/inbox
Dispatch inbound pull request reviews as background sessions that collect in `claude agents`. Use when reviewing multiple PRs, checking the review queue, batch reviews, or managing your review inbox across GitHub and GitLab. Pass --queue to dispatch an already-ordered queue.
$ npx -y skills add bendrucker/claude --skill inbox --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
- Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
- You can call itInvoke it directly when you want it.
- Slash command
/inbox
Context preview
The summary Claude sees to decide when to auto-load this skill.
Dispatch inbound pull request reviews as background sessions that collect in `claude agents`. Use when reviewing multiple PRs, checking the review queue, batch reviews, or managing your review inbox across GitHub and GitLab. Pass --queue to dispatch an already-ordered queue.
SKILL.md
inbox.SKILL.mdname: review:inbox
description: >
Dispatch inbound pull request reviews as background sessions that collect in
`claude agents`. Use when reviewing multiple PRs, checking the review queue,
batch reviews, or managing your review inbox across GitHub and GitLab.
Pass --queue to dispatch an already-ordered queue.
argument-hint: "[--queue]"
disable-model-invocation: true
allowed-tools:
- Monitor
- TaskStop
- Skill(gitlab:merge-request)
- Bash(gh:*)
- Bash(claude:*)
- Bash(jq:*)
- "Bash(bun ${CLAUDE_SKILL_DIR}/scripts/:*)"Review Inbox
Dispatch inbound PR/MR reviews as background sessions. Each review runs as its own `claude --bg --agent review <url>` session, collected in the `claude agents` view. You fetch the pending queue and dispatch. Monitoring, completion, and worktree isolation belong to the agent view and the harness, not to you.
Arguments
- `--queue`: dispatch an already-triaged, ordered queue handed in by a caller, rather than fetching your own. Default: off, the interactive flow that fetches and presents the queue below.
Curated Queue
When `--queue` is set, another skill has already gathered, triaged, and ordered the reviews.
Skip [Fetch Pending Reviews](#fetch-pending-reviews) and [Present Results](#present-results). Take the queue from the invoking context as given and dispatch each review in the order received. Everything from [Dispatch Review Sessions](#dispatch-review-sessions) onward is unchanged.
Fetch Pending Reviews
Both fetches return the `UNREVIEWED` bucket: PRs/MRs awaiting your first review. Approving, requesting changes, or starting a review drops an item. A re-request re-adds it.
GitHub
gh search prs --review-requested=@me --state=open --json number,title,url,repository
GitLab
Load `gitlab:merge-request` and run its `review-queue` command for the UNREVIEWED MRs across all projects. It emits `[{ url, reference, title }]` as JSON. The query and its review-state filter are documented and owned by the gitlab plugin; the inbox only delegates to it.
Present Results
Combine results from both platforms into a summary table. Ask which reviews to start.
Dispatch Review Sessions
For each selected review:
Resolve the Local Repo Path
The `review` agent checks out the PR branch with `gh pr checkout`, so the session runs inside a local clone of the PR's repo. Ask where the repo is cloned. If it is not cloned, clone it first. The path is required to dispatch.
Dispatch
bun ${CLAUDE_SKILL_DIR}/scripts/spawn.ts <pr-url> --repo-path <local-path> --data-dir ${CLAUDE_PLUGIN_DATA}`spawn.ts` runs `claude --bg --agent review <pr-url>` in the repo clone. The session collects in `claude agents`. On its first write (the review agent's `gh pr checkout`) the harness moves it into an isolated worktree, so parallel reviews never share a working tree. `spawn.ts` records the launched session in a dedup set and refuses a URL it has already dispatched, so a re-dispatch is a no-op.
Pass `--permission-mode <default|acceptEdits|plan|bypassPermissions>` to set the dispatched session's permission mode. Omit it to keep Claude's default.
Monitor
Reviews collect in the agent view. Manage them there, not here:
- `claude agents` (or the `ca` alias) opens the live view, grouped into needs-input, working, and completed.
- `claude agents --json` lists sessions for scripting.
- `claude logs <id>`, `claude attach <id>`, and `claude stop <id>` inspect or control one session.
To see what this inbox has dispatched and cross-reference session ids:
bun ${CLAUDE_SKILL_DIR}/scripts/dispatched.ts list --data-dir ${CLAUDE_PLUGIN_DATA}Lifecycle
A dispatched review runs to completion on its own and cleans up its own worktree. The inbox does not track or reclaim it. The dedup set only prevents double-dispatch. Manage it directly when needed:
- `dispatched.ts --reset --data-dir ${CLAUDE_PLUGIN_DATA}` clears the set at the start of a fresh inbox session, so a re-requested review can dispatch again.
