deploy(vera): pin pr-reviewer v0.24.0 — verbatim evidence quotes (grounding recall) - #26
Merged
Merged
Conversation
…iled-read fix Cumulative over v0.23.0 (Vera's current pin): - v0.24.0 (#51): the verify step normalizes a paraphrased evidence quote to the file's exact text before grounding — fixes the ~0%->~30% grounding-downgrade regression on the cloud model (2 of which silently dropped major-severity gates). - v0.23.1 (#50): a failed file read no longer downgrades a finding (protoAgent#2296). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
mabry1985
added a commit
that referenced
this pull request
Aug 11, 2026
The plugin pins are already current (pr-reviewer v0.24.0 and github v0.3.0 are both the newest tags; protopatch 0.6.1 is the newest npm). The core is the only thing that drifted — 24 minor releases behind, so this is the deliberate base bump the Dockerfile asks for, with `verified_against` moved in step. What's in it for Vera specifically (she's headless — most of the range is console/desktop/Windows work that never reaches her): - #2373 `fix(review): findings contract requires verbatim evidence quotes` (0.126.0) — the CORE half of the grounding fix. pr-reviewer v0.24.0 shipped the plugin half in #26 and has been running without its counterpart since. - #2447 `fix(review): build the published review from blocks, never echo model text` (0.130.0) — the core twin of pr-reviewer#53, still open. - #2297 `feat(workflows): opt-in per-step timeout with graceful degradation` (0.116.0) — the mechanism pr-reviewer#47 needs for the slowest-finder floor (live eval: 57 exhaustions, crossfile slowest in 190 reviews, p90 752s). - #2341 `fix(llm): retry a provider that goes silent mid-stream` (0.120.0) and #2349 `fix(a2a): fail a wedged turn instead of leaving it WORKING forever` (0.121.0) — both bear on the 88.6% completion rate. - #2389 `feat(config): merge-on-boot declarative seeding` (0.126.0) — the fleet-wide fix for our seed-once caveat. It is opt-in via PROTOAGENT_SEED_MERGE, left UNSET here: the seed carries `repos: []` while the live volume holds the 9-repo allowlist, so enabling it is a compose change to make deliberately, not a side effect of this bump. Verified by building this image and booting it on a throwaway config volume: all 8 plugins load, pr-reviewer registers its tools + webhook/dispatch/sweep and mounts both routers, the webhook stays auth-exempt (HMAC-authed) while /api/plugins/pr-reviewer/eval is 200 with the bearer and 401 without and returns the unchanged schema, the seeded a2a identity survives (card serves Vera's description + the pr_review skill, version 0.132.0), SOUL seeds, and clawpatch 0.6.1 runs on node 22. No breaking changes in the range. Watchtower rolls her on merge. Claude-Session: https://claude.ai/code/session_0178c2V1CuZy2oQKDRG8zYca Co-authored-by: GitHub CI <ci@example.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
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
1 participant
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.
Rolls pr-reviewer v0.24.0 to Vera (cumulative over her current v0.23.0):
verifystep normalizes a paraphrased evidence quote to the file's exact text before grounding — closes the ~0% → ~30% grounding-downgrade regression on the cloud model, 2 of which were major-severity findings silently losing their gate. Root cause: deepseek-class models reconstruct quotes from memory; grounding correctly rejects non-verbatim ones.Merging triggers GHCR publish → watchtower rolls
vera:latest(~30s). I'll verify the recipe is live and watch the grounding-downgrade rate settle back down.(protoAgent#2373 — the canonical
FINDINGS_CONTRACTverbatim rule — rides the next base bump, not this roll.)🤖 Generated with Claude Code