Workflow Address Feedback

Workflow: Address Feedback

Announce at start: "I'm using the workflow-address-feedback skill to address review feedback on PR #N."

When to invoke

• The user runs /swe-workbench:address-feedback .
• A PR owner wants to systematically work through review threads.
• Phrases: "address the feedback on PR 123", "help me resolve review comments", "triage and fix the review threads on #456".

When NOT to invoke

• The reviewer side of the loop → use swe-workbench:workflow-pr-review or swe-workbench:workflow-pr-review-followup.
• The user just wants to reply to a single comment without the full triage flow.
• The PR is closed/merged.

Composition

This skill orchestrates:

• swe-workbench:ticket-context — prepended context when PR references a ticket.
• swe-workbench:workflow-commit-and-pr — invoked after all ADDRESSED fixes are applied to commit and push.

Phase flow

Phase 1 — Pre-flight + fetch

_RT="${CLAUDE_PLUGIN_ROOT:-$(git rev-parse --show-toplevel)}"
[ -f "$_RT/runtime/clean-state-files.sh" ] || {
  echo "swe-workbench runtime scripts not found under $_RT/runtime — set CLAUDE_PLUGIN_ROOT and retry." >&2
  exit 1
}
JSON="/tmp/swe-workbench-address-feedback/${PR}.json"
eval "$("$_RT/runtime/preflight-pr.sh" "$PR" "$JSON")"
CURRENT_USER=$(gh api /user -q .login)
PR_BRANCH=$(jq -r .headRefName "$JSON")

preflight-pr.sh handles gh auth status, fetches the PR JSON to $JSON, and emits BASE, HEAD_SHA, AUTHOR_LOGIN, OWNER, REPO, STATE as shell assignments. PR_BRANCH is derived from headRefName in $JSON (address-feedback uses it for worktree setup in Phase 2).

Check that the PR is open before proceeding:

[ "$STATE" = "OPEN" ] || { echo "PR #$PR is $STATE — address-feedback only applies to open PRs."; exit 1; }

If CURRENT_USER != AUTHOR_LOGIN, warn: > "You are not the PR author (PR author: @AUTHORLOGIN, you: @CURRENTUSER). Address-feedback flows are typically owner-side. Continue anyway? Reply yes to proceed."

Wait for confirmation before continuing.

Fetch outstanding review threads via GraphQL:

gh api graphql -F number="$PR" -F owner="$OWNER" -F repo="$REPO" -f query='
  query($owner: String!, $repo: String!, $number: Int!) {
    repository(owner: $owner, name: $repo) {
      pullRequest(number: $number) {
        reviewThreads(first: 100) {
          nodes {
            id isResolved path line startLine
            comments(first: 10) {
              nodes {
                id databaseId body
                author { login }
              }
            }
          }
        }
      }
    }
  }' > "/tmp/swe-workbench-address-feedback/${PR}-threads.json"

Fetch PR-level conversation comments (general feedback on the main timeline, not a line comment) via REST, paginated (--paginate on this array endpoint emits one concatenated array per page, so --jq '.[]' | jq -s '...' flattens then re-wraps into one array); exclude bots and the owner ($AUTHOR_LOGIN or $CURRENT_USER, since this phase allows non-author runs — otherwise a non-author's own past replies would resurface as new triage items on every re-run), then flag eligible: false on reviewer comments already handled on a prior run — (a) an owner comment carries their swe-workbench:handled:{id} marker, or (b) an owner comment without any handled marker was posted after them (a manual reply); this dedup is lossy by construction, so Phase 3 always surfaces a transparency note instead of silently dropping:

gh api --paginate "repos/${OWNER}/${REPO}/issues/${PR}/comments" --jq '.[]' | jq -s \
  --arg author "$AUTHOR_LOGIN" --arg me "$CURRENT_USER" '
    (map(select((.user.login // "") == $author or (.user.login // "") == $me))) as $owner
    | map(select(((.user.type // "") != "Bot") and (((.user.login // "") | endswith("[bot]")) | not) and ((.user.login // "") != $author) and ((.user.login // "") != $me)))
    | map(. as $c
        | ($owner | any((.body // "") | contains("swe-workbench:handled:" + ($c.id | tostring) + " "))) as $marker
        | ($owner | any((((.body // "") | contains("swe-workbench:handled")) | not) and (.created_at > $c.created_at))) as $manual
        | $c + {eligible: (($marker or $manual) | not)})
  ' > "/tmp/swe-workbench-address-feedback/${PR}-pr-comments.json"
ELIGIBLE_PR_COMMENTS=$(jq '[.[] | select(.eligible)] | length' "/tmp/swe-workbench-address-feedback/${PR}-pr-comments.json")
SKIPPED_PR_COMMENTS=$(jq '[.[] | select(.eligible | not)] | length' "/tmp/swe-workbench-address-feedback/${PR}-pr-comments.json")

