HozuTrial 0013 — The notes app again, with --with auth (ADR 0030)
Question: trial 0012 built the notes app at 2.79× Nuxt, because the scaffold had no accounts. With
hozu add feature --with auth, what does the Hozu build cost now, and does it stay correct?
Setup
- As in trial 0012: the same spec, prompts, acceptance, model, and
claude -pin the app directory. - Hozu: two runs, with the packed 0.3.0 that includes ADR 0030.
- Nuxt was not re-run: its setup did not change, so the comparison uses trial 0012's two runs (mean 74.8 k).
Results
| Hozu run 1 | Hozu run 2 | Hozu mean | Trial 0012 Hozu mean | Nuxt mean | |
|---|---|---|---|---|---|
| Acceptance | 15 / 15 | 15 / 15 | 15 / 15 | 15 / 15 | |
| Turns | 17 | 20 | 20 | 6 | |
| Output tokens | 6,622 | 5,671 | 14,793 | 8,260 | |
| Weighted | 127.2 k | 120.8 k | 124.0 k | 209.1 k | 74.8 k |
| Cost (USD) | 0.56 | 0.51 | 0.54 | 0.89 | 0.36 |
| vs Nuxt | 1.66× | 2.79× |
Reading the result
- Both runs used
hozu add feature notes --page / --with auth,remove, then edited texts and details. Output tokens fell from 14.8 k to 6.1 k, because the agents no longer wrote the sign-in, session and per-user code. - Every check still passed: isolation, double submit, no-JS forms,
HttpOnly, the sign-in redirect. - The ratio (1.66×) now matches the task board's build (1.64× in trial 0010). What remains is the fixed reading cost of a framework the model does not know.
Conclusion
With --with auth, an account-based app costs 1.66× Nuxt to build instead of 2.79×, with every check passing. The
build-cost gap between simple apps and account-based apps is closed.