Compact GitHub PR dashboard, opener, notification cleaner, and optional collaborative review watcher using Codex, Claude, Kimi, and DeepSeek through Deep Code.
CLI release assets are unpacked, self-contained single-file executables. The
optional Windows desktop companion is a self-contained zip installed beside
the CLI. The app uses GH_TOKEN or GITHUB_TOKEN when set, otherwise it reads
the active token from gh auth token, so install and authenticate GitHub CLI
first with gh auth login. Automatic PR review also requires each enabled
codex, claude, kimi, or deepcode CLI on PATH; the reviewer is disabled
by default.
Linux, bash:
repo=flcl42/pr; arch="$(uname -m)"; asset=pr-linux-x64;case"$arch"in aarch64|arm64) asset=pr-linux-arm64;; esac; curl -fsSL "https://github.com/$repo/releases/latest/download/$asset" -o ./pr; chmod +x ./prmacOS, zsh:
repo=flcl42/pr; arch="$(uname -m)"; asset=pr-macos-arm64; [ "$arch"="x86_64" ] && asset=pr-macos-x64; curl -fsSL "https://github.com/$repo/releases/latest/download/$asset" -o ./pr; chmod +x ./prWindows, PowerShell:
$repo='flcl42/pr'; $zip=Join-Path$env:TEMP'pr-ui-windows-x64.zip'; Invoke-WebRequest"https://github.com/$repo/releases/latest/download/pr-windows-x64.exe"-OutFile ".\pr.exe"; Invoke-WebRequest"https://github.com/$repo/releases/latest/download/pr-ui-windows-x64.zip"-OutFile $zip; Expand-Archive$zip".\pr-ui"-Force; Remove-Item$zippr
pr https://github.com/OWNER/REPO
pr OWNER/REPO
pr 10843
pr OWNER/REPO#10843
pr https://github.com/OWNER/REPO/pull/10843
pr ui
pr --once
pr --cleanup-onceRunning pr with no arguments starts the interactive PR dashboard. Passing a
repository URL or OWNER/REPO adds that repository to the tracked list and saves
it in .pr.yml next to the executable.
Only one interactive dashboard or maintenance command can run for a given
settings file. Starting the console or desktop dashboard takes over the active
instance that uses the same .pr.yml. Repository updates and one-shot PR open
commands exit independently and leave a running dashboard untouched.
If no repositories are tracked, the dashboard stays open with an empty list and shows the add command instead of exiting.
pr ui starts an optional Windows MAUI companion using the same .pr.yml as
the CLI. It provides the grouped Reviewed, Top, and regular PR table; interactive
search; urgency breakdowns; repository addition; refresh and notification
cleanup; weekly stats; ignore and Top actions; manual review queueing; review
enablement and delivery controls; and live status for every configured agent
stage. PR titles open GitHub, while PR numbers open the urgency calculation.
Reviewed titles start with the agents that completed successfully: C for
Codex, c for Claude, K for Kimi, and D for DeepSeek. Failed agents are
omitted, so [CcK] means those three stages completed.
The default solution deliberately excludes MAUI and needs only the .NET SDK:
dotnet build Pr.slnx -c ReleaseBuild the desktop companion explicitly on Windows with the maui-windows
workload. Publish it into a pr-ui directory beside pr.exe, which is one of
the locations discovered by pr ui:
dotnet workload install maui-windows
dotnet publish Pr.Maui/Pr.Maui.csproj -c Release -r win-x64 --self-contained true -p:WindowsPackageType=None -o C:\Programs\pr-uiFor a development output in another location, set PR_UI_PATH to the full
pr-ui.exe path. The launcher passes the CLI settings path through
PR_SETTINGS_PATH, so the companion never creates a separate configuration.
Passing a PR number opens it in the default browser. Bare PR numbers work when a
single repository is tracked. With multiple repositories, use OWNER/REPO#NUMBER
or a full pull request URL to avoid ambiguity.
The dashboard resolves the authenticated gh username and excludes your own
PRs from the list. It polls GitHub every 15 minutes, so the list can lag by up
to 15 minutes. Each refresh scans up to 1000 matching PRs per GitHub search
batch. New matching PRs that appear after the initial load trigger an audible
ding.
GitHub reads share one pooled HTTP/2 connection and are serialized with retries
for transient transport and server failures. Hotness activity is cached by each
PR's updatedAt value and recalculated locally, so an unchanged refresh normally
needs only the paged search request. Notification and ignored-PR status checks
are grouped into GraphQL batches of up to 50 instead of one request per item.
Titles are terminal hyperlinks in --once output. In the interactive
dashboard, click the title column to open a PR directly from the TUI.
Press V to enable or disable automatic agent review. The setting is persisted,
and the configured agent pipeline plus live reviewer state are shown in the
status area. On startup, directly
requested PRs created within the last four days are armed and begin the same
20-minute timer; older existing PRs are baselined without review. After that, a
newly discovered PR or newly added direct review request for the authenticated
user starts the timer while the PR remains open, non-draft, and requested from
that user. A later head change resets the timer. Your own and dashboard-ignored
PRs are skipped.