If all threads are resolved (or no threads exist) and $ELIGIBLE_PR_COMMENTS is zero, print: > "No open threads — nothing to address." Then exit cleanly.

If a prior triage save exists at /tmp/swe-workbench-address-feedback/${PR}-triage.json, offer to resume from it.

Phase 2 — Worktree

First, check whether the current branch already matches the PR head — if so, reuse the current worktree instead of creating a new one:

CURRENT_BRANCH=$(git rev-parse --abbrev-ref HEAD)
if [ "$CURRENT_BRANCH" = "$PR_BRANCH" ] && [ "$CURRENT_BRANCH" != "HEAD" ]; then
  WT=$(pwd)
  REUSED_WT=1
  echo "Already on PR branch '$PR_BRANCH' — reusing the current worktree at $WT (skipping rimba add)."
  DIRTY=$(git status --porcelain)
  [ -n "$DIRTY" ] && echo "Note: working tree has uncommitted changes; the user may stash before Phase 4 commits to avoid sweeping unrelated edits into the feedback commit."
fi

If $WT is not yet set, check whether a worktree for $PR_BRANCH already exists elsewhere on disk (e.g. the session is on main but the branch was checked out previously):

if [ -z "$WT" ]; then
  EXISTING_WT=$(git worktree list --porcelain \
    | awk 'BEGIN{wt=""} /^worktree /{wt=$2} /^branch refs\/heads\/'"$PR_BRANCH"'/{print wt; exit}')
  if [ -n "$EXISTING_WT" ] && [ -d "$EXISTING_WT" ]; then
    WT="$EXISTING_WT"
    REUSED_WT=1
    echo "Found existing worktree for '$PR_BRANCH' at $WT (skipping rimba add)."
  fi
fi

If $WT is set by either check above, skip the rest of Phase 2 and proceed to Phase 3 — when the tree was dirty (first guard only), a non-blocking warning was already emitted; the user may stash before Phase 4 commits. Otherwise create a new durable worktree:

When rimba is available (preferred):

RIMBA_OUT=$(rimba add "pr:$PR" --task "address-feedback-$PR" --skip-deps --skip-hooks 2>&1)
WT=$(echo "$RIMBA_OUT" | awk '/Path:/{print $2}')
[ -d "$WT" ] || { echo "rimba add failed: $RIMBA_OUT"; exit 1; }

When rimba is absent (fallback):

WT="$HOME/.local/share/swe-workbench/address-feedback-${PR}"
mkdir -p "$(dirname "$WT")"
git fetch origin "${PR_BRANCH}"
git worktree add "$WT" "${PR_BRANCH}"

This worktree is disposable — fixes are committed and pushed to the PR branch in Phase 4, so the work lives on the remote, not the worktree. Phase 7 removes it on every exit (success, Q-quit, or error). If removal fails, a fallback is attempted; see Phase 7 for details. If the skill exits with an unrecoverable error at any point after this phase, run Phase 7 before stopping.

Phase 3 — Triage digest

Render outstanding threads and eligible PR-level conversation comments, one by one. Filter out before presenting:

1. Resolved threads — skip any thread where isResolved == true.
2. Already-clarified threads — skip any unresolved thread where at least one reply comment (comments.nodes[1:] onwards — nodes[0] is the thread-opening comment, which in the typical reviewer-opened case belongs to the reviewer, not the PR owner) has author.login equal to $CURRENT_USER. This means the owner replied in a prior pass (e.g. a CLARIFIED reply) but left the thread unresolved. It applies whether that reply was posted by this skill or manually by the user. Detecting via reply comments only prevents false-positive skipping when the current user also authored review threads.
3. Already-handled PR comments — bot/tool comments and the owner's own comments never made it into ${PR}-pr-comments.json (dropped in Phase 1); entries that did but carry eligible == false were deduped there (already marker-replied or manually replied to on a prior run). Only eligible == true entries are presented.

