Repository navigation
Added
- ACP agents are first-class. Devin, Claude Code, Codex, Gemini and
OpenCode over ACP now sign in with the agent's own method (or say "run
devin auth login"), fill the model picker with the agent's models (kept
across ^Tab and reopened threads), and ask permission with the same y/a/n
card agentty's own tools use. ACP CLIs found on PATH show in the provider
picker as "installed - Enter to set up", so no hand-written
acp-agents.jsonis needed. The agent's stderr stays off the TUI
(AGENTTY_ACP_STDERR=1shows it). - A subagent's answer types out live under its calls in the Agent card.
- Open a subagent's call in ^O: Tab / Shift+Tab pick a nested call, Enter
opens its output. - Panel height has four steps: compact 10, normal 14 (default), roomy 20,
tall 28 rows. The ^O / ^U reader and the accounts list follow it too.
Fixed
- ACP: an agent's prose and reasoning after a tool card no longer render
one fragment per line above the card; they continue below it. - ACP: agents' shell commands run. A whole shell line sent to
terminal/create(ls -la | head -3) was quoted as one program name. - ACP: the model picker works before the first turn, and sending never
blocks on "no model selected" for an agent with its own default. - Esc on a running subagent settles its nested calls, and a cancelled
call shows ⊘ instead of a red ✗. - Subagent card polish: "starting…" instead of a guessed role, a
"⋯ N more lines · ^O to read" hint on long reports, a clock-first
progress row, and fixed-width rows in narrow terminals. - ^U / ^O no longer lag on long reasoning. A streaming thinking block is
extended instead of rebuilt each frame (about 25x faster). - ^O / ^U: a long entry opened after a short one no longer keeps the
short one's height.
Full history: CHANGELOG.md