多AI结构化讨论服务

@dmsdc-ai/aigentry-deliberation

Structured multi-AI discussions via MCP.

The only MCP server that lets multiple AI agents debate a question before committing to a decision. Run structured discussions across CLI agents (Claude Code, Codex, Gemini) and browser LLMs (ChatGPT, Claude Web, Gemini Web) with full audit trails, vote tracking, and synthesis.

What it does

• Structured deliberation sessions — pose a topic, route turns to each speaker, collect [AGREE]/[DISAGREE]/[CONDITIONAL] votes, synthesize consensus
• Smart speaker ordering — cyclic, random, or weighted-random strategies to balance participation
• Persona roles — assign critic, implementer, mediator, researcher, or free roles with built-in prompt templates
• Browser LLM integration — CDP-based auto-turn for ChatGPT, Claude Web, Gemini Web and more; clipboard fallback for unsupported providers
• Typed synthesis output — structured ExecutionContractV2 envelopes with decisions, actionable tasks, and optional experiment verdicts for downstream automation

Installation

One-command install — registers the MCP server with Claude Code and Gemini CLI automatically:

npx --yes --package @dmsdc-ai/aigentry-deliberation deliberation-install

This installs the server to ~/.local/lib/mcp-deliberation/ (Windows: %LOCALAPPDATA%/mcp-deliberation/), registers it in ~/.claude/.mcp.json and ~/.gemini/settings.json, and installs the deliberation-gate skill to ~/.claude/skills/.

Restart Claude Code / Gemini CLI after install. That's it.

Global install (alternative):

npm install -g @dmsdc-ai/aigentry-deliberation && deliberation-install

Uninstall:

npx --yes --package @dmsdc-ai/aigentry-deliberation deliberation-install --uninstall

Removes MCP registrations, installed files, and skill files automatically.

Diagnostics:

npx --yes --package @dmsdc-ai/aigentry-deliberation deliberation-doctor

Auto-diagnoses MCP configuration for Claude Code, Codex CLI, and Gemini CLI.

Forum Demo

When a deliberation completes, the synthesized result can be visualized as a Forum view.

Deliberation is the process. Forum is the output. When deliberation ends, the Forum is generated.

open demo/forum/index.html

MCP Tools

Start Flow

Speaker selection is enforced before a session can start. Raw candidate tokens cannot initiate a deliberation.

1. deliberation_speaker_candidates(...)
2. User picks speakers in the TUI
3. deliberation_confirm_speakers(selection_token: "<candidate-token>", speakers: [...])
4. deliberation_start(selection_token: "<confirmed-token>", speakers: [...])

Speaker Ordering Strategies

Persona Roles

Supported CLI Speakers

Supported Browser LLMs

Examples

See examples/ for working session scripts and synthesis output samples.

Experiment Retrospectives

For autoresearch-style keep/discard reviews, inject a compact experiment bundle after the session starts instead of bloating topic.

Guidelines:

• Keep injected JSON around 1.5–2KB
• Include only the last 3–5 relevant experiments
• Keep key_changes to at most 3 scalar before/after pairs
• Reference bulky artifacts (results.tsv, full program.md, JSONL logs) by path only

deliberation_start(...)
deliberation_inject_context(
  session_id: "experiment-review-123",
  speaker: "dustcraw",
  context: "{\"past_experiments\":[{\"experiment_id\":\"dg-20260310-001\",\"signal_kind\":\"INTEREST_DRIFT\",\"patch_summary\":\"Raised relevanceThreshold from 0.30 to 0.35\",\"patch_kind\":\"config\",\"key_changes\":{\"relevanceThreshold\":{\"before\":0.3,\"after\":0.35}},\"score\":0.08,\"score_label\":\"promotion_rate_delta\",\"metric_name\":\"promotion_rate_delta\",\"metric_delta\":0.08,\"verdict\":\"positive\",\"followup_action\":\"kept\",\"reasoning\":\"Threshold raise reduced noise; promotion quality improved 8%\"}],\"experiment_count\":1,\"success_rate\":1.0}"
)

