Skip to content

v0.75.0

Latest

Choose a tag to compare

@cpcloud cpcloud released this 09 Nov 22:11
· 71 commits to main since this release
v0.75.0
7feb72a

New features

  • Reviews now check whether a change does what its commit messages and pull request description say it should. Findings also carry a kind, topic, confidence, and file and line location, and roborev ranks them by severity, panel agreement, and confidence. Rendered reviews look the same as before. Exported reviews use schema version 3, so export clients generated for schema version 2 or lower need to be regenerated.
  • roborev review --type exhaustive asks the reviewer to report every problem it can support, including small concrete ones. It works with CI review_types and panel members too. In our tests it found about 50% more real issues at the same accuracy, but reviews take longer and cost more, so it is opt-in. The name exhaustive is now reserved, and a custom review type with that name no longer loads.
  • Review panels accept synthesis_instructions. When set, the synthesis agent can read the repository with read-only tools, for example to check each finding against the code or label which findings are safe to fix. Commit and range reviews run synthesis against a checkout of the reviewed commit, and reruns keep the instructions the run was queued with.
  • The CI poller can check findings before posting them. With [ci] check = true, roborev-ci tests each synthesized finding in a sandboxed copy of the pull request and posts only the findings that hold up. On a held-out set, precision rose from 56% to 82% and every real finding was kept. Findings that suggest removing a test are posted only when the checker shows another test catches the same changes. The checker can install npm, Bun, and PyPI packages and use installed mise toolchains for Node, Bun, OpenTofu, and Terraform. It needs Linux on amd64, a pinned nono build, Codex, and a few host tools, all listed in the GitHub integration docs. If the checker fails, roborev posts the original review unchanged.
  • With [ci] rewrite = true set alongside check, the checker rewrites kept findings in plain words and names where each fix goes.
  • New [ci] settings control which pull requests the poller reviews:
    • skip_drafts = true skips draft pull requests until they are marked ready.
    • allowed_authors and allowed_author_associations limit reviews to trusted authors, which keeps outside contributors' prompt text away from review agents.
    • allow_codeowners = true also trusts anyone the base branch's CODEOWNERS file names.
    • include_branches and exclude_branches filter pull requests by head branch name, for example to skip dependabot/*.
  • [ci] comment_max_findings caps how many findings a pull request comment shows and links to the full review. comment_when_clean = false skips the comment when a review finds nothing.
  • roborev init gh-action now generates workflows that skip draft pull requests and cancel older runs for the same pull request. Pass --include-drafts to keep reviewing drafts, and regenerate an existing workflow with --force to pick up the change.
  • Automatic reviews can batch commits machine-wide. post_commit_batch_size now works in the global config, and the new post_commit_quiet_period setting queues one review after the branch has gone quiet for that long. Repository settings still override the global ones.
  • roborev refine keeps earlier fix attempts in the next round, with or without --plan, and records why a round made no changes. It also applies your fix_guidelines while making fixes, and its summaries list declined findings with evidence and the commands the agent actually ran.
  • roborev refine warns when three fix rounds in a row keep touching the same files or fail to reduce High and Medium findings. --stop-on-stall stops the loop at that point.
  • roborev refine --skip-trivial-rereview skips the wait for a re-review when a fix commit only changes docs or tests. The branch review still covers those commits.
  • Related CLI commands are grouped under job, agent, hook, report, storage, and signing, and queue controls live under daemon pause and daemon resume. Root help lists 27 commands, down from 47. The old spellings still work for at least one release and print a migration notice.
  • CLI help explains the review workflow, output modes, environment variables, and exit codes. Errors include a hint, and commands with --json write a JSON error object to stderr. Usage mistakes now exit with code 2 and invalid values with code 4, so scripts that check exit codes may need updating.
  • The daemon API accepts a review_prompt with a commit or range and runs it as a normal review in a detached checkout. Callers can also send review_settings to choose file exclusions, tool restrictions, binary attachments, and backup agents without the repository's .roborev.toml changing them. Retries, reruns, and synced copies keep the prompt and settings.
  • roborev can read usage from an authenticated AgentsView server. Set [agentsview] url and an optional token_file in the global config, or use AGENTSVIEW_SERVER_TOKEN. Subagent spend is included, and roborev doctor checks the connection.
  • roborev summary shows CI run health, including panel outcomes, retirement reasons, runs and cost per pull request, and turnaround. roborev doctor warns when one pull request started more than 10 CI runs within an hour.
  • Changing max_workers in config.toml resizes the running daemon's worker pool without a restart. Running reviews finish before extra workers retire.
  • Exported panel reviews mark advisory members with non_voting.

