The console now knows what it can do

What changed

The console sometimes answered ordinary requests with "there is no tool to process …" — while the tool sat right there. The prompt was the root of it: the model did not know a voice note arrives already transcribed (the text in the question is the audio), and it had never seen an example of creating something or pulling a report.

The console's system prompt was rewritten:

• An honest introduction: the console knows it is this business's operations console and sees every product through its existing tools; and it knows a voice note is transcribed before it arrives.
• Short worked examples for the main flows: a voice note ("for each piece of work it calls task_create"), creating a project/task/lead (project_create, task_create, crm_lead_create), and reports and tables (the nearest read tool for that data; records as a table, comparisons as a chart block).
• A routing rule: no tool fits? os_more_tools first, then the nearest tool — never an "it doesn't exist", and never asking the person for a tool or a service.

The rules and the examples live in a class of their own (ConsolePrompt), separate from the code, with a test locking the catalogue of examples and the fact that they ride the prompt — so a quiet return of "I have no tool for that" cannot slip back in.

How to use it

1. Send a voice note — "register two tasks for the shop project": the console reads the transcript like a typed message and offers task_create for each.
2. Say "create a project called New site" or "a new lead named Sara": the same three creating tools come back; all three need an approval card, so nothing is created until you press "Allow".
3. Ask for figures — "what did we sell this month? compare with last month": the answer comes as a table, the comparison as a chart.

Good to know

• The examples ride every turn, which is why each is a single line; answers themselves are unchanged in length.
• Manual/automatic approval modes and the "manual only" per-tool setting are untouched — this change is about what the model knows about itself, not about what it is allowed to do.