agentday-zeroclaw

A forkable, multi-agent GitHub-notification assistant — Track 3: Best Multi-Agent / Skill Composition.

Fork this, point it at your GitHub account, and it works through your whole notification inbox: it plans the inbox, fans out one sub-agent per notification to draft a response, runs an adversarial verifier over the high-stakes drafts, composes with specialist skills, and a summarizer assembles a dated digest. It drafts only — it never posts, reviews, labels, closes, or marks anything read without your explicit say-so.

Run it interactively in Claude Code (parallel fan-out via the Agent/Task tool), or deploy it as scheduled agents on a ZeroClaw instance — a genuine, ZeroClaw-native multi-agent fan-out with no Claude Code dependency: a poll cron runs a sonnet orchestrator that uses ZeroClaw's built-in delegate tool to hand each new notification to one of six model-matched sub-agents (with an opus verifier re-checking PR-review drafts), plus a daily digest to your chat channel.

---

What's in here

Two from-scratch agentskills.io-spec skills:

Skill	What it does
github-notification-orchestrator	The orchestrator: plan → fan out → verify → digest → compose
github-duplicate-check	Read-only duplicate / prior-art check (issues + PRs, by anyone) before you file or draft

The orchestrator bundles 7 model-matched agent profiles + 2 scheduled-agent prompts:

Profile	Model	Role
pr-review-responder	opus	Draft PR review notes
issue-responder	sonnet	Triage issues / assignments
mention-responder	sonnet	Draft replies to @-mentions
author-activity-responder	sonnet	Catch up on your own threads
ci-failure-investigator	sonnet	Root-cause CI failures
verifier	opus	Adversarial quality gate on high-stakes drafts
daily-summarizer	haiku	Build the dated INDEX (a script does the sorting)

(The ZeroClaw poll delegates to six of these as live sub-agents; the seventh, daily-summarizer, is the deterministic build_index.sh step — not a delegated agent.)

Model choices and the reasoning behind them: see model-selection-rationale.md.

The full development lifecycle (scope → design bet → eval/iteration log → deploy → observe) is in the orchestrator's ADLC.md.

How it works (Deep Agents pattern)

plan → fan out → verify → summarize → compose maps onto the four Deep Agents components — planner (plan.md), sub-agents (agents/), virtual filesystem (the dated triage/<date>/ blackboard — agents coordinate only through files), detailed system prompt (SKILL.md + each profile) — plus two novel extensions: skill composition and the adversarial verifier loop. Mapping in deep-agents-mapping.md.

See it work

A real run against a 145-notification inbox is bundled at examples/run-2026-06-27/: the plan, five per-item drafts (with the verifier's verdicts), the cross-skill hand-off, and the final INDEX.md.

Use it on your own repos

It's repo-agnostic — it reads your notification inbox via your gh auth, so there's nothing repo-specific to configure. See SETUP.md for:

• Claude Code — install the skills, authenticate gh, say "orchestrate my notifications". The interactive path fans out sub-agents in parallel via the Agent/Task tool.
• ZeroClaw — install the skills, apply the parameterized cron template deploy/zeroclaw-cron.template.toml (fill in your home path + your chat channel id), seed the state file, and enable. This path runs completely in ZeroClaw — no Claude Code dependency. Piece A is a poll cron that runs the sonnet orchestrator gh_notif: a read-only delta script finds new notifications, then the orchestrator routes each by reason/type and uses ZeroClaw's built-in delegate tool to hand it to one of six model-matched sub-agents (gh_notif_pr_reviewer/opus, gh_notif_verifier/opus, gh_notif_issue/sonnet, gh_notif_mention/sonnet, gh_notif_author/sonnet, gh_notif_ci/sonnet) — synchronously, one at a time, bounded by a per-tick cap, with the opus verifier re-checking each PR-review draft before the script commits state. Piece C delivers a daily digest to your channel — and (optionally) mirrors the drafts to a private GitHub repo so every digest line is a tappable link to that draft's full summary from chat or mobile.

Safety

Every agent is read-only on GitHub and writes drafts to local files. Nothing is posted/reviewed/labeled/closed/marked-read without an explicit, per-action OK. Acting on a draft is opt-in: you set a draft's status: accepted, and a deterministic shipper posts just its marked reply block as a comment (never a review/approval/merge/close/label) — see SETUP.md. See safety.md.

License

MIT — fork it, run it, adapt it. The bundled skills are agentskills.io-spec and declare license: MIT to match.