run-tradingagents-native

name

run-tradingagents-native

description

Run TradingAgents-style multi-agent financial research directly with the current Codex model, without external LLM API keys. Use for stock, international ticker, futures, forex, or crypto analysis; market/sentiment/news/fundamental analyst reports; bull-bear debate; trader planning; aggressive/neutral/conservative risk review; portfolio-manager ratings; historical point-in-time analysis; persistent decision memory; or complete saved research reports.

Run TradingAgents Native

Run the reasoning workflow inside Codex. Do not call an external language-model provider and do not request a model API key.

Resolve files

Resolve the plugin root as two directories above this SKILL.md. Use the bundled scripts with the current python3; they use only the standard library.

Read these references as needed:

• Read roles.md before assigning analyst, debate, trader, risk, or manager roles.
• Read evidence.md before web research or historical analysis.
• Read state-schema.md before writing analysis_state.json.

Start or resume a run

1. Normalize the requested date to YYYY-MM-DD. Reject future dates.
2. Choose stock or crypto. Exclude the fundamentals analyst for crypto.
3. Create a run workspace unless the user supplied an existing run_state.json:

python3 <plugin-root>/scripts/run_workspace.py init <ticker> \
  --date <YYYY-MM-DD> --language <language> --asset-type <stock|crypto> \
  --output-dir <workspace-output-directory>

4. On resume, read run_state.json and continue from the first non-complete stage. Do not redo completed stages unless their evidence changed or the user requests a fresh run.
5. Read prior same-ticker decisions:

python3 <plugin-root>/scripts/memory_log.py context <ticker>

Collect market evidence

Run the no-key market collector and save its JSON output:

python3 <plugin-root>/scripts/market_snapshot.py <ticker> \
  --date <YYYY-MM-DD> --output <run-root>/evidence/market_snapshot.json

Mark market_data complete after validating the symbol, as-of bar, benchmark, returns, indicators, and warnings. If the endpoint fails, use a current callable finance tool or primary market-data source and record the substitution in sources.json; never fabricate missing values.

Use web research for company filings, fundamentals, macro events, news, and sentiment. Financial research is accuracy-sensitive and current facts can change, so browse by default. Enforce the analysis-date cutoff from evidence.md.

Run the agent teams

Use real subagents when subagent tools are available. Give each agent only the raw evidence, cutoff date, role contract, and output path; do not leak another role's conclusion before independent work is complete.

Run in this order:

1. Spawn Market, Sentiment, News, and—when applicable—Fundamentals analysts in parallel.
2. Validate citations and save each report under reports/1_analysts/.
3. Spawn Bull and Bear researchers in parallel using the completed analyst reports.
4. Act as Research Manager to reconcile the strongest evidence and disagreements.
5. Act as Trader to translate the research plan into a conditional action, timing, and sizing proposal.
6. Spawn Aggressive, Neutral, and Conservative risk analysts in parallel.
7. Act as Risk Manager to reconcile exposure, liquidity, volatility, invalidation, and downside controls.
8. Act as Portfolio Manager to issue one five-tier rating: Buy, Overweight, Hold, Underweight, or Sell.

If subagent tools are unavailable, perform separate sequential role passes and disclose that the roles were simulated within one Codex context. Never claim independent agents ran when they did not.

After every stage, update the checkpoint:

python3 <plugin-root>/scripts/run_workspace.py mark \
  --state <run-root>/run_state.json --stage <stage> --status complete

Build and render the report

Write <run-root>/analysis_state.json using state-schema.md. Include source URLs and dates, the snapshot reference close, benchmark, all role outputs, and the final structured portfolio decision.

Render the report tree:

python3 <plugin-root>/scripts/render_report.py \
  --state <run-root>/analysis_state.json --output-root <run-root>

Append the completed decision to persistent memory:

python3 <plugin-root>/scripts/memory_log.py append \
  --state <run-root>/analysis_state.json

Mark report and memory complete. Read reports/run_summary.json and reports/complete_report.md before replying.

Present results

Lead with the rating, conditional action, confidence, and analysis-date cutoff. Then summarize:

• decisive evidence;
• strongest bull and bear points;
• key risks and invalidation conditions;
• monitoring plan;
• report and state paths.

State that the result is research, not financial advice, and that no order was placed.

Guardrails

• Never request or use OpenAI, Anthropic, Google, or other model-provider keys. The current Codex task supplies reasoning.
• Never place trades, connect a broker, or imply simulated approval is real execution.
• Never use evidence published after a historical cutoff.
• Never invent citations, prices, filings, social sentiment, agent independence, or confidence.
• Preserve exchange suffixes and report any symbol normalization such as BTCUSDT to BTC-USD.
• Distinguish measured facts, sourced claims, role arguments, and the final judgment.
• Prefer abstention or a lower-confidence Hold when evidence is incomplete or contradictory.