Bound scheduled PR review intake across OSS repositories
Maintainer thường phản hồi trong vòng 1 ngày
@all-hands-bot đang làm issue này rồi.
Từ ngày 22/9/2026.
Đánh giá
Issue này chưa được đánh giá.
Mô tả
Problem
The GitHub PR reviewer can resume outstanding all-hands-bot requests on a scheduled scan (Extensions #654, stacked on #651), but one scan can start a conversation for every eligible PR it finds. The four main OSS repositories currently have hundreds of open non-draft PRs lacking a review on their current head. Requesting reviews for all of them at once would overload the 15 GiB OSS Agent Canvas VM; four simultaneous agent jobs plus retained Docker runtimes caused an OOM restart on 2026-09-22.
Desired behavior
The scheduled reviewer drains outstanding requests fairly while bounding new agent conversations. It should reuse the existing reviewer delivery and dedupe machinery, not create a separate queue or runtime layer.
Acceptance criteria
- A configurable small maximum (default 2) limits new review conversations per scheduled run across all configured repositories. Scanning CI, reconciling completed reviews, and maintainer handoff remain possible even when the new-conversation limit is reached.
- Eligible PRs are ordered deterministically and fairly by the oldest outstanding reviewer-request event, with repository and PR number as tie-breakers. A failing or pending PR does not consume a conversation slot or prevent later eligible PRs from being considered.
- Repeated scans never create a second conversation or native review for the same request event and head SHA. A later request or changed head can be reviewed under the existing semantics.
- A focused test with at least three eligible PRs across two repositories demonstrates that one scan starts at most two conversations and a later scan reaches the remaining PR. Include a pending-check candidate ahead of a green candidate to verify it does not block the queue.
- The live OSS Agent Canvas canary demonstrates bounded intake without exhausting memory. Include links to the PRs and conversations in the PR's AGENT evidence.
Scope
This is general reviewer scheduling in OpenHands/extensions. Do not add repo-specific policy or VM-specific control to the extension. Stack the PR on #654 if its scheduled-request scan is required; otherwise target the earliest applicable base. Do not merge automatically.
OpenHands AI triage
The following comments and acceptance criteria were added by the OpenHands AI agent.
Triage
The change is bounded scheduling work in the existing reviewer automation, so no further design decision is needed. It belongs in skills/github-pr-reviewer/scripts/worker.py, reusing the machinery #654 already added: the AgentConversationDispatcher.deliver delivery keyed by reviewer-request event and head SHA, _finish_completed_review, the exact-head check gate, and run_repositories. No queue or runtime layer is added. Because the maximum is per scheduled run across repositories while each repository is currently scanned independently, the budget has to be shared across the per-repository loop instead of resetting per repository; and because ordering is by the oldest outstanding request across repositories, the scheduled scan must collect and order candidates before dispatching, while keeping the existing per-repository failure isolation. Configure the maximum through the worker's rendered config, following the max_new_per_run key that github-issue-to-pr, gitlab-issue-to-mr, and github-agents-md-maintainer already use, with a default of 2.
Stack on #654, whose outstanding-request scan supplies the oldest-request ordering; #651 and #654 must merge first. Neither PR is merged or deployed automatically, and the OSS deployment change is out of scope here.
Non-goals: no repository-specific policy, VM tuning, or deployment control in OpenHands/extensions; no second automation record; no new GitHub webhook event types, check_run parsing, or permission scopes; no change to the review prompt, verdict parsing, maintainer handoff, secret selection, the trigger-label scan, or the explicit review_requested event path.
The issue's live canary depends on the OSS deployment accepting a scheduled dispatch; #654 recorded the Canvas automation service as unreachable for create and dispatch mutations on this host, so the PR should state the live evidence it actually obtained.
Acceptance Criteria
- A single scheduled run starts at most a configurable maximum of new review conversations across all configured repositories, defaulting to 2 when unset. The maximum bounds the conversations a run starts rather than the candidates it inspects, so deliveries that only deduplicate or report an already-running conversation do not consume it.
- After the maximum is reached, the same run still completes its non-dispatching work for the remaining eligible candidates: exact-head check evaluation and gate comments, completed-review reconciliation, and maintainer handoff. The scan is not cut short.
- Eligible candidates are ordered by the oldest outstanding
all-hands-botreview_requestedevent, then by repository, then by pull-request number. When more candidates are eligible than the maximum allows, the oldest-request candidates start conversations first and a later scheduled scan reaches the remainder. - A candidate whose exact-head checks are pending or failing posts its waiting or blocked disposition, starts no conversation, consumes none of the maximum, and does not prevent a later eligible candidate from being considered in the same run.
