# JD's Dev Test Kit — 2026-07-10 evening

Target: https://dyus50nds0e0x.cloudfront.net — footer should read v0.6.11 or higher.
Streaming is LIVE when the flags flip (you'll get the word). Until then section 1 will still hit the 30s wall.

---

## 1. STREAMING — long prompts (copy-paste these)

**A. The one that broke Pete (pre-select Sonnet 4.6, then repeat with Opus 4.6):**
```
Write a detailed 2,500-word strategic analysis comparing how sponsorship valuation works in the NBA, the English Premier League, and Formula 1. Cover media-rights spillover, shirt/kit vs venue naming economics, measurement standards, and where valuations most often go wrong. Use section headers and at least two comparison tables.
```
PASS = text starts flowing within a few seconds and keeps flowing well past 30 seconds until a complete answer. No error card, no model badge surprise.

**B. Long coding task (pre-select Sonnet, then try Auto):**
```
Write a complete, production-quality Python module implementing a rate limiter that supports both token-bucket and sliding-window strategies behind one interface. Include full type hints, docstrings, a thread-safety section, unit tests using pytest, and a usage example. Do not abbreviate any code.
```

**C. Long creative (any big model):**
```
Tell me a 2,000-word story about a talent agent in 1970s Los Angeles who accidentally signs the most famous musician in the world under a fake name. Slow build, lots of dialogue, full ending — do not summarize or skip ahead.
```

**D. Auto + long research (tests the classifier under streaming):**
```
Give me a thorough research briefing on how NIL (name, image, likeness) rules have changed college sports since 2021 — the major rule changes, the money flow, the biggest deals, the risks for brands, and what a sports agency should be doing about it. Long form, organized sections.
```

**E. Web search ON + long (tests tool-call heartbeats):**
Toggle web search ON, then:
```
Search for the latest news on the 2028 LA Olympics sponsorship program and write me a detailed summary of the top sponsors announced so far, deal sizes if reported, and categories still open. Cite what you find.
```
Then SAME prompt in a new chat with web search OFF — it must answer from knowledge only, no searching.

**F. STOP button mid-stream:** start prompt A again, let it stream ~10 seconds, hit stop.
PASS = a calm random caption ("Okay, I'll wait." / "I'll sit right here." etc.), partial text stays, no red error, and — new with streaming — the generation actually stops server-side.

**G. Mixed session (the poison test):** in ONE chat: prompt D → then a coding ask → then "generate an image of a stadium at sunset" → then "what's the capital of France?" Every turn should answer normally.

**H. Images:** image mode → generate with each model tier if you feel thorough (Core → SD3.5 Large → Ultra). Multi-image on Ultra is the slowest path.

---

## 2. EVERYTHING ELSE THAT SHIPPED TODAY

| # | What | How to test | PASS looks like |
|---|---|---|---|
| 1 | Random stop captions (#990) | Stop a generation 3+ times | Different friendly message each time (5 variants), persists on reload |
| 2 | Clean version (#991) | Look at admin footer, user sidebar bottom, landing footer | Just "v0.6.11" (no -dev, no SHA); hover shows full build string |
| 3 | Users page search (#992) | /admin/users — type in the search box | Free-floating search below the filter card; dropdown of matching users w/ checkboxes narrows AS YOU TYPE; table live-filters; chips for selected users |
| 4 | Custom-range filters (#992) | /admin/users — Cap Usage + Cost selects | Range inputs appear ONLY when "Custom Range" selected, stay while active; composes with Today/7/30/month presets; dashboard + usage pages unchanged |
| 5 | Category image preview (#987) | /admin/categories — upload image, Save | Grid tile updates IMMEDIATELY, no refresh (you already passed this once — recheck post-deploys) |
| 6 | CSV import + description (#993) | /admin prompts → bulk import → download template | Template now has description + default_agent_id columns; import a row with a description and it persists; export includes it |
| 7 | Workflow prompt content (DB fix) | Open Brand Update Assistant, Deck Doc Sanitizer, AR Collections tiles | No garbage characters (the � are gone), no fake fill-in-the-blank fields from [REVIEW BEFORE SENDING]-style markers |
| 8 | Skeletons (#982) | Click a category tile on slow connection | Shimmer rows, no "Loading prompts…" text (passed already — spot check) |

## 3. WHAT I'M DOING IN PARALLEL (so we don't step on each other)
- I run as lz-tester-1 (+ lz-test-admin for admin checks), sessions labeled "this is a glance test"
- Full matrix: every dropdown model ×3, all image models, Auto across all prompt types, every toggle combo, reasoning display intact, console sweep — screenshots for all of it
- Fires only AFTER you've vetted and said go
- If you stop a stream while mine is running, that's the cross-container cancel test — bonus points

## Known-not-fixed (don't burn time on these)
- Auto can't route an image request to image mode (classifier gap, backlogged) — G's image step will answer as text; that's expected for now
- Image generation itself still has the 30s ceiling on its own endpoint (fast-follow)
- Live "watch it think" reasoning — next phase, after streaming soaks