[trissim]

Toad's settings menu now has two tabs, Interface and Agents and shell. Every tab, group, dropdown, toggle and number box is generated from the settings declarations. A new field or Choice family appears in the menu with no UI code. The input box now has a layout setting, Normal (classic, bordered and padded), Compact (the current look, still the default) or Flush, and the change applies live.

Depends on OpenHCSDev/formbridge#16, now merged. uv.lock is re-locked to formbridge main b67c97e5. The carry PR on agent-comms (work/toad-input-settings-carry) lands after this one (#3568, a draft until this merges). #3567 reverts the copy I pushed to agent-comms main by mistake.

The generator (one owner per fact)

• Field semantics (label, help from the docstring, default, >= 1-style bounds from annotated_types, search text, validation, edit) live in formbridge/forms/declared_fields.py:DeclaredField, which a PyQt backend can also read. toad/forms/declared_views.py, field_path.py and kinds.py are deleted.
• Tabs are FormPage metadata on each top-level group: ui: Annotated[UiSettings, ..., INTERFACE] (src/toad/preferences.py). formbridge.form_pages derives the tabs and their order. FormView.pages(state) (src/toad/forms/form_view.py) builds one form per page, and SettingsScreen (src/toad/screens/settings.py) draws a Tabs row over a ContentSwitcher. Search filters every page.
• Widgets are picked by the DeclaredViews families. Choice families get dropdowns through FamilyChoices, bools get checkboxes, int and float get validated inputs that show their bounds, and nested dataclasses become sub-groups. The closed screen costs nothing because it is built only when opened.
• Hot path: each prompt geometry property reads app.settings.ui.input.box.<attr>. ObjectState returns its cached object, measured at 0.14 µs per app.settings.ui.footer read (timeit, 2000 reads). There is no lookup chain.

Scoped (lazy) inheritance: none

I checked every setting for a real scope (app -> agent kind -> thread or tab). No consumer reads any setting per agent kind or per tab. Every read is app.settings.<group>.<field>, the one app-wide value. Adding ObjectState scope layers would add states nothing reads, so no setting gets scoped inheritance in this PR. The candidates if a scope appears later are agent.thoughts, tools.expand and shell.* (per agent kind). The page split is Interface vs Agents and shell, not scoped vs plain.

Numbers

• Editable settings shown: 40 -> 43. Added ui.input.box, agent_symbol, multiline_symbol and shell_symbol. Moved prompt_message -> ui.input.placeholder. Deleted compact_input.
• Hard-coded answers deleted:
  • 3 if compact_input branches (prompt.py AgentInfo.padding, InfoBar.margin, PromptLabel.padding);
  • 3 glyph constants (PromptLabel.AI/SHELL/MULTILINE);
  • 2 states_default member constants and their branch (NRA 91487a46b2);
  • 2 patch_preview member constants and their branch (NRA a425f587fa);
  • 2 unused FormView parameters (paths= and focus=, used nowhere).
• Bugs fixed on the way:
  • prompt_message was declared and shown in the menu, but nothing read it. It is now the live placeholder.
  • title = "Success only" / "Fail only" / "When there are multiple sessions" / "When app is not focused" were dead: those members lacked TitledChoice, so the dropdowns said "Success", "Fail", "Multiple" and "Blur". They now show their titles (checked in the interpreter).

Inventory (AST scan of src/toad: 16 geometry, 15 glyph and 69 numeric constants, plus property branches)

Promoted:

site	now
widgets/prompt.py:105 AgentInfo padding if compact_input	InputBox.agent_info_padding
widgets/prompt.py:153 InfoBar margin if compact_input	InputBox.info_bar_margin(footer)
widgets/prompt.py:161-167 PromptLabel glyphs ❯ $ ☰ and padding branch	ui.input.agent_symbol / shell_symbol / multiline_symbol, InputBox.label_padding
preferences.py:87 prompt_message (dead)	ui.input.placeholder, read in Prompt.update_prompt
preferences.py:90 compact_input	ui.input.box (InputBox family in setting_choices.py)
PromptContainer had no frame at all	InputBox.frame_border/frame_padding/frame_margin (Normal draws a round border)

Not promoted, with reasons:

• Protocol/safety limits:

  • acp/permission_controller.py:12 PERMISSION_TIMEOUT_SECONDS
  • mcp_decision.py:28-29, mcp_inventory.py:20-21 timeouts and byte caps
  • comms_root.py:89 ROUTE_CHECK_SECONDS
  • acp/agent_session.py:135 reconnect_timeout_seconds
  • file_kind.py:16, clipboard_image.py byte caps

  These guard correctness, not taste.

• Scroll and preparation budgets:

  • widgets/presentation_window.py:60-69 max_items, minimum_widgets, lookahead_seconds, scroll_idle_seconds
  • widgets/conversation.py:809 MAX_LIVE_BLOCKS
  • acp/projection_attachment.py:67 MAX_PREBIND
  • widgets/line_markdown.py:32 RESTYLE_INTERVAL
  • widgets/acp_log.py:35 FOLLOW_INTERVAL
  • fuzzy_index.py:96 MAX_CANDIDATES
  • work_preparation.py:44,50, transcript_preparation.py:226 sizes

  These are measured values on the p99.9 frame path, and CLAUDE.md requires review of scroll-path changes. history_buffer_viewports is already a setting.

