roundtable-adaptive（圆桌会议适应性）

Roundtable v2 — Adaptive Multi-Model Orchestrator

Trigger: roundtable [--mode] [prompt] from any channel your agent monitors. Output: Posted to your configured output channel (set ROUNDTABLE_OUTPUT_CHANNEL in your OpenClaw config, or results are posted back to the triggering channel). Panel agents: Persistent sessions (mode="session", thread=true) — stay alive in the Discord thread for follow-up questions. Meta-panel analysts and synthesis agent are one-shot (mode="run").

The orchestrator = COORDINATOR ONLY. Uses your default model unless overridden in panels.json. Never argues a position, never joins the panel.

Core principle: the Meta-Panel (4 premium models) designs the optimal WORKFLOW for the task — parallel debate, sequential pipeline, or hybrid — then the right agents execute it.

Configuration

Before using, set your output channel in panels.json (or the triggering channel is used):

{
  "output": {
    "channel": "discord",
    "target": "YOUR_CHANNEL_ID_HERE"
  }
}

If using Discord threads (optional — creates one thread per roundtable for clean organization):

{
  "output": {
    "channel": "discord",
    "target": "YOUR_CHANNEL_ID_HERE",
    "useThreads": true
  }
}

Without this config, results are posted directly to the channel where the command was issued.

Cost transparency

--quick flag halves cost (1 round only).

Setup

Minimum (degraded mode — free):

1. Configure anthropic provider in openclaw.json (OAuth or API key)
2. Optionally add openai-codex for GPT-5.3 Codex
3. Done — Grok/Gemini slots fall back to Claude Sonnet

Full panel (adds Grok 4 + Gemini 3.1 Pro via Blockrun):

1. Install Blockrun: openclaw plugins install @blockrun/clawrouter then openclaw gateway restart
2. Fund the Blockrun wallet with USDC on Base (~$5-10). Address shown during install.
3. Full panel costs ~$0.13–$0.50/run; Claude and GPT slots remain free via OAuth.

Results are saved to {workspace}/memory/roundtables/YYYY-MM-DD-slug.json (created automatically).

Trigger Patterns

Optional: auto-trigger a dedicated channel

You can configure a Discord channel as a roundtable-only channel in your AGENTS.md:

Any message in channel [YOUR_CHANNEL_ID] → treat as a roundtable topic automatically.
No prefix needed. Message → auto-detect mode → create thread → spawn orchestrator.

This is entirely optional — the explicit roundtable command works from any channel.

Explicit trigger (any channel)

Explicit trigger (any channel)

• roundtable [prompt] — auto-detect mode, full flow
• roundtable --debate [prompt] — force parallel debate mode
• roundtable --build [prompt] — force build/coding mode
• roundtable --redteam [prompt] — force adversarial mode
• roundtable --vote [prompt] — force decision mode
• roundtable --quick [prompt] — skip meta-panel, use default panel for mode, 1 round only
• roundtable --panel model1,model2,model3 [prompt] — manual panel override, skip meta-panel
• roundtable --validate [prompt] — add Round 3 agent validation of synthesis
• roundtable --no-search [prompt] — skip web search (use only for purely theoretical/abstract topics)

Step -1: Create a Thread (FIRST ACTION)

Before anything else, create a thread in your configured channel and save the thread ID.

-1a) Dedup check (REQUIRED)

Avoid double-spawn if the same topic is triggered twice.

1. Normalize topic string:

- lowercase - trim - collapse multiple spaces - remove trailing punctuation

1. List recent threads in the target channel:

message(action='thread-list', channel='discord', channelId='[CHANNEL_ID]', limit=25)

1. If an existing active thread title matches normalized topic (+ same mode tag like [[DEBATE]]) created in last 24h:

- reuse that thread (THREAD_ID = existing_thread_id) - post: ♻️ Duplicate topic detected — reusing existing thread. - do NOT spawn a new orchestrator/panel

1. If no match: create a new thread.

-1b) Create thread (if no dedup hit)

message(
  action = 'thread-create',
  channel = '[your configured channel]',
  channelId = '[CHANNEL_ID from user config]',
  threadName = '🎯 [topic — max 8 words] [[MODE]]',
  message = '**Panel:** [model list]\
**Mode:** [mode] | **Rounds:** [N]\
⏳ Analysis in progress...'
)

Save the returned thread ID as THREAD_ID.

All subsequent message() calls use target = THREAD_ID, NOT the channel ID.

If thread creation fails or channel is not configured: fall back to posting directly in the active channel.

Step 0: Web Search Grounding (always first)

Run a web search on the topic before anything else — meta-panel and all agents will have current context.

web_search(query = prompt, count = 5)

Timeout policy: If web_search returns no result or errors within ~10s, do NOT block — continue immediately with CURRENT_CONTEXT = "No real-time data available (search failed or timed out).". The roundtable proceeds on model knowledge only.

Caching: If re-running the same topic within the same session, reuse the prior CURRENT_CONTEXT block — do not re-search.

