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 <N>.
• 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 (first-pass or followup mode).
• 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

command -v swe-workbench-address-feedback-fetch >/dev/null 2>&1 || {
  echo "swe-workbench runtime commands not on PATH — reinstall or update the swe-workbench plugin." >&2
  exit 1
}
RESULT=$(swe-workbench-address-feedback-fetch --pr "$PR" \
  | swe-workbench-result-check swb.address-feedback-fetch/1) || exit 1
STATE=$(printf '%s' "$RESULT" | jq -r '.data.state')
JSON=$(printf '%s' "$RESULT" | jq -r '.data.pr_json_path')
[ "$STATE" = "OPEN" ] || { swe-workbench-clean-state-files "$JSON"; echo "PR #$PR is $STATE — address-feedback only applies to open PRs."; exit 1; }
OWNER=$(printf '%s' "$RESULT" | jq -r '.data.owner')
REPO=$(printf '%s' "$RESULT" | jq -r '.data.repo')
AUTHOR_LOGIN=$(printf '%s' "$RESULT" | jq -r '.data.author_login')
CURRENT_USER=$(printf '%s' "$RESULT" | jq -r '.data.current_user')
PR_BRANCH=$(printf '%s' "$RESULT" | jq -r '.data.pr_branch')
THREADS_PATH=$(printf '%s' "$RESULT" | jq -r '.data.threads_path')
PR_COMMENTS_PATH=$(printf '%s' "$RESULT" | jq -r '.data.pr_comments_path')
TRIAGE_PATH=$(printf '%s' "$RESULT" | jq -r '.data.triage_path')
RESUME_TRIAGE_PATH=$(printf '%s' "$RESULT" | jq -r '.data.resume_triage_path')
ELIGIBLE_THREADS=$(printf '%s' "$RESULT" | jq -r '.data.eligible_threads')
SKIPPED_THREADS_CLARIFIED=$(printf '%s' "$RESULT" | jq -r '.data.skipped_threads_clarified')
ELIGIBLE_PR_COMMENTS=$(printf '%s' "$RESULT" | jq -r '.data.eligible_pr_comments')
SKIPPED_PR_COMMENTS=$(printf '%s' "$RESULT" | jq -r '.data.skipped_pr_comments')
RUN_DIR=$(swe-workbench-new-run-dir address-feedback "$PR")

swe-workbench-address-feedback-fetch handles gh auth status, fetches the PR JSON to $JSON (via swe-workbench-preflight-pr), and — when the PR is OPEN — paginates review threads, PR-level conversation comments and submitted-review bodies, projecting eligible/skip_reason onto each entry. A valid generated PR batch — in a conversation comment, or in a review body (the posting core folds PR-level findings into the review: summary first, then the batch marker on its own line) — becomes one projected PR feedback item per finding, each with .parent_comment_id (the comment id, or review-<review-id> for a review body), .finding_id, .triage_key (prcomment:<id>:finding:<n> or review:<id>:finding:<n>), .handled_marker, and the extracted .body (the trailing finding-key bookkeeping line is stripped); unmarked third-party comments stay one legacy item, a plain review summary is not feedback, and a marked batch that fails strict parsing surfaces as one legacy item. Bot exclusion remains absolute; valid batches from the PR author/current runner are retained. It writes those projections to $THREADS_PATH/$PR_COMMENTS_PATH. The [ "$STATE" = "OPEN" ] gate runs immediately after the fetch and before $RUN_DIR is allocated, so a rejected PR reaps $JSON inline via swe-workbench-clean-state-files rather than leaking it — $RUN_DIR never exists on this path, so there is nothing else to reap. new-run-dir.sh allocates $RUN_DIR — a mode-0700 scratch directory under /tmp/swe-workbench-run/ for this run's own ad-hoc bash artifacts, distinct from the deliberate PR-keyed state files above (including $TRIAGE_PATH, which is a cross-invocation resume point and must never move here). All four state paths are repo-scoped: the fetch resolves the owner-repo slug itself — explicit --repo when the invocation carried a full PR URL, else the checkout's origin remote, else legacy un-scoped names — so same-numbered PRs in different repositories never collide.