TirOSMAN Demo Mode#

Multi-agent demo powered by the real NVIDIA adapter (meta/llama-3.1-405b-instruct). Five concurrent projects × 30 tasks each × agent roles (pm, dev, frontend, qa, ceo), stored in Postgres with RLS, ticked through a Jira-like lifecycle: backlog → analysis → in_progress → code_review → qa → done.

When to use#

• User says "start the TirOSMAN demo" / "kick off the 150-task run" / "show the Jira board."
• User wants a cost preview before burning NVIDIA tokens.
• User needs to approve or reject QA-blocked tasks.
• User wants to wipe every demo-generated task and start over.

Call order#

1. demo_estimate first — returns projected tokens + cost, zero DB writes. Use it to confirm intent.
2. demo_start(company_id, auto_run=True) — creates tasks AND kicks off the batch executor. Pass auto_run=False if you want a confirm step before burning tokens.
3. demo_run(company_id, projects?) — resume execution when auto_run=False, after QA approvals, or after an interruption. Idempotent.
4. Poll demo_status(company_id) until all tasks are done/ failed/ qa.
5. If any task lands in qa, use demo_qa_approve or demo_qa_reject, then demo_run again to push it through.
6. demo_board(company_id, project?) for the visual column layout.
7. demo_reset(company_id) wipes everything the demo created.

Notes#

• demo_start only creates the rows. The batch executor runs inside TirOSMAN via FastAPI BackgroundTasks, triggered by POST /api/companies/{id}/demo-mode. If you called demo_start directly over MCP, the UI "advance" button or the HTTP endpoint has to drive execution.
• Task-status events are also POSTed to $TIROSMAN_DEMO_WEBHOOK_URL if set. Downstream webhook plugins (n8n, Zapier, Slack) see every transition in real time.
• Estimates are static — they do not change as tasks run.

Failure modes#

• No agents on the company → demo_start returns started=false with an explanation.
• Invalid company_id → tools surface a 404 from the HTTP layer.
• MCP bearer missing when TIROSMAN_MCP_API_KEY is set server-side → 401.