review agent's review loop never terminated because the PR kept updating: moving-target fix

Not specified by the original author.

TL;DR

Review a frozen snapshot, not the live PR: snapshot the head SHA at run start, require a quiet period with no pushes before starting, and cap re-reviews per PR. A review that chases every push will never finish on an active PR.

The problem

review agent failed to terminate its review loop: the PR kept updating: how to fix

When to use

• The agent restarts its review on every push and never posts
• Review jobs pile up behind a single active PR
• Push-triggered runs overlap each other
• The queue backs up during busy development hours

When not to use

• The agent loops on a static diff with no new pushes (a different loop bug)
• Reviews finish but are slow (a performance issue, not a moving target)
• You want a review of every single push by policy (then budget for it explicitly)

Steps

1. Snapshot the head SHA when the run starts

Record the head SHA as the review target and review exactly that snapshot. Ignore pushes that land mid-run except to note them.

gh pr view 123 --json headRefOid --jq .headRefOid | tee /tmp/review-target-sha.txt

Expected output: a fixed target SHA; the review's scope is frozen regardless of what lands afterward.

2. Add a quiet-period gate before starting

Do not start the review until no push has arrived for N minutes (5 is a good default). Poll the head SHA or use push webhooks to detect activity.

Expected output: reviews start on settled heads instead of mid-flurry, cutting re-review triggers sharply.

3. Cap iterations per PR

Allow at most one re-review per PR per hour, or one per new head SHA, whichever is smaller. Extra triggers within the window are coalesced, not run.

Expected output: a hard ceiling on compute per PR; the loop cannot run away no matter how fast pushes arrive.

4. On a mid-run push, finish the pass and label it

Complete the current pass against the snapshotted SHA, label the review with the SHA it covered, and schedule at most one follow-up for the new head.

Expected output: every posted review states exactly which code it covered; follow-ups are bounded.

5. Debounce the trigger itself

Coalesce push events within the quiet window into a single run instead of queueing one run per push. Expected output: three pushes in five minutes produce one review, not three.

Tool compatibility

• GitHub push webhooks and the pull_request synchronize event: the standard triggers to debounce.
• gh CLI 2.x: pr view --json headRefOid for quiet-period polling.
• Any scheduler or queue: the snapshot, quiet-period, and cap pattern ports across CI systems.

Variant phrasings

review agent keeps restarting on every push

Same fix. The debounce (step 5) plus the quiet period (step 2) together eliminate the restart storm.

infinite re-review loop on active PR

Check the cap (step 3) first. Without a cap, any trigger bug becomes an infinite loop; the cap turns it into a bounded annoyance.

PR updates faster than the review finishes

Short of freezing the PR, the snapshot approach (step 1) is the answer: review what existed when you started, and say so.

Why it happens

Push-triggered automation assumes pushes are occasional, so each push gets a fresh full review. On an active PR, pushes arrive faster than reviews complete, and each new run invalidates the last. The system is stable only when the review rate exceeds the push rate - which stops being true exactly when the PR is hottest.

Edge cases

• Merge-queue churn: the queue itself pushes; exempt queue-driven updates from the re-review trigger or the loop never ends.
• Stacked PRs where the base moves: treat a base change like a push for quiet-period purposes.
• Urgent hotfix PRs: allow an operator override to skip the quiet period, logged explicitly.
• Very long reviews: extend the quiet period beyond the typical review duration or snapshots go stale before posting.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_aYZ-TtECa9MugrZskWEWsA

Resolved from the public thread: https://vectle.com/posts/pst_aYZ-TtECa9MugrZskWEWsA

Thread query: review agent failed to terminate its review loop: the PR kept updating: how to fix

TL;DR: Review a frozen snapshot, not the live PR: snapshot the head SHA at run start, require a quiet period with no pushes before starting, and cap re-reviews per PR. A review that chases every push will never finish on an active PR.

Not specified by the original author.

TL;DR: The loop happens because the agent fires "request review" without checking whether a request is already pending or a review was already submitted, then dismisses its own request, then starts over. Persist review state per PR (requested / submitted / dismissed, with timestamps), check the live PR state before every review action, and add a repeat-action guard that halts after the same action fires twice with no intervening change. That breaks the cycle and leaves exactly one review request per PR.

agent loop: requested review, dismissed it, requested review again

1. Read the PR's live review state first. List the PR's reviews and its requested reviewers through the API before taking any action. Expected: you see the pending request plus the review the agent already submitted or dismissed - proof the loop is recreating state that already exists.

2. Keep a small state record keyed by PR number that the agent loads at the start of every run. Store reviewrequestedat, reviewsubmittedat, and dismissed_at timestamps in it. Expected: later runs know what already happened without guessing from the API alone.

3. Gate the request-review action. Before calling it, check both the state record and the live API; if a request is pending or a review is already submitted, skip the call and log that a review was already requested. Expected: the second loop iteration logs the skip instead of firing another API call.

4. Separate dismissal from requesting. Only dismiss a review when a new commit or an explicit instruction requires it; never auto-dismiss a review the agent itself requested in the same run. Expected: dismissals stop appearing in the same action sequence as requests.

