feat(spa): skills mode UX — mode toggle, skills picker, request wiring (skills-mode PR-3)

[philmerrell]

Summary

PR-3 of skills mode (spec; follows #473 + #474). The user-facing half: a Skills/Tools mode toggle in the settings slide-over, a per-skill picker, and chat-request wiring against the PR-2 backend.

Changes

Frontend

• SkillService (mirrors ToolService): loads GET /skills/, optimistic per-skill toggles persisted via PUT /skills/preferences with revert-on-error.
• ChatModeService: loads the admin policy from GET /system/chat-settings (compiled-in defaults on failure). Effective mode precedence: policy lock > local selection > user preferred mode > admin default. Toggling persists preferredAgentMode (user settings) and the active session's agentType.
• Model-settings panel: a "Capabilities" Skills/Tools segmented control (hidden entirely when the admin disallows toggling — the server enforces regardless); in skills mode a Skills section that is the visual twin of the Tools section (toggle list, enabled-count badge, bound-tool counts, role-based empty state); the existing Tools section shows only in tools mode.
• Chat request: every non-assistant turn sends agent_type; skills mode sends enabled_skills + enabled_tools: [] (capabilities come exclusively from enabled skills' bound tools). Assistant turns are excluded and keep their exact pre-skills-mode payload.
• Session hydration: a conversation reopens in the mode it was using (preferences.agentType), mirroring the lastModel pattern; sessions without a stored mode inherit the current selection so a home-page choice survives into the first message.

Backend (additive)

• SessionPreferences.agent_type (validated skill|chat, merge-safe in the metadata PUT — added to both create/merge branches).
• preferredAgentMode on user settings (model + repository read whitelist).

Testing

• Frontend (ng test): 19 new tests (SkillService load/toggle/revert, ChatModeService precedence/lock/persistence/hydration/fallback, chat-request payload matrix incl. tools mode + assistant exclusion). Full suite 1215 green; production build clean.
• Backend: 7 new tests (agentType merge/create/422, preferredAgentMode aliases + repo whitelist). Full suite green — two pre-existing Hypothesis PBT tests (session-manager truncation, untouched by this PR) flaked under load and pass clean in isolation and on file re-run.

Manual test script (alpha)

1. As an admin, leave chat-mode settings unconfigured (defaults: skill default, toggling allowed).
2. Open the settings slide-over in a chat → a Capabilities segmented control shows Skills selected; a Skills section lists your RBAC-granted active skills with toggles and an enabled count.
3. Toggle a skill off → refresh → it stays off. Send a message → the request payload (network tab) carries agent_type: "skill", enabled_skills without the disabled skill, enabled_tools: [].
4. Switch to Tools → the Tools section replaces Skills; send a message → payload carries agent_type: "chat", your tool selection, and no enabled_skills.
5. Reopen the conversation from the sidenav → it comes back in Tools mode; a brand-new chat starts in your last-used mode.
6. PUT /admin/settings/chat with {"defaultMode": "skill", "allowModeToggle": false} → after a reload the segmented control is gone, only Skills shows; a hand-crafted agent_type: "chat" request is overridden server-side (verify via the turn running through the SkillAgent).
7. Open a conversation with an assistant attached → no mode control behavior change; payload has no agent_type.

🤖 Generated with Claude Code