Skip to content
Agentmart
Submit a tool

regent-workflow

Complete Regent feature work directly, in Hermes/Astra + Claude pairing or Claude-only mode, with relevant specialist skills.

Skill by regents-ai

IndexedMIT
Write a review

Install

Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.

npx agentmarket install skill/regents-ai-regent-workflow

Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.

Get the weekly diff

The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.

We use your email only for this newsletter. Privacy

About

Complete Regent feature work directly, in Hermes/Astra + Claude pairing or Claude-only mode, with relevant specialist skills.

Use cases: productivity

No reviews yet. Installed it? Reviews come from verified installers only. How reviews work

Permissions

What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.

Tools
No tools found
Secrets
None asked for
Install scripts
None found
Runs as
Python package

Network hosts it talks to 1

  • codeload.github.com · found in install command

Files it touches

  • ~/.claude/skills/regent-workflow

What's inside

Skill name
regent-workflow
Description
Complete Regent feature work directly, in Hermes/Astra + Claude pairing or Claude-only mode, with relevant specialist skills.
License
MIT
Instructions
433 lines in SKILL.md · view on GitHub

Before installing

  • Ships executable scripts

Agent Skills spec issues

  • body: contains an absolute path from the author's machine

Files (4)

  • SKILL.md
  • design-system.md · references/
  • worktrees.md · references/
  • worktrees.py · scripts/

SKILL.md (first lines)

# Regent workflow

There are two development modes; the founder's request says which one applies.

- **Pairing:** Astra in the Hermes desktop app and a Claude thread work one lane
  together, following `docs/agent-pairing-template.md` in the workspace root.
- **Claude only:** a single Claude thread is the sole engineering agent for its lane.
  It reviews its own work and the founder reviews the product. It does not wait for
  a peer, votes, locks or channels.

In both modes the implementing agent owns the requested working result and implements,
reviews and verifies it directly, using ash-stack and the relevant specialist skills.
Every agent, Claude or Codex, loads regent-workflow, ash-stack and elixir-stack before
any Elixir, Phoenix, Ecto, Oban or Ash work, including reviews and pairing.
Every design or animation task also loads animejs and vgpu-island and follows the
Layers section of `repos/design-system/STYLE.md` (see below).
Delegation is optional, not a prerequisite; an unavailable peer or external coding
agent must not block direct implementation. This skill is workflow guidance, not a launcher.

## Product terminology

Agent token launches with revenue splitting for token stakers are **Revstake tokens**,
both internally and externally. Use Revstake in new plans, documentation and product
copy. Historical code identifiers may still say revshare, revenue-share or SUBJECT;
preserve their literal spelling when referencing source. This naming convention does
not by itself authorize a broad source rename or a change to contract economics.

Revstake token creation is **Base-only**. Robinhood permits **memestock auction
creation only**. This creation policy does not authorize deleting existing auctions
or disabling their withdrawals and claims. For Base Revstake, the approved LP locker
permanently locks the launch-owned position but lets anyone collect its TOKEN and
REGENT fees and deposit both in kind into the launch's fixed splitter. Keep the
existing swap hook and all splitter lanes, including the 2% skim on all three
recognized assets; no fee conversion or denomination redesign is wanted.

## Wallet identity and transaction state — all five sites

Apply the same rule in Regents, Autolaunch, Techtree, Patchbay and Keyfleet:

- Signed in, only the account's own wallets act: the verified session's
  `wallet_address` and `wallet_addresses` (founder, 21 Oct 2026: "signing out zerion
  should 100% drop that account"). When Privy's active wallet is one of them, panels
  show that wallet's figures and the server builds its steps for it. A wallet the
  account is not linked to, such as a wallet app left connected from before a Google
  sign-in, is ignored: panels stay on the account's own wallet, presses use the
  account's wallet this tab has connected, and no note asks the person to switch.
  Signing out disconnects every wallet Privy holds, wallet apps included, so the next
  sign-in starts with none. Signed out, panels show the active wallet's figures and a
  button's first press opens Privy's sign-in. A short address is always `0x`, the first
  four and the last four characters, joined by `..`. `onchain-buttons` shows the
  pattern.
- A signed-in customer never sees a wallet panel ask them to sign in again. Each panel
  reads the account's wallet from the server session when it mounts (balance,
  positions, form) and shows it straight away, after page jumps and reloads, then
  moves to Privy's active wallet when the page reports one that is linked. Privy's
  browser selection is often empty after navigation, so never wait on it. When no
  wallet is active, a press sends nothing and a note names the account's wallets and
  asks the person to open one; it never opens Privy's connect step.
- Never preserve pending transactions. Do not add browser-storage transaction
  queues, database pending-operation recovery, reload restoration, replay reports,
  recovery inboxes or resend orchestration. The wallet/blockchain owns transaction
  submission and pending state. This supersedes older preservation/recovery advice.
- Show transient feedback for the current interaction and refresh website state
  from verified receipts, canonical events and chain reads where available. Reload
  reads current state; it does not restore pending transactions or resend them.
  Confirmed chain-event indexing and ordinary saved creator drafts are separate
  from pending-transaction persistence and remain legitimate product data.
- Bind each action to the displayed record, chain and authenticated signer. Every
  press reaches the wallet; never block, defer, queue, serialize or deduplicate one.
  The one exception (founder decisions, 2026-09-28 and 2026-10-06): the site may block
  a press that is certain to fail. When the chain's own current state makes a
  transaction certain to fail, such as "144/144 Keys sold" or a Buy whose approval is
  not on the chain yet, the button stays visible but disabled with the reason beside it.
  A pending transaction, a repeat press, or one that only might revert is never a
  reason to disable.
- Remove existing pending-preservation machinery through scoped product changes;
  do not delete historical database records or change other deployments implicitly.

## An action only a person can do is a bug (founder rule, 28 Oct 2026)

Sean, 28 Oct 2026 08:49Z, verbatim: "this was in keyfleet/ , but it must be a stronger
rule: 'Taken as a rule: any action only a human can do on keyfleet.ai is a bug'
prepare for an agent being the owner of a privy address, and so it must be able to
either 1) browse the website and interact 2) webmcp 3) mcp 4) CLI -- to do things ."

Sean, 29 Oct 2026 05:30Z, verbatim: "1. a - except I would define it as: web ui, webmcp,
mcp, CLI, and claude 'connector' , so it is 5. All of this has to start in ash-template
as the golden blueprint 2. a"

- On all five sites and the template, an action only a person can do is a bug. An
  agent may own a Privy account itself, so everything an account's owner can do, an
  agent can do too, through each of five surfaces: the website in a browser, WebMCP,
  hosted MCP at `/mcp` for a signed agent, the Claude connector (the same site added
  to a person's or agent's Claude account as a custom connector), and the regents CLI.
  "Either" is not enough: every action needs all five.
- An agent owns its account through its own wallet (Sean, 2 a): its SIWA signing wallet
  is the account's Privy login wallet. The site makes that account from the agent's
  signed request, and the agent can also sign in on the website with the same wallet.
  A request signed by an account's login wallet acts as that account's owner.
- Template first: every surface of every shared action is built and checked in
  ash-template, the golden blueprint, before any site takes it ("Shared changes").
- One operation, five surfaces. Each surface calls the same Ash action with the same
  inputs, results, words and refusals; none rebuilds the rule. The agent access spec
  (`handoffs/regents-agent-access-spec-2026-10-28.md` in the project files) sets the
  identity model: the transport proves who is acting, and model-written arguments
  never choose the actor.
- The browser surface counts only when an agent can finish the action in a headless
  Chromium with its own wallet: no captcha, no step that needs a person's eyes (a
  picture with no words, a drag with no other way), and every control reachable by
  its name.
- Value transfer stays signed by the owner. When the owner is an agent, its own key
  signs; the site prepares the step and never holds or uses the agent's key.
- The Claude connector, an app a person connects with a token, counts like an agent
  signing its own calls. Once paired, the connected app runs the same operations as
  its paired agent, under the same policies and with the same answers: the site
  resolves the app's token to its paired agent's actor and runs the same Ash actions,
  with no sign-in at the sign-in service per call (the template's
  `AshTemplateWeb.Plugs.McpAgent`). A connected app that can only pair and check in is
  a gap row.

From regents-ai/ash-template, MIT.

More from regents-ai/ash-template

Signals

Stars
0
Forks
0
Open issues
0
Last commit (this folder)
Oct 24, 2026 (5 days ago)

From GitHub (fetched Oct 24, 2026).

Details

License
MIT
First seen
Oct 24, 2026
Last checked
Oct 24, 2026

More from regents-ai

Cite this page

regent-workflow skill. Agent Mart. https://agentmart.dev/skills/regents-ai-regent-workflow (last updated 2026-10-24)