- `dispatched.ts remove --url <pr-url> --data-dir ${CLAUDE_PLUGIN_DATA}` untracks a single review.
When you have dispatched the selected queue, present a summary of what you sent and point at `claude agents` for progress.
Monitor Loop (Hands-Off)
The interactive flow above is the default: fetch once, present, ask, dispatch. The loop is the opt-in hands-off mode. It polls the UNREVIEWED queues on an interval and dispatches a session for each newly-arrived review without prompting, so the queue drains itself while you work elsewhere.
Wire the Review-Queue Sources
`poll.ts` knows no platforms. Each `--queue` is a command that emits an UNREVIEWED queue as `[{ url }]` JSON. Pass one per platform you want polled, and omit the rest:
- GitHub: `gh search prs --review-requested=@me --state=open --json url`.
- GitLab: load `gitlab:merge-request` and pass `bun <ABS>`, where `<ABS>` is the absolute path the gitlab docs resolve to for `scripts/review-queue.ts` (for example `/Users/you/.claude/plugins/cache/.../gitlab/skills/merge-request/scripts/review-queue.ts`). Write the resolved path out in full. Do not reuse `${CLAUDE_SKILL_DIR}` from the examples below: it points at this inbox skill, not gitlab, so it resolves to the wrong directory.
Arm the Monitor
Pass this command to `Monitor` with `persistent: true` and a descriptive label. `watch.ts` runs a single long-lived bun process: each iteration it reads the dedup set, fetches all `--queue` sources, and prints one URL per line per newly-arrived review. Each printed URL is one event. `${CLAUDE_SKILL_DIR}` below is the inbox's own directory, where `watch.ts` lives:
bun ${CLAUDE_SKILL_DIR}/scripts/watch.ts \
--data-dir ${CLAUDE_PLUGIN_DATA} \
--interval 300 \
--queue "gh search prs --review-requested=@me --state=open --json url" \
--queue "bun <ABS>"A failed source emits a `{"type":"source-error",...}` line to stderr a
Read more
name: review:inbox
description: >
Dispatch inbound pull request reviews as background sessions that collect in
`claude agents`. Use when reviewing multiple PRs, checking the review queue,
batch reviews, or managing your review inbox across GitHub and GitLab.
Pass --queue to dispatch an already-ordered queue.
argument-hint: "[--queue]"
disable-model-invocation: true
allowed-tools:
- Monitor
- TaskStop
- Skill(gitlab:merge-request)
- Bash(gh:*)
- Bash(claude:*)
- Bash(jq:*)
- "Bash(bun ${CLAUDE_SKILL_DIR}/scripts/:*)"Review Inbox
Dispatch inbound PR/MR reviews as background sessions. Each review runs as its own `claude --bg --agent review <url>` session, collected in the `claude agents` view. You fetch the pending queue and dispatch. Monitoring, completion, and worktree isolation belong to the agent view and the harness, not to you.
Arguments
- `--queue`: dispatch an already-triaged, ordered queue handed in by a caller, rather than fetching your own. Default: off, the interactive flow that fetches and presents the queue below.
Curated Queue
When `--queue` is set, another skill has already gathered, triaged, and ordered the reviews.
Skip [Fetch Pending Reviews](#fetch-pending-reviews) and [Present Results](#present-results). Take the queue from the invoking context as given and dispatch each review in the order received. Everything from [Dispatch Review Sessions](#dispatch-review-sessions) onward is unchanged.
Fetch Pending Reviews
Both fetches return the `UNREVIEWED` bucket: PRs/MRs awaiting your first review. Approving, requesting changes, or starting a review drops an item. A re-request re-adds it.
GitHub
gh search prs --review-requested=@me --state=open --json number,title,url,repository
GitLab
Load `gitlab:merge-request` and run its `review-queue` command for the UNREVIEWED MRs across all projects. It emits `[{ url, reference, title }]` as JSON. The query and its review-state filter are documented and owned by the gitlab plugin; the inbox only delegates to it.
Present Results
Combine results from both platforms into a summary table. Ask which reviews to start.
Dispatch Review Sessions
For each selected review:
Resolve the Local Repo Path
The `review` agent checks out the PR branch with `gh pr checkout`, so the session runs inside a local clone of the PR's repo. Ask where the repo is cloned. If it is not cloned, clone it first. The path is required to dispatch.
