-- Fix max_tokens default: 32768 is the full context window, not a reasonable
-- output limit.  vLLM rejects requests where max_tokens >= context_window
-- because there's no room left for the input prompt.
--
-- Change the column default to 4096 (sufficient for structured JSON extraction
-- output) and update any existing rows still at the old default.

ALTER TABLE ai_agents ALTER COLUMN max_tokens SET DEFAULT 4096;
ALTER TABLE agent_variants ALTER COLUMN max_tokens SET DEFAULT 4096;

UPDATE ai_agents SET max_tokens = 4096, updated_at = NOW()
WHERE max_tokens = 32768;

UPDATE agent_variants SET max_tokens = 4096
WHERE max_tokens = 32768;