homerail-pr-review

# HomeRail PR Review Use the concrete `pr-review` orchestration. Do not generate a replacement DAG and do not substitute a single-agent review. ## Manager Agent Path When the user says `审查 PR #25`, `review PR 25`, or equivalent: 1. Resolve the repository from the current project or ask only when no repository is available. 2. Call `run_pr_review` with only `repo` in `owner/name` form and integer `pr`. This tool resolves immutable `.base.sha` and `.head.sha` plus trusted base and head clone URLs in code. Never copy, infer, reverse, or manually substitute those fields. Do not use `gh`, `curl`, `instantiate_dag_pattern`, or generic `create_and_run` when this tool exists. 3. Return the real run id immediately and use `get_run_status` for progress. 4. On completion, report the final published handoff and artifact paths. Never claim findings that are not present in the DAG evidence. The workflow is read-only. It must never create a commit, approve a GitHub PR, submit a GitHub review, or merge anything. Its internal Quorum is evidence validation, not GitHub approval. ## CLI Path ```bash hr dag run-template pr-review \ --input '{"repo":"xiaotianfotos/homerail","pr":25}' ``` The CLI resolves trusted GitHub metadata, preserves explicitly pinned SHAs, validates the final input contract, syncs the tracked Asset, and starts the run. Clone URLs always come from GitHub metadata, never logical input. Add `--wait` for CI or unattended operation, then retrieve fixed outputs with: ```bash hr dag artifacts <run-id> hr dag artifact <run-id> pr-review.json --output pr-review.json hr dag artifact <run-id> pr-review.md --output pr-review.md ``` ## Evidence Contract A useful run contains: - three complete reviewer handoffs from Qwen, Kimi, and GLM; - a deduplicated report that preserves every complete reviewer finding; - three independent approve/request-changes/abstain votes; - a deterministic gate requiring two approvals and zero retained findings; - declared `pr-review.json` and `pr-review.md` artifacts; - Manager audit and metrics records for the run. Treat a cancelled run with a published report as `inconclusive`, not as a clean review. Infrastructure, contract, or artifact failures are errors and must not be rendered as zero findings. Do not reconstruct a report by scraping raw handoffs; consume only the Manager-validated declared artifacts.