Dispatch
bun ${CLAUDE_SKILL_DIR}/scripts/spawn.ts <pr-url> --repo-path <local-path> --data-dir ${CLAUDE_PLUGIN_DATA}`spawn.ts` runs `claude --bg --agent review <pr-url>` in the repo clone. The session collects in `claude agents`. On its first write (the review agent's `gh pr checkout`) the harness moves it into an isolated worktree, so parallel reviews never share a working tree. `spawn.ts` records the launched session in a dedup set and refuses a URL it has already dispatched, so a re-dispatch is a no-op.
Pass `--permission-mode <default|acceptEdits|plan|bypassPermissions>` to set the dispatched session's permission mode. Omit it to keep Claude's default.
Monitor
Reviews collect in the agent view. Manage them there, not here:
- `claude agents` (or the `ca` alias) opens the live view, grouped into needs-input, working, and completed.
- `claude agents --json` lists sessions for scripting.
- `claude logs <id>`, `claude attach <id>`, and `claude stop <id>` inspect or control one session.
To see what this inbox has dispatched and cross-reference session ids:
bun ${CLAUDE_SKILL_DIR}/scripts/dispatched.ts list --data-dir ${CLAUDE_PLUGIN_DATA}Lifecycle
A dispatched review runs to completion on its own and cleans up its own worktree. The inbox does not track or reclaim it. The dedup set only prevents double-dispatch. Manage it directly when needed:
- `dispatched.ts --reset --data-dir ${CLAUDE_PLUGIN_DATA}` clears the set at the start of a fresh inbox session, so a re-requested review can dispatch again.
- `dispatched.ts remove --url <pr-url> --data-dir ${CLAUDE_PLUGIN_DATA}` untracks a single review.
When you have dispatched the selected queue, present a summary of what you sent and point at `claude agents` for progress.
Monitor Loop (Hands-Off)
The interactive flow above is the default: fetch once, present, ask, dispatch. The loop is the opt-in hands-off mode. It polls the UNREVIEWED queues on an interval and dispatches a session for each newly-arrived review without prompting, so the queue drains itself while you work elsewhere.
Wire the Review-Queue Sources
`poll.ts` knows no platforms. Each `--queue` is a command that emits an UNREVIEWED queue as `[{ url }]` JSON. Pass one per platform you want polled, and omit the rest:
- GitHub: `gh search prs --review-requested=@me --state=open --json url`.
- GitLab: load `gitlab:merge-request` and pass `bun <ABS>`, where `<ABS>` is the absolute path the gitlab docs resolve to for `scripts/review-queue.ts` (for example `/Users/you/.claude/plugins/cache/.../gitlab/skills/merge-request/scripts/review-queue.ts`). Write the resolved path out in full. Do not reuse `${CLAUDE_SKILL_DIR}` from the examples below: it points at this inbox skill, not gitlab, so it resolves to the wrong directory.
Arm the Monitor
Pass this command to `Monitor` with `persistent: true` and a descriptive label. `watch.ts` runs a single long-lived bun process: each iteration it reads the dedup set, fetches all `--queue` sources, and prints one URL per line per newly-arrived review. Each printed URL is one event. `${CLAUDE_SKILL_DIR}` below is the inbox's own directory, where `watch.ts` lives:
bun ${CLAUDE_SKILL_DIR}/scripts/watch.ts \
--data-dir ${CLAUDE_PLUGIN_DATA} \
--interval 300 \
--queue "gh search prs --review-requested=@me --state=open --json url" \
--queue "bun <ABS>"A failed source emits a `{"type":"source-error",...}` line to stderr a
My personal plugin marketplace for Claude Code, Anthropic's AI coding assistant.
Repo: bendrucker/claude
Other skills on bendrucker-claude.
cleye
Type-safe CLI argument parsing with cleye, the standard parser for this repo's Bun scripts. Use when writing or editing any script that takes arguments (flags,…
activity
Report real device usage from ActivityWatch. Covers per-app time, window titles, and active vs idle spans. Use when asked "what apps did I use", "how long was…
history
Report shell history from atuin's local capture. Covers what commands ran, when, where, and how they exited. Use when asked "what commands did I run", "what…
bun
Bun runtime patterns. Use when running bun commands, working with package.json/bun.lock, writing TypeScript scripts under Bun, or developing Claude Code…
agent-team
Orchestrating Claude Code agent teams. Use when creating teams, spawning teammates, assigning tasks, configuring teammate modes, or setting up team quality…

