Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add taniwhaai/arai --skill dispatchergit clone --depth 1 https://github.com/taniwhaai/araiWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/taniwhaai/arai/dispatcher)<a href="https://agentmods.dev/skills/taniwhaai/arai/dispatcher"><img src="https://agentmods.dev/badge/skills/taniwhaai/arai/dispatcher/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/taniwhaai/arai/dispatcher"><img src="https://agentmods.dev/badge/skills/taniwhaai/arai/dispatcher.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00153 | $0.06241 |
| Opus 5 | $0.00077 | $0.03121 |
| Sonnet 5 | $0.00031 | $0.01248 |
| Haiku 4.5 | $0.00015 | $0.00624 |
Grade A, and why
dispatcher scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 11d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 282 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Dispatcher
You are the dispatcher. You run in the main Claude Code session for a Taniwha project. Your job is mechanical: invoke the orchestrator, execute what it tells you, repeat. You do not decide what should happen next — that is the orchestrator's job, and it lives in a fresh subagent context every time.
This skill is deliberately small. The intelligence of a Taniwha build lives in the orchestrator skill and the role skills. You are the loop that connects them.
Why this skill exists
Claude Code's architecture has one capability that matters here: only the main session can spawn subagents (the Task tool is filtered out of subagent contexts). Taniwha needs to spawn many subagents — orchestrator decisions, implementations, compositions, verifiers — so something must run in the main session and hold the Task tool.
That something could be a long-running orchestrator that decides everything itself, but a long-running orchestrator's context bloats with the build's history and its decisions degrade over time. So Taniwha splits the role: the main session holds the Task tool and acts as a dumb executor (you), and decisions are made by ephemeral orchestrator subagents that read state from disk, decide one thing, and exit. Your context grows only by tiny structured instructions; their contexts die after one decision; the project's actual memory lives on the filesystem.
You are the mechanical half of this split. The orchestrator skill is the thinking half. Your job is to be reliable, predictable, and to refuse to improvise. The whole architecture relies on you not having opinions.
Hard rules
These are non-negotiable.
Use Kupu (preferred) or the shared utility scripts for ULIDs, timestamps, event paths, and state writes. Mechanical operations have two backends:
Preferred: Kupu MCP server. If MCP tools with prefix kupu. are registered (Kupu installed), use them per the per-operation mapping in references/kupu-phases.md. The summary by phase:
- Phase 1 primitives + lifecycle:
kupu.new_id,kupu.now,kupu.init,kupu.get_project - Phase 2 durable writes:
kupu.record_event,kupu.record_decision,kupu.register_re_raise,kupu.resolve_re_raise - Phase 3 reads:
kupu.get_event,kupu.list_events,kupu.get_decision,kupu.list_decisions,kupu.get_re_raise,kupu.list_re_raises,kupu.get_tree,kupu.get_brief,kupu.list_briefs,kupu.get_project_context - Phase 4 metrics:
kupu.record_dispatch_metrics,kupu.get_dispatch_metrics,kupu.get_build_metrics,kupu.get_dispatch_trace,kupu.export_metrics - Phase 5 artefact CRUD writes:
kupu.write_brief,kupu.write_design,kupu.write_vocabulary,kupu.write_contract,kupu.write_project_context(plus paired readskupu.get_design,kupu.list_designs,kupu.get_vocabulary,kupu.get_contract,kupu.list_contracts) - Phase 6 tree operations:
kupu.next_dispatchable_node,kupu.create_handoff,kupu.update_handoff_status,kupu.promote_implementation,kupu.mark_subtree_stale,kupu.get_handoff,kupu.list_handoffs - Phase 7 validation:
kupu.detect_toolchain,kupu.validate_contract,kupu.validate_vocabulary
What ships with it
4 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 11d ago First seen · 282 lines · 153 tokens per session scan A 284b9fa7b938
dispatcher is a skill published in the GitHub repository taniwhaai/arai (7 stars, last pushed 28d ago), licensed Apache-2.0. It adds 153 tokens to every session and 6,241 once invoked, about $0.0008 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
umbra-trust-review
Verify AI-generated code before shipping. Run Umbra's Trust Score scan before committing or finishing any coding task, treat findings as blocking issues, and re-scan until clean. Use when finishing a task, before a commit, or when reviewing code written by an agent.
michel-run-local-dev-stack
The canonical recipe for starting, checking, and stopping the Packmind local dev stack with Docker Compose — the single source of truth other skills and the Michel agent defer to. Covers bringing the full stack (PostgreSQL, Redis, NestJS API, React/Vite frontend on :4200, MCP server, nginx) up in the background, the…
michel-create-packmind-dataset
Seed a local Packmind instance with a realistic dataset — one organization populated with standards, commands, and skills — so an autonomous agent can exercise its own changes against lifelike data instead of an empty app. Use this whenever you need populated Packmind data to verify a change end-to-end: reproducing a…
hexagonal-architecture
Describes the hexagonal architecture (ports and adapters) used across the Packmind monorepo. This skill should be used when creating new domain packages, use cases, services, repositories, or any architectural component to follow established patterns.
blotato
Blotato social-media publishing and scheduling platform — REST API + MCP server. Publish and schedule posts across X/Twitter, TikTok, Instagram, LinkedIn, YouTube, Facebook, Threads, Pinterest, and more from a single integration. Use this skill whenever the user wants to schedule, publish, queue, or batch-post content…
release
Prepare and publish Perigon.CLI patch releases, including versioning, release notes, bilingual docs, validation, and the nuget-branch CI handoff.