Summarize results into a CURRENT_CONTEXT block (max 250 words):

• Key facts, recent developments, relevant data points
• Date of search
• If no useful results found: note "No relevant real-time data found" and continue

This block is injected into:

1. The meta-panel prompt (so they design the workflow with current context)
2. Every Round 1 agent prompt (so all panelists argue from the same updated baseline)

Step 0b: Meta-Panel — Workflow Design

Skip if: --panel flag used, OR --quick flag used.

Spawn 4 premium meta-analysts in parallel

Read panels.json → meta.models. For each:

sessions_spawn(
  task = filled prompts/meta-panel.md,
  model = model_id,
  mode = "run",
  label = "rt-meta-[A/B/C/D]",
  runTimeoutSeconds = 90
)

0b. Synthesize workflow from 4 recommendations

After collecting all meta responses, the orchestrator synthesizes the final workflow:

1. Workflow type: majority vote among 4 recommendations

- Tie → prefer hybrid (more flexible)

1. Stage composition: tally model recommendations per stage

- For each stage position, pick the most-recommended model - If a model is not in agents.defaults.models allowlist → skip, use next - If a model is your orchestrator's model → skip (reserved for the orchestrator, never a panelist)

1. Rounds: median of recommendations (round up if tie) — hard cap at 3 max, always

1. Synthesis model: most-recommended premium model not on the main panel

1. Log the decision (include in output header):

> "Meta-panel designed workflow: [type]. Stages: [N]. Panel: [models]. Synthesis: [model]."

0c. Workflow types explained

parallel_debate — classic roundtable

• All agents in Stage 1 work independently, same prompt
• Round 2: cross-critique
• Best for: debates, opinions, risk analysis, decision-making

sequential — output chains between stages

• Stage 1 agents produce outputs (drafts, code, research)
• Stage 2 agents receive Stage 1 outputs and review/validate/improve
• Best for: coding (write → review), research (collect → synthesize), creative (draft → refine)
• Round 2 within Stage 1 still possible; Stage 2 is a separate pass

hybrid — parallel within stages, sequential between

• Stage 1: N agents work in parallel on different aspects
• Stage 2: 1-2 premium agents receive ALL Stage 1 outputs and produce integrated output
• Best for: complex analysis (parallel research → premium synthesis)

0d. Panel degradation rule

If any agent fails and fallback is SAME MODEL FAMILY → log: ⚠️ PANEL DEGRADED — [role] substituted [original] with [fallback] (same family: [family])

Always surface this in META section of final output with actionable guidance:

• If degraded due to missing blockrun → "Action: Start Blockrun at localhost:8402 for full panel, or use --panel budget for stable 2-model run"
• If degraded due to model not in allowlist → "Action: Add [model] to agents.defaults.models in openclaw.json"
• If degraded due to API error → "Action: Check provider API key / quota, then retry"

Step 1: Detect Mode (if no flag given)

Step 2: Execute Workflow

parallel_debate (standard)

Round 1: Spawn all panel agents in parallel as persistent thread-bound sessions.

sessions_spawn(
  task = filled prompts/round1.md,
  model = model_id,
  mode = "session",        ← persistent — stays alive in the thread
  label = "rt-[role]",
  thread = true            ← bound to the thread from Step -1
)

• Save session keys: { "attacker": sessionKey, "defender": sessionKey, ... }
• Each agent writes their full response + SELF-DIGEST (last section)
• Collect all self-digests
• ⚠️ Agents stay alive — users can address them directly for follow-up questions

Round 2 (if rounds ≥ 2): Send cross-critique prompt to each existing session via sessions_send.

• Do NOT re-spawn — reuse session keys from Round 1
• [SELF_DIGEST] = this agent's own digest from Round 1
• [PEER_DIGESTS] = other agents' digests (labeled with role)
• Extract AGREEMENT SCORES from each response

Round 3 (if --validate): See Step 4.

sequential

Stage 1: Spawn agents in parallel as persistent sessions (mode="session", thread=true).

• Use standard prompts/round1.md.
• Round 2 cross-critique via sessions_send to existing sessions (no re-spawn).
• Collect full Stage 1 outputs for Stage 2.

Stage 2: Spawn new persistent sessions (mode="session", thread=true).

• Build prompt: prompts/round1.md base + prepend Stage 1 outputs as context
• Label: "STAGE 1 OUTPUT from [Role]: [full output]"
• Stage 2 agents review/validate/improve Stage 1 work and write SELF-DIGESTs

hybrid

Stage 1: Parallel persistent sessions (mode="session", thread=true), each with a different sub-task.

• Customize Round 1 prompt to specify each agent's sub-task:

> "Your specific task for this stage: [task from workflow design]"

• Agents write SELF-DIGESTs

Stage 2: 1-2 new persistent sessions (mode="session", thread=true) with all Stage 1 outputs embedded.

