Trial 0019 — 0.7 (ADR 0041) on the widget task and the notes task

Question: ADR 0041 set targets before the runs. Does 0.7 meet them?

• What 0.7 changes:
  • seed (a machine's context from the URL);
  • shared fn helpers;
  • modules as declarations;
  • machine({ on });
  • a shorter changing.md.
• Targets:
  • widget build ≤ 1.8× Nuxt;
  • widget change ≤ 2.5× Nuxt;
  • notes no worse than trial 0016.

Setup

• Unchanged:
  • the widget task and prompts of trials 0017 and 0018;
  • the notes task and prompts of trial 0016;
  • the hidden acceptance scripts (widgets as corrected in trial 0017);
  • the model, claude-opus-5-5.
• Hozu: 0.7.0, packed from the repository, two Claude runs per task, all four running in parallel.
• Nuxt: not re-run.
  • Widget task: trial 0017, build 130.3 k and change 56.3 k.
  • Notes task: trial 0012, build 74.8 k and change 61.1 k.

Results

Correctness: every run passes everything.

• Widgets: build 15/15, change 8/8 new plus 15/15 regression.
• Notes: build 15/15, change 6/6 new plus 15/15 regression.

Cost (weighted tokens):

Reading the result

The widget apps used every new form without being asked.

• Both runs wrote a seed and a machine({ on }), shared at least one module helper between fns, and listed [model, views, widgets].
• The ctx.typed ? ctx.search : search.q pattern went from 19–23 per app to 0.
• The apps are 660–782 lines, against 852–1100 in trial 0018.
• A change adds 138–204 lines, against 214–529.
• The change step fell 41 %, more than the build. This is what ADR 0041 expected: the URL filter and the mode are now one place to edit, not seven.

The notes change got worse, and the cause is E.

• Both 0.7 runs read the data and contracts topics as well, and one also read machine. Trial 0016 read at most patterns, data and testing.
• The pin request is a per-item server action. The old changing.md answered it inline (the event, the mutation, a busy state, "no contract needed"). After E moved the recipes to hozu docs recipes, neither run opened that topic; they reconstructed the answer from three topics.
• Trial 0016's four runs spread from 74 k to 101 k, so two runs at 97 k and 108 k are partly noise. The extra docs calls are not noise.
• Fix: changing.md's table has a row for a per-item action stored on the server, with the steps and the contract answer. The file is 3.4 KB, against 5.6 KB before 0.7.

Re-run of the notes change with the fix

• Setup: both notes apps were reset to their build commit, changing.md was replaced, and the same change prompt ran again.
• Result: 91.1 k and 77.1 k, mean 84.1 k = 1.38× Nuxt (trial 0016: 88.8 k, 1.45×). Every check passes (6/6 new, 15/15 regression).
  • Calls: 12 and 10.
  • Neither run read the contracts topic. One read data and the other machine, besides SKILL, changing and patterns.

Conclusion

• The widget targets are met, build at 1.75× and change at 2.03× Nuxt (from 2.65× / 3.48× at 0.5.1), with every check passing.
• The notes build improved to 1.14×.
• The notes change first regressed to 1.68×, because a recipe left the file every change reads.
• With that row back in changing.md, the re-run is 1.38×. That is below trial 0016, so every ADR 0041 target is met: