[zrmarley]

Summary

• Hold an agent card with mouse or pen to pick it up, slightly shrink it, and preview its new position as neighbours move. Drop commits; Escape cancels.
• Keep the arrangement on this device, scoped to the viewer and community. New agents append; agent settings, runtime, teams, and main-agent choice are unchanged.
• Reordering is drag-only, as requested in product review: no Move earlier / Move later menu items. Ordinary card keyboard navigation and existing touch long-press behaviour remain unchanged; there is no keyboard reorder command in this scope.

Implementation and design

Reuses the existing AgentCard, shared Select/Menu, @dnd-kit/core, Motion, and scoped shared/view-state owner. No dependencies or shared design-system APIs added. Stable agent keys preserve mounted controls and profile drafts.

Consulted the read-only Buzz design pack: AGENTS.md, DESIGN.md (Motion, Accessibility, Surface and depth), tokens/motion.json (duration-settle, easing-settle), tokens/elevation.json (shadow-sm), and components/registry.json. No proposed pack component is introduced. Current repository src/shared/design-system/DESIGN.md governs collection composition, contribution boundaries, focus, shared controls, and direct manipulation. Its semantic-role guidance supersedes the older pack's ramp-first guidance. The subtle card shrink is the explicitly requested local treatment. Reduced motion disables shrink and positional/drop animation.

Review

Code review and Slopaganda Panda ran independently, followed by bounded rechecks. Both report no remaining code blockers, with minimalness/elegance/correctness scores of 9/10.

The user subsequently requested removal of both menu reorder actions. This is an explicit drag-only product choice, not a replacement keyboard interaction. Its absence assertion failed before removal; persistence/error coverage is retained at the existing hook owner. Follow-up review found no blockers (9/10 minimalness, elegance, correctness); TypeScript, 77 AgentsPage tests plus 1 ordering-hook test, and all 4 Chromium/WebKit checks passed after removal.

Two review findings were reproduced before repair in both Chromium and WebKit:

• Holding a portalled Model option started a background card drag. The row now rejects pointer targets outside its own DOM.
• Escape followed by delayed mouse release selected the original card after the drag overlay had unmounted. A row-local capture guard suppresses that gesture's click while preserving keyboard activation and the next ordinary pointer click. No timer or document listener was added.

Validation

• bin/pnpm typecheck
• bin/pnpm exec vitest run src/bundled/agents/AgentsPage.test.tsx src/bundled/agents/agent-order.test.tsx — 78 passed (77 page + 1 hook). Covers absent reorder menu actions, stable mounts, scoped read/restore/appending, arrow navigation, malformed storage, save failure/retry, and removal reconciliation.
• bin/pnpm test:browser tests/browser/agent-reorder.spec.mjs --no-deps --project chromium --project webkit --workers=1 — 4 passed in both engines after menu removal.
• bin/pnpm test:browser tests/browser/agent-reorder.spec.mjs --no-deps --project chromium --project webkit --repeat-each=10 --workers=2 — 40 passed (1.6 minutes locally) after the animation-unmount barrier; focused follow-up review found no blockers (9/10).
• Biome on changed files, git diff --check, and mandatory pre-commit hooks.
• Mandatory pre-push types/related-unit tests, design-system checks, and applicable native-input gate passed again; remote head verified as 96e6ab3b9d3b3d2a9d38c50920bd7f67c437c9c5.
• Hosted CI run 40737759860 passed on 96e6ab3b9d3b3d2a9d38c50920bd7f67c437c9c5, including functional shards and repeated Chromium/WebKit checks. Required CI required and DCO Check are both successful. Previous run 40734373756 exposed a real-time Web Animations vs. test-clock synchronization gap; the test now observes the cancelled overlay unmount. No app behaviour changed, no timeout/retry was added, and no animation was disabled.
• Human confirmed the main drag interaction in the configured desktop development build before the two review fixes.
• Human confirmed the repaired popup-hold and Escape/delayed-release edges work perfectly.
• Human confirmed the menu reorder actions are gone.

Browser coverage

Adds 2 cases, removes none. Real browser hit testing, portalled controls, grid geometry, focus, and native release/click ordering cannot be established with jsdom. Persistence/error matrices remain in colocated React tests. After menu removal, failed-save error/retry is asserted at the real hook rather than via the deleted menu action; the simple page alert wiring is not independently exercised by a failed drag. Both review regression scenarios failed before the repair and pass afterward in both engines. Controlled clocks exercise the activation/cancel boundaries; Select closure is checked through aria-expanded before advancing its exit transition.

Remaining gates / limits

Ready for review after agent reviews, local checks, hosted CI, and explicit human acceptance of the drag interaction, both repaired edge cases, and the final menu removal. Theme/narrow-width/enlarged-text/reduced-motion visual acceptance remains outside that human acceptance scope. Browser fixtures exercise the real Agents page with synthetic service records, not packaged desktop persistence or relay/runtime execution. The configured desktop development build was used for human acceptance. Broad/native lanes passed in hosted CI; no routine local just scan was run. Required code-owner approval is still a merge gate; this update does not merge the PR.

Latest fetched main at this update: a0ea099b5; incoming files were inspected and do not overlap the feature owners. Human acceptance is complete. Required CI and DCO must pass before ready-for-review; required code-owner approval remains a merge gate. No merge is requested by this readiness update.

Screenshots

Actual Agents page rendered with synthetic fixture identities only; no live community/account information.

Before dragging	During drag: Ada previews the last slot

Card actions after removing menu reordering

Agent review, agent exercise, and explicit human acceptance complete: buzz-review-completed. Final edit only synchronizes the browser test with the existing drop animation; accepted app behaviour is unchanged.

[kalvinnchau]

🤖 One persistence defect: reordering a temporarily incomplete inventory removes saved positions for absent agents.

[kalvinnchau]

🤖 [P2] Preserve saved positions for temporarily absent agents

Saving only the currently visible keys discards hosted/elsewhere agents while relay listings are unavailable. The agent service clears listings on session changes, but dragging remains enabled whenever connection.scope exists, including cached/connecting startup states. Reordering local agents during startup or reconnect therefore makes returning hosted agents append at the end.

A hook reproduction starts with saved [H, A, B], renders [A, B], saves [B, A], then restores inventory [A, H, B]: storage becomes [B, A] and display becomes [B, A, H]. Preserve absent saved keys while rewriting the reordered subset, or gate saving on an explicit complete-inventory signal. Add a regression covering save while an agent is temporarily absent and its subsequent return.