• Build prompt: prompts/round1.md base + "You are integrating and synthesizing the work of multiple agents. Their outputs: [all Stage 1 outputs]"
• Stage 2 produces the integrated output

Step 3: Consensus Scoring

After Round 2 (parallel_debate) or Stage 2 (sequential/hybrid):

Extract AGREEMENT SCORES from each agent's Round 2 response. Build score matrix: { agent_role: { peer_role: score_1_to_5 } } Consensus % = (sum of all scores / (n_scores × 5)) × 100 If no Round 2 scores (quick mode / sequential): omit consensus %, mark as "N/A"

Note on Round 3: Round 3 validation uses ACCURATE/PARTIALLY/INACCURATE — this is a separate metric from consensus %. Round 3 checks synthesis fidelity, not inter-agent agreement. Do NOT mix these two metrics. Consensus % comes only from Round 2 scores; Round 3 result appears separately in the META block as Validated: yes/no/partial.

Step 4: Round 3 — Validation (--validate flag only)

When to recommend --validate to the user:

• Consensus % < 40% (high disagreement — synthesis risks distortion)
• Redteam mode (adversarial stakes — synthesis must be bulletproof)
• Build mode with 3+ Stage 2 models (complex integration, easy to misrepresent)
• User explicitly mentions "high-stakes", "final decision", or "publishing this"

When NOT to use it: Quick mode, debate on subjective topics, or when time matters more than precision.

Draft synthesis first (Step 5 below), but do NOT post.

Spawn validation agents:

sessions_spawn(
  task = filled prompts/round3-validation.md,
  model = original agent model,
  label = "rt-r3-validate-[role]",
  runTimeoutSeconds = 60
)

Tally:

• 2+ INACCURATE → rewrite synthesis incorporating corrections
• 1 INACCURATE → note in META: ⚠️ [Role] flagged misrepresentation: [correction summary]
• All ACCURATE/PARTIAL → mark Validated: yes or Validated: partial in META

Step 5: Synthesis — Spawned Neutral Model

Never write synthesis yourself.

sessions_spawn(
  task = filled prompts/final-synthesis.md,
  model = [synthesis model from meta-panel recommendation, or anthropic/claude-opus-4-6 as default],
  label = "rt-synthesis",
  mode = "run",
  runTimeoutSeconds = 180
)

Fill prompts/final-synthesis.md placeholders:

• [ROUND1_SUMMARIES] → all self-digests: "[ROLE] ([model]): [digest]"
• [ROUND2_SUMMARIES] → critiques: "[ROLE] criticized [peer]'s [claim] because [reason]"
• [CONSENSUS_SCORES] → full score matrix + calculated %
• [DISCORD_THREAD_ID] → the THREAD_ID from Step -1 (synthesis agent posts here)

Post to Discord using THREAD_ID from Step -1 (not the channel ID). All round outputs and the final synthesis go into the same thread.

Step 6: Persist Results

Save to {workspace}/memory/roundtables/YYYY-MM-DD-[topic-slug].json:

{
  "date": "YYYY-MM-DD",
  "topic": "[prompt]",
  "mode": "[mode]",
  "workflow_type": "parallel_debate|sequential|hybrid",
  "stages": [{ "model": "...", "role": "...", "task": "..." }],
  "meta_panel_recommendation": "[summary of meta votes]",
  "panel_degraded": false,
  "panel_degradation_notes": "",
  "consensus_pct": "XX% or N/A",
  "synthesis_model": "[model]",
  "validated": "yes|no|partial",
  "elapsed_time_sec": 0,
  "synthesis": "[final synthesis text]"
}

Also append one JSONL line to {workspace}/memory/roundtables/scorecard.jsonl with: ts, topic, mode, workflow_type, elapsed_time_sec, consensus_pct, validated, panel_degraded.

Edge Cases

Placeholder Contract

When filling prompt templates, apply this rule for every [PLACEHOLDER]:

Never leave a [PLACEHOLDER] unfilled in a prompt. Unfilled placeholders confuse models and produce garbage output.

Score Parsing (Round 2)

Agents write scores in free text. Extract scores with this heuristic:

1. Look for the SCORES: block
2. Match pattern: - [Role]: X/5 — extract integer X (1–5)
3. If no clean integer found, scan for digit 1–5 nearest to the role name
4. If still ambiguous → assign 3 (neutral) and note [SCORE INFERRED] in META

Do NOT crash the workflow on a malformed score block.

Quick Reference: Default Panels (fallback if meta-panel fails)

debate:  [opus-4.6, gpt-5.3-codex, gemini-3.1-pro, grok-4] → Advocate / Devil's Advocate / Analyst / Contrarian
build:   [opus-4.6, gemini-3.1-pro, grok-4, gpt-5.3-codex] → Architect / Reviewer / Engineer / Implementer
redteam: [opus-4.6, gemini-3.1-pro, grok-4, gpt-5.3-codex] → Defender / Analyst / Attacker / Red Teamer
vote:    [opus-4.6, gemini-3.1-pro, grok-4, gpt-5.3-codex]  → 4-way vote panel
(all via blockrun/ prefix — see panels.json for exact model IDs and fallbacks)