• Glyph sets:

  • acp/status.py:118,128 ⌛ ✔
  • line_blocks.py:101 −
  • conversation_actions.py:139 ␣
  • prompt_actions.py:41 ⏎
  • sidebar_layout.py:71,76 and widgets/side_bar.py:301,309 <── ──>
  • terminal_attention.py:112 🐸
  • textual_colours.py:126 GLYPHS

  These belong in one future GlyphSet choice (Unicode vs ASCII for fonts that lack them), not as per-glyph strings. Next step.

• Candidate preferences left for a follow-up:

  • widgets/conversation.py:120 AUTO_SESSION_TITLE_MAX_LENGTH
  • widgets/session_details.py:42 DETAIL_ROWS
  • screens/welcome.py:85 LIMIT
  • wire_history.py:297-298 tail/page limits
  • widgets/terminal.py:29 ESCAPE_TAP_DURATION (keyboard timing; I kept clear of the keymap work, P887)

• Design, not preference:

  • forms/* and screens/settings.py form geometry
  • widget_view.py:18-52 base defaults
  • widgets/session_tabs.py:43,179
  • ansi/_ansi.py:1302 terminal defaults

• Colours and motion: left to the UX agent's work/toad-ux-motion. They plug in as Choice fields under ui and need no UI code here.

NRA

I used claude-nra's scan of Toad origin/main 5fa29f3 (NRA eca10589, 105 detectors, complete: true, 0 omitted, 340 findings):

• 91487a46b2 (states_default) and a425f587fa (patch_preview): fixed by overrides on the family.
• 1f92187de2 (hand_rolled_registration on DeclaredViews -> AutoRegisterMeta): declined. AutoRegisterMeta is an ABCMeta and Textual widgets use _MessagePumpMeta. class X(Widget, metaclass=AutoRegisterMeta) raises TypeError: metaclass conflict (checked in the interpreter). DeclaredViews also registers many keys per class, not 1:1.
• The other 31 constant_subtype_answer families are runtime state identities (busy, ready, accepts_updates ...), not user preferences, so promoting them to settings would be wrong.
• ruff --select UP037 on the touched files: nothing to fix.

Proof (headless Textual pilot, isolated XDG and comms root; SVGs and PNGs in /home/ts/.cache/agent-scratch/claude-code/toad-settings/)

• before-main.png, before-settings.png: one long form, no tabs.
• crops.png: the input box, from top to bottom: before, Compact (identical apart from the now-live placeholder), Normal (round bordered padded box), Flush (no spacing).
• after-tabs-settings.png, after-page2-settings-keys.png: the two tabs. The second shows "Fail only" where it used to say "Fail".
• after-search-settings-keys.png: searching "input" shows the Input box group with the Layout dropdown and bounds or default notes.
• crop-live.png: the Layout set to Normal through the editor's own send while the app runs. The prompt re-lays out at once, and toad.json on dismiss holds {"ui":{"input":{"box":"normal"}}}.

Touches outside the core surface

• src/toad/keymap.py: one import line (declared_fields moved). The keymap menu is otherwise untouched (P887 owns it).
• Tests updated for the moved names: app.screen.values replaces app.screen.form.values; the test extension field names a page. I did not run them (no suites here); CI runs them.

Durable state

ui.compact-input and ui.prompt_message are retired keys, and settings decoding refuses unknown keys. The agent-comms carry tools/cutover/carry_toad_input_settings.py (branch work/toad-input-settings-carry) rewrites them once. compact-input: true becomes box: flush, and the message becomes the placeholder. I ran it on a sample file: dry run changed=1, apply changed=1, rerun changed=0, and the result decodes.

Correction after hosted debt (run 40718265703)

That run flagged two FunctionLocalImport hits in setting_effects.py (the new local Prompt imports in prompt_layout and prompt_text) and two UnannotatedParameter hits in setting_choices.py (the status parameter of each expands_preview).

• Hoisting Prompt into setting_effects would create an import cycle (preferences imports setting_effects, and widgets.prompt imports preferences). So the two effects now live on PromptSettingEffects in the leaf module src/toad/setting_widget.py, next to SettingVisibility.
  • apply_input_layout and apply_input_text are static Effect objects. Class access returns the Effect itself, so SavedSettings still applies a change through the same single path: field.metadata(Effect).apply(app, field).
  • Each effect finds open prompts with Effect.open_widgets(app, PromptSettingEffects).
  • Prompt adds the capability to its class bases, and MessagePrompt inherits it.
  • setting_effects.prompt_layout and prompt_text are deleted, with no aliases.
• Both expands_preview status parameters are now annotated ToolCallStatus, imported under TYPE_CHECKING only.
• The three local imports already on main (setting_effects.py:18/41/58) are unchanged and are not counted as fixed.
• Checks run (source probes only):
  • The order's check prints 2 (only the baseline Prompt and PromptTextArea imports remain).
  • A fresh import loads in all three orders: preferences first, prompt first, setting_effects first.
  • All five ui.input metadata entries are Effect objects: box is apply_input_layout, the other four are apply_input_text, and the two callbacks are distinct.
  • PromptSettingEffects appears once in the MRO of both Prompt and MessagePrompt.
  • No attribute name collides with one already on Prompt or Widget.
  • An AST census of src, tests and tools finds zero references to prompt_layout or prompt_text.
  • py_compile passes, and formbridge stays locked at b67c97e5.
• Not run by me: the tests that cover this change (tests/textual/test_settings_screen.py, tests/textual/test_forms.py, tests/model/test_saved_settings.py, tests/textual/test_quit_with_unsaved_settings.py), and no hosted debt re-read. These probes do not prove paint, focus or save behaviour.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RnH6dR85kyFKcc2tQAJLhK

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.