Skip to content
Ariekei

Pull requests and the agent

Connecting a repository, the Inbox, working on a PR with the agent, and reviewing it. Issues, tasks and cleanup: lifecycle.md. The review view in detail: review.md.

What it does

Section titled “What it does”
  • Connect a repo: point Ariekei at a local clone. It finds the GitHub remote (origin, else upstream, else any GitHub remote) and uses your gh login.
  • Inbox: your PRs and issues, polled from GitHub every 60 s (↻ refreshes now), in sections by your role (see design.md), and within a section by what needs you: agent waiting for permission → finished but cleanup blocked → CI failing → new reviews/comments since you last opened the item → agent working → the rest by last update. Each row shows CI, draft, review decision, unresolved threads and agent status. Polling is cheap on GitHub’s GraphQL budget (5000 points an hour, shared with gh and the agents): it asks which PRs exist and when each last changed (1 point per query) and fetches details only for PRs that changed, are new, or have checks running; every tenth poll and ↻ fetch them all. If the budget is used up anyway, the Inbox says so and polling waits for GitHub’s reset.
  • Browse (Ctrl+K): search the repo’s open PRs and issues with GitHub’s search syntax (words, #123, author:name, …) and add them to the Inbox’s Open section; star an item to keep it at the top.
  • Work item tabs: Overview (title, author, branches, labels, description, reviewers, checks) · Comments (review threads by file with their diff snippet, unresolved first; the conversation; reply, resolve/unresolve, new comment) · Review · Terminal (with a worktree). GitHub renders the markdown; Ariekei sanitizes it (DOMPurify) and opens links in the system browser.
  • Reviewing without a worktree: opening a PR fetches GitHub’s refs/pull/<N>/head into the clone as refs/ariekei/pull/<N>/head; no checkout, no agent.
  • Start working: one click per PR. Creates ../<clone>_pr<N> on the PR’s branch, tracking the PR head (fetched from the head repository’s remote, added like gh pr checkout if missing), so a plain git push updates the PR. Reuses an existing worktree that has the PR’s branch, pr-<N>, or a branch tracking the PR head. A reused local branch is fast-forwarded to the PR head; if it has its own commits it is left as it is and the UI says so.
  • Agent: your installed claude in the worktree (see docs/spike-results.md, agent integration v2). A new conversation gets the PR as context via --append-system-prompt-file: description, unresolved review threads with file/line, review summaries, failing checks with the failing part of the Actions job log. The agent waits for your instructions. Claude Code records the system prompt at the conversation’s first request, so the context is frozen there; Send PR feedback to agent sends the current threads, reviews and CI state as a message. Each agent, with what it starts, may use up to half the RAM: a command beyond that fails without taking Ariekei or the other agents along (src-tauri/src/confine.rs: a systemd scope per agent on Linux, a job object on Windows; macOS stops no process group to begin with).
  • Review: all changes vs the PR’s base (merge base → head, like GitHub’s Files tab), including uncommitted ones, or any single commit. Chunks come from git’s own hunks (CodeMirror’s diff only adds character-level detail inside them). When viewing the working tree, the right side is editable and can be saved.
  • Review comments → GitHub or agent: select code on the right to attach a comment; drafts are collected per item and shown in the diff below their lines. On others’ PRs the agent drafts review comments there too (its context and Ariekei’s MCP tools tell it how), and replies in existing threads, which you edit, delete or send back to it with a prompt. Send to agent delivers your drafts on your own work as one message (with a worktree). Submit to GitHub… sends them as a review (Comment, or Approve / Request changes on others’ PRs): drafts on lines of the PR diff at its head commit become line comments, the others are listed in the review text, and the panel says which is which. Replies go into their threads as part of the review: it is made pending, the replies are added to it (GraphQL addPullRequestReviewThreadReply), then it is submitted; if a step fails, the pending review is deleted and nothing is posted.
  • Switching and resume: each item’s tab, review scope, selected file and pending comments are stored with the item. Terminals keep running while you look at other items. A terminal nobody looks at whose agent has been waiting for 45 minutes is closed, unless the agent still runs processes of its own (e.g. background shells, which would end with it and which resuming does not restart); opening the item resumes the conversation (--resume). Restarting the app does the same.
  • Notifications: an OS notification when an agent waits for permission, finishes a turn or stops with an error, when a PR’s CI starts failing, and on new reviews or comments. Nothing for the item you are looking at in a focused window. While an Ariekei window is focused, the notice shows in that window instead (an OS notification is easy to miss with the app in front); clicking it opens the item.

State lives in SQLite (ariekei.db in the app data dir); worktrees, branches and the review refs live in your clone.

Testing

Section titled “Testing”
  • npm run check, npm test (git-hunk → CodeMirror mapping), cargo test in src-tauri (store, GitHub parsing, worktree create/adopt/fast-forward against a fake GitHub remote, inbox ordering, hooks, context rendering, notifications).
  • Live, read-only GitHub check: ARIEKEI_LIVE_REPO=owner/name cargo test -- --ignored live.
  • E2E e2e/agent_scenario.py (run with e2e/run.py; the agent is the fake in e2e/fake_claude.py): connect, Inbox, Start working, slash-command autocomplete, an agent commit after a permission prompt, review since base, a second commit showing up alone, Inbox attention, idle shutdown (not while a background job runs) and resume, restart.
  • E2E e2e/review_scenario.py runs against e2e/fake_gh.py (set ARIEKEI_GH), which serves fixture PRs and records every write; the scenario checks the exact review payload.
  • Verified on besu, read-only: Inbox with 23 PRs, search, Overview, Comments and review of another author’s PR (#11206); Start working on #11577 (worktree created on its branch in ~5 s, agent ready ~1.5 s later, context written).

Not covered yet

Section titled “Not covered yet”
  • Pending (multi-session) GitHub reviews, suggestions, comments on deleted lines (left side), editing or deleting your comments, and a markdown preview for what you write.
  • Accepting or rejecting single hunks; build/test status in the review view (planned).
  • Not exercised end to end: Send PR feedback to agent against real GitHub, and notifications on a real desktop (Xvfb has no notification daemon).
  • The first time Claude Code sees a repository it asks for trust in the terminal, with “No, exit” preselected.
  • WebKitWebDriver screenshots take 40+ s on this setup (an X capture of the display shows only black under automation), and its synthesized clicks miss elements at the window’s bottom edge; the scenarios click those in the DOM.