Skip to content

AGENTS.md

Latest commit

 

History

History
12 lines (10 loc) · 1.7 KB

AGENTS.md

File metadata and controls

12 lines (10 loc) · 1.7 KB
Raw

Wiley agent instructions

This repository implements Wiley, a single voice persona backed by a root Pi orchestrator and Pi subagents.

  • Pi sessions run on the model and thinking level in Settings, defaulting to gpt-5.6-luna at medium thinking. Agent-side models are restricted to the gpt-5.6 family by isAllowedBackendModel in src/main/settings/settings-schema.ts, which is the single knob for the catalog, the allowlist, the approval judge, the relay registration, and the spawn guard. Realtime voice models and the claude/codex worker CLIs are outside that rule. Fast mode, on by default, runs the root session at low thinking. Settings live in settings.json under the app's config directory (WILEY_CONFIG_DIR, else Electron's userData or ~/.wiley); secrets live beside it in secrets.json and never reach the renderer.
  • The Realtime voice session is a conversational frontend only. It cannot mutate the board, filesystem, shell, git, or subagents.
  • Every user task and edit flows through the root orchestrator.
  • Every Pi session receives the canonical voice transcript and may read subsequent conversation deltas.
  • Use the live-excalidraw skill and board tools for canvas work; never mutate renderer state directly.
  • Deliver user changes interrupt-first. After an interruption, verify the state of any command, file edit, or board transaction that may have partially completed.
  • The human experiences one persona. User-facing progress is first-person Wiley language and never mentions agents, subagents, engines, or layers.
  • Ordinary actions are automatic. Always block catastrophic deletion of the home directory, root, system directories, mounted-volume roots, disks, boot configuration, or credential stores.