PR Review Check

TizenFX AI PR Review Feedback Application Pipeline

Overview

For open PRs in samsung/TizenFX that carry the ai-task label:

1. Human reviewer comments → always attempt to apply (ask if ambiguous)
2. AI reviewer comments (🤖 [AI Review] left by pr-code-review) → apply the decision matrix, then apply or respond
3. Only push when the build passes

All comments and commit messages are written in English.

🔁 Infinite-loop convergence mechanism (important)

pr-code-review re-reviews whenever new commits are added. So each time AI feedback is applied and committed, additional AI review can be triggered — a bounded termination mechanism is required.

• Cap: MAX_AI_ROUNDS = 3
• Counter: number of commits on the current PR branch whose message contains an Applied-AI-Comments: trailer
• On exceed: AI comments are neither applied nor responded to — fully skipped (recorded as max-ai-rounds in the report)
• Human comments are not capped (they are not the cause of the loop)

Repository

• Repo: samsung/TizenFX (GitHub)
• Local clone: a pre-cloned local working directory (e.g., ~/src/TizenFX)
• CLI: gh, git, dotnet (all authenticated/installed)

Stage ①: List Target PRs

gh pr list --repo samsung/TizenFX --label "ai-task" --state open \
  --json number,title,headRefName,baseRefName,updatedAt,isDraft,url \
  --jq '[.[] | select(.isDraft | not)]
         | sort_by(.updatedAt) | reverse
         | .[] | @json'

• draft PRs are skipped
• Sort by most recently updated, max 5 PRs per run

Stage ②: AI Round Count → AI Comment Handling Mode

Counting via gh api is possible before checkout, but git log after checkout is faster and more accurate. So measurement happens after Stage ⑤.

Handling mode:

• AI_ROUNDS < 3 → active: apply the AI comment decision matrix
• AI_ROUNDS ≥ 3 → skip: bypass AI comment handling entirely (no application, no response)

Stage ③: Collect New Comments (separated: human / AI)

Core rule: consider only new comments after the last commit timestamp.

LAST_COMMIT_AT=$(gh api repos/samsung/TizenFX/pulls/{NUMBER}/commits \
  --jq '.[-1].commit.committer.date')

Outdated-comment exclusion rule: review comments with position == null are outdated — the diff has evolved and the anchor is gone. These are likely unrelated to the current code, so skip them at the collection stage. (Issue comments have no line anchor, so "outdated" does not apply.)

Collect human comments (review + issue):

HUMAN_REVIEW=$(gh api repos/samsung/TizenFX/pulls/{NUMBER}/comments \
  --jq ".[] | select(.created_at > \"$LAST_COMMIT_AT\")
             | select(.body | startswith(\"🤖 [AI Review]\") | not)
             | select(.position != null)
             | {id, path, line, body, user: .user.login, created_at}")

HUMAN_ISSUE=$(gh api repos/samsung/TizenFX/issues/{NUMBER}/comments \
  --jq ".[] | select(.created_at > \"$LAST_COMMIT_AT\")
             | select(.body | startswith(\"🤖 [AI Review]\") | not)
             | {id, body, user: .user.login, created_at}")

Collect AI comments (review only — AI only leaves inline comments):

AI_REVIEW=$(gh api repos/samsung/TizenFX/pulls/{NUMBER}/comments \
  --jq ".[] | select(.created_at > \"$LAST_COMMIT_AT\")
             | select(.body | startswith(\"🤖 [AI Review]\"))
             | select(.position != null)
             | {id, path, line, body, in_reply_to_id, created_at}")

Stage ④: Delta Judgment — Proceed / Skip

• Human comments ≥ 1 → proceed (always)
• Human comments == 0 && AI mode == active && AI comments ≥ 1 → proceed
• Human comments == 0 && (AI mode == skip || AI comments == 0) → skip (no-delta)

Stage ⑤: Branch Checkout + Merge-conflict Detection

cd {TIZEN_FX_LOCAL_PATH}
git fetch origin
gh pr checkout {NUMBER}
git pull --rebase origin {headRefName}

On merge conflict:

if git status --porcelain | grep -qE '^(UU|AA|DD)'; then
  git rebase --abort 2>/dev/null
  git reset --hard HEAD
  # Record reason `merge-conflict` and move to next PR
fi

After checkout, measure AI round count (to finalize Stage ② decision):

AI_ROUNDS=$(git log "origin/{baseRefName}..HEAD" \
  --grep="^Applied-AI-Comments:" --oneline | wc -l)

Stage ⑥: Apply Human Comments

For each human comment:

1. Inspect path, line, body to understand the requested change
2. Evaluate from a .NET / C# / Tizen perspective
3. Action:
   • Valid → modify the code (add the comment id to Applied-Human-Comments)
   • Ambiguous → do not modify; instead ask a question via reply in Stage ⑨
   • Clearly incorrect → do not modify; reply in Stage ⑨ with a factual disagreement

Application principles:

• No public API signature changes
• Re-verify any change that could break the build

Stage ⑦: AI Comment Handling (only when AI mode == active)

If AI mode is skip, skip this entire stage. Do not even leave response comments.

7-1. Priority Filter (first pass)

"Not every AI comment deserves to be applied." Only meaningful, reasonable suggestions qualify. Because pr-code-review leaves a severity marker (🔴/🟡) on each comment, reuse it as the first-pass filter.

Ignored items do not receive a response comment either (to prevent noise). This mirrors pr-code-review's "no-nitpick" policy.

7-2. Decision Matrix (applied only to items passing the first-pass filter)

"Already addressed" judgment guide:

• Check the diff of the last ~3 commits at that path:line
• See whether the flagged pattern has already been removed/improved
• When unsure, do not classify as "valid + unapplied" — prefer "already addressed" for safety (loop prevention)

Applied AI comment IDs are added to the Applied-AI-Comments list.