Decide

Use this capability only on explicit user invocation. It is a one-shot API tool, not a managed role or a replacement for independent review, deterministic guards, profile routing or user approval. Do not dispatch another agent to use it.

Prepare

1. For an explicit $decide request, reuse the operator's existing Decide credential binding. Check it with the installed launcher's status command without displaying secret values. A successful configured: true means the operator has already selected this tool's credential; the explicit request authorizes its reuse for this bounded call. Do not repeat the generic reuse-versus-new question or reopen credential setup for ordinary use, including in a new conversation. Merely finding an environment variable or an arbitrary env file is not a saved credential choice. If the binding is absent, invalid or rejected, stop and report the specific setup/access problem without selecting another key. Apply the host's API credential setup policy when the user requests initial setup, a new key, replacement or a different credential destination. Those changes still need confirmation. Never read Codex OAuth tokens or change the parent runtime's provider, login, permissions, sandbox or approval settings. A configured binding does not authorize implicit workflow calls, retries or broader data collection.
2. Establish the question and evidence. Use only supplied text/JSON or files the user explicitly selected. Show a concise evidence scope before transmission. Do not collect the repository, conversation or private settings automatically. Remove any known secret; ask for sanitized evidence when necessary.
3. Build a request using the example below. The tool supports text or user text/JSON messages, 1–16 uniquely named questions, and a maximum 64 KiB request. Ask for missing choices or ordered score levels before sending anything.

{
  "model": "gpt-6-luna",
  "input": "The supplied task changes one documentation link.",
  "questions": [{
    "name": "route_suggestion",
    "type": "choice",
    "instructions": "Which workflow fits this task's scope? This is a suggestion only.",
    "choices": [
      {"value": "simple", "description": "One bounded, clear task."},
      {"value": "flow", "description": "Several dependent tasks with modest risk."},
      {"value": "pipeline", "description": "Complex work needing formal gates."}
    ]
  }]
}

Question types: predicate returns a probability; choice needs 2–255 distinct string/boolean values; score needs 2–10 ordered levels with label and optional description. Instructions are text. JSON evidence may use user message content parts of type input_json with data as an object or array of objects. Images, tools, streaming and arbitrary endpoints are outside this delivery.

Invoke

Resolve these paths relative to this skill file; installed copies are rewritten to the persistent support tree. Run with Node.js 18+ and Python 3.11+; entrypoints use Bash on Linux/macOS or PowerShell 7+ on Windows.

• Local validation: node ../../tools/decisions.js --input request.json --dry-run. No network or credential lookup occurs. Input may instead be - for stdin.
• If the operator's existing secret manager supplies a dedicated process-local OPENAI_API_KEY, invoke node ../../tools/decisions.js --input request.json.
• Once the operator has bound a private env file, invoke bash ../../scripts/decide.sh run --input request.json, or on Windows pwsh -NoProfile -File ../../scripts/decide.ps1 run --input request.json. These fixed entrypoints work from any workspace and retain the caller's cwd for request paths. They use the per-user saved credential location, not local workspace discovery. An explicit --env-file <confirmed-file> overrides it for one call without changing the binding. An invalid binding never falls back to a different key. The launcher does not execute env-file contents or export the credential to the parent session.

The direct Python entry python3 ../../scripts/decide.py supports the same commands. Without a saved binding or --env-file, run preserves the existing process-environment mode. Dry-run never reads binding metadata or credentials.

For first-time manual setup, local status, replacement and Windows notes, read the setup reference. The optional standalone entrypoint opens a set / status / clear menu when run without arguments in the human's terminal; set completes key-file setup and location binding together. Never automate secret entry or pass a key as an argument. Unrelated tasks, the main installer and workspace profile must not launch this optional setup. Installation and updates do not create keys or modify credential files.

Interpret and stop

Report validated named answers, model, usage and safe request ID when present. Describe probabilities/confidence as estimates. Score is a weighted level index, not necessarily an integer. Do not invent a rationale the API did not return.

Exit 0 means completed or dry-run; 2 means local input/credential error; 3 means API/transport/response failure; 4 means at least one refusal. Preserve refusals and errors without inventing answers. A returned suggestion never executes commands or changes workflow/profile settings by itself.

One live invocation sends one request to the fixed production Decisions endpoint. It has a 30-second deadline and 256 KiB response limit. No automatic retry, credential/model fallback or redirects. After timeout/transport failure, server outcome and billing may be unknown: report that before any user-authorized retry. Do not expose upstream error bodies, authorization headers, keys or stack traces.

Live calls may spend the selected project's credits/balance. Local limits are not a monetary cap. Do not alter billing or claim available credit from a successful request. See the official API contract.