The old flow fired OpenAI on wizard open with just ticket + chatter,
so the AI summary was just a paraphrase of what the user originally
reported — your engineering analysis (scope, limitations, recommended
approach) never made it to the owner. Restructure to a two-step flow:
1. Open wizard → empty findings + empty summary, NO OpenAI call
2. You write findings: scope / effort / approach / risk
3. Click 'Generate Summary from Findings' → OpenAI runs with
ticket + chatter + findings, where the prompt explicitly tells
the model to weight findings MORE THAN the original report
4. Review/edit, then Send
Bulk wizard mirrors the flow per line: each row gets its own
findings + summary, one 'Generate All Summaries' button fans out
parallel OpenAI calls using each line's own findings.
Updated SUMMARY_PROMPT to:
- Tell the model the support engineer's findings are authoritative
- Emit a bullet structure that leads with the recommendation, not
the user's restated ask
- Side with findings over the original report when they conflict
New tests cover:
- default_get does NOT fire OpenAI (regression guard for auto-AI)
- Findings text actually reaches the OpenAI prompt
- Send works with a manually-typed summary (no AI in the loop)
- Existing bulk + validation paths still pass with the new shape
Also folds in the deferred code-review #7: ThreadPoolExecutor now
explicitly cancels pending futures on timeout via
shutdown(wait=False, cancel_futures=True) so a slow OpenAI day can't
hold the wizard open for ceil(N/workers)*15s.
Bumps fusion_helpdesk_central to 19.0.2.3.0.
Smoke-tested live on nexa: opening the wizard makes zero OpenAI calls;
clicking Generate with findings='My findings: scope is XL, ~8h' makes
exactly one call and the findings text is verifiably in the prompt
body received by call_openai_chat.
c520803c84