Streamline agent creation and library controls - #966

Return focus after closing agent creation

When a keyboard user chooses “Create a new agent,” the focused opener is the transient menu item, which disappears as the dialog opens. AgentCreateDialog is mounted without a finalFocus target pointing to newAgentButton—unlike the import/marketplace dialog—so cancelling or closing creation can leave focus on the document body rather than the durable “New agent” trigger. Pass that trigger through the shared dialog finalFocus contract, as required by the menu-to-dialog focus rule in DESIGN.md.

AGENTS.md reference: AGENTS.md:L46-L50

Useful? React with 👍 / 👎.

Restore focus to the card menu after delete cancellation

When deletion is opened from the card overflow, the focused opener is the temporary “Delete agent” menu item. The new trigger callback only opens AgentDelete; its confirmation dialog has no finalFocus reference to the surviving “Actions for …” button, so Cancel, Close, or Escape can strand keyboard focus after the menu item has unmounted. Extend this trigger contract to retain the overflow-button ref and return focus through the shared dialog contract required by DESIGN.md.

AGENTS.md reference: AGENTS.md:L46-L50

Useful? React with 👍 / 👎.

Keep focus anchored while Claude sign-in is replaced

When the Claude sign-in dialog is open and the login process completes, ClaudeSetup.run() starts the follow-up check, causing this branch to replace ClaudeSignIn immediately with the status text and then the model picker. That unmounts both the focused nested dialog and its opener, leaving keyboard focus on the document rather than a durable control. Keep the dialog mounted through the check and explicitly hand focus to the resulting model picker when it closes, as required by DESIGN.md’s modal focus-restoration contract.

AGENTS.md reference: AGENTS.md:L46-L50

Useful? React with 👍 / 👎.

Avoid consuming unapproved proposed dialog variants

This feature immediately consumes stable-tall and chromeSpacing="padded" even though the same commit explicitly documents both as proposed and expands the shared Dialog owner to support them. Repository policy forbids silently promoting proposed guidance for a local treatment; keep creation on an adopted dialog composition, or obtain scope approval and promote the shared recipe before using these variants.

AGENTS.md reference: AGENTS.md:L42-L45

Useful? React with 👍 / 👎.

Show recovery for Claude setup errors

When checkClaude returns an error state—for example, when the executable exists but claude --version fails—this condition routes to ClaudeSignIn, where missing is false and the UI misleadingly offers only “Sign in” while discarding status.message. In the creation flow there is no alternate recovery control, so another login attempt cannot repair or explain the failed check; render the error and a “Check again” action, and reserve sign-in for signed-out. This conflicts with DESIGN.md’s states guidance to keep failed-loading feedback and retry beside the failed operation.

AGENTS.md reference: AGENTS.md:L46-L50

Useful? React with 👍 / 👎.

Preserve the selected card when deleting another agent

When agent A is selected in the master/detail view and the user deletes agent B from B’s overflow menu, this unconditional callback calls browse() without a key and closes A’s still-valid detail panel. Only clear the selection when the removed pubkey is the selected one; otherwise retain the current card and panel, as required by the collection recipe’s selection-preservation contract.

AGENTS.md reference: AGENTS.md:L22-L25

Useful? React with 👍 / 👎.

Hide team editing when native agents are unavailable

In the browser host, snapshot.status is unavailable, but this unconditional tab still exposes Your teams and its New team form. Submitting that form calls AgentsService.saveTeam(), which rejects whenever native is absent with “Agents run only in the desktop app,” so the newly exposed workflow can never succeed; existing teams can even expose a destructive Delete action because removeTeam() lacks the same guard. Hide or disable team management in the unavailable state while retaining the independently available Import/Marketplace actions, consistent with DESIGN.md’s states rule that setup actions are offered only when available.

AGENTS.md reference: AGENTS.md:L46-L50

Useful? React with 👍 / 👎.

