Stage 7 — Консилиум

⚠️ These checkboxes are not a progress report. They were never ticked during the web-first migration, so an unticked box does not mean the work is outstanding. As of 2026-10-06: fan-out, the structured LLM-judge ({best_index, rationale}, winner highlighted, rationale shown, manual fallback), manual pick, notifications and 7.8 history are done. “Voting + Merge” appears only in the ROADMAP summary line, not in this spec, so it was not invented. Per-item breakdown: VERIFICATION.md.

Epic: Hercules Studio Status: delivered — fan-out, structured LLM-judge with rationale, manual pick, notifications, consensus history Estimate: 2-3 недели Dependencies (backend): нет (Studio orchestrates parallel /api/chat) Dependencies (Studio): Stage 2, Stage 1

Goal

Параллельный chat с N агентами + агрегация (LLM-judge + Manual pick). Notifications для async operations. Voting + Merge = next gen (future).

Tasks

7.1 — Consensus store

• stores/consensus.ts:
  • selectedAgents: Connection[] — agents selected for consensus
  • prompt: string — shared prompt
  • responses: Map<connectionId, ChatResponseDto> — per-agent responses
  • aggregationMode: "manual" | "llm-judge" — current mode
  • judgeAgent: Connection | null — selected judge for LLM-judge
  • aggregatedResult: string | null — final result
  • status: "idle" | "querying" | "aggregating" | "done" — pipeline status
  • send() — Promise.all(selectedAgents.map(a => client.chat(prompt))) parallel
  • aggregateManual(pickedConnectionId) — user picks best
  • aggregateLlmJudge(judgeAgent) — send all responses to judge, get best
  • reset()

7.2 — Consensus view

• views/Consensus.vue:
  • Top: agent selector (multi-select from connections, chips), prompt input, “Send to all” button
  • Middle: response columns (one per selected agent), side-by-side:
    • Column header: agent name, status (querying/done/error)
    • Column body: response markdown, badges (mode, confidence, skill, provider)
    • “Best” button per column (for Manual pick)
  • Bottom: aggregation panel:

    • Mode selector: Manual

      LLM-judge

    • If LLM-judge: judge agent selector (one of selected or separate)
    • “Aggregate” button → runs aggregation
    • Result: aggregated answer in markdown
  • Empty state: “Select agents and enter a prompt to start a consensus”

7.3 — Agent selector

• components/consensus/AgentSelector.vue:
  • Multi-select from connections (chips with remove)
  • Filter: online only (default)
  • Show: agent name, status, capabilities count
  • Min 2 agents required for consensus
  • Max N agents (configurable, default 10)

7.4 — Response columns

• components/consensus/ResponseColumn.vue:
  • Props: connection, response, status
  • Markdown rendering (markdown-it + highlight.js)
  • Badges: mode, confidence, skill, provider
  • “Best” button (Manual pick mode) → highlights column, sets as aggregated
  • Error state: red border, error message
  • Loading state: spinner, “Querying {agent name}…”

7.5 — LLM-judge aggregation

• components/consensus/LlmJudgePanel.vue:
  • Select judge agent (dropdown from connections, can be one of selected or external)
  • “Run judge” → send all responses + prompt to judge agent:

    POST /api/chat to judge agent
    body: { message: "You are a judge. Select the best answer to: '{original prompt}'. Here are {N} responses: [1: {response1}, 2: {response2}, ...]. Return JSON: {best_index, rationale}" }
    

  • Parse judge response: best_index, rationale
  • Highlight winning column, show rationale
  • If judge fails → fallback to Manual pick

7.6 — Manual pick

• User clicks “Best” on a response column:
  • Column highlighted with gold border
  • Aggregated result = that response
  • “Confirm” button → set as final

7.7 — Notifications (async operations)

• stores/notifications.ts (расширить):
  • Native notifications via Electron Notification API (IPC to main process)
  • Triggers:
    • Consensus query complete: “Consensus complete: {N} responses received”
    • Long-running chat (if response > 5s): “Agent {name} responded”
    • Workflow step complete (Stage 8): “Workflow step {name} done”
    • Escalation pending: “Escalation requires approval”
  • In-app notification center (bell icon in StatusBar)
  • Settings: enable/disable native notifications

7.8 — Consensus history

• Save consensus sessions to SQLite (new table or extend chat_history):
  • id, prompt, selectedAgents (JSON), responses (JSON), aggregationMode, result, createdAt
  • “Load past consensus” → list → click → view (read-only)

7.9 — Sidebar: Consensus activity

• Sidebar for “Consensus” activity:
  • New consensus button
  • Past consensus list (click → view)
  • Quick agent select (favorites?)

7.10 — Tests

• Unit: consensus store send/aggregate, LLM-judge parsing, manual pick
• E2E: select 3 agents → send prompt → see 3 columns → pick best → confirm

Acceptance criteria

• Agent selector: multi-select chips, min 2, max 10, online filter
• Send: parallel POST /api/chat to all selected → responses in columns
• Response columns: markdown, badges, “Best” button, loading/error states
• LLM-judge: select judge → run → parse best_index → highlight winner + rationale
• Manual pick: click “Best” → highlight → confirm
• Aggregated result shown in bottom panel
• Notifications: native + in-app, consensus complete, long chat, escalation
• Consensus history: save/load past sessions
• npm run build + tests pass

Scope / Likely files

• renderer/src/views/Consensus.vue
• renderer/src/components/consensus/ (AgentSelector, ResponseColumn, LlmJudgePanel)
• renderer/src/stores/consensus.ts, stores/notifications.ts

Dependencies

• Backend: нет (Studio orchestrates parallel /api/chat)
• Studio: Stage 1 (connections), Stage 2 (chat)

Risks / Rollback

• Parallel request timeout: some agents may be slow. Mitigation: per-agent timeout, show partial results.
• LLM-judge parsing: judge may not return valid JSON. Mitigation: fallback regex parse, then manual pick.
• Voting + Merge: future features, not in MVP. Documented in roadmap.

Links

• Epic: ../README.md
• Stage 1: stage_01_agent_scanner.md
• Stage 2: stage_02_chat_skills.md
• Stage 8: stage_08_workflow.md (workflow = multi-step consensus)