Trial 0018 — the widget task on 0.6 (hozu browse, HZ047)

Question: trial 0017 found that on a widget-heavy app, Hozu runs spent part of their extra cost verifying code that only runs in the browser. Does 0.6 (ADR 0040) lower that cost? It adds:

• hozu browse;
• HZ047;
• no favicon request;
• the add widget fix.

Setup

• Identical to trial 0017: the spec, change request, prompts, model (claude-opus-5-5) and the corrected hidden acceptance.
• Hozu: two Claude runs on the packed 0.6.0, whose generated CLAUDE.md lists hozu browse in the loop.
• Nuxt: trial 0017's two runs, not re-run.

Results

Correctness: every run passes everything: build 15/15, change 8/8 new and 15/15 regression, in both runs.

Cost (weighted tokens):

Reading the result

Build fell 19 %, because browser verification is now one command.

• Both runs used hozu browse 3–5 times per step:
  • fill a search;
  • click a station;
  • read the widget list and the errors;
  • one run also used --screenshot.
• No run wrote its own headless-Chrome or CDP script. In trial 0017 both builds did.
• No favicon detour. No HZ047 was hit: neither run put a helper outside impl.

Change did not move.

• h1's change spent about 5 calls on a detour unrelated to Hozu: it reformatted its client files with Prettier and reverted them.
• Both changes read the lock file and several topics.
• h2's change (158 k) is the cheapest Hozu change on this task so far.
• With two runs the spread (±19 %) is as large as the effect.

What is left is structural:

• Output: about 20 k output tokens per build, which is 98–100 k weighted and 35 % of the cost. Nuxt writes about 13–14 k. A widget is still a declaration, a client module and a ui.use, plus the machine and contracts.
• Docs carried: 74–79 k per build. The agents read the widgets, patterns, views, machine, data, pages, forms and contracts topics before writing.
• Starting the server: every run still ends with npm start, as the prompt allows. browse replaced the browser script, not this final habit.

Conclusion

• 0.6 cut the build on the widget task from 2.65× to 2.15× Nuxt, with every check passing. hozu browse was adopted without being asked for, and replaced the hand-written browser scripts.
• The change step is unchanged at about 3.5×. Its cost is reading and writing, not verification.
• The remaining gap is code volume and the guide. These are the next targets if the task class matters. Both are questions of design (fewer pieces per widget; less to read per change) rather than defects.