Agents/Skills/Tasks — Follow-ups (post-PR-#1019)

Lands on top of develop (which already contains PR #1019). Branch: feat/agents-skills-tasks-followups.

Tier 1 — High-impact

FU-1 · LLM dispatch in AgentRunService.execute

• New AGENT_AI_DISPATCH_FACADE injection token + contract (packages/agent/src/agents/agent-ai-dispatch-facade.ts). Same posture as the other agent tokens — agent-side declares the interface; api-side AgentsModule binds a thin adapter over AiFacadeService.createChatCompletion.
• AgentRunService gains two optional deps (AgentToolService + AgentAiDispatchFacade). When both unbound, the assemble-only branch still fires (backward compat with the existing 7-arg unit-test constructor calls).
• New runToolLoop private drives up to 10 round-trips: resolves tools via AgentToolService, feeds tool results back as tool messages, surfaces cap-hits + provider exceptions as ERROR run-log rows.
• Virtual transitionTask tool exposed only on task runs — captures the model's transition intent without doing the flip. finalize() runs the actual transition via AGENT_RUN_TASK_FINISHER.
• agent-run-execute.spec.ts covers chat reply, task transition capture, heartbeat, AI exception, tool-loop cap, and the assemble-only fallback.

FU-2 · 6 missing Agents API controller routes

Wires agents/plan.md §4's runtime surface:

Endpoint	Behaviour
POST /api/agents/:id/run-now	dispatches agent-heartbeat immediately via new AgentScheduleDispatcherService.dispatchOne
GET /api/agents/:id/runs?limit=&offset=	paginated AgentRun history + total
POST /api/agents/:id/runs/:runId/cancel	flips queued/running → cancelled via new AgentRunRepository.cancel
GET /api/agents/:id/skills	active SkillBinding rollup
GET /api/agents/:id/budget	30-day spend rollup via PluginUsageRepository.getTotalSpendCentsForOwner('agent', …)
POST /api/agents/:id/assign-task	pre-creates AgentRun + enqueues agent-task-execute

Writes are throttled 30/60s per plan §7.1, log AGENT_RUN_TRIGGERED / AGENT_RUN_CANCELLED / AGENT_TASK_ASSIGNED on success. assign-task de-dups against in-flight runs for the same (taskId, agentId).

New supporting surface: AgentRunRepository.countByAgent / cancel / findByIdAndUser, two new ActivityActionType values, and re-exported activity types from @ever-works/agent/activity-log.

agents.controller.runtime.spec.ts covers all 6 endpoints.

FU-3 · Mission/Work/Idea-scoped Agent creation entry points

• NewAgentDialog accepts a pinned prop ({ scope, missionId/workId/ideaId, parentLabel }). When set: step 1 is skipped, scope is locked, parent id forwards into createAgent(…), and a scope chip appears above step 2.
• Three new page routes mount the dialog with the matching pin: /missions/[id]/agents/new, /works/[id]/agents/new, /ideas/[id]/agents/new.
• Entry points: Agents tab on MissionTabs, "+ New Agent" button in WorkHeader, Bot affordance on IdeaCard next to the Build CTA.

FU-4 · 3 placeholder Agent-detail tabs

activity, skills, budgets now ship full surfaces:

• AgentActivityClient — paginated runs from agentsAPI.listRuns, status chip + trigger-kind tag, cancel affordance on queued/running rows.
• AgentSkillsClient — bound Skills list with priority + targetType, remove-binding affordance via existing DELETE /api/skill-bindings/:id.
• AgentBudgetsClient — Intl.NumberFormat money display, progress bar with cap-aware color tier (primary < amber < danger). No-cap case hides the bar.

agentsAPI gains listRuns / listSkills / getBudget / runNow / cancelRun.

FU-5 · Attachment UI on Task detail

• TaskAttachmentsSection mounted between transitions and conversation on TaskDetailClient.
• Drag-drop dropzone + multi-file picker. Uploads route through new /api/uploads Next proxy (mirrors works/[id]/kb/uploads/route.ts pattern: cookie → Authorization Bearer, body streamed with duplex:'half', upstream status surfaced verbatim).
• After upload, calls attachUploadAction → POST /api/tasks/:id/attachments.
• List rows show filename + size (captured client-side) + attached-at timestamp + per-row detach button.
• attachUploadAction / detachAttachmentAction server actions revalidate /tasks/:id.

Tier 2 — UX polish

FU-6 · i18n keys for Tasks/Skills/Templates

• New dashboard.tasksPage.{list, newDialog, detail, recurring, status}, .skillsPage.{list, detail}, .templatesPage.* namespaces in en.json.
• Threaded useTranslations through TasksList (filter labels + status select + empty state) and TaskDetailClient (Move-to heading, transition button labels, conversation heading, draft placeholder, Post button).
• Locale strategy matches the existing pages — en.json holds canonical keys; non-en falls back via next-intl until translation arrives. No es/fr keys added (neither has agentsPage either today).
• Follow-on threading for SkillDetailClient / SkillsPageClient / TaskRecurringSection / NewTaskForm / AstTemplatesBrowser is mechanical now that these keys exist; can ride a polish pass.