Press E to force an immediate agent review of the selected PR. Manual queue
entries are persisted in .pr-review/state.json and run even while automatic
review is off. A forced review bypasses the timer, direct-review-request,
draft, own-PR, ignore, previous-review, approval, and commenter gates. It still
requires a tracked, open PR from an OWNER, MEMBER, or COLLABORATOR, and
cancels if the head commit or authenticated user changes. PRs from external
contributors are never queued, including through E. Forced reviews stage
only actionable findings; a clean result creates nothing. With automatic review
off, the worker resolves only manually queued PRs.
Press A to switch review delivery between drafts and automatic submission. The
choice is persisted as codexReview.autoSubmit; it affects both automatic and
manually forced reviews.
When the timer expires, the watcher counts current approvals and conversation, inline, and non-empty review comments. It reviews only while both counts remain below two distinct humans after excluding the PR author and configured bot patterns. It checks again before publication. An external author association, head change, draft transition, closed PR, removed review request, user change, ignore action, or disabled automatic watcher cancels stale automatic work. A manual queue entry remains eligible when automatic review is off and is removed after completion, closure, head change, account change, or explicit queue loss.
Enabled agents run sequentially against one detached worktree at the exact PR
head with permission to inspect, modify, and test code under a validated
structured-output contract. Edits are investigative only. After every agent,
including a failed, timed-out, or canceled agent, the coordinator force-checks
out the original PR head, removes untracked and ignored outputs, verifies the
worktree is clean, and restores the shared ledger before starting the next
agent. A reset failure stops the pipeline.
Every agent process starts in that worktree. Its absolute path and the separate
repository-cache path are written into the prompt, agent metadata, and initial
shared-ledger record, so repository commands and any tests allowed by the
configured tool policy have an unambiguous working directory. Repository
context can be stored inline under
codexReview.contexts in .pr.yml; use the full OWNER/REPO slug to avoid
collisions. Inline context takes precedence over <repo>.md next to the
executable and the embedded Nethermind fallback. Every finding is validated
against the local PR diff before a pooled GitHub API connection either creates a
server-side PENDING review or immediately submits a COMMENT review. If GitHub
reports a conflict with an existing pending review at the same head, missing
inline threads are appended without submitting or replacing that draft. A
pending review from an older head must be submitted or discarded first. The
review body is a short severity-and-files summary with no agent attribution or
review-state text. Finding titles describe the observed problem rather than
directing the author, and finding bodies keep remedies conditional and
outcome-focused. Before each stage, the agent reads pr-<number>.review.jsonl
from the worktree. The coordinator appends validated findings and agent failures
to that shared ledger, instructs later agents not to repeat the same underlying
defect, and publishes the accumulated findings once after the pipeline finishes.
An agent error advances to the next enabled agent; a job fails and retries only
when every enabled agent fails. A clean result creates nothing by default.
Completed agent results appear in a section above Top PRs. Each row indicates
whether the result was sent, drafted, clean, or kept local. Clicking its title
opens the PR, acknowledges the item, and moves it back to Top or the
regular section. Review state, per-agent prompts/results/logs, and a preserved
copy of the collaborative JSONL ledger live under dataDirectory.
Cached repositories and temporary worktrees live under workspaceDirectory,
which falls back to dataDirectory when it is omitted.
The last column is an in-process mouse target for the ignore action; it does not
use an OS URL protocol handler. Startup removes the old pr-ignore:// handler
if a previous build registered it. Press I for the same action from the
keyboard. Ignored PRs are hidden from the dashboard and saved in .pr.yml.
Press Ctrl+I to reveal ignored PRs that still need review; press I or click
unignore on a revealed ignored PR to unignore it.
Press T to move the selected PR into or out of the top section. Top PRs are
saved in .pr.yml, stay otherwise normal, and are separated from the rest by a
single row when visible. Closed and merged entries are removed by the regular
cleanup pass.
Press F1 to search PR titles, numbers, or authors in a filter row above the
table. Filtering is interactive, ignored PRs are included while search is
active, and Esc cancels the search.
Press S to open this week's tracked-repository stats for your GitHub user:
non-draft PRs you created that are still open or merged, and review submissions
you made. Press B or Esc to return to the dashboard.
While the dashboard runs, it also cleans stale GitHub notification threads for
tracked repositories only: closed, merged, draft, or inaccessible PR
notifications and closed or inaccessible issue notifications are marked as read.
Cleanup runs shortly after startup and then once per hour. Press C in the
dashboard to run cleanup on demand. Cleanup also checks ignored PRs and removes
them from the ignored list after they are closed or merged. The same batched
state check removes closed or merged PRs from the top list.
