agent-bear-researcher — Supabase Edge Function

This is a Supabase Edge Function that implements the Bear Researcher agent in TradingGoose's multi-agent stock analysis system. It receives analysis requests from a coordinator, gathers insights from peer agents (market analyst, social media, news, fundamentals), constructs a position-aware prompt, calls an AI provider, and saves bearish analysis results back to the database. It participates in a multi-round debate with a Bull Researcher agent.

Architecture

The agent runs as a Deno-based Supabase Edge Function (serve()). On invocation: 1. Validates the POST request (requires analysisId, ticker, userId, apiSettings) 2. Initializes Supabase client and sets up a self-retry timeout (3 retries, 180s timeout) 3. Checks for analysis cancellation 4. Fetches insights from peer agents (agent_insights: marketAnalyst, socialMediaAnalyst, newsAnalyst, fundamentalsAnalyst) 5. Builds position context tailored to portfolio state (profit-taking vs stop-loss vs no-position) 6. Constructs a debate-round-aware prompt incorporating previous bull/bear rounds 7. Calls AI provider with retry logic 8. Atomically saves results via updateAgentInsights and updateDebateRounds 9. Notifies coordinator to check if more debate rounds are needed

Position-Aware Prompting

The agent's prompt construction is the most interesting part. It tailors its bearish stance based on portfolio context:

• Profitable position (above profit target): argues for taking profits before reversal, warns about greed, suggests reducing to target allocation
• Losing position (below stop loss): argues for cutting losses, warns about catching a falling knife, questions the investment thesis
• Within-range position: presents risk/reward analysis, suggests reduction or exit
• No position: argues why staying out is correct, suggests alternatives for available cash

Position sizing is evaluated against configurable min/max thresholds with near-threshold warnings.

Debate Pattern

The Bear Researcher engages in a structured debate with a Bull Researcher across multiple rounds. Each round: - The bear receives the bull's arguments from the current round - Must directly counter specific bull claims - Must add NEW risks and evidence (not repeat previous points) - References specific overly optimistic assumptions in the bull case

Error Handling

Categorizes AI provider errors into: rate_limit, api_key, ai_error, data_fetch, database, timeout, other. On error, calls setAgentToError which notifies the coordinator.

Imports (shared modules)

• atomicUpdate.ts — updateAgentInsights, updateWorkflowStepStatus, updateAnalysisPhase, updateDebateRounds, setAgentToError
• cancellationCheck.ts — checkAnalysisCancellation
• aiProviders.ts — callAIProviderWithRetry, SYSTEM_PROMPTS
• coordinatorNotification.ts — notifyCoordinatorAsync
• agentSelfInvoke.ts — setupAgentTimeout, clearAgentTimeout, getRetryStatus
• types.ts — AgentRequest, getHistoryDays

Technology

Deno, TypeScript, Supabase (PostgreSQL, Edge Functions, Auth), pluggable AI providers (OpenAI, Anthropic, Google, DeepSeek, etc.).