Feed the PR diff to Qwen3.8-flash; it gives a merge verdict and blocking / suggestion / nit findings, each with concrete impact and fix.
Hand the diff to the model; it returns a merge verdict and blocking / suggestion / nit findings.
This walkthrough uses a 6-file PR diff (new search endpoint + add caching + change payment settlement): hand the diff to qwen3.8-flash, ask for merge verdict, tiered findings (each with concrete consequence and fix), and a false-positive self-check list. The diff includes 7 test points, one of which is a decoy to test for false positives. The same input was run with thinking off and on to compare the two modes.
Call chain: qwen3.8-flash (6-file diff in one input → merge verdict + tiered findings + false-positive checklist). Thinking off suits pre-commit second-level screening; thinking on for a final pass before merge.
Send the diff along with the prompt below to qwen3.8-flash. The diff below is excerpted for readability (the tested input is the full diff across 6 files). After excerpting, the input is shorter, so the number and wording of output items may differ from what is shown on this page. For reference only
Key parameters:
Under OpenAI-compatible mode,
Disclaimer: All content on this page is AI-generated, for reference only, and does not constitute any performance guarantee. Actual results and Credits consumption are subject to the console "Usage Details".
Steps: hand the PR diff to the model for review
Send the diff along with the prompt below to qwen3.8-flash. The diff below is excerpted for readability (the tested input is the full diff across 6 files). After excerpting, the input is shorter, so the number and wording of output items may differ from what is shown on this page. For reference only
enable_thinking goes at the top level of the request body; flash defaults to thinking on.
Notes
- Consumption: This playbook uses text generation (qwen3.8-flash). Thinking off: no reasoning cost; thinking on: output ~10× larger, mostly reasoning chain—cost estimation can't just look at the review report length.
- Measured result: Thinking-off read 6 files, verdict
request_changes, findings 4 blocking + 1 suggestion + 1 nit, hit 5 of 7 test points—SQL injection (gave parameterized fix),except Exception: passswallowing charge failure (wrote "goods delivered without payment" consequence), cross-file type mismatch (caller passes string, callee signature is int); decoy not false-positived (hashlib.md5for ETag put innot_issues). - Note: Thinking-off missed "cache key uses only uid without tenant_id, causing cross-tenant data bleed" (needs combining two changes to spot); switching to thinking-on hit all 7, flagged that one as blocking, but time and output volume rose significantly (this behavioral difference was consistent across an independent reproduction). Recommend two passes: thinking-off for pre-commit second-level screening, thinking-on for a final merge gate.
- Note: Thinking-off summary mentioned items not in the issues list; when integrating or citing, use issues as source of truth; both configs are assistive only, not a replacement for human review.
- Data boundary: Code in context means content goes to the server. Replace real secrets, internal addresses, and customer data with placeholders before submitting; keep code structure, the model can still spot issues.