If any threads or PR comments were skipped (rule 2 / rule 3), print transparency notes before the digest: > "(N thread(s) skipped — already clarified.)" > "(N PR comment(s) skipped — already handled.)"

If no threads and no eligible PR comments ($ELIGIBLE_PR_COMMENTS == 0) remain after filtering:

• When any items were skipped under rule 2 or rule 3: print "No new items to triage — N already clarified/handled."
• When nothing was skipped (only resolved threads filtered, and no PR comments existed): print "No new items to triage."

Then run Phase 7 — Cleanup and exit cleanly.

For each remaining thread:

─────────────────────────────────────────────────
Thread #ID — {path}:{line}  by @{author}  [{Severity if parseable}]
─────────────────────────────────────────────────
> {first 200 chars of comment body}

[A]ddressed — fix + commit + reply + resolve
[C]larified — reply only (no resolve)
[D]eferred — skip this thread
[Q]uit — save progress and exit

Parse severity from Severity: prefix in comment body if present; otherwise label Unknown.

Capture: triage[] = A|C|D.

For each remaining PR comment (no path:line, no resolve state), key as triage["prcomment:"] (namespaced against review-thread node IDs in the same flat map; carries the id needed for the Phase 5 marker; both key kinds round-trip through Q-quit save/resume unchanged):

PR comment by @{author}
> {first 200 chars of comment body}

[A]ddressed — fix + commit + reply (PR comments have no thread to resolve)
[C]larified — reply only
[D]eferred — skip this comment
[Q]uit — save progress and exit

If the owner replies Q at any point in either loop, save triage state to /tmp/swe-workbench-address-feedback/${PR}-triage.json, then run Phase 7 — Cleanup, and exit. Re-invocation resumes from this file (Phase 2 re-creates the worktree).

Phase 4 — Implement + commit

For each ADDRESSED review thread or PR comment (in order — both sources share this loop, since the commit step is source-agnostic):

1. Show the finding and the relevant file/line context (PR comments have no path:line; show the comment body instead).
2. Ask the owner for the fix approach (free-text). If the comment already contains a ### Suggested fix block, offer to apply it automatically via the Edit tool.
3. Apply edits using the Edit tool.

After all ADDRESSED fixes are applied, invoke swe-workbench:workflow-commit-and-pr with the prompt: > "commit and push these fixes addressing review feedback on PR #N"

This reuses the existing [type] commit format, branch-naming check, and push logic. After the skill returns, capture the resulting commit SHA:

FIX_SHA=$(git -C "$WT" rev-parse HEAD)

Phase 5 — Reply + resolve

For each ADDRESSED or CLARIFIED review thread, post a reply via REST then conditionally resolve (DEFERRED threads skip this call entirely). Use comments.nodes[0].databaseId (the thread root comment) as $COMMENT_DATABASEID — replies must target the first comment in the thread, not a subsequent reply.

Reply body templates by triage classification:

• ADDRESSED: "Addressed in ${FIX_SHA}: ." — pass both $REPLY_BODY and $THREAD_ID.
• CLARIFIED: free-text owner-authored reply (asked interactively) — pass $REPLY_BODY with empty $THREAD_ID (reply only, no resolve).
• DEFERRED: pass empty $REPLY_BODY and empty $THREAD_ID (neither reply nor resolve).

bash "$_RT/runtime/reply-and-resolve.sh" \
  "$OWNER" "$REPO" "$PR" "$COMMENT_DATABASEID" "$THREAD_ID" "$REPLY_BODY"

