Repository navigation
Restore the pre-terse review prompt; keep parallel fetches and a runaway output bound - #36
Merged
Merged
Conversation
…way output bound Live benchmarks across five models show the terse-output prompt from the latency change costs 10 to 27 points of detection and multiplies false positives (mistral-small: 100% detection and 0 false positives before, ~73% and 15 after; similar shifts for kimi, qwen, and gemma) regardless of output-cap size or contract position. The system and user prompts return to their prior wording and structure, byte-identical rendering guarded by a regression test. The parallel metadata and diff fetches stay, and the review output cap stays at 16384 tokens purely as a runaway-generation bound: no observed legitimate review approaches it, and a truncated response would be salvaged into low-quality findings by the JSON repair path.
![postil-dev[bot]](https://avatars.githubusercontent.com/in/3455612?s=60&v=4){kind=small}
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Live benchmarks across five models show the terse-output prompt introduced for latency costs 10 to 27 points of detection and multiplies false positives (mistral-small: 100% detection and 0 false positives before, ~73% and 15 after; kimi, qwen, and gemma shift similarly) regardless of output-cap size or where the response contract sits. The system and user prompts return to their prior wording and structure, with a regression test guarding byte-identical rendering. The parallel metadata and diff fetches stay, and the review output cap stays at 16384 tokens purely as a runaway-generation bound: no observed legitimate review approaches it, and a truncated response would be salvaged into low-quality findings by the JSON repair path. A confirmation benchmark on this branch restores baseline detection (deepseek 97.0%, mistral 93.9%, qwen 93.9%). Verified with fmt, clippy -D warnings, 87 library tests, and 38 wiremock e2e tests on top of the re-run reconciliation change.