Synthesis output with an explicit experiment verdict:

{
  "summary": "Lower the blast radius and re-run with stricter constraints.",
  "decisions": [
    "Keep the experiment loop bounded to one editable file",
    "Retry after restoring the failing test baseline"
  ],
  "actionable_tasks": [
    { "id": 1, "task": "Tighten editable globs", "priority": "high" }
  ],
  "experiment_outcome": {
    "verdict": "modify",
    "suggested_action": "iterate",
    "confidence": 0.78,
    "measurement_window_hours": 24
  }
}

deliberation-gate (Superpowers Integration)

Installs a skill that inserts multi-AI verification gates at key superpowers workflow decision points.

Scenarios:

• brainstorming — multi-AI design validation before writing plans
• code-review — multi-AI review via deliberation_request_review
• debugging — multi-AI hypothesis verification when stuck

Trigger: Semi-automatic — skill recommends deliberation, user approves.

Fallback: Without MCP installed, falls back to self-criticism-based verification. With MCP installed, runs full multi-AI discussion.

Auto-installed by deliberation-install. Manual install:

cp skills/deliberation-gate/SKILL.md ~/.claude/skills/deliberation-gate/SKILL.md

RFC: Prerequisites header for tool-dependent skills

Telepty Transport (Advanced)

For teams using telepty to manage AI sessions, deliberation supports routing turns through the telepty bus — enabling cross-machine and cross-session deliberation flows.

• deliberation_route_turn publishes typed turn_request envelopes on ws://localhost:3848/api/bus
• deliberation_run_until_blocked continues across cli_respond, browser_auto, and telepty_bus speakers until a manual block
• Transport delivery tracked with a 5-second inject_written ack window
• Semantic completion tracked with a 60-second self-submit window
• deliberation_synthesize validates and emits typed deliberation_completed envelopes for downstream automation

Cross-Machine Event Catalog

• Guaranteed (daemon-emitted): inject_written, session_health, session_register, session.replaced, session.idle, thread.opened, thread.closed, handoff.*, message_routed
• Best-effort (bus relay only): turn_request, turn_completed, deliberation_completed
• kind is the canonical event discriminator
• target identifies the telepty session target
• payload.prompt is the canonical prompt field for turn_request
• source_host is optional transport metadata for cross-machine tracing

Remote Reply Ingress

For distributed setups where a remote participant cannot call local MCP tools directly, use the deliberation-owned semantic ingress instead of proxying raw bus events:

deliberation_ingest_remote_reply(
  session_id: "...",
  speaker: "...",
  turn_id: "...",
  content: "...",
  source_machine_id: "peer-01",
  source_session_id: "remote-gemini-001",
  transport_scope: "remote_mcp",
  artifact_refs: ["results.jsonl"]
)

This preserves explicit provenance rather than inferring semantics from raw bus events.

What's New

v0.0.39

• Entitlement layer — Free/Pro/Team tier gating for deliberation features
• Gemini CLI model flags and Codex GPT-5.4 routing improvements

v0.0.36

• Clarified mandatory TUI speaker-selection flow and confirmed-token handoff before deliberation_start
• Documented active telepty session discovery with lightweight session_id + host/pid locators
• Cross-project delivery: active session lookup now resolves across project state directories

v0.0.35

• Manual selection enforcement: deliberation_confirm_speakers binds a fresh candidate snapshot to the exact user-picked speaker set before deliberation_start
• Telepty session candidates: active telepty sessions appear in speaker discovery with lightweight host/pid locators
• Telepty bus routing: telepty speakers route via typed turn_request envelopes with 5s transport and 60s semantic timeout tracking
• Structured synthesis envelopes: deliberation_synthesize validates typed payloads before telepty bus publication
• Codex CLI hardening: reduced prompt budgets, lower-friction exec profile, clearer timeout diagnostics

Ecosystem

@dmsdc-ai/aigentry-deliberation is one module of the aigentry ecosystem — each is independently published and independently useful.

Licenses: all MIT except @dmsdc-ai/aterm (UNLICENSED).

License

MIT