trader-room

Prompt file imported from McCabeAI/market-watch-public-dashboard (.cursor/commands/trader-room.md). Copyright stays with the author.

Run the on-demand Market Watch Trader Room debate. If the user says go or supplies no extra topic, that is sufficient.

Read docs/TRADER_ROOM_PROTOCOL.md, docs/TRADER_ROOM_EVIDENCE_CONTRACT.md, docs/TRADER_RESEARCH_METHOD.md, and docs/TRADER_ROOM_ON_DEMAND.md first and follow them exactly. The evidence contract and on-demand execution contract are mandatory. This command is advocacy/orchestration only. You are not the arbiter. ChatGPT in the Market Watch Trader Room is the sole arbiter.

The single repository entrypoint is:

PYTHONPATH=. python scripts/trader_room_go.py go --topic "<user text or go>"

Default go freeze-validates the packet and runs the complete 14-advocate -> conflict aggregator -> one-pass rebuttal -> final aggregator workflow as a dry-run. It does not consume the production Trader Room model budget and must not launch the real 14-trader research run. Refuse --live unless a later authenticated human has armed TRADER_ROOM_LIVE=1.

User topic/hypothesis: use the text supplied after /trader-room or go. If none is supplied, use go and let the frozen packet define G10 FX scope, with emphasis on CAD/AUD/NZD and USD as benchmark, without inventing a trade.

Execution:

1. Preflight.

   • Confirm all 14 standing advocates and both aggregators are available as grok-4.6[] seats.
   • Write TRADER_ROOM_MODEL_POLICY from scripts/trader_room/models.py into the run context. Advocates and aggregators are exact model grok-4.6. Internal advocate subagents allowlist only composer-2.5.
   • Confirm the Market Watch Supabase MCP connection is available and read-only when assembling live evidence. If the cloud environment exposes the project-scoped connection under a different tool name, use that connection rather than inventing a replacement.
   • Record the current UTC cutoff.
   • Attempt and status all four mandatory evidence families from docs/TRADER_ROOM_EVIDENCE_CONTRACT.md: temperature gauges + hard-input drill-down; central-bank/article/news research feed; deterministic market-state packet; durable empirical research method. Record each as available, partial, stale, or unavailable before launching advocates.
   • If a missing mandatory family is essential to the user question, stop before spending the 14-agent run. Otherwise preserve the gap explicitly and continue with reduced confidence. The mechanical orchestrator treats all four families as essential by default.

2. Build and freeze one common evidence packet.

   • Query the relevant Market Watch Supabase operational state and provenance when assembling a live packet.
   • Load the current US/CA/AU/NZ 1–100 Inflation, Labor, Activity and Consumer temperature gauges, their actual hard inputs/weights or contributions where available, and their corroborating/contextual evidence under the current score contracts. Preserve vintages, provenance and staleness.
   • Load the existing Market Watch central-bank research and normalized article/news feed relevant to the run. Preserve official-vs-commentary distinction, country/category tags, publication times, provenance and retained summaries/synthesis.
   • Load the freshest valid deterministic market-state snapshot under docs/MARKET_STATE_FEED_V1.md / scripts/market_state.py, including relevant sovereign rates, curves, cross-market RV, G10 FX, returns, realized-vol/distribution context, provenance and staleness. Preserve explicit unavailable/null states.
   • Load docs/TRADER_RESEARCH_METHOD.md in full into the common packet as research_method.
   • Use available private research/synthesis only while assembling the packet and only when genuinely accessible and lawful. Do not republish paywalled text.
   • The packet must include the named sections temperature_gauges, central_bank_research, news_and_research, market_state, research_method, source_index, and known_gaps.
   • Freeze this packet. Every advocate receives the exact same packet and cutoff. After freeze, no web, search, or new evidence acquisition.

3. Round 1.

   • Launch all 14 standing advocates concurrently, each exact model grok-4.6.
   • Give each the exact same frozen common evidence packet and user topic.
   • Instruct each advocate to apply research_method before its archetypal bias.
   • Each advocate may make at most two internal subagent calls, model composer-2.5 only, on the same frozen packet.
   • Do not show agents one another's Round 1 work.
   • Every advocate except no-trade-skeptic must return one cogent actionable trade. The skeptic may submit no-trade.
   • Collect exactly one TRADER_ROOM_CONTRIBUTION object from each.
   • Do not retry in a way that silently exceeds the Grok ceiling of 30 or Composer ceiling of 28.

4. Deterministic validation.

   • Verify 14 unique expected agent names and remits.
   • Verify required trade schema fields and packet evidence_refs.
   • Verify confidence is 0-100.
   • Verify the run ID and packet hash match across all contributions.
   • Verify the common packet contains all mandatory evidence sections and four-family availability statuses.
   • Reject executable price levels that lack evidence/source support.
   • Preserve nulls rather than filling gaps.

5. Conflict map.

   • Launch conflict-aggregator on exact model grok-4.6 with all 14 originals and the frozen packet.
   • Identify opposite directions, opposite currency exposures, incompatible macro/rates/regime assumptions, and trade-versus-no-trade conflicts.
   • Do not vote, rank, synthesize a winner, or treat confidence as a vote.

6. Round 2.

   • Each conflicted advocate gets exactly one rebuttal pass. Max 14 rebuttal Grok calls.
   • Give the unchanged frozen packet, the advocate's own original, and the opposing original trade(s).
   • No additional subagent calls. No new evidence.
   • The advocate may defend, amend, or withdraw and must explicitly shoot holes in the opposing case.
   • Validate each output against TRADER_ROOM_REBUTTAL.

7. Final aggregation and artifact persist.

   • Launch final-aggregator on exact model grok-4.6 with all originals, the conflict map, and every rebuttal.
   • Produce the structured PM handoff: all proposed trades, agreement clusters, conflicts, strongest evidence on each side, rebuttals, amendments/withdrawals, shared assumptions, unresolved questions/gaps, and durable artifact references.
   • Do not select a winner or house view.
   • Persist complete run artifacts under trader-room/runs/<run_id>/ with immutable run ID and evidence cutoff.
   • End the packet itself with exactly STATUS: AWAITING_CHATGPT_ARBITRATION.
   • Google Drive folder ID 1NS6Qb6vNGKM18_PW0zPl4NOIJZOLyfUD is an optional private Markdown copy, not a required preflight for go.
   • Never create an official trade-decision record. The packet is unarbitrated input only.
   • Do not put project output in ACP.

8. Return only a handoff receipt.

   • State the run ID and evidence cutoff.
   • State ARTIFACT_HANDOFF: trader-room/runs/<run_id>/ when persist and verification succeeded.
   • State any hard evidence-access failures.
   • End with exactly STATUS: AWAITING_CHATGPT_ARBITRATION.

Ceilings: baseline main Grok invocations = 16 plus only conflict rebuttal Grok calls; max 14 rebuttals; total Grok ceiling 30. Composer subagent ceiling = 28. No retries/model reroutes may silently exceed these ceilings.

This workflow is designed to run entirely in Cursor Cloud/Automations. Do not require a local checkout, local terminal, or local Cursor session for the user to say go.