agentleFS
metratonmetraton/gaia/skills/orchestrator-present-approval/SKILL.md

orchestrator-present-approval

Skill4 starsChanged 5 months agoMITView on GitHub

Use when presenting a returned APPROVAL_REQUEST for informed user consent

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Run gaia approvals question <approval_id> [<approval_id> ...] with the pending ids -- up to 4 questions in all: one per signature in Claude …

    line 51
  2. Distinctive:

    Call your question tool (AskUserQuestion in Claude Code, question in OpenCode) with that output unchanged, and print nothing about the …

    line 54
  3. Very specific:

    Read the decision with gaia approvals show <approval_id> (its State).

    line 82

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Present Approval — Orchestrator Branch

Gaia builds every signature. In Claude Code a signature is one English question whose first line is its title, Signature N/M: (the desktop app does not show the header), then who asks and how many commands (and the folder, when every command is git -C of the same one), each command numbered on its own line, and a closing What it does: / What they do: line with what the set does. In OpenCode each command is its own one-line question, marked [ GAIA-SECURITY ]. Details adds what each command does, its impact, the verification and the rollback. The specialist's phrases -- what each command does, its impact, the verification, the rollback -- are sealed with the command when it requests (subagent-request-approval); the frame around them (Signature N/M:, who asks and how many commands, What it does:) is composed by gaia/approvals/surface.py when the question is shown. Neither part is yours to write or change. You never print, copy, summarise or translate any part of it; you only run gaia approvals question and open what it gives you, unchanged. You can run it again at any time for the approvals still pending -- for example when several specialists return signatures.

A signature is approved only when its question gets Approve (in OpenCode, every one of its questions); a Reject rejects the whole signature. Every signature in a call is decided on its own. One call asks at most 4 questions: 4 signatures in Claude Code, 4 commands in OpenCode. A signature carries at most 4 commands either way, enforced when it is requested.

In Claude Code, any question you type that reads as a signature -- a Signature N/M, Firma N/M or Aprobación header, Approve / Reject / Details options, a [ GAIA-SECURITY ] or Signature N/M: opening, a <agent> requests <k> commands opening (or the earlier <agent> pide <k> comandos) -- is refused unless it is exactly what gaia approvals question printed.

Presenting

The flow is the same in Claude Code and OpenCode. Only you open the question: a specialist that asks for a signature, or is blocked on a command, asks the user nothing. It ends its turn and returns APPROVAL_REQUEST with the approval id in its contract. You decide when to ask.

A T3 command you run yourself is denied the same way, never sent to the host's permission prompt. Run the request line its denial prints, which carries --agent-id so the request is yours, then present the approval_id it prints like any other. A protected write you attempt is refused outright: dispatch a specialist for it.

  1. Run gaia approvals question <approval_id> [<approval_id> ...] with the pending ids -- up to 4 questions in all: one per signature in Claude Code, one per command in OpenCode.
  2. Call your question tool (AskUserQuestion in Claude Code, question in OpenCode) with that output unchanged, and print nothing about the signature. Gaia checks the call before the questions open -- the hook in Claude Code, the plugin in OpenCode, which fills the call with every signature's questions -- and ties each answer to its own signature. Only approvals your own specialists requested are accepted. In OpenCode, ask two signatures of the same specialist in separate calls: that specialist runs one approval at a time.
  3. Details: Gaia names the command to run -- the hook's message in Claude Code, your call's result in OpenCode -- gaia approvals question --details <approval_id> .... Pass its output to the question tool the same way. If Gaia refuses a question, run gaia approvals question again for the ids it names.

The user signs what the command does -- a push, an apply, a delete -- in the plain English every phrase was sealed in, seeing the whole route before the first signature; the tier vocabulary never reaches them. Each grant is dispatched while it is still alive.

What is not a signature

A denial that carries NO approval_id is categorical, not a tier decision: there is nothing to present and nothing to approve, and calling it T3 invites the user to sign a boundary no signature lifts. Name which boundary fired in plain terms -- a blocked command, a protected path, a DB-write guard -- and reroute the work through the governed surface.

After the answer

Read the decision with gaia approvals show <approval_id> (its State); confirm it left pending before resuming the requester. Approved: resume the specialist that requested it -- the same agent, not a new one -- with execution; the signature is bound to that agent and session. When a grant's TTL is running out mid-verification, re-mint it: verification after the mutation is narration, and "execute now, nothing before" costs more than a fresh signature. Rejected: nothing runs; resume the same requester so it closes its contract and removes what it created, instead of dispatching a new one. A typed answer is not a decision. You may withdraw a pending approval (pending-approvals), but the decision to approve is the user's alone.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.