refactor(agents): list supported agents in one registry

[Ziinc]

Each agent picker and check listed the supported agents by hand. The repo-level Default Agent and Review Agent selects had fallen behind and left out Copilot.

Change

• New src/lib/agents.ts:
  • AGENTS lists each supported agent with its id, label, hasPlanMode and hasModelPicker;
  • AgentKind is derived from that list;
  • also exported: DEFAULT_AGENT, isAgentKind, toAgentKind (narrows a stored setting) and agentInfo.
• New <AgentOptions> renders one <option> per agent. It is used by all four agent selects: the task input, app Default Agent, repo Default Agent and Review Agent. Copilot now appears in the two repo selects.
• Replaced the duplicated types and checks:
  • the AgentType union (agentDeepLink) and the ReviewAgent union with normalizeReviewAgent (agent-review-launch);
  • the AgentKind union in agentCommand, which now lives in agents.ts;
  • inline "claude" | "codex" | "cursor" | "copilot" unions in Dashboard, TaskInput, ShowWorkspace, ResolveConflictsDialog, terminal/types, types/sessions and api-types-checks;
  • hand-written === "codex" || === "cursor" … checks, now toAgentKind.
• Behaviour now read from the registry instead of name comparisons:
  • session names (Claude 1, Copilot 2) use agentInfo(...).label;
  • the Plan menu shows only for agents with hasPlanMode;
  • the terminal model picker shows only for agents with hasModelPicker.
• AgentIcons is a Record<AgentKind, Icon>, so adding an agent without an icon fails the typecheck. An unknown stored agent still falls back to the Claude icon.
  Adding an agent now means one entry in AGENTS, an icon, and its launch command in agentCommand.ts.

Not changed

• A few "claude" | "codex" | … unions remain in WorkspaceTerminalPane / workspace-terminal-pane/types.ts, plus the Reset toast label in AgentTerminalPanel. fix(terminal): keep focus on close, relaunch without the task on model change, drop Reset #628 and refactor: remove Mission Control and dead terminal code; key agent columns "agent-<id>" #629 delete those lines, so I left them to avoid merge conflicts.
• The backend has its own lists for binary detection (commands/binary.rs), the remote supervisor allowlist (cursor-agent binary names) and chat log parsing, and RemoteAgentId covers remote agents. Those are separate from agent selection.

Tests

• New src/lib/agents.test.ts.
• New integration test in default-agent.test.tsx: every agent picker (task input, repo Default Agent, Review Agent, app Default Agent) lists exactly the AGENTS ids. With the old repo settings file it fails on the missing copilot.
• Removed the normalizeReviewAgent tests; toAgentKind covers the same cases.
• Checks: tsc, lint and npm run check are clean. Unit tests pass (799/799). The agent integration tests pass (15/15): default agent, default model, plan split button, fake agent terminal.
• Screenshot specs: copilot-agent-option, agent-review-settings and agent-prompt-dialog-options pass. I checked the task toolbar with Copilot selected: it shows Run and no Plan chevron.
  Generated by Claude Code