Repository navigation
Releases
Crewreview 0.14.2
Answers that report what they couldn't do and retry safely, and a website that says who makes it.
Fixed
- A thread the attendant couldn't resolve is reported instead of silently skipped; the answer still counts as done.
- Retrying an answer whose publishing failed halfway no longer marks it stale because of its own push, and never posts a reply twice.
- The inbox shows why an answer failed and offers to retry it.
Changed
- The website shows that Crewreview is a Pururucos product: the footer links to the company, and a band above it lists the other Pururucos apps.
Crewreview 0.14.1
Reviews that read the whole change, fair to every lens, and posted once on GitLab.
Fixed
- A minified bundle in a change could hide the rest of it from the crew. One megabyte-long line in
public/app.jsfilled the diff, the agents couldn't read past it, and the files after it (the server code among them) went unreviewed. Minified, generated and oversized files now stay out of the diff the agents read, with only their header and a note. The review lists them as left out, and their lines don't count towardlimits.maxDiffLines. - A whole lens could be thrown away as invented over one file. If a lens cited two files and one of them was missing, every finding it had was dropped. A symlink the change adds counted as missing when it pointed outside the repository. Now a lens is set aside only when two or more of the files it cites are missing, a link counts as the one-line file git stores, and a link is never followed outside the tree.
- A retry on GitLab could post the same review twice. GitLab can answer 500 to a batch publish and still publish it, and the retry posted everything again. It now skips the comments and the summary that are already on the merge request. The footer and markers carry the version that ran the review, not the one that posted it, and a post that goes through clears the error from the earlier attempt.
- Each agent's raw answer is kept in
agents/next to the review's report, so a bad review can be looked into.
Crewreview 0.14.0
Tasks in the app: give the crew a task, approve it from the dashboard, let the daemon carry it, and choose who opens the pull request.
Added
- A Tasks page in the dashboard: give the crew a task (the repository, what to do, who does it, whether it waits for you), read the plan and approve it, follow each step, the self-review and the handoffs, and open the pull request. The nav shows how many tasks wait for you.
- The daemon carries tasks forward in the background, one at a time: a task you create or approve starts right away, without waiting for the review cycle, and reviews don't wait for a task. Its planner and builder show on the Now page as agents of their own, and its milestones (plan ready, step committed, self-review, pull request opened, waiting, needs you) in the activity.
- Who does a task: you (your identity, signature and account), or the crew's account, the repository's second account. Then the commits carry its name, it pushes, and it opens the pull request, so you're not the author and can approve and merge it yourself. It's the default when the repository has a review account;
crewreview task --as you|crewpicks it from the command line.
Crewreview 0.13.0
Crew mode, preview: give the crew a task, and it plans, builds, reviews itself and opens the pull request.
Added
- Crew mode, preview:
crewreview task "…" --repo <id>takes a task to a pull request. A planner proposes small steps with their files, checks and risks, and you approve. A builder does one step per turn in Crewreview's own checkout; the repository's validate commands run after each step, never commands from the plan, and each finished step is committed with your identity. The crew reviews its own branch and fixes what's blocking. After you approve again, it opens the pull request with the plan, the risks and the self-review's findings.task list,show,approve,runandcancelfollow it, and--autoskips both approvals. - Every agent's report lands in a ledger (
tasks/<id>/in the data folder: the task, the plan, numbered handoffs, the agents' raw answers, a log), so any seat can continue the task. A builder cut off at its usage limit gets a handoff written by Crewreview from what it left, and the next seat of thebuildrole takes the step over. A blocker, too many turns, or three turns that finish nothing stop the task and ask you. planandbuildroles for seats, and[task]settings:maxTurns,maxReviewRounds,branchPrefix.
Crewreview 0.12.1
Posting reviews on GitLab, fixed, and an activity log that says what it means.
Fixed
- Posting a review on GitLab could fail with "500 Internal Server Error" on
draft_notes/bulk_publish(seen on GitLab 19.5). A comment on an unchanged line went without its old line number and the file's old path, so GitLab saved the draft but couldn't place it, and then failed the whole batch. Comments now carry both. A draft GitLab still can't place goes into the summary instead, and if the batch fails anyway, each comment is published on its own. - Publishing the batch also published drafts you had written yourself on that merge request. The crew now publishes only its own drafts; when you have some pending, it publishes its own one by one and leaves yours alone.
- Retrying a post that failed left the first attempt's drafts behind, and would have posted them twice. The crew now removes its own leftover drafts first, recognized by its hidden marker.
- A merge request larger than
limits.maxDiffLinesshowed "failed: diff-too-large", and then "failed too many times on this commit" on every check. It now says, once, that it's too large, with its size and the limit, and leaves it until a new commit. - The same skip ("it's a draft") was logged again on every check. Each one is now said once per commit and reason.
- The check line said "pending" for pull requests queued for the next check, which read like the inbox. It now says "queued for the next check".
Crewreview 0.12.0
Seats and roles: a model and an effort for each job, shared across the plans you pay for, with a Team page.
Added
- Seats and roles, the first part of crew mode (plan). A seat is an agent with a model and an effort (
[seats] fast = { harness = "codex", model = "…", effort = "low" }).[roles]says which seats review, verify and answer reviews, and how they share the work:order(the next takes over when one is paused or out of runs),spread(each plan used evenly, lenses running at once land on different seats) orcheapest. A lens or verifier that hits its usage limit carries on at once on the next seat, and the seat pauses until its limit resets, with any other seat on the same harness and model. Seats can have their owndailyRuns. Without[roles]nothing changes. - A Team page in the dashboard: each seat as a member of the crew, with its jobs, model and effort, what it's doing now, today's runs and its pause; the jobs and how they share their seats; and how to set it up.
doctorlists the jobs and their seats, and warns about a seat whose harness isn't installed. The terminal shows which seat ran each lens.