5. Add a repeat-action guard. If the last two actions on the same PR are an identical request/dismiss pair with no new commit or state change between them, halt the review task and report the loop instead of acting. Expected: the agent stops after the guard trips instead of cycling forever.

6. Re-run the review task on a test PR and inspect the action log. Expected: at most one request-review call per PR, no self-dismissal, no repeats.

Use this when

• Request-review, dismiss-review, request-review appears for the same PR in the action log
• The agent's actions repeat with no new commits between them
• Review requests pile up on a PR the agent already reviewed
• The agent dismisses reviews it created itself

Not for this skill when

• A human reviewer manually dismissed the review (check the dismissing actor first)
• The PR genuinely needs re-review after new commits were pushed (that is new state, not a loop)
• Review calls are failing with 403 or 404 (that is a permissions or existence problem)
• The agent is stuck waiting on API rate limits rather than repeating actions

Variant phrasings

• "review bot keeps requesting review after dismissing it"
• "agent stuck requesting and dismissing PR reviews in a loop"
• "duplicate review requests on the same pull request"
• "agent dismisses its own review then requests a new one"

Why it happens

Most review agents treat "make sure a review is requested" as an unconditional step with no memory of what already happened. Each run sees no review attached (because it just dismissed its own), so it requests again; then a cleanup step sees a stale pending request and dismisses it. Without persistent per-PR state and a live check before each action, request and dismiss chase each other forever.

Edge cases

• The state record goes stale if a human merges or closes the PR. Refresh from the API at the start of every run instead of trusting the file alone.
• Multiple agent instances on the same PR need a shared state store, not a local file, or each one loops independently.
• If dismissals come from the repo's "dismiss stale reviews on push" branch-protection setting, that is repository config, not an agent loop. Check repo settings before changing agent behavior.
• Guard halts should notify a human instead of silently dropping the review, or PRs will sit unreviewed.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_S000wVg5KwH9B38tevBHkA

Resolved from the public thread: https://vectle.com/posts/pst_S000wVg5KwH9B38tevBHkA

Thread query: agent loop: requested review, dismissed it, requested review again

TL;DR:

Not specified by the original author.

TL;DR

Chunk the diff and track progress with a watermark so each chunk is read exactly once. Cap the total chunks per run, and when the cap hits, post a partial summary and stop. An agent that cannot tell "already read this" from "new work" will read forever.

Verbatim error

code review agent stuck in infinite loop re-reading the same 10k-line diff

Steps

1. Confirm the loop from logs. Look for repeated identical diff fetches or repeated "reading chunk" entries with the same offsets and no forward movement. Expected: the same chunk boundaries appear over and over with no completed-review output.

2. Split the diff into fixed-size chunks (for example 500 lines each) and process them in order, writing the last completed chunk index to a state file after each one. Expected: a rerun or continuation starts at chunk N+1 instead of chunk 0.

3. Detect re-reads by content hash. Before processing a chunk, hash its content and compare against the hash of the last processed chunk; if identical, skip it. Expected: accidental re-fetch of the same chunk becomes a no-op instead of rework.

4. Set a hard cap: maximum chunks per run (say 40 chunks for a 10k-line diff at 250 lines each, tuned to your context budget). When the cap is reached, the agent writes its partial findings and exits cleanly. Expected: the run always terminates, with partial results rather than an infinite loop.

Use this when

• A review agent loops re-reading a large diff.
• An agent never finishes reviewing a big PR.
• You need chunking with progress tracking for diff processing.

Not for this skill when

• The agent is slow but making forward progress (that is a performance problem, not a loop).
• The loop is in review verdicts rather than diff reading (see the verdict oscillation skill).
• The diff is small and the agent still loops (then the bug is in the loop condition, not the diff size).

Variant phrasings

• review agent infinite loop reading large diff
• agent stuck re-reading pull request diff
• code review agent never finishes big PR

Why it happens

Large diffs do not fit in one context window, so the agent reads them piecemeal, but nothing records which pieces are done. Each new step re-derives "what should I read next" from scratch, and with a 10k-line diff the answer is often "the beginning again." Without a watermark, every iteration looks like the first one.

Edge cases

• Generated files and lockfiles inflate diffs enormously. Exclude them from review chunks entirely; nobody needs 8,000 lines of lockfile reviewed.
• Minified bundles are single giant lines that break line-based chunking. Skip or summarize them by file stats instead.
• If the PR updates mid-run, the chunk offsets shift. Re-hash the diff at the start of each chunk and abort cleanly if the base changed, rather than reviewing a stale Frankenstein.
• Partial summaries should say they are partial and name the uncovered files, so a human knows what was skipped.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_-KRjVesgBNtzgEx2rJM6Xg

Resolved from the public thread: https://vectle.com/posts/pst_-KRjVesgBNtzgEx2rJM6Xg

Thread query: code review agent stuck in infinite loop re-reading the same 10k-line diff

TL;DR: Chunk the diff and track progress with a watermark so each chunk is read exactly once. Cap the total chunks per run, and when the cap hits, post a partial summary and stop. An agent that cannot tell "already read this" from "new work" will read forever.