Skip to content

Dashboard

Open the address set in config.webUrl and sign in. The dashboard shows review progress, findings and costs for the accounts you can access.

Use the forge to discuss findings. Use the dashboard to inspect reviews and manage the service. The application configuration lives in the configuration file.

Sign in as an admin and open All accounts → Accounts. The setup checklist shows what is missing:

  1. A GitHub App declared under apps, a Forgejo bot under forgejo, a GitLab service account under gitlab, or a Gitea bot under gitea.
  2. An installed App, or a bot with access to the account, and at least one enabled repository.
  3. A default model in review.model.
  4. An optional embedder in embedding for finding similar code.

Newly discovered repositories start disabled. Open the account’s Repositories page and enable the ones that you want reviewed.

The account switcher selects one account or All accounts. Press Ctrl+K or ⌘K to find a page or a recently updated pull request.

Page Use it to…
Overview See running work and problems that need attention.
Pull requests Read reviews, inspect findings and open their threads on the forge.
Queue Find waiting, running, retrying or failed jobs.
Repositories Enable repositories and inspect their configuration and code index.
Analytics Compare review activity, findings and spending over time.
Usage & costs Inspect costs by account, repository, model or review.
Follow-ups Read questions and the bot’s replies.
Findings Search findings across reviews.
Rules See which review rules apply and where they come from.
Audit log Inspect management actions, as an admin.

Some pages are available only within an account. All accounts combines activity and usage across accounts you can read.

Open a pull request for its summary, findings and activity. Expand a finding for its explanation and suggested fix. A finding that an agent raised carries the agent’s name. A finding without a name is from the default pass, which runs kritika’s own instructions.

Run details contains the timeline, model conversation, usage and raw data for troubleshooting. When agents ran, the timeline lists each agent’s parts with their cost, and the cost of each agent. The default pass is listed as default.

Search accepts plain text and filters:

List Example filters
Pull requests repo:owner/name, author:login, status:failed, is:paused, is:p0
Findings repo:owner/name, severity:p1, category:security, status:open, dismissed:conceded, rule:wrap-errors, agent:security, path:internal/, pr:merged, is:downvoted

The dashboard suggests filters as you type. It offers agent: once a finding on the page names an agent. agent:default finds the findings of the default pass. Expand a list row to preview its latest review without leaving the list.

Common findings counts the findings that the filters keep. It also counts how many of them a later review found addressed, were dismissed, or drew a 👎. Group them by category, rule, agent or directory. Select a group to narrow the list to it. A directory then shows the directories below it.

A group that keeps coming back is worth an instruction in AGENTS.md or a skill, so the change’s author gets it right the first time. A group that is often dismissed or drew a 👎 is worth a rule that tells reviews what not to report. Like the analytics, 👎 counts come from GitHub alone and need polling enabled.

path: keeps the findings under a path, and is:downvoted keeps those that drew a 👎. pr: keeps those whose pull request is open, merged, or closed without merging. For example, pr:merged status:open lists the findings that were still open when their pull requests merged. Those are the findings that their authors left for later or never saw. A rule or an instruction pays off most there. dismissed: keeps those dismissed one way:

  • command, with @<bot> dismiss.
  • resolved, by resolving the thread.
  • conceded, when kritika agreed in the thread that the finding was wrong.

Hover a dismissed finding’s status to see which.

Export Markdown and Export JSON download the findings the filters keep, newest first, up to 1,000. Give the Markdown to a coding agent and ask it to propose those instructions and rules. Reviews read AGENTS.md only when it is listed under context.

Start with Needs attention to find retrying or discarded jobs and workers whose heartbeat is overdue. Expand a row to see attempts, timing and the full error. An overdue heartbeat can recover. It does not always mean a failed job.

Retryable jobs resume automatically. Admins can cancel active reviews, re-run canceled or discarded reviews, and reindex after a failed index job.

The job kind menu narrows the queue to Index jobs, or to Reviews and follow-ups, the runs. Click a column header to sort the jobs by it, and again to reverse the order.

Choose a period to compare reviews, findings, duration, tokens and API spending. The page also shows addressed findings and reactions to inline comments. Reactions are counted on GitHub alone and need polling enabled.