Improvements

  • Agent Hook reminders now ask agents to triage findings: fix them, decline them, mark them invalid, or ask you. The roborev-fix skill works the same way, and the reminder puts your latest instruction first. Design reviews no longer count toward the failed-review threshold, and a design review of a change with no design document passes.
  • Fix agents get clearer findings. Each finding describes what triggers the problem, the wrong result, the expected behavior, and the cases a fix must keep working.
  • Low severity now means a real problem with minor impact or a named code-quality fix. Ideas with no concrete problem go in the review summary instead.
  • roborev insights reads each panel's combined review, with its comments and closed state, and no longer counts member reviews separately.
  • The web Reviews page loads older jobs as you scroll, and sorting a column sorts the whole job list on the daemon, so a cost sort no longer misses expensive jobs you haven't scrolled to yet.
  • New users see how to get their first review in an empty queue in both the web UI and the TUI.
  • roborev job list explains where matching jobs are when the current directory or branch hides them. fix and compact give similar hints outside a repository.
  • An empty multi-word roborev search --lexical explains that every word must match. JSON and MCP output include the same hint.
  • roborev list --json and status --json no longer include stored prompts for finished jobs, which cut 50 jobs from 448 KB to 37 KB. roborev job show --prompt prints the prompt, including for failed and canceled jobs.
  • The post-commit hook tells the committer when a commit was not queued for review and points to roborev doctor. Existing hooks update when the daemon next starts.
  • On macOS and Linux the daemon adds common tool directories, such as mise and asdf shims, Nix profiles, Homebrew, and ~/.local/bin, to its PATH, so it finds agents even when a service manager starts it. If a configured agent still can't run, roborev status and roborev doctor report it.
  • On macOS, roborev daemon restart, roborev update, and automatic recovery restart a LaunchAgent-managed daemon through launchd, so you end up with one daemon.
  • The MCP server reports roborev's website and icon, and the web UI has a favicon.
  • macOS users can install roborev with brew install roborev from Homebrew core.
  • PostgreSQL sync is much faster on large histories:
    • The first push of 20,000 reviews takes about 3 s, down from 100 s.
    • A new machine's first pull takes 35% less time.
    • roborev sync status takes 0.05 s at 1M jobs, down from 6 s.
    • roborev sync now shows progress at once while a background push runs.
    • Status commands stay responsive when the PostgreSQL connection stalls.
    • Teams with 100,000 central reviews can connect again without hitting the connect timeout.
  • Writes during a sync pull, such as enqueueing a review, commenting, or closing a review, no longer stall for seconds or fail with "database is locked".
  • Large local databases are faster across the board. On 100,000-job databases:
    • Branch-filtered and closed-review queue pages return in milliseconds.
    • roborev export of 94,000 reviews takes 8 s, down from 90 s.
    • summary and doctor time windows use indexes.
    • The TUI cost header refreshes in 24 ms.
    • insights --since queues in milliseconds.
    • Daemon startup no longer rereads every review.
    • The Agent Hook's failed-review lookup stays near 1.5 ms.
  • Several schema migrations build new indexes once on the first daemon start after upgrading. At 1M jobs this can take up to about 30 seconds.
  • Review search stays fast as history grows. A search that matches nothing takes a few milliseconds, common-word searches run about 5x faster, the idle background sweep no longer re-renders every review, and sync no longer forces a full search rescan.
  • The CI poller uses far less GitHub API quota. It revalidates pull request lists with ETags and stops rereading CODEOWNERS for heads it already reviewed. It also publishes each skip once per commit, posts "Review deferred" once, and backs off failed repository lookups.
  • After a GitHub secondary rate limit, roborev pauses every client on that token for as long as GitHub asks, or one minute if GitHub doesn't say. During the pause, errors read rate limit cooldown.
  • Range reviews build their prompt in milliseconds in repositories with thousands of past range reviews.
  • Agents that print very long lines no longer stall the daemon.
  • The daemon's memory use no longer grows with core count. A 256-core machine saves about 300 MB.
  • Status stays fast with many worktrees or deleted repositories registered.
  • The Agent Hook drops sessions idle for 30 days from its state file, which keeps each tool call fast.
  • The web UI's DOMPurify dependency is updated to 3.4.17, which fixes three security advisories.

Bug fixes

  • job_timeout_minutes now counts wall-clock time, so a review that spans a laptop sleep times out on schedule.
  • roborev analyze runs with read-only agent permissions. --fix still edits in a separate step.
  • Agents that roborev launches for reviews, synthesis, and fixes no longer trigger the Agent Hook.
  • A launchd or systemd daemon that briefly stops answering HTTP keeps running. Commands no longer drain and restart it, and roborev status --json reports it as running.
  • roborev daemon restart and post-upgrade restarts finish in containers without an init process.
  • Pi reviews report the provider's real error message. Amazon Bedrock's rejection of image input for text-only models moves the job to the backup agent without retrying.
  • Linked worktrees of a bare repository that keeps core.bare in config.worktree work again.
  • With the TUI's "(none)" branch filter active, the queue stays responsive, scrolling loads more jobs, and the listed jobs match the picker's count.
  • Scrolling TUI release notes is no longer slow.
  • Ctrl-C now stops roborev refine while it waits for a review.
  • roborev ui prints the URL before it opens the browser, so you can open it yourself if the browser doesn't appear.
  • On Windows, the daemon no longer flashes console windows during background work.
  • Web UI content no longer jumps sideways when a scrollbar appears on the Analytics page, the job table, the repository picker, release notes, or the login page.