Repository navigation
Releases
v0.78.1
GitHub Plugin
- The lane guard now judges omitted
repoarguments against the active round's scope when using github-plugin >= 0.12.0, preventing valid round-scoped calls from being marked as incomplete. - The guard only trusts a GitHub scope if it was successfully entered for the current round, ignoring scopes that failed to initialize due to degradation.
- Round scope marks are cleared at the start of each new review session to ensure stale state from previous rounds does not incorrectly authorize current operations.
v0.78.0
Correctness
- Review rounds now scope
github-plugin's default repository to the round's target repository, preventing finder lanes from accidentally reviewing unrelated repos whenrepois omitted. - This fix requires
github-plugin >= 0.12.0; on older versions or if the plugin is absent, the default repo behavior remains unscoped, and the existing lane guard acts as the backstop. - Explicitly specified repositories still take precedence, so
lane_repo_violationcontinues to catch lanes that name the wrong repository.
v0.77.0
v0.77.0 adjusts runner independence accounting and QA panel failure handling.
Runner Independence
- The dispatcher now waives seed-origin gaps for steps it authored when re-running the
synthesizestep, so that specific origin no longer falsely marks a live round as "Runner independence unverified". - Any other gap on dispatcher-authored steps, and every gap on execution steps, still count against runner independence.
- The accounting wrapper does not pass the
authoredflag to the host runner, preserving the isolation check for external runners.
QA Panel
- Finder lanes that fail after a retry now register as a coverage Gap rather than exhausting the round.
v0.76.0
Finding Accounting Gate
The finding_accounting_gate default changes to losses, holding a round only for provably dropped findings.
Previously, raw accounting gaps set verified=false, blocking merges even for honest structural features; on 40 live rounds, losses holds 0 while all holds 20.
Set the gate to all to restore the previous behavior, or off to disable gating.
QA Panel Behavior
An exhausted unverified hold now concludes the QA panel check as neutral instead of in_progress.
Auto-approve remains withheld, but the merge is no longer blocked, preventing the "Verifier failed twice" stall observed in v0.75.0.
Integrity and Classification
Transformation and disposition gaps now hold the round as incomplete and unverified.
Finder-lane gaps mark the round as incomplete only, since a verify re-run cannot recover a missing lane.
Merges explaining severity or location changes no longer require a reason, while downgrades or unknown moves still do.
Stability
reconcile() no longer crashes on unhashable model-emitted severity values.
Unknown severity values rank as no upgrade, preserving the requirement for a reason.
v0.75.0
This release stabilizes live structural code review, resolving reconciliation failures that previously capped verdicts at WARN.
Reliability
- Live structural rounds now reconcile correctly; previous versions held verdicts at WARN due to accounting mismatches.
- A survivor without a
finding_idis identified by its first raw id. - Only a change in disposition kind counts as a changed disposition; severity, file, or line changes still require a reason.
- Whitespace-only evidence is treated as empty for explained-change checks.
Gap:lines emitted by the machinery (structural outages, partial passes, time-budget cut-offs) correctly mark raw input as incomplete, while finder-specificGap:notes no longer do.
Data Integrity
- Dispositions are now read from json/bare fences and ```accounting blocks; trailing accounting blocks are not misidentified as findings arrays.
- A survivor's evidence cannot be blanked without a stated reason, preventing findings from passing grounding checks without supporting quotes.
v0.74.0
Reliability
search_pr_repotruncates matched lines to 8 KiB instead of failing when a line exceeds 64 KiB, preventing search errors caused by minified bundles or large JSON files.- Structural plan analysis now ignores context file changes that do not mention or affect the target feature, reducing false-positive "touched" flags in pull requests.
- Fixed route mention detection to handle routes registered with a trailing slash, ensuring relevant handler tests are correctly identified as affecting the feature.
Code Review Accuracy
- The accounting barrier now seeds finder outputs with the exact byte-for-byte data from the first run, restoring
originverification and preventingcode-review-structuralrounds from incorrectly reporting "Runner independence unverified."
v0.73.0
The structural plan now treats features whose code is untouched by the diff as "siblings," ranking them after touched features to prevent valid handlers from being incorrectly flagged as coverage gaps. Use structural_symbol_scope: false to disable this behaviour.
Review Accuracy
- The symbol scope now analyses the base side of the diff to catch removed names, changed headers, and enclosing definitions, preventing false "sibling" rankings when code is modified indirectly.
- Incremental review is restricted by default to repos listed in the
incremental_review_reposallowlist. - Posted review summaries explicitly state carried findings when present, preventing a "clean" status from appearing alongside a WARN.
QA Logic
- The QA panel now fails to post a PASS verdict if synthesis detected findings that were subsequently lost in the verify or report stages.
v0.72.0
v0.71.0
v0.70.0
Panel Model Isolation
- Correctness and removed-behavior review lanes now run as
review-finder-deepon a model specified bypr_reviewer.deep_finder_model, allowing distinct frontier models for these lanes without affecting the host review-finder. - Replay now pins deep lanes to
pr_reviewer.deep_finder_model(with its circuit breaker) per trial, ensuring A/B tests read the intended model from run output instead of whatever model the last live attempt left on the registry. - A failure in a deep lane triggers a retry and a 30-minute cooldown on the default model, preventing rate limits or token issues from blocking the entire review round.
min_protoagent_versionis now read from the registry entry at dispatch, requiring protoAgent >= 0.186.0 for this model pinning to function.
Review Gate Behavior
pr_reviewer.reviewers_of_recordallows a designated reviewer’s verdict to stand in for the panel on labeled PRs, eliminating redundant panel runs when the reviewer passes; pending, failed, or unreadable statuses hold the head without posting a verdict.- A pending reviewer-of-record status older than
pending_grace_s(3600s) is treated as a dead gate, triggering an immediate head review to prevent indefinite holds. - Draft PRs are dropped before entering the panel queue, freeing slots for non-draft reviews;
ready_for_reviewevents still queue normally.