -- note them, surface to the user, move on. Do not make drive-by fixes here.

5. Push, then re-trigger reviewers

After pushing the fix commit(s):

bash .agents/skills/pr-reviews/scripts/trigger-cubic.sh      <PR_NUMBER>
bash .agents/skills/pr-reviews/scripts/trigger-coderabbit.sh <PR_NUMBER>

cubic and coderabbit re-review when mentioned in a new top-level PR comment. Copilot is deliberately absent: see rule 8. If you add a reviewer to PR_AI_BOT_RE, add its re-trigger here in the same change -- a classified reviewer with no re-trigger is silently skipped every iteration.

6. Wait for new activity

bash .agents/skills/pr-reviews/scripts/wait-for-activity.sh <PR_NUMBER>

Default timeout 30 min, poll every 30 s. Returns 0 on new activity, 124 on timeout. Both bots typically post a "no new findings" comment when they have nothing left, so the loop ends naturally on a clean PR.

What counts as "new activity":

• New issue comment / review comment / review on the PR.
• New commit pushed to the PR head.
• A review thread getting resolved or unresolved (often by a bot saying "addressed; resolving" -- without this signal we'd miss thread state flips and time out spuriously).

7. Loop

Go back to step 1. Continue until ALL of these are true:

• fetch-all.sh reports all review threads resolved.
• fetch-sonar-findings.sh reports zero open issues / hotspots that this PR introduced (or the remaining ones are explicitly marked FP / WontFix).
• The re-triggered AI bots (cubic-dev-ai, coderabbitai) have posted a "no new findings" or equivalent comment after their most recent re-trigger. Copilot is not re-triggered and never gates this.
• ci-status.sh reports no failures caused by this PR (failures unrelated to the PR are noted, surfaced to the user, but not fixed).

When wait-for-activity.sh times out (30 min):

• Re-run ci-status.sh. If checks are still running, surface to the user and stop -- the PR is in a clean intermediate state.
• If there are CI failures attributable to the PR, treat them as a new finding and iterate (commit -> ci-check -> push -> retrigger -> wait).
• If the failures are unrelated, note them in the final report.

8. Final report

When the loop ends, summarize for the user:

• Findings addressed (count by source: review threads, Sonar, CI).
• Any unrelated CI failures observed but not fixed (with check name + URL).
• Any human comments that need their attention.
• Current PR state (mergeable / blocked, decision, head SHA).

Commit message hygiene

Commit messages on the address-the-comments cycle should describe the change, not the reviewer or the cycle:

• BAD: "address copilot comments"
• BAD: "fix bot review feedback"
• GOOD: "scripts: fix dry-run env var name and printf format-string usage"

Never reference an AI tool by name in commit messages or PR bodies. The work matters; the tool that flagged it does not.

(Comments on the PR are an exception when they're operational mentions required by the bot itself: @cubic-dev-ai please review again is a direct trigger for that bot, and the trigger script enforces it. Outside operational triggers, the same rule applies to comments.)

Replying to bots -- tone

Be substantive but brief. The bot's prompt-text is verbose; your reply doesn't have to be. Examples:

• For a valid fix: "Fixed in <sha-or-paragraph>: <one-sentence what changed>."
• For a false positive: "False positive -- <one-sentence why>: <evidence citation>." Add a code comment if it'll help future reviewers.
• For a partial fix: "Partial -- fixed the immediate case at <line>, but the related <other-line> is intentional because <reason>."

Replying to humans -- consult the user

Maintainer / dev / community comments go to the user FIRST. Your message should:

1. Quote the relevant part of their comment.
2. State your read of what they're asking for.
3. Propose 1-3 options if it's a design call, or one option if obvious.
4. Wait for the user's decision.

Then act per their direction. Do not respond to humans on the user's behalf without explicit direction.

Bot directory

If a new AI reviewer appears in the project, classify it by adding to PR_AI_BOT_RE in _lib.sh so the skill recognizes it.

Reviewer-specific notes

• The mention is @coderabbitai, never @coderabbit. @coderabbit is a real, unrelated GitHub user; mentioning it pings a stranger on every iteration and never reaches the bot. trigger-coderabbit.sh hardcodes the correct handle -- do not hand-write the mention. The same care applies to @cubic-dev-ai. Before posting any comment containing an @, check the handle against the bot directory above.

• coderabbitai[bot] posts line-level findings, not just summaries. It was originally classified here as informational; it is an AI reviewer and its threads need the same verify-reply-resolve treatment as cubic's.

• coderabbit cites external URLs (learn.netdata.cloud, upstream GitHub) as evidence. Those citations are often directionally right but not authoritative for the branch under review -- verify against the source in the checkout before acting. Example: it correctly flagged that a Function needs a signed-in identity, but the proof is the HTTP_ACCESS_* flags in the producer, not the doc page it linked.

• Both reviewers will flag a generated page's content when the real defect is in the generator or the shared template. Fix the producer, regenerate, and say so in the reply -- otherwise the same finding returns on the next vendor page.

Failure modes -- quick diagnosis

MANDATORY -- keep this skill alive

If you (the agent) discover a new pattern, gotcha, working flow, correction, or any piece of knowledge while running this skill -- update this SKILL.md AND commit it BEFORE proceeding. Knowledge that isn't committed is lost.

Examples of things to capture:

• A new AI reviewer bot that appears in the project (add to the directory + PR_AI_BOT_RE)
• A new common false-positive pattern that warrants a clarifying source comment
• A new GitHub API quirk (rate limits, undocumented response shapes, pagination edge cases)
• A retrigger mechanism that changed (e.g. copilot's re-request behavior)