- A dispatch that raises is reported and does not abort the run or the other repositories, and does not prevent the remaining candidates from being considered.
- Repeated scheduled scans never start a second conversation or publish a second native review for the same reviewer-request event and head SHA; a later request event or a changed head is reviewed under the existing delivery-key semantics.
- The trigger-label scan and the explicit
review_requestedevent path keep their current behavior; only the scheduled scan's new conversations are bounded. - No new secret, token scope, or GitHub permission is required, and no secret value appears in logs, comments, or the PR.
- A focused automated test driving the shipped reviewer entrypoint through the existing catalog-bundle test helper asserts that, with at least three eligible PRs across two repositories, one scheduled scan starts at most two conversations and a later scan reaches the remaining PR; that a pending-check candidate ordered ahead of a green candidate consumes no slot and does not block the green candidate; and that a repeated scan creates no duplicate conversation.
- The reviewer
SKILL.mdandREADME.mddocument the per-run maximum, its default, and that it applies per scheduled run across repositories, and the catalog entry's bundle and entry versions are bumped with generated catalogs regenerated sopython scripts/sync_extensions.py --checkandnpm run buildreport no drift. - Live evidence in the PR's AGENT section links the OSS Agent Canvas scheduled run's pull requests and conversations and shows the run started no more than the configured number of conversations without exhausting the VM.
- Ngôn ngữ chính
- Python
- Star
- 184
- Fork
- 137
- Merge trung bình
- 1 ngày 19 giờ
- Pull request đã merge (30 ngày)
- 46
Chuẩn bị môi trường
- Không có Dockerfile hay tệp Docker Compose
- Có mẫu pull request
- Không có hướng dẫn đóng góp
Bắt đầu từ đâu
- Đọc hết issue, rồi đọc hướng dẫn đóng góp của dự án.
- Bình luận trên issue rằng bạn sẽ nhận — tránh hai người làm cùng một việc.
- Fork repository và làm thay đổi trên một nhánh.
- Mở pull request có tham chiếu số hiệu của issue.
Issue khác của OpenHands/extensions
-
priority:low ready-for-dev
Độ khó 2/5 1-3 giờ Mức phù hợp với người mới 68/100
OpenHands/extensions#975 ·
Maintainer thường phản hồi trong vòng 1 ngày
-
priority:low ready-for-dev
Độ khó 2/5 1-3 giờ Mức phù hợp với người mới 74/100
OpenHands/extensions#973 ·
Maintainer thường phản hồi trong vòng 1 ngày
-
Độ khó 2/5 1-3 giờ Mức phù hợp với người mới 76/100
OpenHands/extensions#970 ·
Maintainer thường phản hồi trong vòng 1 ngày
-
Độ khó 2/5 1-3 giờ Mức phù hợp với người mới 78/100
OpenHands/extensions#969 ·
Maintainer thường phản hồi trong vòng 1 ngày
-
Add a `maven` skill: agents install Debian's `maven` package, which can't load plugins built for Java 17+Có thể đã có người làm @sumitvairagar đã nhận 3 ngày trước. Đang mởpriority:medium ready-for-dev
Độ khó 2/5 1-3 giờ Mức phù hợp với người mới 80/100
OpenHands/extensions#963 ·
Maintainer thường phản hồi trong vòng 1 ngày
Tất cả issue của OpenHands/extensions
Issue tương tự
-
Độ khó 2/5 1-3 giờ Mức phù hợp với người mới 78/100
huggingface/datasets#9128 ·
Maintainer thường phản hồi trong vòng 1 ngày
-
Function/dict state handlers can't supply provider-state injected values (`generate.provider_state`)Đang mở
Độ khó 2/5 1-3 giờ Mức phù hợp với người mới 78/100
pact-foundation/pact-python#1789 ·
Maintainer thường phản hồi trong vòng 9 ngày
-
documentation
Độ khó 2/5 1-3 giờ Mức phù hợp với người mới 82/100
juandresrodca/AdminForge#66 ·
Maintainer thường phản hồi trong vòng 1 ngày
-
Độ khó 2/5 1-3 giờ Mức phù hợp với người mới 68/100
Maintainer thường phản hồi trong vòng 1 ngày
-
backend/local comp/tools P2 tool/terminal type/bug
Độ khó 2/5 1-3 giờ Mức phù hợp với người mới 72/100
NousResearch/hermes-agent#158788 ·
Maintainer thường phản hồi trong vòng 1 ngày