Give the Today card's reasoning room: 1600 output tokens (#484)

OP OpenVibers committed 1 days ago · b26e7bab9f · 1 file, +1 −1

The provider answered with 1082 input tokens and 0 output: gpt-5-mini's reasoning used the old
600-token cap before it wrote the JSON. The workflow asks for 1600 and 60 seconds, and stays on the
JSON step so the provider is given the output schema.

Checked: ov test test/workflows.test.js test/live-templates.test.js test/latency-classes.test.js.

Changes

ddiff --git a/server/workflows/live.js b/server/workflows/live.js
iindex 5ae8b68..51f5c07 100644
---- a/server/workflows/live.js
++++ b/server/workflows/live.js
@@@ -432,7 +432,7 @@ const structured = [
              headline: STR(100), summary: STR(700),
              highlights: { type: 'array', maxItems: 3, items: { type: 'object', required: ['take_id', 'line'], additionalProperties: false, properties: { take_id: { type: 'integer' }, line: STR(240) } } },
          }, ['headline', 'summary', 'highlights']),
--        steps: [{ kind: 'llm', template: 'live.rundown.today', operation: 'generate', output: 'json', prepare: 'live.rundown.today', postprocess: 'live.rundown.today', params: { max_tokens: 600, temperature: 0.9, timeout_ms: 40000 } }],
++        steps: [{ kind: 'llm', template: 'live.rundown.today', operation: 'generate', output: 'json', prepare: 'live.rundown.today', postprocess: 'live.rundown.today', params: { max_tokens: 1600, temperature: 0.9, timeout_ms: 60000 } }],
          cache_mode: 'none',
      },
      {