Settings are stored next to the executable in .pr.yml:
requiredApprovals: 2repositories:
- https://github.com/OWNER/REPOtopPullRequests:
- https://github.com/OWNER/REPO/pull/10844ignoredPullRequests:
- https://github.com/OWNER/REPO/pull/10843priority:
superHotThreshold: 40hotThreshold: 25noComments: 30oneCommenter: 15twoOrMoreCommenters: -30tenDaysNoReviews: 15tenDaysOneReview: 10reviewRequestedFromUser: 30fiveDaysNoReviews: 10fiveDaysOneReview: 5noReviewsNoComments: 10twentyDays: 35lessThanThreeHoursNoComments: -30ignoredCommentAuthors:
- "[bot]"
- bot
- codex
- claude
- kimi
- deepseek
- deepcode
- copilotcodexReview:
enabled: falsepollIntervalSeconds: 60readyDelayMinutes: 20startupScanDays: 4eligibilityCheckSeconds: 15maxOpenPullRequests: 1000agents:
codex:
enabled: truemodel: gpt-5.6-soleffort: maxclaude:
enabled: falsemodel: opuseffort: maxkimi:
enabled: falsemodel: kimi-code/k3deepseek:
enabled: falsemodel: deepseek-v4-proeffort: maxsandbox: workspace-writeephemeral: trueignoreUserConfig: truemaxFindings: 25skipWhenApprovalCountAtLeast: 2skipWhenUniqueCommentersAtLeast: 2skipOwnPullRequests: truetimeoutMinutes: 90failureRetryMinutes: 2postNoFindingsComment: falseautoSubmit: falseprocessExistingOnFirstRun: falsedryRun: falsedataDirectory: .pr-reviewworkspaceDirectory: .pr-reviewcontextDirectory: .contexts:
"OWNER/REPO": |- Review this repository using its architecture and contribution rules. Treat generated files as outputs and trace findings to their source.ignoredAuthorPatterns:
- "[bot]"
- bot
- codex
- claude
- kimi
- deepseek
- deepcode
- copilotThe priority values are additive. superHotThreshold marks PR numbers red when
the score is above that value, hotThreshold marks PR numbers yellow at or
above that value, and lower scores stay green. Comment counts use distinct
human commenters and approvers after excluding the PR author and any
ignoredCommentAuthors pattern.
The review watcher state is intentionally separate from .pr.yml so frequent
poll updates do not rewrite user settings. startupScanDays controls the
startup catch-up window and 0 disables it. processExistingOnFirstRun: true
still opts into every already-requested PR after a fresh state file, regardless
of age. dryRun: true runs the enabled pipeline and keeps local results without
posting to GitHub.
postNoFindingsComment: true creates a summary-only review only when every
enabled agent completed cleanly; it follows the current autoSubmit delivery
mode.
agents is an ordered map. Its file order is the review order. Every agent block
accepts enabled, model, effort, and an optional command override. Omit
model to use that CLI's configured default. Legacy top-level agent, model,
reasoningEffort, and command settings remain readable and are migrated to a
single enabled pipeline entry the next time settings are saved.
Codex receives model and effort through its CLI and uses the global
sandbox, whose default is workspace-write. Claude receives both values,
runs in auto permission mode, and exposes only local read, search, shell,
edit, and write tools. Kimi receives model, but its CLI does not expose an
effort control; any configured Kimi effort is ignored. It runs with --auto
and an explicit profile exposing Read, Grep, Glob, Write, Edit, and
Bash, plus a pre-generated PR diff. DeepSeek runs through the deepcode CLI
in a cross-platform pseudo terminal because Deep Code requires a TTY. Deep Code
receives the model through DEEPCODE_MODEL; its effort may be high or max
and is passed through DEEPCODE_REASONING_EFFORT. A temporary project policy
allows reads, writes, deletion, tests, and Git inspection inside the worktree
while denying access outside it, network access, MCP, and Git-history mutation.
The original project policy is restored after the stage.
Each agent returns the same summary plus findings JSON object. The
coordinator validates the schema and diff anchors, removes exact repeated
findings, and appends accepted entries to the shared JSONL ledger.
timeoutMinutes applies separately to each agent. The worktree reset also runs
after timeouts and cancellation. Ephemeral sessions and ignored user
configuration apply where the selected CLI supports them.
workspaceDirectory is the root for repository clones and PR worktrees. The
directory is created when a review first needs it. It accepts an absolute path,
a path relative to .pr.yml, environment variables, or a ~/ home-relative
path. Cached clones use <workspaceDirectory>/repositories/<owner>-<repo>;
agents run in <workspaceDirectory>/worktrees/<owner>-<repo>-pr-<number>.
When omitted, it uses dataDirectory for backward compatibility.
Context lookup first checks the full case-insensitive repository slug, then the
short repository name for compatibility, then contextDirectory, and finally
an embedded context. Inline values use YAML literal blocks (|-), so Markdown,
comments, colons, and nested indentation are preserved.
Tagged commits build and publish these release assets:
pr-linux-x64pr-linux-arm64pr-windows-x64.exepr-windows-arm64.exepr-macos-x64pr-macos-arm64pr-ui-windows-x64.zippr-ui-windows-arm64.zip
Push a tag such as v1.0.0 or release/1.0.0 to create a GitHub release.
MIT.