Fixed
- Right after linking a repository or group, every open pull request was skipped as "waiting for pushes to settle", even ones pushed days ago with a green pipeline: the wait counted from when the crew first saw the commit. It now counts from the host's last update, and the message says what it means (pushed a moment ago).
- An attendant run that hit its usage limit didn't pause its harness, so the next cycle tried it again.
Crewreview 0.11.0
A second account on GitLab with one pasted token, and every first-run flow walked through and fixed.
Fixed
doctorand the setup said glab wasn't signed in, or hid a company GitLab, whenever another configured host didn't answer:glab auth statuschecks hosts one after another and gave up after the first slow one. Each host in glab's config is now checked on its own, at the same time, for 8 seconds at most.- Answers that ask the git hosts first (who you are on each, a large GitLab group) were cut after 10 seconds, leaving the page waiting forever. The dashboard's server now waits up to two minutes, and the setup keeps the hosts it showed while it checks again.
- Linking a repository, or resuming a paused one, now starts a check right away instead of at the next cycle.
- The Docker image named the version about to be released, so building it from main between a version bump and its release failed. It now installs the latest published release unless you pin one with
--build-arg VERSION=. - The second-account setup was a dead end without a linked GitHub repository, and had nothing for GitLab: every button stayed disabled. It now starts with where the account will review, links a repository, organization or group right on the page when nothing is linked, and then takes the path of that host.
Added
crewreview status: whether the crew runs on this computer, what it's doing, when it checks next and what waits in the inbox. Exits 3 when it isn't running.- Prove any agent not proven yet from Harnesses & limits, the built-in ones included, not only those that need the proof to review.
- When a preview finds only your own pull requests, it says how to have them reviewed (Whose pull requests: everyone's) and approved (a second account), with a link to the setup.
- A second account on GitLab with an access token: a group or project access token creates a bot account by itself, so nobody signs up. The app opens the token page of your group or project, lists what to fill in, and you paste the token. It checks who the token belongs to, refuses your own, asks each place what the account can do there (can approve, read only, no access) and uses it only where it reaches. A group token covers its subgroups.
- On GitHub, the second account can sign in by pasting a classic token of it, instead of the
ghsign-in code: the crew accepts its invitations, checks the access and uses it. - Pasted tokens live in the system keychain (macOS Keychain, Windows Credential Manager, libsecret on Linux), never in a file; a repository refers to one as
keychain:<host>/<user>.
Crewreview 0.10.0
More agents (Antigravity, OpenCode, Gemini CLI), proven from the app, and a setup that knows your company GitLab.
Fixed
- The setup showed "no login" for gitlab.com while glab was signed in to a company GitLab. It now lists the hosts gh and glab are actually signed in to (a configured host whose token fails doesn't count), and shows gitlab.com only when no other GitLab is known.
- Previewing a GitLab group by its address (
gitlab.example.com/group/subgroup) treated it as a project and failed with "404 Project Not Found". The preview now asks GitLab whether the path is a group, as linking does, and then covers its subgroups. For an organization or group it also says how many repositories it holds and names the ones with open pull requests.
Added
- Antigravity (
agy), OpenCode and Gemini CLI built in, over the Agent Client Protocol: they show up when installed, with no config. Crewreview answers their permission requests, so read-only holds, and they review once they pass the read-only proof. Antigravity 1.5.13 passes read-only and write scope, so it can answer reviews too. - Prove an agent's read-only lock from the app (setup and Harnesses & limits), instead of
doctor --conformancein a terminal. It runs in the background and the list shows the result.
Changed
- The website counts visits with GoatCounter (no cookies, nothing personal; a line in the footer says so), including downloads and copied install commands as events, so it's clear where people come from and what they do next. Links to it from the README, the Claude Code plugin, Homebrew, Scoop, the app and the launch posts carry
?ref=for that.
Crewreview 0.9.0
Run it from the app: settings, open at login, pausing repositories, and approvals from the menu bar.
Added
- Settings in the app: everything under Limits and schedule is now a form (how often to check, peak hours, reviews and answers per check, keep-awake, agents at once, runs per day, time limits, the agent for reviews and the verifier, waits after a push or new comments, automatic approvals, update checks). Saving changes only those lines of config.toml, your comments included, and the crew uses the new values at once.
- Open at login, from the app's settings as well as its menu.
- Choose the agent for reviews right in the setup: a "Use for reviews" button on each agent that's ready. The settings have it too.
- Approve from the menu bar (macOS) or the tray: Inbox lists what waits for you, each item with Approve (after a confirmation that names the pull request, since it posts on the host), Open (that item in the dashboard) and Discard. A notification says how it went.
- Pause a repository or a whole organization without unlinking it (the Active switch, or
crewreview repo set <id> --paused on): no reviews or answers until it's back on. A repository paused on its own stays paused inside an active organization.
Crewreview 0.8.0
A first contact that looks the part: a disk image that says what to do, a native app icon, and the next steps on the website right after the download.
Changed
- The macOS disk image says what to do: "Drag Crewreview into Applications", an arrow from the app to the Applications folder, and a sharp background on Retina screens. The icon labels stay readable in light and dark mode, and hidden files stay out of view even when Finder shows them. The image is now built with dmgbuild, which writes the layout into it, because Tauri's skipped its layout on CI.
- A new app icon: transparent corners instead of white ones, and on macOS the margins and shadow of a native icon.
- The website: right after you click download, it shows the next three steps for your system, with a short animation (dragging into Applications on macOS, the SmartScreen warning on Windows, chmod on Linux). Reduced motion is respected.