agentloop

Plugin agentloop from marketplace arcblock-agent-skills in ArcBlock/agent-skills. Commands read from the repo's manifest — 3 days ago.

Skills (24)

• build-phases

  Execute a phased implementation plan (tasks.md) autonomously, one phase at a time: implement, then verify (static + dynamic + adversarial) before advancing. Use for a multi-phase build of one issue, attended or unattended.

• codex-review-backlog

  Scan merged (optionally open) PRs for unhandled Codex / bot review comments, re-validate each against main, fix the easy ones, track the hard ones. Use for "codex backlog", "scan Codex reviews", late bot findings after a merge wave, or a daily routine.

• design-review

  Iterative clean-context review of a design / plan document or issue: independent reviewers per round, scored and synthesized without prior-round bias. Use to review an implementation plan, architecture doc or design proposal before building.

• epic-conductor

  Attended end-to-end implementation of a whole epic: decompose into sub-issues, one isolated worker each, package tests, one clean-context review, one batched fix, PR, merge, next wave. Use to build an epic with a reachable human.

• factory-dispatch

  Staged factory dispatch (epic #6896, superseding the retired #4289 topology). Stage 1 — current dispatch path — is one local ARC instance scheduling locally installed code-agents via work-board Schedule → /dev/code-agents/.actions/dispatch-work (work + engine only; the host creates the run's worktre

• factory-planner

  The factory's own sprint planner (scrum master): wake only the work whose recorded wake condition fired, judge each into a closed verdict set, and advance self-plannable tracks stage by stage (design → build → accept) via the trusted plan-submit entry. State lives in work objects; GitHub issues are

• factory-portfolio

  The factory's portfolio planner, one layer above factory-planner: daily (full re-rank weekly) it clusters idea / planning / research work into ranked epic plans with value, blast radius, cost, order, release and per-task-type acceptance; cuts trivia with a wakeOn. Reversible calls apply, irreversibl

• factory-reviewer

  Data, not a workflow: the text of the factory's reviewer prompts (correctness contract, security contract, juror, jury-scoped re-review), read by the aos-factory review runner. Not invoked by hand.

• factory-steward

  Data, not a workflow: a factory node steward's reading rules and own-node engine rule, read by the code-agents provider. Not invoked by hand.

• fleet-report

  Read a fleet's telemetry (fleet.jsonl): rounds run vs skipped and why, output, per-repo×skill duration, leftover processes. Use for "is the fleet healthy", "what did it accomplish", "why hasn't this repo moved", or the HTML dashboard.

• fleet-setup

  Set up or update an agentloop fleet on a machine or in the cloud: generate/reconcile deployment.json + repos.json and install the schedule (local crontab block or cloud routines). Idempotent, interactive; never unattended.

• git-hygiene

  After a merge or multi-branch work: switch the primary checkout to the default branch, fast-forward, prune worktrees and branches with no unique work (safe by default); on request ("这些是不是该补 PR") investigates kept branches and opens belated PRs. Triggers: "切回 main 清理", "清理 worktree", "prune branches"

• impact-check

  Find related code that may also need changing for uncommitted changes, a commit, a range or a branch. Use before committing or opening a PR to catch missed updates.

• issue-graph

  Deterministic issue graph over GitHub sub-issues + dependencies: ready set, parent-rollup candidates, close-kicks, edge writing for spin-offs, and claim-comment fencing across parallel agents. Used by issue-sweep and issue-review.

• issue-review

  Process one GitHub issue end-to-end: read it with its docs, code and comments, verify against the landed implementation, and post evidence-backed findings. Covers doc-review, doc-audit, system-audit, research and idea issues; issue-sweep calls it per issue.

• issue-sweep-batch

  Form dispatchable epics from unassigned open issues: cluster by defect layer, prove file-level disjointness against in-flight epics (disjoint / overlap / unproven), keep an incremental ledger. Produces epics only; never dispatches or merges.

• issue-sweep

  Sweep open issues for unprocessed human input (comment or body) and act via issue-review: doc delete/update PRs, bug fixes, feature pipelines, closes after merge, comment-only for security or real decisions. --autofix-green also fixes unambiguous verifiable issues. Manual or scheduled.

• land

  Use when asked to "finish this" or land ONE target (work DID, issue, PR, or topic) all the way to merged, or a batch of explicit targets via /agentloop:land.

• media-upload

  Upload an image, video or a directory of them to the shared public asset repo and get raw.githubusercontent.com/main URLs (images embed inline, videos link). Binary-safe via the plugin's gh-upload scripts. Used by screenshot and recording flows.

• pr-review

  Independent clean-context review of ONE open PR or branch range: verify every claim against live code, run the changed package's tests, check cross-cutting impact, bot threads and sibling conflicts, then a verdict (MERGE / COMMENT / SUPERSEDE / BLOCK / CLOSE). --post writes it. Never merges.

• pr-sweep

  Batch-review all open PRs to a terminal state: clean-context pr-review per PR, dedup-close same-issue twins, auto-merge reviewed non-breaking PRs whose package tests pass; escalate only security, breaking or direction calls. Flags: --merge, --dry-run. Built for scheduled runs.

• repo-setup

  One-time onboarding of a repo to the agentloop skills: scaffold repo-profile, create the controlled labels, preflight the environment. Does not schedule routines.

• setup-routines

  Interactive setup/update of a runner's repo-declared cloud routines and local cron jobs from the repo's .claude/routines/ catalog (RemoteTrigger reconcile, crontab marker block, one-time prerequisites). The issue/pr-sweep loop belongs to fleet-setup.

• test-audit

  Find tests that cannot fail (no assertions, swallowed by catch, weakened / deleted / disabled coverage) via the TypeScript AST, each with a witness. Use on PRs touching *.test.*, test-quality sweeps, "vacuous test", "空跑的测试", "测试有没有用", "did this PR weaken the tests".

Similar plugins