For each ADDRESSED or CLARIFIED PR comment (triage["prcomment:"]), post a reply on the PR's top-level conversation instead of a thread reply — PR comments have no thread, so resolve is always suppressed and KIND=issue is passed explicitly. Compose $REPLY_BODY as @{comment author} re: + a single-line blockquote of the original (first ~100 chars, newlines collapsed) + the addressed/clarified body (same wording as the review-thread templates above) + a hidden marker on its own line — ` — which Phase 1's dedup filter matches on re-runs (omitting it makes the comment look unhandled forever). The original comment body is attacker-controlled: extract the blockquote via jq -r into a shell variable (e.g. QUOTE=$(jq -r '.[] | select(.id==ID) | .body' ... | head -c 100 | tr '\n' ' ')) and reference "$QUOTE" when building $REPLY_BODY — never retype the raw comment text into a double-quoted bash literal, since $(...)`/backticks in the source text would execute at assignment time. DEFERRED PR comments skip the call entirely, same as DEFERRED threads:

bash "$_RT/runtime/reply-and-resolve.sh" \
  "$OWNER" "$REPO" "$PR" "" "" "$REPLY_BODY" "issue"

After all replies and resolutions land, emit the follow-up CTA:

> "Want me to ping the reviewer to re-check? Reply yes to run /swe-workbench:review --check-followup ."

On the Phase 5 success path, delete the address-feedback state files. The reap runs foreground; failures surface (no 2>/dev/null):

bash "$_RT/runtime/clean-state-files.sh" \
  "/tmp/swe-workbench-address-feedback/${PR}.json" \
  "/tmp/swe-workbench-address-feedback/${PR}-threads.json" \
  "/tmp/swe-workbench-address-feedback/${PR}-pr-comments.json" \
  "/tmp/swe-workbench-address-feedback/${PR}-triage.json"
for f in "/tmp/swe-workbench-address-feedback/${PR}.json" \
         "/tmp/swe-workbench-address-feedback/${PR}-threads.json" \
         "/tmp/swe-workbench-address-feedback/${PR}-pr-comments.json" \
         "/tmp/swe-workbench-address-feedback/${PR}-triage.json"; do
  [ -e "$f" ] && echo "⚠ state file NOT reaped: $f" >&2 || echo "✓ state file reaped: $f"
done

Then run Phase 6 — Sync PR metadata.

Phase 6 — Sync PR metadata (when fixes were committed)

Skip this phase entirely if $FIX_SHA is unset (no fixes were committed in Phase 4).

Fetch: gh pr view "$PR" --json title,body; commit subjects via git -C "$WT" log "$BASE"..HEAD --format='%s'; diff stat via git -C "$WT" diff "$BASE"..HEAD --stat. No new state file — the reap already ran in Phase 5 (covers ${PR}-pr-comments.json too).

Judge drift by comparing the live title and ## Summary section of the PR body against the commit subjects and diff stat. If aligned, emit "PR metadata is up to date — no changes needed." and fall through to Phase 7 — Cleanup.

If drift is detected, draft a revised $NEW_TITLE and $NEW_SUMMARY. Rewrite only the ## Summary section of the body; preserve ## Test Plan, the Closes #/Fixes #/Issue: N/A trailer, and all other collaborator sections. Preview old→new for title and summary, then:

> Reply yes to apply these changes to PR #N.

On yes:

NEW_BODY_FILE=$(mktemp)
trap 'rm -f "$NEW_BODY_FILE"' EXIT
printf '%s' "$NEW_BODY" > "$NEW_BODY_FILE"
bash "$_RT/runtime/sync-pr-metadata.sh" "$PR" "$NEW_TITLE" "$NEW_BODY_FILE" \
  || echo "Warning: PR metadata update failed — continuing to cleanup." >&2

Then run Phase 7 — Cleanup.

Phase 7 — Cleanup (always)

Run on every exit that occurs after a worktree was created in Phase 2 (success, Q-quit, or error). Skip on Phase 1 early-exits (before any worktree exists) and when the reuse-guard fired (REUSED_WT=1) — the reuse path sets $WT to an existing checkout, never creates a worktree, so there is nothing to remove.

if [ "${REUSED_WT:-0}" = "1" ]; then
  echo "Reused existing worktree at $WT — skipping cleanup (nothing was created)."
else
  # positional arg matches the --task label set in Phase 2 ("address-feedback-$PR")
  if rimba remove "address-feedback-$PR" --force 2>/dev/null; then
    echo "Cleaned up worktree address-

…

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [lugassawan](https://github.com/lugassawan)
- **Source:** [lugassawan/swe-workbench](https://github.com/lugassawan/swe-workbench)
- **License:** MIT

Install and usage instructions live in the source repository linked above.