Skip to main content
Provider accounts, API keys on every provider, Custom presets, a proxy for subscriptions in other harnesses, usage per account, Codex async questions and reset credits, and a stale-run sweep for the chat actions pane.

​
Providers and accounts

  • Claude and Codex accounts are first-class. Each account has its own config home, label, and accent. The pre-existing login becomes the Default account with no action. Add an account from Settings and sign in through an embedded terminal.
  • Smart balance and auto-start per provider. With two or more accounts, ADE picks the account with the most headroom at chat creation. Auto-start sends one trivial request when a 5-hour window closes, so the next window is already open.
  • API keys on every provider that can take one. Keys are multi-credential, injected per harness, and models from a key appear under their provider. The old two-button custom provider form is gone.
  • Custom presets pair any harness with any model source: a native account, a stored key, or a subscription borrowed through the CLIProxyAPI supervisor, which is downloaded on first use, hash-verified, and supervised by the brain. Presets carry a name, accent, logo, subagent model, and permission mode, and sync through the account.
  • Usage per account. The usage popup shows one row per account in the provider brand color, with 5-hour and weekly bars, resets, and a “Use reset” action on Codex rows. The same rows reach iOS and the TUI.
  • The Custom mark is a gear-and-wrench glyph in the ADE purple.

​
Chat

  • Codex async questions render above the open composer and can be dismissed. Codex reset credits show as a chat notice once per window, re-armed on reset.
  • Claude resume_return dialogs render through the question card, with “Don’t ask again” per machine.
  • Chat runtime ownership is persisted. A second brain never adopts a live chat. Stale subagent and background rows in the chat actions pane are terminalized by a bounded sweep and healed in the renderer, so rows stop reading “running” a day after their runtime died.
  • OpenCode gains inline steer, honest full-auto permissions, and image rows. Grok ACP graduates from preview. Kimi exposes mode, model, and thinking controls. Qwen aligns with 0.24.0. Copilot moves to 1.0.86.
  • Workflow activity details and Ultra tiers appear in the transcript.
  • Orchestration mode is removed. Spawned chats, lineage, and the CTO spawn path are unchanged.

​
Fixes

  • Slash-command discovery reads YAML 1.1 booleans, so user-invocable: no hides a skill.
  • The turn diff on an uncommitted tree shows only the turn’s files.
  • A shared note on a multi-question card reaches the answer.
  • The Codex rate-limit warning re-arms after a reset. Usage tracking honors CLAUDE_CONFIG_DIR. Droid uses FACTORY_HOME_OVERRIDE.
  • The Work composer scopes lanes to the picked machine, and “Update & restart” targets the right release tag.

​
CLI and TUI

  • ade providers accounts and ade proxy manage accounts and subscription sign-ins. ade chat models prints a real table.
  • Every launch surface accepts --instance, --preset, and --credential, and rejects bad --arg and --input-json overrides before they reach the brain.
  • The TUI carries the launch identity through selection, resume, and launch, and spends a reset credit from /usage.

​
iOS

  • Per-account usage rows, the reset-credit notice with “Use reset”, and the reset action on the Usage page.