Preserve Claude's catalog during routine health checks

When at least one Claude model control is mounted, opening More → Provider settings invokes provider.check(), which reaches this method without an argument. Because refreshModels now defaults to true, even an unchanged readiness check empties the cached models and resets modelsLoaded; useSetupCatalog then starts another Claude discovery process and disables every mounted picker. Reopening the submenu repeatedly causes unnecessary subprocess churn, so routine health checks should preserve the catalog and reserve invalidation for completed install or sign-in operations.

Useful? React with 👍 / 👎.

Transfer focus after successful Codex sign-in

When a keyboard user activates Sign in from a Codex creation or library control and login succeeds, the provider clears SIGN_IN_ERROR, so this branch unmounts the focused CodexSignIn button and replaces it with the model picker without moving focus. Focus therefore falls back to the document body; keep the control mounted through the transition or explicitly focus the resulting model trigger, consistent with the design guide's focus-restoration contract.

AGENTS.md reference: AGENTS.md:L46-L50

Useful? React with 👍 / 👎.

Focus the URL field when upload is unavailable

When agent creation is opened while the relay is disconnected, community is undefined, so choosing Use image removes this trigger and leaves the image editor's first button—the upload dropzone—disabled. AgentCreateHeader still tries to focus that first button after the previous collection button unmounts, which strands keyboard focus even though the Picture URL field remains usable. Keep a focusable trigger mounted or explicitly focus the first enabled control, such as the URL input.

AGENTS.md reference: AGENTS.md:L46-L50

Useful? React with 👍 / 👎.

🤖 Low: lowering this selector to (0,1,0) fixed focus, but now .buzz-input:disabled (0,2,0, line 103) wins over background: transparent. While an agent is being created, the name field shows the grey --affordance-disabled box with zero inline padding. At the base, the title rule won through source order.

Fix: add a scoped :disabled override for the title appearance that keeps background: transparent.

🤖 Low: CreateModelChoices never passes allowCustomModel. A library card for an agent whose model was set to a custom ID in Settings (still supported, per the earlier product decision) therefore lists that model as "Unavailable model" (line 148), although the agent is using it.

Fix: pass allowCustomModel through for the description only, or use neutral copy such as "Custom model" when the value isn't in the catalog. Keep the finite list.

🤖 Low: each agent is its own one-cell row, so assistive tech announces an N×1 grid. ArrowLeft/Right move between rows, and ArrowUp/Down jump by columns() (lines 697-700), so ArrowDown skips agents relative to the announced structure.

Fix: group the rows by visual row, or make Up/Down move one row. Non-blocking.

Expose installation from the Codex model control

When Buzz's managed Codex binary is missing, provider.check() sets error to NOT_INSTALLED, but this branch special-cases only SIGN_IN_ERROR; the creation dialog therefore offers only “Try again,” which repeats the same check and cannot install Codex. Because the modal blocks access to the page-level Provider settings menu, users must abandon creation to recover. Handle NOT_INSTALLED here by exposing the setup install action, consistent with the design guide's requirement to keep recovery beside the failed operation.

AGENTS.md reference: AGENTS.md:L46-L50

Useful? React with 👍 / 👎.

Restore hover feedback on selectable agent cards

Pointer users now get only a cursor change when hovering a selectable card: the previous .card[role="option"]:hover fill was removed while the cards moved to role="gridcell", and no gridcell hover treatment replaces it. This leaves the card's primary click target visually indistinguishable from its resting state, contrary to DESIGN.md's collection/state rules that hover, selection, and focus remain distinct; restore an appropriate card hover state without changing selected-card feedback.

AGENTS.md reference: AGENTS.md:L46-L50

Useful? React with 👍 / 👎.

Clear unsupported effort when changing Claude models

When an existing Claude agent has an effort supported only by its current model—for example, xhigh on Sonnet—and the new library-card picker switches it to Haiku, this callback preserves the old current.effort. The settings picker explicitly clears levels absent from effortsFor, while the runtime passes every nonempty saved effort to the Claude process, so this shortcut can save an invalid model/effort pair and break subsequent runs. Apply the same effort normalization used by the settings picker before saving the compact choice.

Useful? React with 👍 / 👎.