Files
DroidDeck/app
Kurt Himebauch eb190f2dd7 feat(agent): Agent tab refinements: one-state composer, no snapping, notifications, choices, subagents, dictation, game frame, copy, History search (#2626)
* fix(agent): one-state composer with model and usage on top; reading back stays detached while the agent streams

- The composer no longer has a compact reading bar that a first tap expands: the editor is
  always the editor, and the model picker and context ring move to the top beside the menu.
- The pet chat drops its DeeDee title so that top bar gets the whole width.
- Scrolling back stops following new output; new entries and streamed text only scroll the
  list while it is at the bottom, and sending a message resumes following.
- When the editor slides back in or the keyboard rises, a list at the bottom keeps its last
  lines above the editor, even mid-drag.

* fix(agent): a revise while reading back shows the editor it focuses

* feat(agent): dictate a message, and attach what the running game shows

- With nothing typed, the editor's send button dictates instead, through Android's speech
  recognizer; what is heard follows the text already there. It asks for the microphone the
  first time.
- While a game session is open, a camera button attaches the game's current frame. It is
  copied from the game's own surface, as the bottom screen's screenshots are, so the chat or
  drawer over it is not in the picture.

* feat(agent): a notification when a turn ends unseen, and replies that end in choices offer them as buttons

- A turn that ends while neither the Agent page nor the pet is in view posts a notification:
  finished, failed or asking, with the reply's first line. Opening the Agent page clears it.
- Claude Code's question tool is not available in print mode, so the brief asks the agent to
  end a reply with a choices block instead. The chat shows its options as buttons, and the one
  picked is sent as the next message.

* feat(agent): a subagent's work shows under the call that started it

Claude Code tags a subagent's messages with the call that started it. Its text, thinking and
streamed deltas used to show as the main agent's reply, and its tool calls mixed into the
turn. Now only its tool calls are kept, stored with their parent, and shown inside the
parent's card: a live line with the step count and the latest call, and every call when
the card is opened. Its token counts no longer overwrite the context reading.

* feat(agent): copy a reply or a tool's text, and open a turn's changes in one place

- A turn's footer has a copy button for everything the agent said in it, and tool input and
  output boxes have one each.
- The files-changed chip opens every edit the turn made, file by file.

* feat(agent): search conversations in History

A field above the list keeps the conversations whose title, messages or replies hold the
text, ignoring case. Chat files are read off the main thread, a quarter second after the
last keystroke.

* fix(agent): sending while reading back jumps to the latest

* feat(agent): manage Claude Code's MCP servers and skills from the Agent menu

A new panel lists the MCP servers claude mcp list reports, adds one for every project from a
URL (over HTTP) or a start command (over stdio), and removes it; connectors from the
claude.ai account are listed but managed there. It also lists the skills in
~/.claude/skills with their descriptions, and deletes one after a confirming second tap.

* fix(agent): review follow-ups for the Agent tab refinements

- Choices only come from one fenced block that ends the reply.
- The Agent page counts as on screen only while its activity is started, so a turn ending
  during a game still notifies.
- Picking a choice or Go jumps to the latest like sending does.
- Dictation stops when the chat is hidden or the keyboard sends; stopping it by hand keeps
  the recognizer until its final result.
- A failed JPEG encode is not attached.
- The copied reply is read from the turn each time.
- MCP start commands keep their quoted arguments; the CLI is read beside its 90 s deadline;
  deleting a skill runs off the main thread and removes links without following them.

* fix(agent): scroll the shrinking list after layout, not during measure

* fix(agent): fence-aware choices parsing; dictation stops overwriting edited drafts

* fix(agent): ignore choices blocks nested in a code fence

* fix(agent): treat tilde fences like backtick fences when finding choices

* test(agent): cover a tilde-fenced choices block
2026-11-14 17:35:54 -05:00
..