Refused hosts shows destinations blocked by the egress gateway, the forward proxy for runner Jobs. Review why a host was blocked before adding it to egress.allow.

Dependency ratings lists, on an account’s page, each dependency that more than one pull request updated in the period, with the risk the scorer gave each update before any floor, where the pull request moved that one version alone, and the reach it gave the dependency. A dependency rated at more than one level drifted; the most drifted come first. Use it to see whether the scorer answers the same way twice for the same dependency, and which dependencies it does not.

In Usage & costs, select a repository, model or review to inspect spending. Unpriced means a call has no known price and totals are incomplete. API spending excludes subscription fees. Cache details depend on transcript retention.

The Rules page lists configured rules and context files, with their source and scope. It also lists findings that cite each rule, with how many of those were addressed, dismissed, or drew a 👎. A rule whose findings are often dismissed can need rewording. Edit rules in the configuration file or repository settings.

Use Settings in the user menu to change theme, time zone and clock format. Daily charts use UTC. If the top bar says Reconnecting…, live data can be out of date until the connection returns.

Under API tokens, create a personal access token for an agent that reads through kritika’s MCP endpoint. Name it, and choose whether it reads everything you can read or only some of those accounts. Then choose when it expires: in 7, 30 or 90 days, a year, or never.

The dashboard shows its value once. A token never changes anything. It stops reading once it expires, is revoked, or the sign-in it came from changes. The list then marks it expired or stale. See personal access tokens.

Admins can enable or disable repositories individually or in bulk, and request reindexing. Enablement changes are saved and audited.

Use Resync after changing the App’s or the bot’s access, or after unarchiving a repository. It names the account’s forge, such as Resync from Forgejo. Then enable the repository if it is still off. The next poll can pick up open pull requests created while it was disabled.

Click a column header to sort the repositories loaded so far by it, and again to reverse the order. The Index column sorts what is indexing first, then what is queued or waiting.

The leader, which is the elected replica, queues repositories’ first indexes a few at a time. It keeps onboardWindow of them queued or running at once. Index jobs run one at a time on each replica unless indexWorkers says otherwise. So after you turn on many repositories, most of them wait. The Index column shows each one’s place in line, such as Waiting · #3. The leader queues the next as each first index ends.

A reindex goes ahead of the first indexes still waiting.

A repository’s page shows its index jobs queued or running, and keeps them current while there are any. The Reindex button reads Reindex queued while a reindex is queued. Expand Repository settings to inspect the effective configuration. Expand Index runs to investigate indexing failures. Expand Index contents to see what the similar-code index holds: the commit and embedder it was built at, and its files and their chunks.

Only admins see a chunk’s text. The paths that the indexer skipped, such as those an ignore rule matches, are not recorded, so they are not listed. See which repositories run for how dashboard choices and configured defaults interact.

Admins run the instance from All accounts, under Configuration:

Page Use it to…
Accounts Follow the setup checklist while something is missing, and see the accounts the instance serves.
Connections Inspect GitHub Apps and forge bots and their installations, and uninstall an App from an account nobody serves.
Users See everyone who signed in, how they sign in, what their newest session reads and when they last signed in. Sign someone out everywhere, and revoke anyone’s API token.
Settings Find each effective configuration value and where it is set, the egress gateway’s Download cache configuration and the Most used tools of recent runs.
Audit log Inspect management actions across accounts.

These pages do not edit the configuration file. Change the file and roll out the deployment. External ConfigMap changes and secret rotations require a restart or a controller such as Reloader.

Admins can:

  • Re-run a pull request, or select several open pull requests to re-run.
  • Cancel a review.
  • Reindex a repository.
  • Enable or disable repositories.
  • Resync repository access from the account’s forge.
  • Uninstall an App from an account that the configuration does not serve.
  • Sign a user out everywhere, which ends every session of theirs, an admin’s own included.
  • Revoke anyone’s API token.

Re-run, cancel and reindex requests queue work. A duplicate review or index request can be refused while one is already queued or running.

A configuration that fails validation prevents startup. If the leader cannot apply it to the database, kritika keeps the last applied configuration and sets kritika_config_error. Fix the problem and restart to try again.

See monitoring for alerts and security for access controls.