FU-7 · Recurring picker friendly controls

• Time-of-day picker (<input type="time">) → emits BYHOUR=H;BYMINUTE=M.
• Day-of-week chip strip for Weekly → emits BYDAY=MO,TU,…. Save blocked when Weekly has no days selected.
• Day-of-month numeric picker for Monthly (1-31) → emits BYMONTHDAY=X.
• Timezone free-text with datalist of common IANA names. Defaults to Intl.DateTimeFormat().resolvedOptions().timeZone, falls back to UTC.
• Client-side validation: regex on FREQ token + Weekly-needs-BYDAY. Server's full rrule parse remains authoritative (agent-side).

FU-8 · Skills binding picker UI

• New SkillBindingTargetPicker replaces the raw UUID textbox on the bindings panel:
  • tenant → disabled input (auto-fills server-side)
  • agent → loads via agentsAPI.list({ limit: 100 })
  • mission → missionsAPI.list()
  • idea → workProposalsAPI.list(['pending','accepted'])
  • work → workAPI.getAll({ limit: 100 })
• Search filter above the listbox (case-insensitive substring), sized to min(5, options+1).
• Falls back to a paste-uuid input when the endpoint errors or returns nothing.
• Added explanatory copy: "Lower priority numbers win when multiple bindings overlap — priority 1 = highest precedence."

FU-9 · Kanban drag-and-drop transitions

• HTML5 drag-drop without library dependency. Cards draggable; onDragStart stashes the task id (React state + dataTransfer fallback).
• Columns evaluate NEXT_STATUS[src.status] on each dragOver — only legal targets preventDefault() and light up with a primary ring. Illegal columns stay inert.
• Drop routes through the existing handleMove → transitionTaskAction flow (optimistic update + revert on rejection).
• Click-popover preserved as the keyboard-accessible fallback.

FU-10 · GitHub-sync v2 toggles UI

• 4 new toggles (Include Agents / Skills / Tasks / Task chat) added to GitHubSync.tsx, matching DataManagement.tsx's shape exactly.
• handlePush forwards all 4 into the existing pushToGitHub action (HIGH-pri PASS-4 fix had already widened its signature).
• includeTaskChat is gated on includeTasks (toggling Tasks off resets the chat flag).
• Persistence to UserSyncConfig is server-side; the UI defaults all-off so v1 syncs keep the v1 payload until the user opts in.

FU-11 · Templates browser content swap to ADR-010 catalog

• listAstTemplates(entity) gains an env-flag gate (NEXT_PUBLIC_AGENT_TEMPLATES_CATALOG=1). When on, lazy-imports serverFetch and hits /api/agent-templates?entity=<entity>. When off (default) or on with fetch failure, returns the hand-curated fallback constants.
• Lazy import keeps getAuthAccessCookie (server-only) out of the client bundle when the flag is off.
• agent-templates.spec.ts gains a 3-test describe block covering flag-on success, flag-on error → fallback, flag-off → no fetch.
• When ADR-010 ships the catalog endpoint, an operator flips NEXT_PUBLIC_AGENT_TEMPLATES_CATALOG to '1' and the swap is automatic; no further code change in this repo.

Tier 3 — Operator decisions (deferred)

These genuinely need operator input before they can land:

• FU-12 · Transition lattice done → in_progress divergence from spec FR-8 — soft re-open path needs either a spec carve-out or a tightening of TaskTransitionService.ALLOWED[DONE].
• FU-13 · AGENT_GIT_FACADE binding — needs operator preference on committer identity (Agent name? user email?) + git provider resolution.
• FU-14 · Phase 4 Git-mode AgentFileService writes — needs GitFacadeService scope-repo helpers that don't exist yet.

FOLLOWUP-PROGRESS.md carries the running log.

Test posture

Per the FU prompt's operating rules: tests were written as files but NOT run. The operator runs the suite (agent jest + web vitest + build) at review time. Type-checks pass for both the agent package (packages/agent) and apps/web; the API typecheck is clean for the agents/* surfaces (unrelated package-not-built errors for @ever-works/monitoring etc. are pre-existing in this worktree without a full root build).

What didn't ship

Beyond Tier 3 (which is intentionally deferred):

• Full i18n threading for SkillDetailClient / SkillsPageClient / TaskRecurringSection / NewTaskForm / AstTemplatesBrowser. Keys exist in en.json; threading is a mechanical follow-up.
• es/fr translations for the new namespaces — matching the existing convention of en-only for dashboard.agentsPage.* etc.
• ADR-010 unified catalog ENDPOINT itself (FU-11 wires the client; the server endpoint is operator-led work).